Anne E CarpenterpersName, Thouis R JonespersName, Michael R LamprechtpersName, Colin ClarkepersName, In Han KangpersName, Ola FrimanpersName, David A GuertinpersName, Joo Han ChangpersName, Robert A LindquistpersName, Jason MoffatpersName, Polina GollandpersName, David M SabatinipersName (2006date). “CellProfiler: image analysis software for identifying and quantifying cell phenotypes.” Genome Biologytitle·j.
Object of study — automated identification and quantitative measurementterm of cellular phenotypes in microscopyterm images
Research question — Can a free, modular, open-source system make accurate, flexible, high-throughput cell-image analysis accessible across cell typesterm and assays?
Contribution — the CellProfilerterm software system, its modular pipelineterm architecture, validation across diverse phenotypes, cytological profiling example, and cluster-scale workflow
Cross-disciplinary analysis
content-tags & facets:
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | automated single-cell image analysis exposes quantitative, heterogeneous, and morphologic phenotypes that visual or population-level scoring can miss — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | |
| claim-type | method — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | |
| method — evidence | — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | |
| object of study — evidence | — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | |
| research question — evidence | — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | |
| contribution — evidence | — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | |
| mechanism | PIPELINE COMPOSITION sequences reusable modules so an analysis can be configured, repeated, and shared — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | mechanisms[0] |
| mechanism | ILLUMINATION NORMALIZATION removes field-of-view variation before identification and measurement, reducing downstream noise — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | mechanisms[1] |
| mechanism | THREE-STAGE DECLUMPING recognizes clumps, locates dividing lines, and removes or merges implausible objects — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | mechanisms[2] |
| mechanism | PRIMARY-TO-SECONDARY PROPAGATION uses already identified nuclei to find surrounding cell boundaries — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | mechanisms[3] |
| mechanism | SINGLE-CELL RETENTION preserves heterogeneous responses that population averages would mask — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | mechanisms[4] |
| mechanism | MULTIFEATURE PROFILING combines intensity, texture, shape, and compartment measures to distinguish sample effects — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | mechanisms[5] |
| mechanism | BATCH PARTITIONING divides a large image set into groups submitted to a computing cluster — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | mechanisms[6] |
| supports | CellProfiler as a free flexible high-throughput image-analysis system — “CellProfiler, the first free, open-source system for flexible and high-throughput cell image analysis is described.” | claim_relations[0] |
| supports | need for an open-source high-throughput cell-image-analysis platform — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | claim_relations[1] |
| supports | modular pipelines for reproducible image processing, identification, and measurement — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | claim_relations[2] |
| supports | accurate measurement across varied cellular phenotypes — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | claim_relations[3] |
| supports | cytological profiles can distinguish drug-specific phenotypic effects — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | claim_relations[4] |
| supports | cluster execution keeps analysis ahead of acquisition throughput — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | claim_relations[5] |
| sense | `image cytometry` means automated cell image analysis in this paper — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | senses[0] |
| sense | `object identification` is used synonymously with `segmentation` — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | senses[1] |
| sense | `primary objects` are the first detected objects, commonly nuclei — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | senses[2] |
| sense | `cytoplasm` means the portion of a cell excluding its nucleus — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | senses[3] |
| sense | `pipeline` means a sequentially ordered set of image-processing modules — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | senses[4] |
| sense | `cytological profiles` are the broad spectrum of measurements recorded for screened samples — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | senses[5] |
| note | CellProfiler as a free flexible high-throughput image-analysis system — The abstract directly describes the software contribution. — “CellProfiler, the first free, open-source system for flexible and high-throughput cell image analysis is described.” | claim_relations[0] |
| weight | CellProfiler as a free flexible high-throughput image-analysis system — 0.98 — “CellProfiler, the first free, open-source system for flexible and high-throughput cell image analysis is described.” | claim_relations[0] |
| note | need for an open-source high-throughput cell-image-analysis platform — The rationale identifies a clear unmet need and introduces the project. — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | claim_relations[1] |
| weight | need for an open-source high-throughput cell-image-analysis platform — 0.98 — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | claim_relations[1] |
| note | modular pipelines for reproducible image processing, identification, and measurement — The ordered module architecture and saved settings are described. — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | claim_relations[2] |
| weight | modular pipelines for reproducible image processing, identification, and measurement — 0.98 — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | claim_relations[2] |
| note | accurate measurement across varied cellular phenotypes — The validation section compares counts, size, DNA content, localization, and morphology. — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | claim_relations[3] |
| weight | accurate measurement across varied cellular phenotypes — 0.98 — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | claim_relations[3] |
| note | cytological profiles can distinguish drug-specific phenotypic effects — Wortmannin-specific shape changes distinguish two drugs with overlapping targets. — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | claim_relations[4] |
| weight | cytological profiles can distinguish drug-specific phenotypic effects — 0.98 — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | claim_relations[4] |
| note | cluster execution keeps analysis ahead of acquisition throughput — The reported cluster rate is stated to exceed image-acquisition rates. — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | claim_relations[5] |
| weight | cluster execution keeps analysis ahead of acquisition throughput — 0.98 — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | claim_relations[5] |
| note | visual analysis —contrasted-with→ image cytometry — Explicit preference/contrast between analysis by eye and image cytometry. — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concept_edges[0] |
| relation | visual analysis —contrasted-with→ image cytometry — Expert visual inspection is contrasted with automated image cytometry. — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concept_edges[0] |
| note | multiparametric single cell data —contrasted-with→ whole-population data — Explicit more-powerful-than comparison; categorical contrast is retained. — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concept_edges[1] |
| relation | multiparametric single cell data —contrasted-with→ whole-population data — Single-cell data are contrasted with aggregated whole-population measurements. — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concept_edges[1] |
| note | object identification —same-as→ segmentation — Directly stated relation. — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concept_edges[2] |
| relation | object identification —same-as→ segmentation — Object identification is explicitly also called segmentation. — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concept_edges[2] |
| audit | cytoplasm —part-of→ cell — Audit 2026-09-28: RESTORED - the full span was re-read and supports the edge; the full span names the cytoplasm as 'the part of each cell excluding the nucleus', so cytoplasm part-of cell is stated; the withdrawal said 'cytoplasm' was not named. The earlier withdrawal is kept in the record. — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concept_edges[3] |
| note | cytoplasm —part-of→ cell — Directly stated relation. — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concept_edges[3] |
| relation | cytoplasm —part-of→ cell — The cytoplasm is the part of each cell excluding the nucleus. — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concept_edges[3] |
| audit | illumination correction —causes→ fluorescence intensity measurements — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but illumination variation prevents accurate intensity measurements, which correction addresses; causes is loose. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concept_edges[4] |
| note | illumination correction —causes→ fluorescence intensity measurements — Directly stated relation. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concept_edges[4] |
| relation | illumination correction —causes→ fluorescence intensity measurements — Illumination correction reduces noise in quantitative fluorescence measurements. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concept_edges[4] |
| weight | illumination correction —causes→ fluorescence intensity measurements — 0.95 — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concept_edges[4] |
| relation | raw images → intensity measurements — Using raw images degrades intensity measurements. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[0] |
| relation | raw images → cell identification — Using raw images can preclude accurate cell identification. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[1] |
| relation | cell cycle stage → biological variability — Cell-cycle stage is a stated cause of variability in biological samples. — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | unhoused_relations[2] |
| relation | commercial software → new biological questions — Most commercial software cannot readily calculate new features for new biological questions. — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | unhoused_relations[3] |
| relation | individual cell measurements → subpopulation differences — Individual-cell measurements reveal differences confined to a subpopulation. — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | unhoused_relations[4] |
| relation | uneven illumination → quantitative differences — Uneven illumination obscures real quantitative differences. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[5] |
| relation | crowded cell samples → existing software — Existing software often fails to identify cells accurately in crowded samples. — “Commercial software has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell types and cellular features of pharmaceutical interest, including protein translocation, micronucleus formation, neurite outgrowth, and cell count [ 25 ]. The high cost and the bundling of commercial software with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cells, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysis has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cell identification is required to extract meaningful measures from images, but even for mammalian cell types, existing software often fails on crowded cell samples, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection.” | unhoused_relations[6] |
| relation | open-source code → method modification — Open-source code lets others inspect, modify, and improve the methodology. — “Overview of the software system The following can be freely downloaded from the CellProfiler website [ 31 ]: CellProfiler for Windows, Mac, and Unix (compiled, not requiring MATLAB); CellProfiler's MATLAB source code; a full technical description of CellProfiler's algorithms and measurements in an extensive PDF formatted manual (Additional data file 1), identical to the information found in help buttons within CellProfiler; and pipelines to identify the various cell types in this paper (see Additional data file 2 for a list of the modules in each pipeline). CellProfiler is freely available modular image analysis software that is capable of handling hundreds of thousands of images. The software contains already-developed methods for many cell types and assays and is also an open-source, flexible platform for the sharing, testing, and development of new methods by image analysis experts. CellProfiler meets the needs discussed in the introduction, in that it contains: advanced algorithms for image analysis that are able to accurately identify crowded cells and non-mammalian cell types; a modular, flexible design allowing analysis of new assays and phenotypes; open-source code so the underlying methodology is known and can be modified or improved by others; a user-friendly interface; the capability to make use of clusters of computers when available; and a design that eliminates the tedium of the many steps typically involved in image analysis, many of which are not easily transferable from one project to another (for example, image formatting, combining several image analysis steps, or repeating the analysis with slightly different parameters).” | unhoused_relations[7] |
| relation | cytological profiles → potential pathways — Drug-specific profiles narrow the potential pathways involved. — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | unhoused_relations[8] |
| relation | 100 CPU cluster → image analysis bottleneck — At the stated rate, image analysis is not a bottleneck relative to acquisition. — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | unhoused_relations[9] |
| relation | time-lapse movie → extra context information — Dedicated algorithms are needed to exploit context across time-lapse slices. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | unhoused_relations[10] |
| relation | rich data sets → gene annotation — Fully exploiting rich datasets is expected to enable rich annotation of many genes. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | unhoused_relations[11] |
| no type for this | raw images → intensity measurements — Using raw images degrades intensity measurements. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[0] |
| no type for this | raw images → cell identification — Using raw images can preclude accurate cell identification. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[1] |
| no type for this | cell cycle stage → biological variability — Cell-cycle stage is a stated cause of variability in biological samples. — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | unhoused_relations[2] |
| no type for this | commercial software → new biological questions — Most commercial software cannot readily calculate new features for new biological questions. — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | unhoused_relations[3] |
| no type for this | individual cell measurements → subpopulation differences — Individual-cell measurements reveal differences confined to a subpopulation. — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | unhoused_relations[4] |
| no type for this | uneven illumination → quantitative differences — Uneven illumination obscures real quantitative differences. — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | unhoused_relations[5] |
| no type for this | crowded cell samples → existing software — Existing software often fails to identify cells accurately in crowded samples. — “Commercial software has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell types and cellular features of pharmaceutical interest, including protein translocation, micronucleus formation, neurite outgrowth, and cell count [ 25 ]. The high cost and the bundling of commercial software with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cells, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysis has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cell identification is required to extract meaningful measures from images, but even for mammalian cell types, existing software often fails on crowded cell samples, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection.” | unhoused_relations[6] |
| no type for this | open-source code → method modification — Open-source code lets others inspect, modify, and improve the methodology. — “Overview of the software system The following can be freely downloaded from the CellProfiler website [ 31 ]: CellProfiler for Windows, Mac, and Unix (compiled, not requiring MATLAB); CellProfiler's MATLAB source code; a full technical description of CellProfiler's algorithms and measurements in an extensive PDF formatted manual (Additional data file 1), identical to the information found in help buttons within CellProfiler; and pipelines to identify the various cell types in this paper (see Additional data file 2 for a list of the modules in each pipeline). CellProfiler is freely available modular image analysis software that is capable of handling hundreds of thousands of images. The software contains already-developed methods for many cell types and assays and is also an open-source, flexible platform for the sharing, testing, and development of new methods by image analysis experts. CellProfiler meets the needs discussed in the introduction, in that it contains: advanced algorithms for image analysis that are able to accurately identify crowded cells and non-mammalian cell types; a modular, flexible design allowing analysis of new assays and phenotypes; open-source code so the underlying methodology is known and can be modified or improved by others; a user-friendly interface; the capability to make use of clusters of computers when available; and a design that eliminates the tedium of the many steps typically involved in image analysis, many of which are not easily transferable from one project to another (for example, image formatting, combining several image analysis steps, or repeating the analysis with slightly different parameters).” | unhoused_relations[7] |
| no type for this | cytological profiles → potential pathways — Drug-specific profiles narrow the potential pathways involved. — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | unhoused_relations[8] |
| no type for this | 100 CPU cluster → image analysis bottleneck — At the stated rate, image analysis is not a bottleneck relative to acquisition. — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | unhoused_relations[9] |
| no type for this | time-lapse movie → extra context information — Dedicated algorithms are needed to exploit context across time-lapse slices. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | unhoused_relations[10] |
| no type for this | rich data sets → gene annotation — Fully exploiting rich datasets is expected to enable rich annotation of many genes. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | unhoused_relations[11] |
| limitation | CellProfiler initially has very limited support for time-lapse and three-dimensional image-stack analysis. — “CellProfiler was designed and optimized for the most common high-content screening image format, that is, two-dimensional images. It has very limited support for time-lapse and three-dimensional image stack analysis, although researchers interested in these areas could build compatible modules.” | limitations[0] |
| limitation | Initial setup can take several days for a first-time user. — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | limitations[1] |
| limitation | Processing time varies widely with image size, object count, and feature count. — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | limitations[2] |
| limitation | Direct software comparisons are confounded by tuning, expense, and proprietary algorithms. — “Direct comparison of image analysis software is difficult because results from image analysis can be heavily skewed by how the software is tuned and commercial software packages are numerous and expensive. Furthermore, the algorithms in commercial software are proprietary and so cannot be directly compared apart from the entire software package, including preprocessing methods. The best practical comparison, therefore, is for software developers to release the results of their software on standard image sets or versus gold standards (visual inspection, Coulter particle counters, and so on). In subsequent sections, therefore, we present such comparisons. Note that once validated, any of these experiments could be expanded to a large-scale genome-wide RNAi screen or chemical library screen.” | limitations[3] |
| limitation | DNA-content measurement is readily degraded by illumination anomalies and object-identification errors. — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | limitations[4] |
| limitation | The biological basis and RNAi specificity of observed morphology changes remained undetermined. — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | limitations[5] |
| limitation | Most experiments using time-lapse or three-dimensional data require algorithms that exploit their additional context. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | limitations[6] |
| limitation | The downstream bottleneck shifts to visualization, exploration, and statistical analysis of rich measurements. — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | limitations[7] |
Concepts
terms this paper carries, each with the sentence it was cut from — the layer that joins one paper to another. Terms with no quote yet are on CONCEPTS.html.
| concept | CellProfiler — not defined in this paper's own words; free modular software system for automated cellular image analysis — “CellProfiler, the first free, open-source system for flexible and high-throughput cell image analysis is described.” | concepts[0] |
| concept | cell image analysis — not defined in this paper's own words; computational identification and measurement of cellular phenotypes in images — “CellProfiler, the first free, open-source system for flexible and high-throughput cell image analysis is described.” | concepts[1] |
| concept | microscopy — not defined in this paper's own words; imaging method used to study cellular function — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concepts[2] |
| concept | cellular function — not defined in this paper's own words; biological activity studied through cellular images — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concepts[3] |
| concept | visual analysis — not defined in this paper's own words; expert human inspection and scoring of microscopy images — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concepts[4] |
| concept | image cytometry — not defined in this paper's own words; automated quantitative analysis of cells in images — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concepts[5] |
| concept | high resolution images — not defined in this paper's own words; microscopy images collected in large numbers by advanced microscopes — “Examining cells by microscopy has long been a primary method for studying cellular function. When cells are stained appropriately, visual analysis can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution images of cells from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometry (automated cell image analysis) is strongly preferable to analysis by eye. In fact, in some cases image cytometry is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here.” | concepts[6] |
| concept | subcellular compartment — not defined in this paper's own words; labeled region within a cell whose properties are measured — “First, while human observers typically score one or at most a few cellular features, image cytometry simultaneously yields many informative measures of cells, including the intensity and localization of each fluorescently labeled cellular component (for example, DNA or protein) within each subcellular compartment, as well as the number, size, and shape of those subcellular compartments. Image-based analysis is thus versatile, inherently multiplexed, and high in information content. Like flow cytometry, image cytometry measures the per-cell amount of protein and DNA, but can more conveniently handle hundreds of thousands of distinct samples and is also compatible with adherent cell types, time-lapse samples, and intact tissues. In addition, image cytometry can accurately measure protein texture and localization as well as cell shape and size.” | concepts[7] |
| concept | flow cytometry — not defined in this paper's own words; comparison method measuring per-cell protein and DNA amounts — “First, while human observers typically score one or at most a few cellular features, image cytometry simultaneously yields many informative measures of cells, including the intensity and localization of each fluorescently labeled cellular component (for example, DNA or protein) within each subcellular compartment, as well as the number, size, and shape of those subcellular compartments. Image-based analysis is thus versatile, inherently multiplexed, and high in information content. Like flow cytometry, image cytometry measures the per-cell amount of protein and DNA, but can more conveniently handle hundreds of thousands of distinct samples and is also compatible with adherent cell types, time-lapse samples, and intact tissues. In addition, image cytometry can accurately measure protein texture and localization as well as cell shape and size.” | concepts[8] |
| concept | protein texture — not defined in this paper's own words; spatial pattern of protein staining — “First, while human observers typically score one or at most a few cellular features, image cytometry simultaneously yields many informative measures of cells, including the intensity and localization of each fluorescently labeled cellular component (for example, DNA or protein) within each subcellular compartment, as well as the number, size, and shape of those subcellular compartments. Image-based analysis is thus versatile, inherently multiplexed, and high in information content. Like flow cytometry, image cytometry measures the per-cell amount of protein and DNA, but can more conveniently handle hundreds of thousands of distinct samples and is also compatible with adherent cell types, time-lapse samples, and intact tissues. In addition, image cytometry can accurately measure protein texture and localization as well as cell shape and size.” | concepts[9] |
| concept | protein localization — not defined in this paper's own words; position of protein staining relative to cellular compartments — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concepts[10] |
| concept | cell shape — not defined in this paper's own words; quantitative geometry of a cell — “First, while human observers typically score one or at most a few cellular features, image cytometry simultaneously yields many informative measures of cells, including the intensity and localization of each fluorescently labeled cellular component (for example, DNA or protein) within each subcellular compartment, as well as the number, size, and shape of those subcellular compartments. Image-based analysis is thus versatile, inherently multiplexed, and high in information content. Like flow cytometry, image cytometry measures the per-cell amount of protein and DNA, but can more conveniently handle hundreds of thousands of distinct samples and is also compatible with adherent cell types, time-lapse samples, and intact tissues. In addition, image cytometry can accurately measure protein texture and localization as well as cell shape and size.” | concepts[11] |
| concept | human-scored image analysis — not defined in this paper's own words; qualitative categorization of samples by observers — “Second, human-scored image analysis is qualitative, usually categorizing samples as 'hits' (where normal physiology is grossly disturbed) or 'non-hits'. By contrast, automated analysis rapidly produces consistent, quantitative measures for every image. In addition to uncovering subtle samples of interest that would otherwise be missed, systems-level conclusions can be drawn directly from the quantitative measures for every image. Measuring a large number of features, even features undetectable by eye, has proven useful for screening as well as cytological/cytometric profiling, which can group similar genes or reveal a drug's mechanism of action [ 3 , 8 - 14 ].” | concepts[12] |
| concept | quantitative measures — not defined in this paper's own words; consistent numeric outputs produced for every image — “Second, human-scored image analysis is qualitative, usually categorizing samples as 'hits' (where normal physiology is grossly disturbed) or 'non-hits'. By contrast, automated analysis rapidly produces consistent, quantitative measures for every image. In addition to uncovering subtle samples of interest that would otherwise be missed, systems-level conclusions can be drawn directly from the quantitative measures for every image. Measuring a large number of features, even features undetectable by eye, has proven useful for screening as well as cytological/cytometric profiling, which can group similar genes or reveal a drug's mechanism of action [ 3 , 8 - 14 ].” | concepts[13] |
| concept | cytological/cytometric profiling — not defined in this paper's own words; multifeature profiling used to group genes or reveal drug action — “Second, human-scored image analysis is qualitative, usually categorizing samples as 'hits' (where normal physiology is grossly disturbed) or 'non-hits'. By contrast, automated analysis rapidly produces consistent, quantitative measures for every image. In addition to uncovering subtle samples of interest that would otherwise be missed, systems-level conclusions can be drawn directly from the quantitative measures for every image. Measuring a large number of features, even features undetectable by eye, has proven useful for screening as well as cytological/cytometric profiling, which can group similar genes or reveal a drug's mechanism of action [ 3 , 8 - 14 ].” | concepts[14] |
| concept | mechanism of action — not defined in this paper's own words; biological action inferred from measured phenotypic profiles — “Second, human-scored image analysis is qualitative, usually categorizing samples as 'hits' (where normal physiology is grossly disturbed) or 'non-hits'. By contrast, automated analysis rapidly produces consistent, quantitative measures for every image. In addition to uncovering subtle samples of interest that would otherwise be missed, systems-level conclusions can be drawn directly from the quantitative measures for every image. Measuring a large number of features, even features undetectable by eye, has proven useful for screening as well as cytological/cytometric profiling, which can group similar genes or reveal a drug's mechanism of action [ 3 , 8 - 14 ].” | concepts[15] |
| concept | multiparametric single cell data — not defined in this paper's own words; multiple measurements retained separately for individual cells — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concepts[16] |
| concept | whole-population data — not defined in this paper's own words; measurements aggregated across an entire cell population — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concepts[17] |
| concept | causal networks — not defined in this paper's own words; networks inferred using multiparametric single-cell measurements — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concepts[18] |
| concept | subpopulation — not defined in this paper's own words; subset of cells whose response can be masked by population summaries — “Third, image cytometry individually measures each cell rather than producing a score for the entire image. Because individual cells' responses are inhomogeneous [ 15 ], multiparametric single cell data from several types of instruments have proven much more powerful than whole-population data (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networks, classifying protein localization, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cell measurements can reveal samples that differ in only a subpopulation of cells, which would otherwise be masked in whole-population measures.” | concepts[19] |
| concept | DNA staining intensity — not defined in this paper's own words; image-derived signal used to distinguish cell-cycle states — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[20] |
| concept | cell cycle — not defined in this paper's own words; progression whose G1 and G2 states differ in DNA staining — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[21] |
| concept | nucleus size — not defined in this paper's own words; nuclear area whose small changes are difficult to score by eye — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[22] |
| concept | texture — not defined in this paper's own words; smoothness pattern of protein or DNA staining — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[23] |
| concept | cell physiology — not defined in this paper's own words; cell state reflected in otherwise subtle texture changes — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[24] |
| concept | disease state — not defined in this paper's own words; pathological condition revealed by image features — “Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensity that reveals whether a cell is in G1 or G2 phase of the cell cycle are measurable by computer but are difficult for the human eye to observe in cell images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus size, are not noticeable by eye. Other features, for example, the texture (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changes in DNA or protein texture can correlate to profound and otherwise undetectable changes in cell physiology, a fact used in diagnosis of disease [ 17 , 19 ]. Even changes not visible to the human eye can reveal disease state [ 20 ].” | concepts[25] |
| concept | large-scale experiment — not defined in this paper's own words; high-volume image experiment processed automatically — “Fifth, image cytometry is much less labor-intensive and higher-throughput. Appropriate software produces reliable results from a large-scale experiment in hours, versus months of tedious visual inspection. This improvement is more than an incremental technical advance, because it relieves the one remaining bottleneck to routinely conducting such experiments.” | concepts[26] |
| concept | ImageJ/NIH Image — not defined in this paper's own words; earlier flexible open-source image-analysis package oriented to individual images — “Prior to the work presented here, the only flexible, open-source biological image analysis package was ImageJ/NIH Image [ 21 ]. This package has been successfully used by many laboratories. Its design, however, is geared more towards the analysis of individual images (comparable to Adobe Photoshop) rather than flexible, high-throughput work. Macros can be written in ImageJ for high-throughput work but adapting macros to new projects requires that biologists learn a programming language. While not creating a general, flexible software tool, many groups have benefited from automated cell image analysis by developing their own scripts, macros, and plug-ins to accomplish specific image analysis tasks. Custom programs written in commercial software (for example, MetaMorph, ImagePro Plus, MATLAB) or Java have been used to identify, measure, and track cells in images and time lapse movies [ 10 , 22 , 23 ]. Such studies clearly show the power of automated image analysis for biological discovery. However, most of these custom programs are not modular, so combining several steps and changing settings requires interacting directly with the code and is simply not practical for routinely processing hundreds of thousands of images or sending jobs to a cluster. The effort expended by laboratories in creating an analysis solution with a particular software package is often lost after the initial experiment is completed; other laboratories rarely use the methods because they are customized for a particular cell type, assay or even image set. Furthermore, although developing a routine for a new cell type or assay usually requires testing multiple algorithms, it is impractical to implement and test several published methods for a particular project.” | concepts[27] |
| concept | programming language — not defined in this paper's own words; skill required to adapt ImageJ macros to new projects — “Prior to the work presented here, the only flexible, open-source biological image analysis package was ImageJ/NIH Image [ 21 ]. This package has been successfully used by many laboratories. Its design, however, is geared more towards the analysis of individual images (comparable to Adobe Photoshop) rather than flexible, high-throughput work. Macros can be written in ImageJ for high-throughput work but adapting macros to new projects requires that biologists learn a programming language. While not creating a general, flexible software tool, many groups have benefited from automated cell image analysis by developing their own scripts, macros, and plug-ins to accomplish specific image analysis tasks. Custom programs written in commercial software (for example, MetaMorph, ImagePro Plus, MATLAB) or Java have been used to identify, measure, and track cells in images and time lapse movies [ 10 , 22 , 23 ]. Such studies clearly show the power of automated image analysis for biological discovery. However, most of these custom programs are not modular, so combining several steps and changing settings requires interacting directly with the code and is simply not practical for routinely processing hundreds of thousands of images or sending jobs to a cluster. The effort expended by laboratories in creating an analysis solution with a particular software package is often lost after the initial experiment is completed; other laboratories rarely use the methods because they are customized for a particular cell type, assay or even image set. Furthermore, although developing a routine for a new cell type or assay usually requires testing multiple algorithms, it is impractical to implement and test several published methods for a particular project.” | concepts[28] |
| concept | custom programs — not defined in this paper's own words; project-specific scripts, macros, and plug-ins for image analysis — “Prior to the work presented here, the only flexible, open-source biological image analysis package was ImageJ/NIH Image [ 21 ]. This package has been successfully used by many laboratories. Its design, however, is geared more towards the analysis of individual images (comparable to Adobe Photoshop) rather than flexible, high-throughput work. Macros can be written in ImageJ for high-throughput work but adapting macros to new projects requires that biologists learn a programming language. While not creating a general, flexible software tool, many groups have benefited from automated cell image analysis by developing their own scripts, macros, and plug-ins to accomplish specific image analysis tasks. Custom programs written in commercial software (for example, MetaMorph, ImagePro Plus, MATLAB) or Java have been used to identify, measure, and track cells in images and time lapse movies [ 10 , 22 , 23 ]. Such studies clearly show the power of automated image analysis for biological discovery. However, most of these custom programs are not modular, so combining several steps and changing settings requires interacting directly with the code and is simply not practical for routinely processing hundreds of thousands of images or sending jobs to a cluster. The effort expended by laboratories in creating an analysis solution with a particular software package is often lost after the initial experiment is completed; other laboratories rarely use the methods because they are customized for a particular cell type, assay or even image set. Furthermore, although developing a routine for a new cell type or assay usually requires testing multiple algorithms, it is impractical to implement and test several published methods for a particular project.” | concepts[29] |
| concept | commercial software — not defined in this paper's own words; proprietary packages developed chiefly for pharmaceutical screening — “Commercial software has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell types and cellular features of pharmaceutical interest, including protein translocation, micronucleus formation, neurite outgrowth, and cell count [ 25 ]. The high cost and the bundling of commercial software with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cells, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysis has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cell identification is required to extract meaningful measures from images, but even for mammalian cell types, existing software often fails on crowded cell samples, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection.” | concepts[30] |
| concept | protein translocation — not defined in this paper's own words; cellular feature targeted by commercial screening packages — “Commercial software has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell types and cellular features of pharmaceutical interest, including protein translocation, micronucleus formation, neurite outgrowth, and cell count [ 25 ]. The high cost and the bundling of commercial software with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cells, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysis has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cell identification is required to extract meaningful measures from images, but even for mammalian cell types, existing software often fails on crowded cell samples, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection.” | concepts[31] |
| concept | crowded cell samples — not defined in this paper's own words; images in which touching cells challenge identification algorithms — “Commercial software has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell types and cellular features of pharmaceutical interest, including protein translocation, micronucleus formation, neurite outgrowth, and cell count [ 25 ]. The high cost and the bundling of commercial software with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cells, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysis has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cell identification is required to extract meaningful measures from images, but even for mammalian cell types, existing software often fails on crowded cell samples, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection.” | concepts[32] |
| concept | flexible, open-source platform — not defined in this paper's own words; needed extensible platform for high-throughput cell-image analysis — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | concepts[33] |
| concept | cell types — not defined in this paper's own words; biological classes across which the system is evaluated — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | concepts[34] |
| concept | real-world biological problems — not defined in this paper's own words; diverse practical assays used to validate the system — “In summary, while existing software enables particular assays for particular cell types, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell types. There exists a clear need for a powerful, flexible, open-source platform for high-throughput cell image analysis. Here we describe the open-source CellProfiler project, our effort to develop such a software system for the scientific community. CellProfiler simultaneously measures the size, shape, intensity and texture of a variety of cell types in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problems, demonstrate the breadth of its utility (including on various cell types and assays), and hope to stimulate ideas within the biological community for future applications of the software.” | concepts[35] |
| concept | modular image analysis software — not defined in this paper's own words; extensible software assembled from reusable processing modules — “Overview of the software system The following can be freely downloaded from the CellProfiler website [ 31 ]: CellProfiler for Windows, Mac, and Unix (compiled, not requiring MATLAB); CellProfiler's MATLAB source code; a full technical description of CellProfiler's algorithms and measurements in an extensive PDF formatted manual (Additional data file 1), identical to the information found in help buttons within CellProfiler; and pipelines to identify the various cell types in this paper (see Additional data file 2 for a list of the modules in each pipeline). CellProfiler is freely available modular image analysis software that is capable of handling hundreds of thousands of images. The software contains already-developed methods for many cell types and assays and is also an open-source, flexible platform for the sharing, testing, and development of new methods by image analysis experts. CellProfiler meets the needs discussed in the introduction, in that it contains: advanced algorithms for image analysis that are able to accurately identify crowded cells and non-mammalian cell types; a modular, flexible design allowing analysis of new assays and phenotypes; open-source code so the underlying methodology is known and can be modified or improved by others; a user-friendly interface; the capability to make use of clusters of computers when available; and a design that eliminates the tedium of the many steps typically involved in image analysis, many of which are not easily transferable from one project to another (for example, image formatting, combining several image analysis steps, or repeating the analysis with slightly different parameters).” | concepts[36] |
| concept | graphical user interface — not defined in this paper's own words; point-and-click interface for configuring analyses — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[37] |
| concept | pipeline — not defined in this paper's own words; ordered sequence of modules applied to image groups — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[38] |
| concept | modules — not defined in this paper's own words; reusable processing units arranged sequentially in a pipeline — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[39] |
| concept | image processing — not defined in this paper's own words; initial pipeline stage preparing images — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[40] |
| concept | object identification — not defined in this paper's own words; pipeline stage detecting cells or subcellular objects — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[41] |
| concept | measurement — not defined in this paper's own words; pipeline stage quantifying identified objects — “Most image analysis projects, even for new cell types or assays, can be accomplished simply by pointing and clicking using CellProfiler's graphical user interface (Figure 1a ). The software uses the concept of a 'pipeline' of individual modules (Figure 1b ; Additional data file 2). Each module processes the images in some manner, and the modules are placed in sequential order to create a pipeline: usually image processing, then object identification, then measurement. Over 50 CellProfiler modules are currently available (Additional data file 3). Most modules are automatic, but the software also allows interactive modules (for example, the user clicks to outline a region of interest in each image). Modules are mixed and matched for a specific project and each module's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipeline and is processed by each module in order.” | concepts[42] |
| concept | illumination correction — not defined in this paper's own words; normalization of spatial lighting variation before measurement — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concepts[43] |
| concept | field of view — not defined in this paper's own words; image region across which illumination varies — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concepts[44] |
| concept | fluorescence intensity measurements — not defined in this paper's own words; numeric fluorescence signals distorted by uneven lighting — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concepts[45] |
| concept | raw images — not defined in this paper's own words; uncorrected inputs whose illumination variation degrades measurements — “Image processing, including illumination correction One of the most critical steps in image analysis is illumination correction. Illumination often varies more than 1.5-fold across the field of view, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurements (for example, DNA content of a nucleus, which only varies by two-fold during the cell cycle). CellProfiler contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measures (Figure 1c,d ). We use these illumination correction methods for every high-throughput image set we process, because using raw images degrades intensity measurements and, less obviously, can preclude accurate cell identification. This adversely affects all types of measurements, from intensity based measures (for example, DNA content histograms [ 1 ]) to area and shape measurements (TRJ, AEC, DMS, and PG, unpublished data). CellProfiler's other image processing modules perform other needed adjustments prior to identifying cells in images, for example, aligning or cropping (Additional data file 3).” | concepts[46] |
| concept | segmentation — not defined in this paper's own words; alternate name for object identification — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[47] |
| concept | primary objects — not defined in this paper's own words; first objects detected, commonly nuclei — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[48] |
| concept | nuclei — not defined in this paper's own words; DNA-stained primary objects commonly identified first — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[49] |
| concept | three-step strategy — not defined in this paper's own words; recognize and split clumps, find boundaries, then remove or merge objects — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[50] |
| concept | clumped objects — not defined in this paper's own words; touching objects requiring separation during segmentation — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[51] |
| concept | secondary objects — not defined in this paper's own words; objects such as cell outlines found around primary objects — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[52] |
| concept | watershed method — not defined in this paper's own words; common method that can fail at borders between clumped cells — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[53] |
| concept | Propagate algorithm — not defined in this paper's own words; improved algorithm for secondary-object identification — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[54] |
| concept | cell — not defined in this paper's own words; whole cellular object whose cytoplasm excludes the nucleus — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[55] |
| concept | cytoplasm — not defined in this paper's own words; part of each cell excluding its nucleus — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[56] |
| concept | nuclear membrane — not defined in this paper's own words; edge of a nucleus represented as a compartment — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[57] |
| concept | test mode — not defined in this paper's own words; side-by-side comparison mode for identification algorithms — “Cell identification Object identification (also called segmentation) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cell measurements. CellProfiler's object identification modules contain a variety of published and tested algorithms for identifying cells based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cells touch each other, causing the simple, fast algorithms used in some commercial software packages to fail. The first objects identified in an image (called primary objects) are often nuclei identified from DNA-stained images, although primary objects can also be whole cells, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfiler for cases where primary objects are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objects, CellProfiler contains a modular three-step strategy based on previously published algorithms [ 33 - 37 ]. First, clumped objects are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurements, for example, their size or shape. After primary objects (often nuclei) are identified, the edges of secondary objects that surround each primary object (often cell edges) can be found more easily. Measuring cell size in Drosophila was not previously feasible because the commonly used watershed method [ 37 ] often fails to find the borders between clumped cells. We have, therefore, implemented in CellProfiler an improved Propagate algorithm [ 38 ], in addition to several standard methods of secondary object identification. Other subcellular compartments can also be identified, including the cytoplasm (the part of each cell excluding the nucleus) and the cell or nuclear membrane (the edge of the cell or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modules include a 'test mode' for comparing several algorithms side by side in order to choose the best approach. We have found that these cell identification methods are flexible to various cell morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cells with unusual morphologies within a population of normal cells.” | concepts[58] |
| concept | Zernike shape features — not defined in this paper's own words; complex descriptors of object shape — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | concepts[59] |
| concept | Haralick and Gabor texture features — not defined in this paper's own words; complex descriptors of image texture — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | concepts[60] |
| concept | tab-delimited spreadsheet format — not defined in this paper's own words; one supported export form for measurements — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | concepts[61] |
| concept | database — not defined in this paper's own words; structured destination for exported measurements — “Measurements and data analysis CellProfiler measures a large number of features for each identified cell or subcellular compartment, including area, shape, intensity, and texture (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurements like Zernike shape features [ 41 ], and Haralick and Gabor texture features [ 42 - 44 ], which are described in detail in the online help and manual. There are also modules to measure various features (for example, intensity, texture, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial software is the inability to adapt to new biological questions by calculating new features from identified cells [ 5 ]. By contrast, CellProfiler's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurements are accessible in several ways: using CellProfiler's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet format that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a database like Oracle or MySQL; or directly in MATLAB.” | concepts[62] |
| concept | non-programmers — not defined in this paper's own words; users enabled by the point-and-click modular design — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | concepts[63] |
| concept | quality control step — not defined in this paper's own words; testing sample images and adjusting identification settings — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | concepts[64] |
| concept | MATLAB — not defined in this paper's own words; language in which the source code was written — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | concepts[65] |
| concept | analysis times — not defined in this paper's own words; per-image processing duration under typical pipelines — “Usability Like most new software in the laboratory, the process of setting up a CellProfiler analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test mode' for the Identify modules that show results from various options side by side, and built-in image and data tools to interact with processed images and cell measurements (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmers. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell type or unusual measurement scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modules. This takes less than an hour and is essentially a quality control step. Once a pipeline is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfiler's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLAB because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmers. Computationally intensive tasks use either MATLAB's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis times vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelines require from 20 seconds to five minutes per image on standard desktop computers.” | concepts[66] |
| concept | Drosophila Kc167 cells — not defined in this paper's own words; challenging cell type used in validation experiments — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | concepts[67] |
| concept | cell count — not defined in this paper's own words; number of cells detected in an image — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | concepts[68] |
| concept | cell area measurements — not defined in this paper's own words; image-derived cell-size estimates — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | concepts[69] |
| concept | Coulter particle counter — not defined in this paper's own words; gold-standard instrument used to compare cell size — “Validation of CellProfiler for many phenotypes We first demonstrated that CellProfiler's methods could accurately measure many different biologically important features of cells using several cell types, including Drosophila Kc167 cells because these cells are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cell microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cells for experiments shown in Figures 1 and 2 by pretreating the cells with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cells in 10 ml medium. We prepared human HT29 cells (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfiler for many cellular phenotypes. (a) Cell count: for a set of 6 images of wild-type human HT29 cells (left), two researchers' counts varied by 11%, and CellProfiler's counts were within 6% of their average. For images of Drosophila Kc167 cells with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfiler's counts were within 17% of their average. Example images and CellProfiler outlines for these cell types are shown in Figure 1e. (b) Cell size: CellProfiler's cell area measurements are comparable to those of a Coulter particle counter for Drosophila Kc167 cells, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA content in cell populations: measurements are shown for human HT29 cell populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cell populations (1,750 images for each RNAi condition were combined, right). The cell cycle distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nuclei are known to have 4N DNA content [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycle distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-protein content: human HT29 cells treated with RNAi reagent against Polo kinase have an increased percentage of nuclei with high phospho-H3 staining compared to wild-type cells, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nuclei that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cells (right). (f) Protein localization: the mean intensity of NFκB staining in the cytoplasm and the nucleus is shown in response to TNFα in human MCF7 cells (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cells disappear at timepoints as the cells recover. Scale bar = 10 μm. The SEM is too small to show error bars.” | concepts[70] |
| concept | DNA content — not defined in this paper's own words; per-cell DNA quantity inferred from fluorescence intensity — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | concepts[71] |
| concept | cell cycle stage — not defined in this paper's own words; G1, G2, or mitotic state inferred from DNA and morphology — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | concepts[72] |
| concept | phospho-protein — not defined in this paper's own words; phosphorylated protein amount measured from antibody staining — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | concepts[73] |
| concept | subcellular structures — not defined in this paper's own words; intracellular objects whose number and morphology can be measured — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | concepts[74] |
| concept | non-Gaussian distributions — not defined in this paper's own words; shape of many wild-type morphology-feature distributions — “Cell count (used to probe cell proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell types due to the poor ability of existing software to separate clumped nuclei. For human cells (Figure 2a , left), CellProfiler's accuracy compared to manual counting is twice that reported for a commercial software package [ 25 ]. CellProfiler also counted the more difficult-to-identify Drosophila Kc167 cells (Figure 2a , right). Cell size was not previously measurable for many cell types, but CellProfiler's measurements were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell type [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cell size in Drosophila Kc167 cells and the results indicate that such screens are now feasible. Cell count and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changes in DNA content, are impossible to discern by eye. Unlike whole population-based methods, image cytometry measures individual cell fluorescence intensities so that the DNA content of DNA-stained cells can be determined [ 46 , 47 ]. These measurements are very easily degraded by anomalies in the illumination of the field of view and poor identification algorithms (the most common errors are counting two nearby nuclei as one nucleus with twice the DNA content and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfiler produced the expected DNA content distributions for both human and Drosophila cell populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cell decreases by roughly half when a cell divides into two daughter cells during mitosis (Figure 2d ). Classification of cells based on 2N, 4N, and 8N DNA content is, therefore, possible based on an image of DNA-stained nuclei (Additional data file 6). This is useful not only in studies of cell cycle per se : cell cycle stage is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycle eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cells, which have 2N DNA content). Further, image cytometry adds an additional level of information about cell cycle distribution. Whereas flow cytometry based on a DNA stain alone cannot distinguish cells in G2 and M phase (both having the same 4N DNA content), image cytometry reveals that these two populations differ in that mitotic cells have smaller nuclei on average (Figure 2e , right). The total amount of a protein or phospho-protein per cell can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometry can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cells can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structures such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and texture/smoothness. Cell morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and texture measurements for wild-type cells show non-Gaussian distributions (Additional data file 7). Therefore, independently measuring every cell by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changes in cell shape and actin texture induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies.” | concepts[75] |
| concept | Forkhead-EGFP — not defined in this paper's own words; reporter used in the cytoplasm-to-nucleus translocation assay — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | concepts[76] |
| concept | cytological profiles — not defined in this paper's own words; broad multifeature measurement vectors for screened samples — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | concepts[77] |
| concept | assay quality — not defined in this paper's own words; screenability evaluated using Z-prime and V factors — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | concepts[78] |
| concept | wortmannin — not defined in this paper's own words; drug producing translocation and distinctive nuclear-shape changes — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | concepts[79] |
| concept | LY294002 — not defined in this paper's own words; comparison drug producing translocation without the same shape change — “Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfiler's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFP cytoplasm to nucleus translocation assay in human cells grown in multi-well plates (Figure 4a ). First, we ran a CellProfiler pipeline (Additional data file 2, part E) to calculate an illumination correction function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfiler pipeline (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nuclei, identify cell edges, and use the nucleus and cell outlines to define the cytoplasmic region of each cell, thereby defining three compartments for each cell: nucleus, cell, and cytoplasm. For each slide, we tested the pipeline on several random test images and fine-tuned settings in the identification modules as needed. Modules in the pipeline were included to measure: multiple features describing the area and shape of each compartment for each cell; multiple features describing the intensity and texture of each channel within each compartment, including several scales of texture; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfiler analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cells, showing no treatment (left) and 150 nM Wortmannin (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cells whose ratio of GFP in the cytoplasm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysis using Z' and non-logistic-fit V factors, which are standard measures of assay quality (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nuclei change shape in response to wortmannin but not LY294002, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay quality yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial software [ 50 , 51 ]. Existing commercial software for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurements recorded by CellProfiler (the cytological profiles) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmannin but, interestingly, not the other positive control drug in this assay, LY294002 (Figure 4d ). These subtle changes were seen at all doses at and above the EC50 of wortmannin but at none of the doses of LY294002, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmannin and LY294002 target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changes using primary screening data.” | concepts[80] |
| concept | computing cluster — not defined in this paper's own words; parallel computing resource for large image sets — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | concepts[81] |
| concept | 3,000 images/hour — not defined in this paper's own words; stated throughput of a one-hundred-CPU cluster — “Cluster computing We routinely run CellProfiler on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modules to the end of a pipeline to enable processing batches of images on the cluster and exporting data into a database. We then process the first image on a desktop computer, after which CellProfiler automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing cluster. We then use simple commands, outside of CellProfiler, to submit the jobs to our cluster of computers and export the resulting measurements to a database. Each of these steps is described in the CellProfiler help for batch processing, and researchers without a computing cluster can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hour. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck.” | concepts[82] |
| concept | high-throughput biology laboratories — not defined in this paper's own words; laboratories presented as beneficiaries of broad applicability — “Broad applicability Here we have shown that CellProfiler is useful for measuring a number of cell features, including cell count, cell size, cell cycle distribution, organelle number and size, cell shape, texture, and the levels and localization of proteins and phospho-proteins. Unlike previously existing software, CellProfiler is effective in a number of cell types and organisms, such that new avenues of research in both standard and high-throughput biology laboratories can now be pursued. CellProfiler is already being used by laboratories worldwide studying a variety of biological processes in other cell types and organisms, including Drosophila (S2R+ cells, epithelial tissue), human (TOV21G, biopsied prostate gland tissue, adult mesenchymal stem cells, H1299 lung carcinoma), mouse (NIH/3T3, neural precursor cells derived from embryos, lung tissue sections, isolated germ cells), and rat (H9c2 cells) [ 1 , 26 , 48 , 52 - 54 ] (KA Hartwell, personal communication). We have also modified CellProfiler to measure yeast colonies, yeast growth patches, wounds in scratch assays, and tumors [ 55 ]. Importantly, the only successfully completed Drosophila screen using automated image analysis, to date, has been a cell-count/object-count screen in the S2 line whose appearance is comparable to human cell lines [ 56 ]. We are currently using CellProfiler to analyze screens using the clumpy Drosophila Kc167 cell type (AEC, TRJ, MRL, DB Wheeler, PG, DMS, unpublished data). Given the power of RNA interference and genetic tools in Drosophila and the demand for screening in its community [ 57 ], this is an area that can now move past tedious visual analysis of thousands of images, accelerating the rate of discovery.” | concepts[83] |
| concept | computer vision researchers — not defined in this paper's own words; researchers invited to contribute new algorithms — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[84] |
| concept | three-dimensional image set — not defined in this paper's own words; image-stack input needing algorithms that exploit contextual information — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[85] |
| concept | Open Microscopy Environment — not defined in this paper's own words; open-source infrastructure being integrated with CellProfiler — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[86] |
| concept | data visualization — not defined in this paper's own words; downstream activity becoming a bottleneck — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[87] |
| concept | statistical analysis — not defined in this paper's own words; downstream activity needed for rich genome-wide measurements — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[88] |
| concept | cellular signaling networks — not defined in this paper's own words; systems expected to be revealed by exploiting rich datasets — “Future development We hope that computer vision researchers will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfiler can currently analyze each slice of a time-lapse movie or three-dimensional image set independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfiler is currently being integrated with the open-source Open Microscopy Environment project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualization, exploration, and statistical analysis in order to accommodate the number and richness of measurements that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networks and lead to the unprecedented rich annotation of hundreds of genes in parallel.” | concepts[89] |
Full text — tagged
The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.
Abstract Short abstract CellProfilerterm, the first free, open-source system for flexible and high-throughput cell image analysisterm is described. Rationale Examining cellsterm by microscopyterm has long been a primary method for studying cellular functionterm. When cellsterm are stained appropriately, visual analysisterm can reveal biological mechanisms. Advanced microscopes can now, in a single day, easily collect thousands of high resolution imagesterm of cellsterm from time-lapse experiments and from large-scale screens using chemical compounds, RNA interference (RNAi) reagents, or expression plasmids [ 1 - 5 ]. However, a bottleneck exists at the image analysis stage. Several pioneering large screens have been scored through visual inspection by expert biologists [ 6 , 7 ], whose interpretive ability will not soon be replicated by a computer. Still, for most applications, image cytometryterm (automated cell image analysisterm) is strongly preferable to analysis by eye. In fact, in some cases image cytometryterm is absolutely required to extract the full spectrum of information present in biological images, for reasons we discuss here. First, while human observers typically score one or at most a few cellular features, image cytometryterm simultaneously yields many informative measures of cellsterm, including the intensity and localization of each fluorescently labeled cellular component (for example, DNA or protein) within each subcellular compartmentterm, as well as the number, size, and shape of those subcellular compartmentsterm. Image-based analysis is thus versatile, inherently multiplexed, and high in information content. Like flow cytometryterm, image cytometryterm measures the per-cell amount of protein and DNA, but can more conveniently handle hundreds of thousands of distinct samples and is also compatible with adherent cell typesterm, time-lapse samples, and intact tissues. In addition, image cytometryterm can accurately measure protein textureterm and localization as well as cell shapeterm and size. Second, human-scored image analysisterm is qualitative, usually categorizing samples as 'hits' (where normal physiology is grossly disturbed) or 'non-hits'. By contrast, automated analysis rapidly produces consistent, quantitative measuresterm for every image. In addition to uncovering subtle samples of interest that would otherwise be missed, systems-level conclusions can be drawn directly from the quantitative measuresterm for every image. Measuring a large number of features, even features undetectable by eye, has proven useful for screening as well as cytological/cytometric profilingterm, which can group similar genes or reveal a drug's mechanism of actionterm [ 3 , 8 - 14 ]. Third, image cytometryterm individually measures each cellterm rather than producing a score for the entire image. Because individual cellsterm' responses are inhomogeneous [ 15 ], multiparametric single cell dataterm from several types of instruments have proven much more powerful than whole-population dataterm (for example, western blots or mRNA expression chips) for clustering genes, deriving causal networksterm, classifying protein localizationterm, and diagnosing disease [ 10 , 16 - 18 ]. In addition, individual cellterm measurementsterm can reveal samples that differ in only a subpopulationterm of cellsterm, which would otherwise be masked in whole-population measures. Fourth, quantitative image analysis is able to detect some features that are not readily detectable by a human observer. For example, the two-fold difference in DNA staining intensityterm that reveals whether a cellterm is in G1 or G2 phase of the cell cycleterm are measurable by computer but are difficult for the human eye to observe in cellterm images. Furthermore, small but biologically significant differences, for example, a 10% increase in nucleus sizeterm, are not noticeable by eye. Other features, for example, the textureterm (smoothness) of protein or DNA staining, are observable but not quantifiable by eye. Pathologists have known for years that changespersName in DNA or protein textureterm can correlate to profound and otherwise undetectable changespersName in cell physiologyterm, a fact used in diagnosis of disease [ 17 , 19 ]. Even changespersName not visible to the human eye can reveal disease stateterm [ 20 ]. Fifth, image cytometryterm is much less labor-intensive and higher-throughput. Appropriate software produces reliable results from a large-scale experimentterm in hours, versus months of tedious visual inspection. This improvement is more than an incremental technical advance, because it relieves the one remaining bottleneck to routinely conducting such experiments. Prior to the work presented here, the only flexible, open-source biological image analysis package was ImageJ/NIH Imageterm [ 21 ]. This package has been successfully used by many laboratories. Its design, however, is geared more towards the analysis of individual images (comparable to Adobe Photoshop) rather than flexible, high-throughput work. Macros can be written in ImageJ for high-throughput work but adapting macros to new projects requires that biologists learn a programming languageterm. While not creating a general, flexible software tool, many groups have benefited from automated cell image analysisterm by developing their own scripts, macros, and plug-ins to accomplish specific image analysis tasks. Custom programsterm written in commercial softwareterm (for example, MetaMorph, ImagePro Plus, MATLABterm) or Java have been used to identify, measure, and track cellsterm in images and time lapse movies [ 10 , 22 , 23 ]. Such studies clearly show the power of automated image analysis for biological discovery. However, most of these custom programsterm are not modular, so combining several steps and changing settings requires interacting directly with the code and is simply not practical for routinely processing hundreds of thousands of images or sending jobs to a cluster. The effort expended by laboratories in creating an analysis solution with a particular software package is often lost after the initial experiment is completed; other laboratories rarely use the methods because they are customized for a particular cell typeterm, assay or even image set. Furthermore, although developing a routine for a new cell typeterm or assay usually requires testing multiple algorithms, it is impractical to implement and test several published methods for a particular project. Commercial softwareterm has also been developed, mainly for the pharmaceutical screening market, by companies including Cellomics, TTP LabTech, Evotec, Molecular Devices, and GE Healthcare [ 24 ]. Development of these packages has been guided mainly by mammalian cell typesterm and cellular features of pharmaceutical interest, including protein translocationterm, micronucleus formation, neurite outgrowth, and cell countterm [ 25 ]. The high cost and the bundling of commercial softwareterm with hardware makes it impractical to test several programs for a new project. The proprietary nature of the code prevents researchers from knowing the strategy of a given algorithm and it cannot be modified if desired. As is the case with many laboratories, we have found commercial packages useful for some screens in mammalian cellsterm, but in other cases limiting [ 1 , 5 , 26 , 27 ]. Furthermore, key challenges remain in image analysis algorithm development itself [ 28 ]. Cell image analysisterm has been described as one of the greatest remaining challenges in screening [ 5 , 29 ], and as a field is "very much in its infancy" [ 30 ] and "lag [s] behind the adoption of high-throughput imaging technologies" [ 10 ]. Accurate cellterm identification is required to extract meaningful measures from images, but even for mammalian cell typesterm, existing software often fails on crowded cell samplesterm, which has severely limited screens thus far. Screens in most non-mammalian organisms have been limited to visual inspection. In summary, while existing software enables particular assays for particular cell typesterm, high throughput image analysis has, to this point, been impractical unless an image analysis expert develops a customized solution, or unless commercial packages are used with their built-in algorithms for a limited set of cellular features and for a limited set of cell typesterm. There exists a clear need for a powerful, flexible, open-source platformterm for high-throughput cell image analysisterm. Here we describe the open-source CellProfilerterm project, our effort to develop such a software system for the scientific community. CellProfilerterm simultaneously measures the size, shape, intensity and textureterm of a variety of cell typesterm in a high throughput manner. Note that we focus in this paper not on the technical details of the software (which are described in the manual), nor computational validation of the mostly published algorithms, nor on a mechanistic study of any particular biological finding. Rather, we describe the system, validate the software for a variety of real-world biological problemsterm, demonstrate the breadth of its utility (including on various cell typesterm and assays), and hope to stimulate ideas within the biological community for future applications of the software. Overview of the software system The following can be freely downloaded from the CellProfilerterm website [ 31 ]: CellProfilerterm for Windows, Mac, and Unix (compiled, not requiring MATLABterm); CellProfilerterm's MATLABterm source code; a full technical description of CellProfilerterm's algorithms and measurementsterm in an extensive PDF formatted manual (Additional data file 1), identical to the information found in help buttons within CellProfilerterm; and pipelinesterm to identify the various cell typesterm in this paper (see Additional data file 2 for a list of the modulesterm in each pipelineterm). CellProfilerterm is freely available modular image analysis softwareterm that is capable of handling hundreds of thousands of images. The software contains already-developed methods for many cell typesterm and assays and is also an open-source, flexible platform for the sharing, testing, and development of new methods by image analysis experts. CellProfilerterm meets the needs discussed in the introduction, in that it contains: advanced algorithms for image analysis that are able to accurately identify crowded cellsterm and non-mammalian cell typesterm; a modular, flexible design allowing analysis of new assays and phenotypes; open-source code so the underlying methodology is known and can be modified or improved by others; a user-friendly interface; the capability to make use of clusters of computers when available; and a design that eliminates the tedium of the many steps typically involved in image analysis, many of which are not easily transferable from one project to another (for example, image formatting, combining several image analysis steps, or repeating the analysis with slightly different parameters). CellProfilerterm was designed and optimized for the most common high-content screening image format, that is, two-dimensional images. It has very limited support for time-lapse and three-dimensional image stack analysis, although researchers interested in these areas could build compatible modulesterm. Most image analysis projects, even for new cell typesterm or assays, can be accomplished simply by pointing and clicking using CellProfilerterm's graphical user interfaceterm (Figure 1a ). The software uses the concept of a 'pipelineterm' of individual modulesterm (Figure 1b ; Additional data file 2). Each moduleterm processes the images in some manner, and the modulesterm are placed in sequential order to create a pipelineterm: usually image processingterm, then object identificationterm, then measurementterm. Over 50 CellProfilerterm modulesterm are currently available (Additional data file 3). Most modulesterm are automatic, but the software also allows interactive modulesterm (for example, the user clicks to outline a region of interest in each image). Modulesterm are mixed and matched for a specific project and each moduleterm's settings are adjusted appropriately. Upon starting the analysis, each image (or group of images if multiple wavelengths are available) travels through the pipelineterm and is processed by each moduleterm in order. CellProfilerterm overview and features. (a) Main CellProfilerterm interface, with an analysis pipelineterm displayed. (b) Schematic of a typical CellProfilerterm pipelineterm. (c) Image processingterm example: uneven illumination from the left to the right within each field of viewterm is noticeable in this three row by five column tiled image (left). CellProfilerterm's illumination correctionterm modulesterm correct these anomalies (right). Images were contrast-enhanced to display this effect. (d) These corrections reduce noise in quantitative measurementsterm, demonstrated here in DNA contentterm measures (middle) from images of Drosophila Kc167 cellsterm that are improved over the raw imagesterm (left). The results are comparable to those produced by white referenced images (right), but they do not require the error prone and often omitted step of collecting a white reference image immediately before image acquisition. (e) Outlines show the identification of nucleiterm and identification of cellterm edges made by CellProfilerterm in human HT29 (left) and Drosophila Kc167 (right) cellsterm. Cellsterm touching the border are intentionally excluded from analysis and images were contrast stretched for display. Scale bars = 15 μm. The pipelineterm's modulesterm and their settings are saved and can be used to reproduce the analysis or share with colleagues. Many example pipelinesterm are provided at the CellProfilerterm website [ 31 ] to provide a starting point for new analyses. To explain some features of CellProfilerterm, we describe in the subsequent sections the general steps in a typical pipelineterm. Image processingterm, including illumination correctionterm One of the most critical steps in image analysis is illumination correctionterm. Illumination often varies more than 1.5-fold across the field of viewterm, even when using fiber optic light sources, and occasionally even when images are thought to be already illumination-corrected by commercial image analysis software packages (TRJ, AEC, DMS, and PG, unpublished data). This adds an unacceptable level of noise, obscures real quantitative differences, and prevents many types of biological experiments that rely on accurate fluorescence intensity measurementsterm (for example, DNA contentterm of a nucleus, which only varies by two-fold during the cell cycleterm). CellProfilerterm contains standard methods plus our new methods [ 32 ] to address illumination variation, allowing various methods to be compared side by side and, ultimately, providing less noisy quantitative measuresterm (Figure 1c,d ). We use these illumination correctionterm methods for every high-throughput image set we process, because using raw imagesterm degrades intensity measurementsterm and, less obviously, can preclude accurate cellterm identification. This adversely affects all types of measurementsterm, from intensity based measures (for example, DNA contentterm histograms [ 1 ]) to area and shape measurementsterm (TRJ, AEC, DMS, and PG, unpublished data). CellProfilerterm's other image processingterm modulesterm perform other needed adjustments prior to identifying cellsterm in images, for example, aligning or cropping (Additional data file 3). Cellterm identification Object identificationterm (also called segmentationterm) is the most challenging step in image analysis and its accuracy determines the accuracy of the resulting cellterm measurementsterm. CellProfilerterm's object identificationterm modulesterm contain a variety of published and tested algorithms for identifying cellsterm based on fluorescence, including work from our own group and others (Figure 1e ). In most biological images, cellsterm touch each other, causing the simple, fast algorithms used in some commercial softwareterm packages to fail. The first objects identified in an image (called primary objectsterm) are often nucleiterm identified from DNA-stained images, although primary objectsterm can also be whole cellsterm, beads, speckles, tumors, and so on. Several simple algorithms are built into CellProfilerterm for cases where primary objectsterm are well-dispersed, non-confluent, and bright relative to the background. More importantly, to effectively identify clumped objectsterm, CellProfilerterm contains a modular three-step strategyterm based on previously published algorithms [ 33 - 37 ]. First, clumped objectsterm are recognized and separated; second, the dividing lines between objects are found; and third, some of the resulting objects are either removed or merged together based on their measurementsterm, for example, their size or shape. After primary objectsterm (often nucleiterm) are identified, the edges of secondary objectsterm that surround each primary objectterm (often cellterm edges) can be found more easily. Measuring cellterm size in Drosophila was not previously feasible because the commonly used watershed methodterm [ 37 ] often fails to find the borders between clumped cellsterm. We have, therefore, implemented in CellProfilerterm an improved Propagate algorithmterm [ 38 ], in addition to several standard methods of secondary objectterm identification. Other subcellular compartmentsterm can also be identified, including the cytoplasmterm (the part of each cellterm excluding the nucleus) and the cellterm or nuclear membraneterm (the edge of the cellterm or nucleus). The technical description of these algorithms is omitted here but is available in the online help and manual, in addition to previously published references cited therein (Additional data file 1). The identification modulesterm include a 'test modeterm' for comparing several algorithms side by side in order to choose the best approach. We have found that these cellterm identification methods are flexible to various cellterm morphologies. This flexibility is convenient but, more importantly, often allows accurate identification of cellsterm with unusual morphologies within a population of normal cellsterm. Measurementsterm and data analysis CellProfilerterm measures a large number of features for each identified cellterm or subcellular compartmentterm, including area, shape, intensity, and textureterm (each feature is described in Additional data file 4). This includes many standard features [ 39 , 40 ], but also complex measurementsterm like Zernike shape featuresterm [ 41 ], and Haralick and Gabor texture featuresterm [ 42 - 44 ], which are described in detail in the online help and manual. There are also modulesterm to measure various features (for example, intensity, textureterm, saturation, blur, area occupied by a stain) of an image in its entirety. A severe limitation of most commercial softwareterm is the inability to adapt to new biological questions by calculating new features from identified cellsterm [ 5 ]. By contrast, CellProfilerterm's modular design and open-source code allows quickly measuring new cellular phenotypes as needed. Measurementsterm are accessible in several ways: using CellProfilerterm's built-in viewing and plotting data tools (Additional data file 5); exporting in a tab-delimited spreadsheet formatterm that can be opened in programs like Microsoft Excel or OpenOffice Calc; exporting in a format that can be imported into a databaseterm like Oracle or MySQL; or directly in MATLABterm. Usability Like most new software in the laboratory, the process of setting up a CellProfilerterm analysis may take several days if the user is learning the software for the first time. Several resources help at this stage: the built-in help, the manual (Additional data file 1), the online discussion forum [ 31 ], the 'test modeterm' for the Identify modulesterm that show results from various options side by side, and built-in image and data tools to interact with processed images and cellterm measurementsterm (Additional data file 5). The flexible, modular design and point-and-click interface make setting up an analysis feasible for non-programmersterm. Over time, experienced users typically require less than a day to set up an entirely new experiment (for example, a new cell typeterm or unusual measurementterm scheme). When performing the same analysis on different image sets where sample preparation is the only variable, we test the analysis on a few sample images and sometimes change one or two settings in the Identify modulesterm. This takes less than an hour and is essentially a quality control stepterm. Once a pipelineterm is satisfactory, analysis can be performed on the local computer or automatically divided into smaller batches to be sent to a cluster of computers, described in more detail in later sections. CellProfilerterm's code is open-source under the GNU public license. Its image handling is flexible: there is no requirement for images to have a certain naming structure and many standard image formats plus some movie formats are supported. Its modular structure allows experts to expand the software to new file formats or add new algorithms. The source code was written in MATLABterm because it is a powerful, easy to learn language, commonly used for scientific applications, including prototyping image analysis routines. Because the source code is well-documented, it is understandable even by non-programmersterm. Computationally intensive tasks use either MATLABterm's native compiled functions or our own compiled C++ implementations to improve the speed. Analysis timesterm vary widely depending on the image size, the number of objects found per image, and the number of features measured, but typical pipelinesterm require from 20 seconds to five minutes per image on standard desktop computers. Validation of CellProfilerterm for many phenotypes We first demonstrated that CellProfilerterm's methods could accurately measure many different biologically important features of cellsterm using several cell typesterm, including Drosophila Kc167 cellsterm because these cellsterm are particularly challenging to identify by automated image analysis [ 5 , 27 ], and they enable rapid genome-wide screens using living cellterm microarrays [ 26 ]. Using the basic cell-culture methods described previously [ 26 ], we prepared Drosophila Kc167 cellsterm for experiments shown in Figures 1 and 2 by pretreating the cellsterm with double-stranded RNA (dsRNA) against the noted genes for two days prior to plating on plain glass slides and growing for a further 3 days in the presence of dsRNA. Specifically, 50 μg dsRNA plus 30 μl fugene in 1 ml serum-free medium were transfected into a 10 cm plate containing 20 million cellsterm in 10 ml medium. We prepared human HT29 cellsterm (Figures 1 , 2 , 3 ) as previously described [ 1 ]. Validation of CellProfilerterm for many cellular phenotypes. (a) Cell countterm: for a set of 6 images of wild-type human HT29 cellsterm (left), two researchers' counts varied by 11%, and CellProfilerterm's counts were within 6% of their average. For images of Drosophila Kc167 cellsterm with various genes knocked down by RNAi (right), the two researchers' counts varied by 16% and CellProfilerterm's counts were within 17% of their average. Example images and CellProfilerterm outlines for these cell typesterm are shown in Figure 1e. (b) Cellterm size: CellProfilerterm's cell area measurementsterm are comparable to those of a Coulter particle counterterm for Drosophila Kc167 cellsterm, for wild-type (no dsRNA) and RNA-interference induced samples. The SEM is too small to show error bars. (c) DNA contentterm in cellterm populations: measurementsterm are shown for human HT29 cellterm populations (1 image for each RNAi condition, left) and for Drosophila Kc167 cellterm populations (1,750 images for each RNAi condition were combined, right). The cell cycleterm distributions are as expected, with the 2N peak being predominant in the wild-type human sample, whereas most wild-type Drosophila nucleiterm are known to have 4N DNA contentterm [62]. RNAi-targeted samples were also as expected for Aurora kinase B (polyploid), Mad2 (fairly normal cell cycleterm distribution), String (4N-enriched), and Anillin and Cyclin A (both polyploid). (d) Chromatin content in time lapse movies: GFP-histone H4 ( S. cerevisiae ) or GFP-histone H2B (HeLa and C. elegans ) content is shown near each nucleus in arbitrary intensity units. The histone content is decreased by roughly half in each daughter nucleus after division. For C. elegans , only the boxed region of interest was analyzed. Scale bars: C. elegans , unknown; human HeLa = 20 μm; S. cerevisiae = 10 μm. (e) Phospho-proteinterm content: human HT29 cellsterm treated with RNAi reagent against Polo kinase have an increased percentage of nucleiterm with high phospho-H3 staining compared to wild-type cellsterm, consistent with a mitosis-stalled phenotype (left). Wild-type human HT29 nucleiterm that stain positively for phospho-histone H3 tend to be smaller than phospho-H3-negative cellsterm (right). (f) Protein localizationterm: the mean intensity of NFκB staining in the cytoplasmterm and the nucleus is shown in response to TNFα in human MCF7 cellsterm (top). Totals do not equal 100% due to slight overlap between compartments. (g) Speckles: fluorescent foci of phospho-Histone2AX induced by 2 Gy of irradiation in human U2OS cellsterm disappear at timepoints as the cellsterm recover. Scale bar = 10 μm. The SEM is too small to show error bars. Identifying mutant shapes and texturesterm. In each case, four images of each sample were quantitatively analyzed and images were adjusted using Adobe Photoshop auto levels for display only. Scale bars = 15 μm. (a) The unusual cell shapeterm induced by an RNAi reagent against Myo3A in human HT29 cellsterm is quantitatively distinguishable from wild-type control cellsterm. (b) The unusual cell shapeterm induced by an RNAi reagent against PTPN21 in human HT29 cellsterm is quantitatively distinguishable from wild-type control cellsterm. (c) The unusual actin textureterm induced by an RNAi reagent against DUSP19 in human HT29 cellsterm is quantitatively distinguishable from wild-type control cellsterm. The images are pseudocolored to show the actin staining textureterm. The biological basis of these morphological changespersName and the specificity of the RNAi reagents remain to be determined. Direct comparison of image analysis software is difficult because results from image analysis can be heavily skewed by how the software is tuned and commercial softwareterm packages are numerous and expensive. Furthermore, the algorithms in commercial softwareterm are proprietary and so cannot be directly compared apart from the entire software package, including preprocessing methods. The best practical comparison, therefore, is for software developers to release the results of their software on standard image sets or versus gold standards (visual inspection, Coulter particle countersterm, and so on). In subsequent sections, therefore, we present such comparisons. Note that once validated, any of these experiments could be expanded to a large-scale genome-wide RNAi screen or chemical library screen. Cell countterm (used to probe cellterm proliferation/apoptosis/death) is a straightforward phenotype that has, nonetheless, proved challenging for many cell typesterm due to the poor ability of existing software to separate clumped nucleiterm. For human cellsterm (Figure 2a , left), CellProfilerterm's accuracy compared to manual counting is twice that reported for a commercial softwareterm package [ 25 ]. CellProfilerterm also counted the more difficult-to-identify Drosophila Kc167 cellsterm (Figure 2a , right). Cellterm size was not previously measurable for many cell typesterm, but CellProfilerterm's measurementsterm were consistent with the gold standard, a Coulter particle size counter (Figure 2b ). While an automated routine has been developed for this cell typeterm [ 45 ], this is, to our knowledge, the first report on the quantitative accuracy of any software to count and measure cellterm size in Drosophila Kc167 cellsterm and the results indicate that such screens are now feasible. Cell countterm and size can at least be observed by eye even if quantitative high throughput screening is impractical. In contrast, certain phenotypes, like changespersName in DNA contentterm, are impossible to discern by eye. Unlike whole population-based methods, image cytometryterm measures individual cellterm fluorescence intensities so that the DNA contentterm of DNA-stained cellsterm can be determined [ 46 , 47 ]. These measurementsterm are very easily degraded by anomalies in the illumination of the field of viewterm and poor identification algorithms (the most common errors are counting two nearby nucleiterm as one nucleus with twice the DNA contentterm and incorrectly splitting a nucleus into two half-nuclei). This is, therefore, a very demanding phenotype to measure from images. Image analysis with CellProfilerterm produced the expected DNA contentterm distributions for both human and Drosophila cellterm populations (Figure 2c ). As another test, we confirmed that the green fluorescent protein (GFP)-histone content per cellterm decreases by roughly half when a cellterm divides into two daughter cellsterm during mitosis (Figure 2d ). Classification of cellsterm based on 2N, 4N, and 8N DNA contentterm is, therefore, possible based on an image of DNA-stained nucleiterm (Additional data file 6). This is useful not only in studies of cell cycleterm per se : cell cycle stageterm is a known cause of variability in biological samples, so analyzing a phenotype of interest with respect to the cell cycleterm eliminates a confounding variable (for example, the phenotype of interest could be assessed only in G1 phase cellsterm, which have 2N DNA contentterm). Further, image cytometryterm adds an additional level of information about cell cycleterm distribution. Whereas flow cytometryterm based on a DNA stain alone cannot distinguish cellsterm in G2 and M phase (both having the same 4N DNA contentterm), image cytometryterm reveals that these two populations differ in that mitotic cellsterm have smaller nucleiterm on average (Figure 2e , right). The total amount of a protein or phospho-proteinterm per cellterm can be measured by analysis of fluorescent antibody staining (Figure 2e , left), amounting to single-cell western blots. Furthermore, image cytometryterm can determine the localization of staining relative to other labeled cellular compartments. The change in localization of the nuclear factor (NF)κB transcription factor in response to tumor necrosis factor (TNF)α in MCF7 cellsterm can be monitored (Figure 2f ). We have previously used the software to confirm the localization of a protein predominantly at the membrane [ 48 ]. The software can also identify, count, and measure the shape, size, and intensity of subcellular structuresterm such as nuclear speckles (Figure 2g ). Finally, image analysis can probe other phenotypes that are not otherwise easily measured, such as shape and textureterm/smoothness. Cellterm morphology has not often been quantitatively measured, despite its importance in normal cellular physiology and in disease diagnosis [ 6 , 17 , 19 , 49 ]. Many of the shape and textureterm measurementsterm for wild-type cellsterm show non-Gaussian distributionsterm (Additional data file 7). Therefore, independently measuring every cellterm by image analysis is particularly valuable because the population cannot be accurately described by reduction to a few parameters like mean and standard deviation. We found that changespersName in cell shapeterm and actin textureterm induced by gene-specific RNAi were measurable (Figure 3 ), opening up the possibility for high-throughput screens for these and other morphologies. Cytological profiling to reveal pathways targeted by drugs Having demonstrated CellProfilerterm's ability to measure a large number of relevant phenotypes, we applied it to a publicly available dose-response image set of a Forkhead-EGFPterm cytoplasmterm to nucleus translocation assay in human cellsterm grown in multi-well plates (Figure 4a ). First, we ran a CellProfilerterm pipelineterm (Additional data file 2, part E) to calculate an illumination correctionterm function for each of the five slides and each of the two channels (<10 minutes processing time per slide on a single computer). We then used another CellProfilerterm pipelineterm (Additional data file 2, part A) to load each image, correct its illumination using the pre-calculated functions, identify nucleiterm, identify cellterm edges, and use the nucleus and cellterm outlines to define the cytoplasmic region of each cellterm, thereby defining three compartments for each cellterm: nucleus, cellterm, and cytoplasmterm. For each slide, we tested the pipelineterm on several random test images and fine-tuned settings in the identification modulesterm as needed. Modulesterm in the pipelineterm were included to measure: multiple features describing the area and shape of each compartment for each cellterm; multiple features describing the intensity and textureterm of each channel within each compartment, including several scales of textureterm; and the overall intensity, the percent saturation and the amount of blur for the entire image, for quality control purposes. The analysis was run on a desktop computer at a rate of >1 image/minute. CellProfilerterm analysis of a Forkhead (FOXO1A) cytoplasm-nucleus translocation assay. (a) Example images from the high throughput image set in human U2OS osteosarcoma cellsterm, showing no treatment (left) and 150 nM Wortmanninterm (right) after 1 hour treatment, scale unknown. (b) Translocation scored as the fraction of cellsterm whose ratio of GFP in the cytoplasmterm versus the nucleus was above a threshold. Error bars = SEM. (c) Statistical analysisterm using Z' and non-logistic-fit V factors, which are standard measures of assay qualityterm (>0.4 is considered screenable and 1 is an ideal assay) [63-65]. (d) Nucleiterm change shape in response to wortmanninterm but not LY294002term, as judged by three shape features. Error bars = SEM; * p < 0.05. The translocation was easily quantified by many features (Figure 4b ; Additional data file 8), the best of which achieved the highest published scores for assay qualityterm yet (Figure 4c ), indicating this software's improved ability to identify samples of interest in screens versus algorithms in commercial softwareterm [ 50 , 51 ]. Existing commercial softwareterm for this assay typically measures translocation only, but we wondered whether the broad spectrum of measurementsterm recorded by CellProfilerterm (the cytological profilesterm) could reveal further insights. We noticed that certain features of nuclear shape change in response to increasing doses of wortmanninterm but, interestingly, not the other positive control drug in this assay, LY294002term (Figure 4d ). These subtle changespersName were seen at all doses at and above the EC50 of wortmanninterm but at none of the doses of LY294002term, even those that clearly are sufficient for translocation of Forkhead, the main readout of this screen. Because wortmanninterm and LY294002term target an overlapping set of proteins, this result indicates that using this software in a primary screening assay would allow classification of any positively scoring samples as being wortmannin-specific or not, immediately narrowing down the potential pathways involved. While billions of samples have been scored using translocation assays, this is, to our knowledge, the first report of the ability to sub-classify samples based on morphological changespersName using primary screening data. Cluster computing We routinely run CellProfilerterm on large image sets (more than 45,000 four-color images) using a cluster of computers. To do this, we add modulesterm to the end of a pipelineterm to enable processing batches of images on the cluster and exporting data into a databaseterm. We then process the first image on a desktop computer, after which CellProfilerterm automatically divides the remainder of the large image set into groups and creates the files needed to submit each group to a computing clusterterm. We then use simple commands, outside of CellProfilerterm, to submit the jobs to our cluster of computers and export the resulting measurementsterm to a databaseterm. Each of these steps is described in the CellProfilerterm help for batch processing, and researchers without a computing clusterterm can now rent one remotely and inexpensively. Given that a typical image analysis takes approximately two minutes, a single CPU can process 30 images/hour and a 100 CPU cluster processes 3,000 images/hourterm. This is a much faster rate than existing image acquisition instruments, such that image analysis is not a bottleneck. Broad applicability Here we have shown that CellProfilerterm is useful for measuring a number of cellterm features, including cell countterm, cellterm size, cell cycleterm distribution, organelle number and size, cell shapeterm, textureterm, and the levels and localization of proteins and phospho-proteinsterm. Unlike previously existing software, CellProfilerterm is effective in a number of cell typesterm and organisms, such that new avenues of research in both standard and high-throughput biology laboratoriesterm can now be pursued. CellProfilerterm is already being used by laboratories worldwide studying a variety of biological processes in other cell typesterm and organisms, including Drosophila (S2R+ cellsterm, epithelial tissue), human (TOV21G, biopsied prostate gland tissue, adult mesenchymal stem cellsterm, H1299 lung carcinoma), mouse (NIH/3T3, neural precursor cellsterm derived from embryos, lung tissue sections, isolated germ cellsterm), and rat (H9c2 cellsterm) [ 1 , 26 , 48 , 52 - 54 ] (KA Hartwell, personal communication). We have also modified CellProfilerterm to measure yeast colonies, yeast growth patches, wounds in scratch assays, and tumors [ 55 ]. Importantly, the only successfully completed Drosophila screen using automated image analysis, to date, has been a cell-count/object-count screen in the S2 line whose appearance is comparable to human cellterm lines [ 56 ]. We are currently using CellProfilerterm to analyze screens using the clumpy Drosophila Kc167 cellterm type (AEC, TRJ, MRL, DB Wheeler, PG, DMS, unpublished data). Given the power of RNA interference and genetic tools in Drosophila and the demand for screening in its community [ 57 ], this is an area that can now move past tedious visual analysisterm of thousands of images, accelerating the rate of discovery. Future development We hope that computer vision researchersterm will contribute new algorithms to the project so that their theoretical work can be applied to practical biological problems. For example, while CellProfilerterm can currently analyze each slice of a time-lapse movie or three-dimensional image setterm independently, implementation of algorithms specifically designed to take advantage of the extra context information present in this type of data would be necessary for most experiments using these image types. Furthermore, CellProfilerterm is currently being integrated with the open-source Open Microscopy Environmentterm project (OME) [ 58 ], which would provide a complete open-source infrastructure for organizing and analyzing images from high-throughput experiments. With the successful application of sophisticated image analysis methods, the bottleneck of image-based genome-wide screens is now moving downstream to data visualizationterm, exploration, and statistical analysisterm in order to accommodate the number and richness of measurementsterm that result from image-based genome-wide assays [ 32 ]. Fully exploiting these rich data sets will reveal cellular signaling networksterm and lead to the unprecedented rich annotation of hundreds of genes in parallel. Additional data files The following additional data are available with the online version of this paper. Additional data file 1 is the CellProfilerterm manual. Additional data file 2 shows the CellProfilerterm pipelinesterm for experiments shown in this paper, listing the modulesterm in the order used. Additional data file 3 is a table listing CellProfilerterm modulesterm by category, with their descriptions. Additional data file 4 is a table listing the measurementsterm made by CellProfilerterm modulesterm. Additional data file 5 lists the data and image tools in CellProfilerterm, with their descriptions. Additional data file 6 is a figure showing an example from CellProfilerterm analysis of DNA contentterm (cell cycleterm) in Drosophila Kc167 cellsterm. Additional data file 7 is a figure showing histograms of shape and textureterm features for wild-type cellsterm. Additional data file 8 is a table listing measures for the cytoplasm-nucleus translocation assay (Figure 4 ) for which the Z' factor is above 0.5. Supplementary Material Acknowledgments Acknowledgements We gratefully thank the researchers providing images for this study: Steve N Bailey (Whitehead Institute, Figure 1c ), Scott Floyd (MIT, Figure 2g ), Kirsten Hagstrom ( C. elegans , Figure 2d [ 59 ]), Ruth Brack and Horst Wolff (GSF-Institute for Molecular Virology, Neuherberg, Germany, HeLa, Figure 2d [ 60 ]), Dominic Hoepfner, Arndt Brachat and Peter Philippsen (Universitat Basel, Switzerland, Saccharomyces cerevisiae , Figure 2d ), Ilya Ravkin (Vitra CNT, Figure 2f and BioImage CNT, Figure 4 [ 61 ]). We are grateful to Wayne Rasband for his work on the open-source NIH Image/ImageJ package and for advice from Zach E Perlman, both of whom served to inspire this project. We appreciate Ilya Ravkin's work to provide test images for the community; those who provided technical assistance, Dianne CarpenterpersName, Biao Luo, Nora Taylor, Susan Ma, and James Whittle; those who contributed to the software, Steve Lowe; and those who provided helpful comments, Doug Wheeler, Tamar Resnick and Kimberly Hartwell. This work was supported by a Merck/CSBi postdoctoral fellowship (AEC), a Novartis fellowship from the Life Sciences Research Foundation (AEC), a Society for Biomolecular Screening Academic grant (AEC), the MIT EECS/Whitehead/Broad Training Program in Computational Biology (NIH grant DK070069-01) supporting TRJ, a Damon Runyon Cancer Research Foundation fellowship (DAG), an NSERC postdoctoral fellowship (JM), DOD TSC research program grant W81XWH-05-1-0318-DS (DMS), NIH grant R01 GM072555-01 (DMS), and the Keck Foundation (DMS). Contributor Information Anne E CarpenterpersName, Email: carpenterpersName@wi.mit.edu. Thouis R JonespersName, Email: thouis@csail.mit.edu. Michael R LamprechtpersName, Email: mrl@wi.mit.edu. Colin ClarkepersName, Email: colinc@mit.edu. In Han KangpersName, Email: ikang@mit.edu. Ola FrimanpersName, Email: olafr@imt.liu.se. David A GuertinpersName, Email: guertinpersName@wi.mit.edu. Joo Han ChangpersName, Email: joohan.changpersName@gmail.com. Robert A LindquistpersName, Email: rlindqui@wi.mit.edu. Jason MoffatpersName, Email: moffatpersName@wi.mit.edu. Polina GollandpersName, Email: polina@csail.mit.edu. David M SabatinipersName, Email: sabatinipersName@wi.mit.edu. References Associated Data Supplementary Materials