Murat Cihan SorkunpersName, Abhishek KhetanpersName, Süleyman ErpersName (2019date). “AqSolDB, a curated reference set of aqueous solubility and 2D descriptors for a diverse set of compounds.” Scientific Datatitle·j.
Object of study — AqSolDBterm, its experimental aqueous-solubility values, compound representations, reliability labelsterm, and calculated two-dimensional descriptors
Research question — how to merge disparate public solubility datasets into a standardized, validated, nonredundant resource with statistically selected values and useful descriptors
Contribution — an openly accessible curated reference dataset of 9,982 compounds plus a reproducible algorithm for identifier validation, unit standardization, competing-value selection, reliabilityterm grouping, and descriptor enrichment
phenomenon · claim-type · cross-domain-transfer · mechanism · claim-relation · replication-status · limitation — grounded in the full text| phenomenon | variation and disagreement in experimental aqueous-solubility measurements across compounds, sources, methods, and conditions — “Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data.” | |
| claim-type | existence — “AqSolDB consists of 9,982 unique compounds. AqSolDB data is stored in the comma-separated values (CSV) format and contains representations, experimental aqueous solubility and calculated 2D descriptor data of all compounds, as described in Table 2 . AqSolDB is openly accessible at the Harvard Dataverse Repository 24 .” | |
| method — evidence | — “To curate our dataset we followed three steps. First, we collected nine publicly available aqueous solubility dataset and converted them into a standardized format. Second, we combined datasets into one single dataset by applying a data verification algorithm that selects statistically most reliable experimental value among multiple occurrences. Finally, we added topological and physico-chemical 2D descriptors to the merged dataset.” | |
| object of study — evidence | — “AqSolDB consists of aqueous solubility values of 9,982 unique compounds, along with some relevant topological and physico-chemical 2D descriptors. Additionally, the dataset contains validated representations of each of the compounds.” | |
| research question — evidence | — “The aim of this study is to curate a large experimental aqueous solubility data, AqSolDB, for data-driven model development.” | |
| contribution — evidence | — “AqSolDB is an openly accessible, easy-to-use, and well-structured database of compound. We expect it to serve a broad community as a reference aqueous solubility dataset for the bench-marking of new experimental and physics-based modelling results, and additionally as machine-readable ancillary resource to improve the prediction capability of future machine learning approaches.” | |
| mechanism | three-stage curation workflow Public datasets are standardized, merged with a reliability-selection algorithm, and enriched with two-dimensional descriptors. — “To curate our dataset we followed three steps. First, we collected nine publicly available aqueous solubility dataset and converted them into a standardized format. Second, we combined datasets into one single dataset by applying a data verification algorithm that selects statistically most reliable experimental value among multiple occurrences. Finally, we added topological and physico-chemical 2D descriptors to the merged dataset.” | mechanisms[0] |
| mechanism | identifier generation Missing SMILES values are retrieved from names and CAS numbers, while SLN identifiers are converted with RDKit. — “In external datasets, where SMILES representations were not available, we used the name and the CAS Registry Number of compounds as inputs to retrieve the SMILES strings from the Chemical Identifier Resolver web service of the National Cancer Institute ( https://cactus.nci.nih.gov/chemical/structure ). Lastly, SLN identifiers available from some datasets were converted to SMILES using RDKit open-source cheminformatics software.” | mechanisms[1] |
| mechanism | unit standardization Source measurements in mass-per-volume units are converted to molar LogS using molecular mass. — “Unit Conversion: The chosen unit of solubility in this dataset is LogS, where S is the aqueous solubility in mol/L (or M). Units such as g/L and mg/L were converted to LogS using the molecular mass of the compounds.” | mechanisms[2] |
| mechanism | round-trip identifier validation SMILES are converted to RDKit molecules and InChI, regenerated through the same representations, and checked for consistency and common molecular identity. — “First, SMILES strings were converted to into RDKit mol objects. If an error occurred during the conversion, the input SMILES string was considered to be invalid. Next, the obtained RDKit mol objects were converted to InChI representations. The InChI representations were used to regenerate the RDKit mol objects. Finally, the thus obtained mol objects were converted back to InChI representations. The original and regenerated InChI were checked for consistency to ensure that the generated InChI were reproducible. This step also validated that both SMILES and InChI representations led to the same RDKit mol object, and thus the chemical compound.” | mechanisms[3] |
| mechanism | redundancy classification Repeated compounds are separated into same-value and different-value redundancies using InChI identity and a 0.01-LogS equivalence tolerance. — “The purpose of this step is combining all datasets into the one single repository that contains only unique compounds paired with the most reliable aqueous solubility value. The InChI representation was used to identify compounds uniquely and solubility values within 0.01 LogS units of each other were deemed to be identical. Based on these conditions, a preliminary analysis of the combined repository revealed two different kinds of redundancies - (1) a given compound was found to repeat with a different solubility value, or (2) a given compound was found to repeat with the same solubility value.” | mechanisms[4] |
| mechanism | closest-to-mean selection For compounds with more than one differing measurement, the reported value closest to their mean is selected and dispersion assigns G4 or G5. — “Next, for compounds with occurrence count >1 with different solubility values (819), we used the closest to the mean algorithm to select the value. In this method, the mean value is first calculated then the closest value to mean value among the candidates is selected. If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | mechanisms[5] |
| mechanism | closest-to-reference selection For exactly two competing measurements, the value nearest an external ALOGPS prediction is selected and dispersion assigns G2 or G3. — “For compounds with exactly 2 values (1,417), the closest to the mean method cannot be applied because mean is always at the middle of the two values. For this case, we used an alternative method, which is closest to the reference. We selected the closest value to an external reference value, which is obtained using the solubility prediction tool ALOGPS 21 . ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 . If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | mechanisms[6] |
| mechanism | reliability grouping Occurrence count and a 0.5-LogS standard-deviation threshold divide selected observations into five reliability groups. — “We selected 0.5 as a threshold for degree of agreement between multiple values based on the predictive capabilities of some the state-of-the-art models 3 , 5 , 6 . It must also be noted that the average SD of experimental solubility values for a given compound from different sources has been reported to be 0.5 LogS 7 , 23 . Using this threshold, the grouping of compound into 5 different groups provides a credible way of assessing reliability for data-driven modeling.” | mechanisms[7] |
| mechanism | descriptor enrichment Relevant RDKit topological and physicochemical descriptors extend the dataset's information space for direct machine-learning use. — “The purpose of this step is extending the information space of compounds by adding basic topological and physico-chemical information. For this purpose, we calculated all the relevant 2D descriptors available from RDKit.” | mechanisms[8] |
| builds-on | nine publicly available aqueous-solubility datasets — “Solubility data was first collected from nine publicly available datasets as shown in Table 1 . A set of three pre-processing steps were applied to each of the datasets in order to standardize the representation format and solubility values in the same units.” | claim_relations[0] |
| contrasts-with | disparate non-standardized source datasets — “Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data. To address this problem, we generated a new database of compounds, AqSolDB, by merging a total of nine different aqueous solubility datasets, curating the merged data, standardizing and validating the compound representation formats, marking with reliability labels, and providing 2D descriptors of compounds as a Supplementary Resource.” | claim_relations[1] |
| qualifies | the reliability of directly selected G1 observations — “The G1 group constitutes the largest part of the data and has been encountered only once in all datasets. These compounds are selected directly and it is not possible to comment on their reliability.” | claim_relations[2] |
| sense | AqSolDB as an open reference database for aqueous solubility — “AqSolDB is an openly accessible, easy-to-use, and well-structured database of compound. We expect it to serve a broad community as a reference aqueous solubility dataset for the bench-marking of new experimental and physics-based modelling results, and additionally as machine-readable ancillary resource to improve the prediction capability of future machine learning approaches.” | senses[0] |
| sense | aqueous solubility as maximum dissolved amount under physical conditions — “It is defined as the maximum amount of a compound, i.e., the solute, that can get dissolved in a given volume of water, and depends on physical conditions such as temperature and pressure.” | senses[1] |
| sense | dataset quality as generalizability and fidelity — “In this work, we assess the quality of aqueous solubility datasets under 2 categories: generalizability and fidelity.” | senses[2] |
| sense | generalizability as chemical diversity and dataset size — “Generalizability can be interpreted in terms of the chemical diversity of the dataset, as well as its size.” | senses[3] |
| sense | fidelity as experimental and recording accuracy — “Fidelity can be understood as accuracy of data in terms of the reliability of the experimental technique, human errors in either conducting the experiments and recording the measured values.” | senses[4] |
| sense | LogS as standardized molar aqueous solubility — “Unit Conversion: The chosen unit of solubility in this dataset is LogS, where S is the aqueous solubility in mol/L (or M).” | senses[5] |
| sense | reliability as a joint function of replication and dispersion — “The difference between the results of the independent experiments shows the reliability of this value. Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | senses[6] |
| sense | two-dimensional descriptors as ready-to-use machine-learning inputs — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | senses[7] |
| note | nine publicly available aqueous-solubility datasets — AqSolDB is constructed by curating and merging these prior data resources. — “Solubility data was first collected from nine publicly available datasets as shown in Table 1 . A set of three pre-processing steps were applied to each of the datasets in order to standardize the representation format and solubility values in the same units.” | claim_relations[0] |
| weight | nine publicly available aqueous-solubility datasets — 0.95 — “Solubility data was first collected from nine publicly available datasets as shown in Table 1 . A set of three pre-processing steps were applied to each of the datasets in order to standardize the representation format and solubility values in the same units.” | claim_relations[0] |
| note | disparate non-standardized source datasets — The standardized curated resource is motivated in contrast to source heterogeneity. — “Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data. To address this problem, we generated a new database of compounds, AqSolDB, by merging a total of nine different aqueous solubility datasets, curating the merged data, standardizing and validating the compound representation formats, marking with reliability labels, and providing 2D descriptors of compounds as a Supplementary Resource.” | claim_relations[1] |
| weight | disparate non-standardized source datasets — 0.95 — “Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data. To address this problem, we generated a new database of compounds, AqSolDB, by merging a total of nine different aqueous solubility datasets, curating the merged data, standardizing and validating the compound representation formats, marking with reliability labels, and providing 2D descriptors of compounds as a Supplementary Resource.” | claim_relations[1] |
| note | the reliability of directly selected G1 observations — The authors categorically withhold a reliability interpretation for single-observation G1 values. — “The G1 group constitutes the largest part of the data and has been encountered only once in all datasets. These compounds are selected directly and it is not possible to comment on their reliability.” | claim_relations[2] |
| note | ALOGPS —is-a→ open source online solubility prediction tool — The source makes an explicit categorical class statement. — “ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 .” | concept_edges[0] |
| relation | ALOGPS —is-a→ open source online solubility prediction tool — ALOGPS is an open-source online solubility-prediction tool. — “ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 .” | concept_edges[0] |
| audit | poor aqueous solubility —causes→ high risk of erroneous results — Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span says poor solubility is LIKELY to lead to precipitation, which MAY create a high risk of erroneous results; hedged. — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concept_edges[1] |
| note | poor aqueous solubility —causes→ high risk of erroneous results — Likely and may explicitly hedge the causal chain. — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concept_edges[1] |
| relation | poor aqueous solubility —causes→ high risk of erroneous results — Poor aqueous solubility may lead through precipitation to a high risk of erroneous screening results. — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concept_edges[1] |
| weight | poor aqueous solubility —causes→ high risk of erroneous results — 0.65 — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concept_edges[1] |
| effect direction | sampling —causes→ reliability — increases — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” — “when sampling increases, reliability will increase” | concept_edges[2] |
| note | sampling —causes→ reliability — The source makes a flat increase claim, so the gradable weight is high but not categorical. — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | concept_edges[2] |
| relation | sampling —causes→ reliability — Increased sampling increases reliability in the authors' statistical rationale. — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | concept_edges[2] |
| weight | sampling —causes→ reliability — 0.95 — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | concept_edges[2] |
| relation | experimental calibration data → accurate data-driven estimation models — Accurate data-driven models require high-fidelity, high-quality calibration data. — “To construct accurate data-driven estimation models, it is essential that the underlying experimental calibration data used by these models is of high fidelity and quality.” | unhoused_relations[0] |
| relation | uncertainties and disagreements in the underlying data → development of reliable data-driven models — Data uncertainty and disagreement hinder reliable model development. — “The development of reliable data-driven models, however, has been hindered by uncertainties and disagreements in the underlying data, which are obtained from many disparate sources.” | unhoused_relations[1] |
| weight | uncertainties and disagreements in the underlying data → development of reliable data-driven models — 0.95 — “The development of reliable data-driven models, however, has been hindered by uncertainties and disagreements in the underlying data, which are obtained from many disparate sources.” | unhoused_relations[1] |
| relation | SMILES representation → standard identifier — SMILES is chosen as the standard compound identifier. — “Identifier generation: We chose the SMILES representation as the standard identifier for compounds for our curated dataset.” | unhoused_relations[2] |
| relation | SMILES and InChI representations → same RDKit mol object — SMILES and InChI are validated by convergence on the same RDKit molecule object. — “This step also validated that both SMILES and InChI representations led to the same RDKit mol object, and thus the chemical compound.” | unhoused_relations[3] |
| relation | small size and lack chemical diversity → poor predictive capability on external test sets — Small and chemically narrow training datasets are associated with poor external-test prediction. — “Machine learning models developed using datasets, which have small size and lack chemical diversity, show poor predictive capability on external test sets, as shown in the study by Wang et al . 8 .” | unhoused_relations[4] |
| weight | small size and lack chemical diversity → poor predictive capability on external test sets — 0.9 — “Machine learning models developed using datasets, which have small size and lack chemical diversity, show poor predictive capability on external test sets, as shown in the study by Wang et al . 8 .” | unhoused_relations[4] |
| relation | calculated 2D descriptors → direct use for developing machine learning models — Calculated descriptors enable direct machine-learning use. — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | unhoused_relations[5] |
| weight | calculated 2D descriptors → direct use for developing machine learning models — 0.95 — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | unhoused_relations[5] |
| relation | AqSolDB data → comma-separated values format — AqSolDB is stored in CSV format. — “AqSolDB data is stored in the comma-separated values (CSV) format and contains representations, experimental aqueous solubility and calculated 2D descriptor data of all compounds, as described in Table 2 .” | unhoused_relations[6] |
| relation | AqSolDB → Harvard Dataverse Repository — AqSolDB is openly accessible through Harvard Dataverse. — “AqSolDB is openly accessible at the Harvard Dataverse Repository 24 .” | unhoused_relations[7] |
| relation | highly soluble → Compounds with 0 and higher solubility value — Highly soluble corresponds to LogS zero or above. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[8] |
| relation | soluble → those in the range of 0 to −2 — Soluble corresponds to LogS from zero to minus two. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[9] |
| relation | slightly soluble → those in the range of −2 to −4 — Slightly soluble corresponds to LogS from minus two to minus four. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[10] |
| relation | insoluble → less than −4 — Insoluble corresponds to LogS below minus four. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[11] |
| no type for this | experimental calibration data → accurate data-driven estimation models — Accurate models require experimental calibration data with high fidelity and quality. — “To construct accurate data-driven estimation models, it is essential that the underlying experimental calibration data used by these models is of high fidelity and quality.” | unhoused_relations[0] |
| no type for this | uncertainties and disagreements in the underlying data → development of reliable data-driven models — Underlying data uncertainty and disagreement hinder reliable model development. — “The development of reliable data-driven models, however, has been hindered by uncertainties and disagreements in the underlying data, which are obtained from many disparate sources.” | unhoused_relations[1] |
| no type for this | SMILES representation → standard identifier — The curation workflow assigns SMILES the functional role of standard compound identifier. — “Identifier generation: We chose the SMILES representation as the standard identifier for compounds for our curated dataset.” | unhoused_relations[2] |
| no type for this | SMILES and InChI representations → same RDKit mol object — Round-trip validation checks that SMILES and InChI resolve to the same molecular object. — “This step also validated that both SMILES and InChI representations led to the same RDKit mol object, and thus the chemical compound.” | unhoused_relations[3] |
| no type for this | small size and lack chemical diversity → poor predictive capability on external test sets — Models trained on small, chemically narrow datasets generalize poorly to external tests. — “Machine learning models developed using datasets, which have small size and lack chemical diversity, show poor predictive capability on external test sets, as shown in the study by Wang et al . 8 .” | unhoused_relations[4] |
| no type for this | calculated 2D descriptors → direct use for developing machine learning models — The supplied descriptors enable direct use of AqSolDB in machine-learning model development. — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | unhoused_relations[5] |
| no type for this | AqSolDB data → comma-separated values format — AqSolDB data are serialized and distributed in CSV format. — “AqSolDB data is stored in the comma-separated values (CSV) format and contains representations, experimental aqueous solubility and calculated 2D descriptor data of all compounds, as described in Table 2 .” | unhoused_relations[6] |
| no type for this | AqSolDB → Harvard Dataverse Repository — AqSolDB is openly hosted by Harvard Dataverse. — “AqSolDB is openly accessible at the Harvard Dataverse Repository 24 .” | unhoused_relations[7] |
| no type for this | highly soluble → Compounds with 0 and higher solubility value — The highly-soluble class is assigned to compounds with LogS zero or above. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[8] |
| no type for this | soluble → those in the range of 0 to −2 — The soluble class is assigned to compounds in the stated LogS range. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[9] |
| no type for this | slightly soluble → those in the range of −2 to −4 — The slightly-soluble class is assigned to compounds in the stated lower LogS range. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[10] |
| no type for this | insoluble → less than −4 — The insoluble class is assigned below the stated LogS boundary. — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | unhoused_relations[11] |
| limitation | Aqueous-solubility measurement and prediction remain complex, and existing datasets vary in coverage, method, conditions, representation, size, and access. — “Measurement and prediction of aqueous solubility is a complex and prevailing challenge in chemistry. For the latter, different data-driven prediction models have recently been developed to augment the physics-based modeling approaches. To construct accurate data-driven estimation models, it is essential that the underlying experimental calibration data used by these models is of high fidelity and quality. Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data.” | limitations[0] |
| limitation | Reliable data-driven model development is hindered by uncertainty and disagreement across disparate source measurements. — “The development of reliable data-driven models, however, has been hindered by uncertainties and disagreements in the underlying data, which are obtained from many disparate sources.” | limitations[1] |
| limitation | Unsystematic differences between experimental methodologies can limit model-training and validation accuracy. — “Unsystematic errors between different experimental methodologies potentially limit the accuracy with which the models can be trained and validated.” | limitations[2] |
| limitation | The source datasets did not provide experimental-error information, requiring statistical rather than direct error-based reliability estimates. — “As no information about experimental errors from the original data sources was found to be available, we determined their reliability with a statistical approach.” | limitations[3] |
| limitation | G1 compounds appear only once, so the authors state that their reliability cannot be assessed. — “The G1 group constitutes the largest part of the data and has been encountered only once in all datasets. These compounds are selected directly and it is not possible to comment on their reliability.” | limitations[4] |
| limitation | Exactly two observations cannot use closest-to-mean selection because their mean is always midway between them, requiring an external model-based reference. — “For compounds with exactly 2 values (1,417), the closest to the mean method cannot be applied because mean is always at the middle of the two values. For this case, we used an alternative method, which is closest to the reference. We selected the closest value to an external reference value, which is obtained using the solubility prediction tool ALOGPS 21 .” | limitations[5] |
finding · quantity · variable · definition · key-term — grounded in the full text| finding | AqSolDB combines nine solubility datasets with standardization, representation validation, reliability labels, and calculated two-dimensional descriptors. — “To address this problem, we generated a new database of compounds, AqSolDB, by merging a total of nine different aqueous solubility datasets, curating the merged data, standardizing and validating the compound representation formats, marking with reliability labels, and providing 2D descriptors of compounds as a Supplementary Resource.” | content_tags.findings[0] |
| finding | The final resource contains aqueous-solubility values and descriptors for 9,982 unique compounds. — “AqSolDB consists of aqueous solubility values of 9,982 unique compounds, along with some relevant topological and physico-chemical 2D descriptors. Additionally, the dataset contains validated representations of each of the compounds.” | content_tags.findings[1] |
| finding | Before redundant values were eliminated, the merged repository held 19,796 instances for 9,982 unique compounds. — “There are a total of 19,796 instances in the merged repository with 9,982 unique compounds before redundant values are eliminated.” | content_tags.findings[2] |
| finding | Direct acceptance of unique or near-identical observations curated 7,746 G1 instances. — “For every compound in the dataset, the number of occurrences is determined. If the compound has a unique value or multiple values which are within 0.01 LogS units, the value is simply accepted. This step leads to the curation of 7,746 unique instances, which were assigned to group G1.” | content_tags.findings[3] |
| finding | For 819 compounds with multiple differing values, closest-to-mean selection produced 183 G4 and 636 G5 records according to the 0.5-LogS dispersion cutoff. — “Next, for compounds with occurrence count >1 with different solubility values (819), we used the closest to the mean algorithm to select the value. In this method, the mean value is first calculated then the closest value to mean value among the candidates is selected. If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | content_tags.findings[4] |
| finding | For 1,417 compounds with exactly two values, closest-to-reference selection produced 235 G2 and 1,182 G3 records. — “For compounds with exactly 2 values (1,417), the closest to the mean method cannot be applied because mean is always at the middle of the two values. For this case, we used an alternative method, which is closest to the reference. We selected the closest value to an external reference value, which is obtained using the solubility prediction tool ALOGPS 21 . ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 . If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | content_tags.findings[5] |
| finding | The dataset is distributed openly as CSV through Harvard Dataverse. — “AqSolDB data is stored in the comma-separated values (CSV) format and contains representations, experimental aqueous solubility and calculated 2D descriptor data of all compounds, as described in Table 2 . AqSolDB is openly accessible at the Harvard Dataverse Repository 24 .” | content_tags.findings[6] |
| finding | The authors infer that G5 is more reliable than G3 and G4 more reliable than G2 because reliability rises with repeated sampling. — “The difference between the results of the independent experiments shows the reliability of this value. Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | content_tags.findings[7] |
| finding | Calculated descriptors allow AqSolDB to be used directly for developing machine-learning models. — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | content_tags.findings[8] |
| finding | Both SMILES and InChI representations were validated and can be converted into RDKit molecule objects. — “We provided both SMILES and InChI representations of compounds which are validated and can be easily converted into the RDKit mol object.” | content_tags.findings[9] |
| quantity | nine — public aqueous-solubility datasets collected and merged — “For this purpose, we searched for and collected nine open source datasets on aqueous solubility.” | content_tags.quantities[0] |
| quantity | 9,982 — unique compounds in AqSolDB — “AqSolDB consists of 9,982 unique compounds.” | content_tags.quantities[1] |
| quantity | 19,796 — instances in the merged repository before redundant values were eliminated — “There are a total of 19,796 instances in the merged repository with 9,982 unique compounds before redundant values are eliminated.” | content_tags.quantities[2] |
| quantity | 0.01 LogS units — maximum separation at which multiple solubility values are treated as identical — “The InChI representation was used to identify compounds uniquely and solubility values within 0.01 LogS units of each other were deemed to be identical.” | content_tags.quantities[3] |
| quantity | 7,746 — unique accepted instances assigned to G1 — “This step leads to the curation of 7,746 unique instances, which were assigned to group G1.” | content_tags.quantities[4] |
| quantity | 819 — compounds with more than one differing solubility value processed by closest-to-mean selection — “Next, for compounds with occurrence count >1 with different solubility values (819), we used the closest to the mean algorithm to select the value.” | content_tags.quantities[5] |
| quantity | 0.5 LogS units — standard-deviation threshold dividing higher- and lower-agreement reliability groups — “We selected 0.5 as a threshold for degree of agreement between multiple values based on the predictive capabilities of some the state-of-the-art models 3 , 5 , 6 .” | content_tags.quantities[6] |
| quantity | 183 — compounds assigned to G4 — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | content_tags.quantities[7] |
| quantity | 636 — compounds assigned to G5 — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | content_tags.quantities[8] |
| quantity | 1,417 — compounds with exactly two competing solubility values — “For compounds with exactly 2 values (1,417), the closest to the mean method cannot be applied because mean is always at the middle of the two values.” | content_tags.quantities[9] |
| quantity | 0.49 RMSE — reported overall ALOGPS error in LogS units — “ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 .” | content_tags.quantities[10] |
| quantity | 235 — two-observation compounds assigned to G2 — “If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | content_tags.quantities[11] |
| quantity | 1,182 — two-observation compounds assigned to G3 — “If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | content_tags.quantities[12] |
| quantity | 17 — descriptor rows listed at the end of the dataset-information table — “The last 17 rows of Table 2 show the name, description and data type of the 2D descriptors.” | content_tags.quantities[13] |
| quantity | 0, −2, and −4 LogS — reported boundaries between highly soluble, soluble, slightly soluble, and insoluble categories — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | content_tags.quantities[14] |
| variable | aqueous solubility (experimental target property standardized and selected for each unique compound) — “The chosen unit of solubility in this dataset is LogS, where S is the aqueous solubility in mol/L (or M).” | content_tags.variables[0] |
| variable | SMILES representation (standard compound identifier and input to representation validation) — “Identifier generation: We chose the SMILES representation as the standard identifier for compounds for our curated dataset.” | content_tags.variables[1] |
| variable | InChI representation (identifier used for unique matching and SMILES round-trip validation) — “The InChI representation was used to identify compounds uniquely and solubility values within 0.01 LogS units of each other were deemed to be identical.” | content_tags.variables[2] |
| variable | molecular mass (conversion factor from mass-per-volume solubility units to molar LogS) — “Units such as g/L and mg/L were converted to LogS using the molecular mass of the compounds.” | content_tags.variables[3] |
| variable | occurrence count (decision variable selecting among direct acceptance, closest-to-mean, and closest-to-reference branches) — “For every compound in the dataset, the number of occurrences is determined.” | content_tags.variables[4] |
| variable | standard deviation (agreement statistic used with the 0.5-LogS threshold to assign reliability groups) — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | content_tags.variables[5] |
| variable | reference value (external model estimate used to choose between exactly two competing observations) — “We selected the closest value to an external reference value, which is obtained using the solubility prediction tool ALOGPS 21 .” | content_tags.variables[6] |
| variable | reliability group (label indicating selection path, occurrence evidence, and cross-source agreement) — “It is recommended for users to consider the group description when using the data as input to other models.” | content_tags.variables[7] |
| variable | 2D descriptors (calculated topological and physicochemical predictors supplied for modeling) — “The purpose of this step is extending the information space of compounds by adding basic topological and physico-chemical information. For this purpose, we calculated all the relevant 2D descriptors available from RDKit.” | content_tags.variables[8] |
| variable | chemical diversity (dimensions through which the authors interpret generalizability) — “Generalizability can be interpreted in terms of the chemical diversity of the dataset, as well as its size.” | content_tags.variables[9] |
| definition | aqueous solubility — “It is defined as the maximum amount of a compound, i.e., the solute, that can get dissolved in a given volume of water, and depends on physical conditions such as temperature and pressure.” | content_tags.definitions[0] |
| definition | generalizability — “Generalizability can be interpreted in terms of the chemical diversity of the dataset, as well as its size.” | content_tags.definitions[1] |
| definition | fidelity — “Fidelity can be understood as accuracy of data in terms of the reliability of the experimental technique, human errors in either conducting the experiments and recording the measured values.” | content_tags.definitions[2] |
| definition | LogS — “Unit Conversion: The chosen unit of solubility in this dataset is LogS, where S is the aqueous solubility in mol/L (or M).” | content_tags.definitions[3] |
| definition | same solubility value — “The InChI representation was used to identify compounds uniquely and solubility values within 0.01 LogS units of each other were deemed to be identical.” | content_tags.definitions[4] |
| definition | closest to the mean algorithm — “In this method, the mean value is first calculated then the closest value to mean value among the candidates is selected.” | content_tags.definitions[5] |
| definition | closest to the reference — “For this case, we used an alternative method, which is closest to the reference. We selected the closest value to an external reference value, which is obtained using the solubility prediction tool ALOGPS 21 .” | content_tags.definitions[6] |
| concept | AqSolDB — not defined in this paper's own words; the curated aqueous-solubility reference database — “AqSolDB is an openly accessible, easy-to-use, and well-structured database of compound.” | concepts[0] |
| concept | aqueous solubility — not defined in this paper's own words; the maximum amount of a compound dissolvable in a given water volume under specified conditions — “It is defined as the maximum amount of a compound, i.e., the solute, that can get dissolved in a given volume of water, and depends on physical conditions such as temperature and pressure.” | concepts[1] |
| concept | generalizability — not defined in this paper's own words; dataset quality in terms of chemical diversity and size — “Generalizability can be interpreted in terms of the chemical diversity of the dataset, as well as its size.” | concepts[2] |
| concept | fidelity — not defined in this paper's own words; dataset accuracy with respect to experimental reliability and human error — “Fidelity can be understood as accuracy of data in terms of the reliability of the experimental technique, human errors in either conducting the experiments and recording the measured values.” | concepts[3] |
| concept | chemical diversity — not defined in this paper's own words; breadth of chemical structures represented in a dataset — “Machine learning models developed using datasets, which have small size and lack chemical diversity, show poor predictive capability on external test sets, as shown in the study by Wang et al . 8 .” | concepts[4] |
| concept | data verification algorithm — not defined in this paper's own words; the rule set selecting statistically reliable values among repeated observations — “Second, we combined datasets into one single dataset by applying a data verification algorithm that selects statistically most reliable experimental value among multiple occurrences.” | concepts[5] |
| concept | SMILES representation — not defined in this paper's own words; the standardized compound identifier selected for AqSolDB — “Identifier generation: We chose the SMILES representation as the standard identifier for compounds for our curated dataset.” | concepts[6] |
| concept | standard identifier — not defined in this paper's own words; the normalized representation used to identify compounds in the curated dataset — “Identifier generation: We chose the SMILES representation as the standard identifier for compounds for our curated dataset.” | concepts[7] |
| concept | Chemical Identifier Resolver — not defined in this paper's own words; the NCI service used to retrieve missing SMILES from names and CAS numbers — “In external datasets, where SMILES representations were not available, we used the name and the CAS Registry Number of compounds as inputs to retrieve the SMILES strings from the Chemical Identifier Resolver web service of the National Cancer Institute ( https://cactus.nci.nih.gov/chemical/structure ).” | concepts[8] |
| concept | RDKit — not defined in this paper's own words; open-source cheminformatics software used for representation conversion and descriptor calculation — “Lastly, SLN identifiers available from some datasets were converted to SMILES using RDKit open-source cheminformatics software.” | concepts[9] |
| concept | LogS — not defined in this paper's own words; the standardized aqueous-solubility unit used by AqSolDB — “Unit Conversion: The chosen unit of solubility in this dataset is LogS, where S is the aqueous solubility in mol/L (or M).” | concepts[10] |
| concept | InChI representations — not defined in this paper's own words; compound representations used for round-trip validation and unique matching — “SMILES Validation: In order to ensure consistency and robustness of the SMILES representations, we used InChI representations in the scheme shown in Fig. 2 13 .” | concepts[11] |
| concept | RDKit mol objects — not defined in this paper's own words; molecular objects regenerated during identifier validation — “First, SMILES strings were converted to into RDKit mol objects.” | concepts[12] |
| concept | closest to the mean algorithm — not defined in this paper's own words; selection of the observed value nearest the mean of competing values — “Next, for compounds with occurrence count >1 with different solubility values (819), we used the closest to the mean algorithm to select the value.” | concepts[13] |
| concept | closest to the reference — not defined in this paper's own words; selection of the observed value nearest an external ALOGPS prediction — “For this case, we used an alternative method, which is closest to the reference.” | concepts[14] |
| concept | ALOGPS — not defined in this paper's own words; the external neural-network solubility predictor used for two-value conflicts — “ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 .” | concepts[15] |
| concept | open source online solubility prediction tool — not defined in this paper's own words; the methodological class assigned to ALOGPS — “ALOGPS is an open source online solubility prediction tool that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogS units 22 .” | concepts[16] |
| concept | reliability labels — not defined in this paper's own words; group labels reflecting occurrence count and agreement among measurements — “After we standardized the datasets, we merged all the datasets into one and further grouped them based on their reliability label and the number of occurrences in the merged dataset.” | concepts[17] |
| concept | G1 group — not defined in this paper's own words; unique or effectively identical observations selected directly — “This step leads to the curation of 7,746 unique instances, which were assigned to group G1.” | concepts[18] |
| concept | G2 group — not defined in this paper's own words; two-observation cases with standard deviation at least 0.5 LogS — “If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | concepts[19] |
| concept | G3 — not defined in this paper's own words; two-observation cases with standard deviation below 0.5 LogS — “If the SD of the two values was >=0.5 LogS units, we assigned the compounds to group G2 (235), else to G3 (1,182).” | concepts[20] |
| concept | G4 — not defined in this paper's own words; multi-observation cases with standard deviation at least 0.5 LogS — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | concepts[21] |
| concept | G5 — not defined in this paper's own words; multi-observation cases with standard deviation below 0.5 LogS — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | concepts[22] |
| concept | standard deviation — not defined in this paper's own words; cross-source dispersion statistic used to assign reliability groups — “If the standard deviation (SD) of the set of values was >=0.5 LogS units, we assigned the compounds to group G4 (183), else to G5 (636).” | concepts[23] |
| concept | degree of agreement — not defined in this paper's own words; cross-source consistency assessed using a 0.5-LogS threshold — “We selected 0.5 as a threshold for degree of agreement between multiple values based on the predictive capabilities of some the state-of-the-art models 3 , 5 , 6 .” | concepts[24] |
| concept | 2D descriptors — not defined in this paper's own words; calculated topological and physicochemical compound features — “The purpose of this step is extending the information space of compounds by adding basic topological and physico-chemical information. For this purpose, we calculated all the relevant 2D descriptors available from RDKit.” | concepts[25] |
| concept | comma-separated values — not defined in this paper's own words; the distribution file format used for AqSolDB — “AqSolDB data is stored in the comma-separated values (CSV) format and contains representations, experimental aqueous solubility and calculated 2D descriptor data of all compounds, as described in Table 2 .” | concepts[26] |
| concept | Harvard Dataverse Repository — not defined in this paper's own words; the repository providing open access to AqSolDB — “AqSolDB is openly accessible at the Harvard Dataverse Repository 24 .” | concepts[27] |
| concept | highly soluble — not defined in this paper's own words; the category for LogS values at or above zero — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[28] |
| concept | soluble — not defined in this paper's own words; the category for LogS values from zero to minus two — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[29] |
| concept | slightly soluble — not defined in this paper's own words; the category for LogS values from minus two to minus four — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[30] |
| concept | insoluble — not defined in this paper's own words; the category for LogS values below minus four — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[31] |
| concept | Compounds with 0 and higher solubility value — not defined in this paper's own words; the numeric condition assigned the highly-soluble class — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[32] |
| concept | those in the range of 0 to −2 — not defined in this paper's own words; the numeric condition assigned the soluble class — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[33] |
| concept | those in the range of −2 to −4 — not defined in this paper's own words; the numeric condition assigned the slightly-soluble class — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[34] |
| concept | less than −4 — not defined in this paper's own words; the numeric condition assigned the insoluble class — “Compounds can be classified according to solubility values (LogS); Compounds with 0 and higher solubility value are highly soluble, those in the range of 0 to −2 are soluble, those in the range of −2 to −4 are slightly soluble and insoluble if less than −4.” | concepts[35] |
| concept | poor predictive capability — not defined in this paper's own words; weak generalization performance on external test sets — “Machine learning models developed using datasets, which have small size and lack chemical diversity, show poor predictive capability on external test sets, as shown in the study by Wang et al . 8 .” | concepts[36] |
| concept | poor aqueous solubility — not defined in this paper's own words; low water solubility that risks precipitation during drug screening — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concepts[37] |
| concept | high risk of erroneous results — not defined in this paper's own words; one adverse screening consequence linked to precipitation from poor solubility — “It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubility is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous results, false leads, and increased costs and formulation difficulties during clinical development.” | concepts[38] |
| concept | sampling — not defined in this paper's own words; number of independent measurements contributing to reliability assessment — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | concepts[39] |
| concept | reliability — not defined in this paper's own words; confidence in a selected solubility value based on replication and agreement — “Statistically, due to the fact that when sampling increases, reliability will increase, it can be concluded that G5 group is more reliable than G3 group and G4 group is more reliable than G2 group.” | concepts[40] |
| concept | machine learning models — not defined in this paper's own words; data-driven solubility models for which AqSolDB supplies inputs — “The availability of the calculated 2D descriptors makes it possible to directly use the data for developing machine learning models.” | concepts[41] |
Full text — tagged
AqSolDBterm, a curated reference set of aqueous solubilityterm and 2D descriptorsterm for a diverse set of compounds Abstract Water is a ubiquitous solvent in chemistry and life. It is therefore no surprise that the aqueous solubilityterm of compounds has a key role in various domains, including but not limited to drug discovery, paint, coating, and battery materials design. Measurement and prediction of aqueous solubilityterm is a complex and prevailing challenge in chemistry. For the latter, different data-driven prediction models have recently been developed to augment the physics-based modeling approaches. To construct accurate data-driven estimation models, it is essential that the underlying experimental calibration data used by these models is of high fidelityterm and quality. Existing solubility datasets show variance in the chemical space of compounds covered, measurement methods, experimental conditions, but also in the non-standard representations, size, and accessibility of data. To address this problem, we generated a new database of compounds, AqSolDBterm, by merging a total of nine different aqueous solubilityterm datasets, curating the merged data, standardizing and validating the compound representation formats, marking with reliability labelsterm, and providing 2D descriptorsterm of compounds as a Supplementary Resource.
Subject terms: Thermodynamics, Cheminformatics Background & Summary Aqueous solubilityterm constitutes a crucial property of chemical substances that governs behavior of phenomena in several areas like geochemistry, climate predictions, biochemistry, drug-design, agrochemical design, and protein ligand binding. It is defined as the maximum amount of a compound, i.e., the solute, that can get dissolved in a given volume of water, and depends on physical conditions such as temperature and pressure. It is of critical importance in especially pharmaceutical drug design, where poor aqueous solubilityterm is likely to lead to precipitation of compounds from screening buffer, which may create a high risk of erroneous resultsterm, false leads, and increased costs and formulation difficulties during clinical development.
Although the aqueous solubilityterm of a compound can be related to its other structural and physico-chemical properties such as shape, polar surface area (PSA), acid dissociation constant (pKa), lipophilicity (logD), and the number of hydrogen bond donors and acceptors, theoretical predictions are often inaccurate. In order to overcome these challenges, several data-driven models have been developed to predict the aqueous solubilityterm of compounds last couple of decades 1 – 6 .
The development of reliable data-driven models, however, has been hindered by uncertainties and disagreements in the underlying data, which are obtained from many disparate sources. Unsystematic errors between different experimental methodologies potentially limit the accuracy with which the models can be trained and validated. To develop generalizable prediction models, accurate datasets are needed that are diverse and large at the same time 7 .
In this work, we assess the quality of aqueous solubilityterm datasets under 2 categories: generalizabilityterm and fidelityterm. Generalizabilityterm can be interpreted in terms of the chemical diversityterm of the dataset, as well as its size. Machine learning modelsterm developed using datasets, which have small size and lack chemical diversityterm, show poor predictive capabilityterm on external test sets, as shown in the study by Wang et al . 8 . Another very important indicator of dataset quality is fidelityterm. Fidelityterm can be understood as accuracy of data in terms of the reliabilityterm of the experimental technique, human errors in either conducting the experiments and recording the measured values. In their review, Wang et al . reported inconsistencies of experimental values in different databases 9 . Balakin et al . also reported the same problem where they found standard deviationterm (SD) of experimental solubility values of the same compounds as large as 0.5 in LogSterm units 7 . These errors may result from experimental noise or unintentional misprints. Data verification is important in order to increase the reliabilityterm of datasets 9 .
The aim of this study is to curate a large experimental aqueous solubilityterm data, AqSolDBterm, for data-driven model development. For this purpose, we searched for and collected nine open source datasets on aqueous solubilityterm. In order to merge the datasets, we followed systematic steps of identifier generation by converting CAS numbers and SLN identifiers into SMILES representationsterm, and validation 10 – 12 . All identifiers were converted to SMILES format and experimental solubility values were all standardized to the LogSterm units. After we standardized the datasets, we merged all the datasets into one and further grouped them based on their reliability labelterm and the number of occurrences in the merged dataset.
In this data descriptor , we provided a general algorithm for selection of the statistically most reliable values from a set of competing values. AqSolDBterm consists of aqueous solubilityterm values of 9,982 unique compounds, along with some relevant topological and physico-chemical 2D descriptorsterm. Additionally, the dataset contains validated representations of each of the compounds.
AqSolDBterm is an openly accessible, easy-to-use, and well-structured database of compound. We expect it to serve a broad community as a reference aqueous solubilityterm dataset for the bench-marking of new experimental and physics-based modelling results, and additionally as machine-readable ancillary resource to improve the prediction capability of future machine learning approaches.
Methods To curate our dataset we followed three steps. First, we collected nine publicly available aqueous solubilityterm dataset and converted them into a standardized format. Second, we combined datasets into one single dataset by applying a data verification algorithmterm that selects statistically most reliable experimental value among multiple occurrences. Finally, we added topological and physico-chemical 2D descriptorsterm to the merged dataset. Figure 1 shows the flow of the curation process.
Process diagram of curating solubility dataset.
Step 1: Collecting and pre-processing datasets Solubility data was first collected from nine publicly available datasets as shown in Table 1 . A set of three pre-processing steps were applied to each of the datasets in order to standardize the representation format and solubility values in the same units. These steps also describe our exclusion criteria on the basis of unique identifier validation. The steps are as follows: Identifier generation: We chose the SMILES representationterm as the standard identifierterm for compounds for our curated dataset. In external datasets, where SMILES representationsterm were not available, we used the name and the CAS Registry Number of compounds as inputs to retrieve the SMILES strings from the Chemical Identifier Resolverterm web service of the National Cancer Institute ( https://cactus.nci.nih.gov/chemical/structure ). Lastly, SLN identifiers available from some datasets were converted to SMILES using RDKitterm open-source cheminformatics software.
Unit Conversion: The chosen unit of solubility in this dataset is LogSterm, where S is the aqueous solubilityterm in mol/L (or M). Units such as g/L and mg/L were converted to LogSterm using the molecular massterm·var of the compounds.
SMILES Validation: In order to ensure consistency and robustness of the SMILES representationsterm, we used InChI representationsterm in the scheme shown in Fig. 2 13 . First, SMILES strings were converted to into RDKit mol objectsterm. If an error occurred during the conversion, the input SMILES string was considered to be invalid. Next, the obtained RDKit mol objectsterm were converted to InChI representationsterm. The InChI representationsterm were used to regenerate the RDKit mol objectsterm. Finally, the thus obtained mol objects were converted back to InChI representationsterm. The original and regenerated InChI were checked for consistency to ensure that the generated InChI were reproducible. This step also validated that both SMILES and InChI representationsterm led to the same RDKit mol objectterm, and thus the chemical compound.
List of datasets used to curate AqSolDBterm.
Validation steps of compound representations. Blue box represents the SMILES values from the dataset and gray boxes represent the generated values using RDKitterm. Red arrows represent the conversion steps and green equal sign represents the validation of consistency.
Table 1 shows the type of information contained in the datasets. Every dataset was processed separately in order to standardize them. The extraction process and standardization methods applied for each dataset, along with the temperature based exclusion criteria, are explained below. We named the datasets from A to I according to the number of instances they have in descending order.
Dataset A (6,110 instances) Dataset A was obtained from eChemPortal 14 , which is an open source chemical property database developed by the Organisation for Economic Co-operation and Development (OECD). Solubility data was extracted after applying the filters “experimental studies” and “water solubility”. This yielded several lines of bulk text which were then parsed to obtain CAS number, name, and experimental results on solubility including temperature and pH conditions. A total of 14,180 instances were thus obtained and these were further filtered by temperature for a range between 25 ± 5 °C. After filtering, 8,419 instances were obtained. In the identifier generation step, 6,183 of 8,419 compounds were successfully converted into SMILES. Finally, after applying SMILES validation 6,110 instances were obtained.
Dataset B (4,651 instances) Dataset B was downloaded from EPI Suite Data website 15 . This open-source dataset consisted of 5,764 liquid and crystalline organic compounds with the following properties: CAS number, name, molecular weight, water solubility, temperature. SMILES identifiers were successfully generated for 5,367 of these compounds. After that, we filtered the data by temperature between 25 ± 5 °C to obtain 5,206 compounds. In the final step, the InChI and InChIKey were validated to obtain 4,651 compounds.
Dataset C (2,603 instances) Dataset C was collected from the work of Raevsky et al . 16 and it contains solubility data measured at 25 ± 5 °C. The dataset consists of solubility of 2,603 crystalline solid compounds along with SMILES strings. All compounds were successfully recreated after pre-processing steps.
Dataset D (2,115 instances) Dataset D was downloaded from EPI Suite Data website 17 . This open-source dataset consisted of 2,267 liquid and crystalline organic compounds, out of which 2,115 compounds remained after applying the pre-processing steps.
Dataset E (1,291 instances) Dataset E was taken from the work of Huuskonen et al . 1 . In this study, the experimental aqueous solubilityterm value measured between 20–25 °C were obtained from the AQUASOL database of the University of Arizona and SCR’s PHYSPROP Database. The extended version of this dataset with 1,291 solubility values and SMILES was downloaded from the Cheminformatics ( http://cheminformatics.org/ ). All compounds were successfully recreated after pre-processing steps.
Dataset F (1,210 instances) Dataset F was taken from the work of Wang et al . 8 . They extracted 1,210 compounds from the Beilstein database and sanitized it. However, the dataset contains compound identifiers in only the SLN format 12 . We converted SLN to SMILES representationterm using RDKitterm SLN parser. During the conversion 93 of 1,210 compounds could not be produced. Using Molview ( http://molview.org/ ) web tool, we obtained valid SMILES for 93 missing compounds. Name information was collected from NCI Chemical Identifier Resolverterm service, Molview and SpyderChem 18 websites. InChI and InChIKey values are produced and validated using the pre-processing steps and all compounds were successfully recreated.
Dataset G (1,144 instances) Dataset G was taken from the work of Delaney et al . 2 . The dataset consists of 1,144 small compounds with experimental solubility measured at 25 °C and SMILES information. All compounds were successfully recreated after applying the pre-processing steps.
Dataset H (578 instances) Dataset H was taken from the work of Wang et al . 8 , who sanitized the dataset used by Jain and Yalkowsky by removing duplicate entries 19 . This dataset consists of 322 liquid and 256 solid compounds. The dataset contained only SLN as the compound identifiers and after applying the pre-processing steps all compounds are successfully recreated.
Dataset I (94 instances) Dataset I was taken from the Goodman Group website ( http://www-jmg.ch.cam.ac.uk/data/solubility/ ) as the corrected version of solubility challenge 20 . The dataset consists of 105 drug-like compounds with name, SMILES, and solubility information. The solubility values were measured at 25 °C. 11 of 105 crystalline data had to be removed because their solubility values were missing. All compounds were successfully recreated after applying the pre-processing steps.
Step 2: Merging datasets The purpose of this step is combining all datasets into the one single repository that contains only unique compounds paired with the most reliable aqueous solubilityterm value. The InChI representationterm·var was used to identify compounds uniquely and solubility values within 0.01 LogSterm units of each other were deemed to be identical. Based on these conditions, a preliminary analysis of the combined repository revealed two different kinds of redundancies - (1) a given compound was found to repeat with a different solubility value, or (2) a given compound was found to repeat with the same solubility valueterm·def. In order to quantify the relative uniqueness of each of these datasets, redundancy matrices are plotted in Fig. 3 , where the rows and columns of these matrices represent the various datasets. Redundant compounds of kind (1) and (2) between any two data sets are represented as fractional values M i j d (Fig. 3a ) and M i j s (Fig. 3b ), respectively, where i and j represent the two datasets in consideration. As an example, the value M B A d = 0.13 represents that 13% of the compounds from dataset B can be found in dataset A, but with a different solubility value. In a similar way, the value M B A s = 0.09 represents that 9% of the compounds from dataset B can be found in dataset A, but with the same solubility valueterm·def. The matrix is not symmetric in the fractional representation because of the different sizes of the datasets. While data of kind (2) can be handled simply by removing identical copies, it can be deduced from Fig. 3a about compounds of kind (1) that the datasets possess a high degree of redundancy, which necessitates a strategy for selecting the most reliable value.
Redundancy matrices showing fractional values for shared compounds between all collected datasets. ( a ) M i j d shows fraction of compounds with differing solubility values, and ( b ) M i j s shows fraction of compounds with the same solubility valuesterm·def.
There are a total of 19,796 instances in the merged repository with 9,982 unique compounds before redundant values are eliminated. To curate this data set with a unique solubility value for every compound, we design an algorithm to select the most reliable experimental value. The selection is performed by first classifying the compounds into five distinct groups which are defined based on the statistics of occurrence of a compound in the dataset. The flow chart of the curation algorithm is shown in Fig. 4 and described as follows: For every compound in the dataset, the number of occurrences is determined. If the compound has a unique value or multiple values which are within 0.01 LogSterm units, the value is simply accepted. This step leads to the curation of 7,746 unique instances, which were assigned to group G1.
Next, for compounds with occurrence countterm·var >1 with different solubility values (819), we used the closest to the mean algorithmterm to select the value. In this method, the mean value is first calculated then the closest value to mean value among the candidates is selected. If the standard deviationterm (SD) of the set of values was >=0.5 LogSterm units, we assigned the compounds to group G4term (183), else to G5term (636).
For compounds with exactly 2 values (1,417), the closest to the mean method cannot be applied because mean is always at the middle of the two values. For this case, we used an alternative method, which is closest to the referenceterm. We selected the closest value to an external reference valueterm·var, which is obtained using the solubility prediction tool ALOGPSterm 21 . ALOGPSterm is an open source online solubility prediction toolterm that is based on artificial neural networks and has an overall error of 0.49 Root Mean Squared Error (RMSE) in LogSterm units 22 . If the SD of the two values was >=0.5 LogSterm units, we assigned the compounds to group G2 (235), else to G3term (1,182).
Flowchart of the curation algorithm. Green box represents the initial state. Blue diamond shapes represent a decision according to the number of occurrences of a compound and the SD of multiple occurrences. Pink boxes represent the reliability groupterm·var. Gray boxes represent the selection method for multiple occurrences. The numbers over the arrows represent the number of unique compound in the corresponding classification path.
We selected 0.5 as a threshold for degree of agreementterm between multiple values based on the predictive capabilities of some the state-of-the-art models 3 , 5 , 6 . It must also be noted that the average SD of experimental solubility values for a given compound from different sources has been reported to be 0.5 LogSterm 7 , 23 . Using this threshold, the grouping of compound into 5 different groups provides a credible way of assessing reliabilityterm for data-driven modeling.
Step 3: Extending the dataset with 2D descriptorsterm The purpose of this step is extending the information space of compounds by adding basic topological and physico-chemical information. For this purpose, we calculated all the relevant 2D descriptorsterm available from RDKitterm. The last 17 rows of Table 2 show the name, description and data type of the 2D descriptorsterm.
List of available information in terms of name, description, and type of each column in the AqSolDBterm.
Data Records AqSolDBterm consists of 9,982 unique compounds. AqSolDBterm data is stored in the comma-separated valuesterm (CSV) format and contains representations, experimental aqueous solubilityterm and calculated 2D descriptorterm data of all compounds, as described in Table 2 . AqSolDBterm is openly accessible at the Harvard Dataverse Repositoryterm 24 .
Technical Validation Analysis of solubility values Compounds can be classified according to solubility values (LogSterm); Compounds with 0 and higher solubility valueterm are highly solubleterm, those in the range of 0 to −2term are solubleterm, those in the range of −2 to −4term are slightly solubleterm and insolubleterm if less than −4term. Figure 5c shows the distribution of solubility values.
Bar charts for analyzing the curated dataset. ( a ) Distribution of instances according to source dataset. ( b ) Distribution of instances according to reliability groupterm·var. ( c ) Distribution of instances according to aqueous solubilityterm ranges (LogSterm).
As no information about experimental errors from the original data sources was found to be available, we determined their reliabilityterm with a statistical approach. As described in the Methods section, each compound was labeled according to the selection process. Figure 5b shows the distribution of compounds to five groups. The G1 groupterm constitutes the largest part of the data and has been encountered only once in all datasets. These compounds are selected directly and it is not possible to comment on their reliabilityterm. G2 and G3term groups are composed of compounds that are found only twice in all datasets. Those with SD values greater than 0.5 were assigned to G2 groupterm and those with small or equal values were assigned to G3term group. G4term and G5term groups are composed of compounds that are found three times or more in all datasets. Using the same process as the previous one, compounds with an SD of greater than 0.5 were included in the G4term group and those with a small or equal value in the G5term group. The difference between the results of the independent experiments shows the reliabilityterm of this value. Statistically, due to the fact that when samplingterm increases, reliabilityterm will increase, it can be concluded that G5term group is more reliable than G3term group and G4term group is more reliable than G2 groupterm.
Usage Notes It is recommended for users to consider the group description when using the data as input to other models. The availability of the calculated 2D descriptorsterm makes it possible to directly use the data for developing machine learning modelsterm. To create a more complex representation of compounds such as graphs or circular fingerprints, we recommend to use RDKitterm. We provided both SMILES and InChI representationsterm of compounds which are validated and can be easily converted into the RDKit mol objectterm. Further methodological notes on data processing can be found in the Code Ocean repository 25 .
ISA-Tab metadata file Download metadata file Acknowledgements M.C.S. and S.E. acknowledge funding from the initiative “Computational Sciences for Energy Research” of Shell and the Netherlands Organisation for Scientific Research (NWO). We acknowledge Elif SorkunpersName for helping in software development. Last but not least, we acknowledge researchers who generated the experimental aqueous solubilityterm data and released them for public use.
Author Contributions M.C.S. collected, analyzed, merged and curated the data, developed codes for these purposes and computed all additional 2D descriptorsterm, A.K. aided in analyzing data. S.E. devised and supervised the project. All authors contributed to writing of the manuscript.
Code Availability The reproducibility of the curation algorithm can be verified by executing the provided scripts on Code Ocean 25 . The code has been developed and tested using Python 3.5 on Linux operating system and is available under the MIT license.
The RDKitterm cheminformatics software is freely available under the BSD licence ( http://www.rdkitterm.org ).
ALOGPSterm 2.1 used for reference valueterm·var generation is freely available online ( http://www.vcclab.org/lab/alogpsterm/ ).
Competing Interests The authors declare no competing interests.
Footnotes ISA-Tab metadata is available for this paper at 10.1038/s41597-019-0151-1.
References Associated Data Data Citations Supplementary Materials Download metadata file Data Availability Statement The reproducibility of the curation algorithm can be verified by executing the provided scripts on Code Ocean 25 . The code has been developed and tested using Python 3.5 on Linux operating system and is available under the MIT license.
The RDKitterm cheminformatics software is freely available under the BSD licence ( http://www.rdkitterm.org ).
ALOGPSterm 2.1 used for reference valueterm·var generation is freely available online ( http://www.vcclab.org/lab/alogpsterm/ ).