Integrated analysis of multimodal single-cell data

Integrated analysis of multimodal single-cell data Yuhan Hao Stephanie Hao Erica Andersen‐Nissen William M. Mauck Shiwei Zheng Andrew Butler Madeline J. Lee Aaron J. Wilk Charlotte A. Darby Michael Zager Paul Hoffman Marlon Stoeckius Efthymia Papalexi Eleni P. Mimitou Jaison Jain Avi Srivastava Tim Stuart Lamar M. Fleming Bertrand Z. Yeung Angela J. Rogers M. Juliana McElrath Catherine A. Blish Raphaël Gottardo Peter Smibert Rahul Satija

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Integrated analysis of multimodal single-cell data 2021 Cell 10.1016/j.cell.2021.04.048 hao2021-integrated-analysis-of-multimodal hao2021-integrated-analysis-of-multimodal.txt, the project's staged copy b0985d6f1a9efbac81f98ba177bd4377b44e5764 cc-by genetics

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them cells measured two ways at once - transcriptome plus surface protein, or plus chromatin accessibility - and the problem of deciding which measurement to believe Quotation withheld because this record is not cleared for redistribution. how to combine several simultaneous measurements of one cell into a single definition of its state, without being told in advance which measurement matters Quotation withheld because this record is not cleared for redistribution. weighted-nearest-neighbor analysis: a per-cell, unsupervised weight for each modality, plus a 211,000-cell multimodal reference atlas and a way to map new datasets onto it Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. cell states that one measurement type cannot separate and another can - CD4 and CD8 T cells blend in the transcriptome and separate in protein; dendritic cells do the reverse Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. the core move: if a cell's neighbours in one modality predict its measured profile well, that modality is capturing its state; if they predict badly, it is not. The weight is the ratio of those two accuracies prediction accuracy as evidence of informativeness Quotation withheld because this record is not cleared for redistribution. why rare cells are not penalised: a fixed neighbour count for the kernel bandwidth would hold common and rare states to the same scale, so the bandwidth is set from the 20 cells with the LOWEST non-zero Jaccard similarity - similar enough to matter, unlikely to be the same state adaptive kernel bandwidth Quotation withheld because this record is not cleared for redistribution. how a spurious edge is told from a real one - the Jaccard metric counts shared nearest neighbours, so local density separates well-supported links from coincidental similarity shared-neighbour filtering Quotation withheld because this record is not cleared for redistribution. what the weighting buys: one data type is allowed to make up for another's weakness, cell by cell, so the combined analysis is not limited by whichever measurement happens to be poor for that cell compensation between modalities Quotation withheld because this record is not cleared for redistribution. WNN needs a dimensional reduction to describe neighbourhood structure, so it does not work on categorical or low-dimensional data Quotation withheld because this record is not cleared for redistribution. IT ASSUMES THE MODALITIES DO NOT DEFINE CONFLICTING SETS OF CELL STATES - and the authors say the assumption may be problematic when integrating morphological, functional and molecular data. An integration method whose premise is that its inputs agree Quotation withheld because this record is not cleared for redistribution. the atlas is built from peripheral blood mononuclear cells, so it contains almost no erythrocytes or granulocytes - and any query containing them will have nowhere to map them Quotation withheld because this record is not cleared for redistribution. demonstrated on exactly that: neutrophil populations present in a COVID-19 query and absent from the reference could not be mapped Quotation withheld because this record is not cleared for redistribution. the enrichment panels achieve high fold-enrichment and still cannot sort a truly homogeneous population - the precision and recall figures are published alongside the enrichment ones Quotation withheld because this record is not cleared for redistribution. heterogeneity in monocyte inflammatory genes was left undivided BECAUSE it varied across volunteers - a cluster the authors declined to call a cell state Quotation withheld because this record is not cleared for redistribution. adding increasing amounts of random Gaussian noise to the protein data, to mimic nonspecific antibody binding, and watching what the learned weights do eliminates: the modality weights being an artefact of the method rather than a measurement of information content - THE WEIGHTS FELL DOSE-DEPENDENTLY AND REACHED ZERO, which is the behaviour the method claims and could have failed to show · outcome: held Quotation withheld because this record is not cleared for redistribution. calculating the modality weights with no knowledge of cell-type labels, then checking whether they line up with known biology eliminates: the weights being read off the answer - CD8 T cells got high protein weights and dendritic cells high RNA weights, and the labels were not available to the calculation · outcome: held Quotation withheld because this record is not cleared for redistribution. counting incorrect CD8/CD4 edges in each graph as a hard number rather than judging the UMAP pictures eliminates: a visual impression of better separation standing in for a measurement - 944 wrong edges in RNA, 12 in protein, 20 in the combined graph · outcome: held Quotation withheld because this record is not cleared for redistribution. varying k from 10 to 50 and re-running the whole procedure eliminates: the result depending on the neighbourhood size, which is the one free parameter a reader would suspect first · outcome: held Quotation withheld because this record is not cleared for redistribution. benchmarking against two named alternative methods, totalVI and MOFA+, on a stated metric - how well each integrated space reconstructs a cell's molecular state from its neighbours eliminates: 'our method is better' resting on the authors' own preferred view of the output; the metric is declared before the comparison and applies equally to all three · outcome: held Quotation withheld because this record is not cleared for redistribution. HOLDING THE QUERY DATASET'S PROTEIN MEASUREMENTS OUT of the reference mapping, then using them to check the annotations the transcriptome alone produced eliminates: the mapping being scored on the data it was given - the protein data is a true held-out test, and the predicted cell types carried the protein markers they should · outcome: held Quotation withheld because this record is not cleared for redistribution. validating a computational prediction on a different instrument - CyTOF on the original samples plus 16 more, against the scRNA-seq-predicted MAIT abundance eliminates: a depletion that exists only in the pipeline; the correlation between the two methods was 0.911 and the CyTOF cohort reproduced the depletion independently · outcome: held Quotation withheld because this record is not cleared for redistribution. sorting cells by the predicted marker panels and sequencing them in bulk, using THE EXACT ANTIBODY CLONES from the CITE-seq panel eliminates: a discrepancy between the two assays caused by different antibodies binding different epitopes - the authors name this as the reason for using the same clones · outcome: held Quotation withheld because this record is not cleared for redistribution. modality weights modality weights Quotation withheld because this record is not cleared for redistribution. modalities modalities Quotation withheld because this record is not cleared for redistribution. spurious edges spurious edges Quotation withheld because this record is not cleared for redistribution. Jaccard metric Jaccard metric Quotation withheld because this record is not cleared for redistribution. cell states cell states Quotation withheld because this record is not cleared for redistribution. transcriptome transcriptome Quotation withheld because this record is not cleared for redistribution. scRNA-seq scRNA-seq Quotation withheld because this record is not cleared for redistribution. CITE-seq CITE-seq Quotation withheld because this record is not cleared for redistribution. CyTOF CyTOF Quotation withheld because this record is not cleared for redistribution. totalVI totalVI Quotation withheld because this record is not cleared for redistribution. sPCA sPCA Quotation withheld because this record is not cleared for redistribution. neutrophils neutrophils Quotation withheld because this record is not cleared for redistribution. modality weights → cell states A NUMBER WHOSE MEANING IS UNDECIDED, AND THE PAPER SAYS SO. The weight is a proxy for the technical quality of a measurement AND may reflect the biological importance of that modality to the cell's identity. Those are different claims about different things and the same float carries both. There is no relation type for 'this measure has two candidate interpretations and we cannot separate them' - the nearest available, 'operationalised-as', would pick one and hide the other Quotation withheld because this record is not cleared for redistribution. modalities → cell states AN ASSUMPTION OF NON-CONFLICT, WHICH IS THE METHOD'S PREMISE AND ITS SECOND STATED LIMITATION. 'WNN assumes that…', and the authors flag that the assumption may fail for morphological and functional data. Every one of the eight relation types asserts something HOLDS between two concepts; none can say 'these two are assumed not to disagree, and the method has no behaviour if they do' Quotation withheld because this record is not cleared for redistribution. modality weight - A PER-CELL NUMBER, NOT A PER-DATASET ONE, AND THE PAPER SAYS IT HAS TWO POSSIBLE MEANINGS. It is a proxy for the TECHNICAL QUALITY of a measurement in that cell, and it `may also reflect the biological importance of each modality in determining cellular identity` - progenitors and differentiated cells came out with divergent weights. The same figure can mean `this instrument worked here` or `this axis is what this cell is`, and nothing in the number distinguishes them Quotation withheld because this record is not cleared for redistribution. weight, as computed here - EARNED BY A PREDICTION TEST, not assigned. Each cell is predicted from its neighbours within each modality and across them, the predictions are compared with the cell's actual measurements, and the RATIO of those accuracies becomes the weight after a softmax. Non-negative, unique to each cell, summing to 1 Quotation withheld because this record is not cleared for redistribution. unsupervised, of this procedure - MEANS UNAWARE OF CELL-TYPE LABELS, AND THAT IS WHAT MAKES THE AGREEMENT EVIDENCE. CD8 T cells got high protein weights and dendritic cells high RNA weights, `despite the fact that the calculation of modality weights was unsupervised and unaware of cell-type labels`. The recapitulation of a known biology is the check, and it only counts because the labels were withheld Quotation withheld because this record is not cleared for redistribution. spurious edge - A LINK BETWEEN CELLS THAT RESEMBLE EACH OTHER WITHOUT BEING IN THE SAME STATE, and the paper counts them rather than describing them: 944 wrong CD8/CD4 edges in the RNA graph, 12 in the protein graph, 20 in the combined one. A quality claim with an integer behind it Quotation withheld because this record is not cleared for redistribution. annotation, in this atlas - THREE NESTED TIERS AND A CELL CARRIES ALL THREE. Level 1 has 8 categories, level 2 has 30, level 3 has 57. A result reported at one tier is not comparable with a result reported at another, and the granularity is a choice the reader has to make rather than one the data settles Quotation withheld because this record is not cleared for redistribution. supervised, in `supervised principal component analysis` - SUPERVISED BY THE WNN GRAPH, NOT BY A HUMAN. sPCA finds the transcriptomic projection that best captures the structure the multimodal analysis already found, so the supervision comes from the other modality rather than from labels. `Supervised` and `unsupervised` are both used in this paper about the same pipeline, meaning different things Quotation withheld because this record is not cleared for redistribution. distinguishes-from → totalVI and MOFA+, two recently introduced multimodal integration methods based on deep neural networks with a hierarchical Bayesian model and on factor analysis respectively Quotation withheld because this record is not cleared for redistribution. distinguishes-from → scArches, a method for mapping single-cell RNA-seq queries onto reference datasets Quotation withheld because this record is not cleared for redistribution. critiques → the practice of defining cell states from the transcriptome alone, which cannot separate functionally distinct T cell populations however sensitive the technology Quotation withheld because this record is not cleared for redistribution. extends → the unsupervised clustering analysis of Wilk et al.'s COVID-19 PBMC dataset, which found six T cell clusters where reference mapping resolved twelve note: and the extension is not one-directional - the same comparison shows what the reference CANNOT do: neutrophils that the original unsupervised analysis found, because they were absent from the reference, could not be mapped at all Quotation withheld because this record is not cleared for redistribution. builds-on → Barshan et al.'s supervised principal component analysis, applied here with the WNN graph as the supervising structure Quotation withheld because this record is not cleared for redistribution. transcriptome —contrasted-with→ modalities transcriptome modalities strength of the relation: 0.8 · note: the paper's whole framing: the transcriptome is one modality among several and has been treated as the definition of cell state Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span questions defining cell states from the transcriptome alone; contrasted-with 'modalities' is loose. Quotation withheld because this record is not cleared for redistribution. CITE-seq —distinct-from→ CyTOF CITE-seq CyTOF strength of the relation: 0.9 · note: and the distinctness is what makes the agreement worth something - the MAIT depletion was found in one and confirmed in the other at R = 0.911 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but CITE-seq provides a 'unique opportunity' beside flow cytometry and CyTOF; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. per-cell modality weights, learned unsupervised, recover known biology: CD8 T cells weight protein highly, dendritic cells weight RNA highly Quotation withheld because this record is not cleared for redistribution. The RNA KNN graph had 944 incorrect CD8+/CD4+ edges, the protein graph 12. Quotation withheld because this record is not cleared for redistribution. adding Gaussian noise to the protein data drove protein weights down dose-dependently and to zero, correctly redirecting the analysis to RNA alone Quotation withheld because this record is not cleared for redistribution. The final multimodal dataset has 210,911 cells covering resting and post-vaccination immune states. Quotation withheld because this record is not cleared for redistribution. MAIT-cell fractions predicted by scRNA-seq agreed with CyTOF measurements at R = 0.911. Quotation withheld because this record is not cleared for redistribution. WNN matched or beat alternative approaches in reconstructing each of 25 proteins. Quotation withheld because this record is not cleared for redistribution. modality weights = per-cell values reflecting the relative information content of each data type in that cell; non-negative, unique per cell, summing to 1 modality weights Quotation withheld because this record is not cleared for redistribution. the four steps: build a KNN graph per modality, predict within and across modalities, compute per-cell weights, build the WNN graph weighted nearest neighbor analysis Quotation withheld because this record is not cleared for redistribution. spurious edge = a link between cells that share some molecular similarity but are not in a matched molecular state spurious edges Quotation withheld because this record is not cleared for redistribution. sPCA = the transcriptomic projection that maximally captures the structure defined in the WNN graph, rather than the one maximising total variance sPCA Quotation withheld because this record is not cleared for redistribution. the weights come from a softmax over the ratio of within-modality to cross-modality prediction affinities Quotation withheld because this record is not cleared for redistribution. the kernel bandwidth is set from the 20 cells with the LOWEST non-zero Jaccard similarity - similar enough to matter, unlikely to be the same state Quotation withheld because this record is not cleared for redistribution. results held across k from 10 to 50, and the protein information did not cost transcriptomic neighbourhood coherence Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1016/j.cell.2021.04.048.