Integrated analysis of multimodal single-cell data

Multimodal analysis allows identification and validation of previously unreported lymphoid subpopulations.

“Multimodal analysis substantially improves our ability to resolve cell states, allowing us to identify and validate previously unreported lymphoid subpopulations.”

Multimodal analysis substantially improves the ability to resolve cell states.

“Multimodal analysis substantially improves our ability to resolve cell states, allowing us to identify and validate previously unreported lymphoid subpopulations.”

WNN integration substantially increased the ability to resolve hematopoietic cell states in a CITE-seq dataset of human bone marrow.

“Consistent with our previous example, WNN integration substantially increased our ability to resolve hematopoietic cell states (Figures 2A and S2).”

For CD8+ T cells, protein neighbors were predominantly correctly identified as CD8+ T cells (12 CD8+/CD4+ edges in protein KNN graph).

“By contrast, protein neighbors were predominantly correctly identified as CD8+ T cells (in the protein KNN graph, 12 CD8+/CD4+ edges were identified).”

Increasing ADT noise led to a decrease in protein weights for all cell types, in a dose-dependent manner.

“We found that the increasing ADT noise led to a decrease in protein weights for all cell types, in a dose-dependent manner.”

WNN was the only method where CD25 and CD57 T cell populations were not intermixed with other groups in UMAP visualization.

“This was consistent with UMAP visualization, in which WNN was the only method where these populations were not intermixed with other groups (Figure S2).”

Each cluster exhibited distinct molecular patterns and biomarkers for both RNA and protein modalities.

“We found that each cluster exhibited distinct molecular patterns and biomarkers for both modalities (Figure 4A; additional heatmaps in Figure S4).”

WNN analysis applied to ASAP-seq produced highly concordant results.

“We obtained highly concordant results when applying WNN analysis to ASAP-seq (Mimitou et al., 2020), a third multimodal technology, that pairs measurements of surface protein abundance with ATAC-seq profiles in single cells (Figure S3).”

• cites Mimitou et al., 2020Not checked yet

  Not checked in the free run.

Unsupervised transcriptome-focused clustering was unable to separate these T cell states, even with focused re-clustering using only T cells.

“We found that unsupervised transcriptome-focused clustering was unable to separate these cell states, even if we performed a focused re-clustering using only T cells (Figure S2).”

Markers identified by differential expression are repeatedly identified across samples.

“Heatmap displays pseudobulk averages where cells are grouped by cell type, human volunteer, and vaccination time point and demonstrates that markers are repeatedly identified across samples.”

For CD8+ T cells, the most similar RNA neighbors often reflected a mix of CD8+ and CD4+ T cells (944 incorrect edges in RNA KNN graph).

“We found that for CD8+ T cells, the most similar RNA neighbors often reflected a mix of CD8+ and CD4+ T cells (in the RNA KNN graph, there are a total of 944 incorrect edges that connect CD8+ to CD4+ T cells).”

Protein modality weights were assigned to 0 after a sufficient amount of protein noise was added, correctly instructing downstream analyses to focus only on scRNA-seq data.

“Moreover, protein modality weights were assigned to 0 after a sufficient amount of protein noise was added, correctly instructing downstream analyses to focus only on scRNA-seq data.”

WNN analysis exhibited significant improvements in speed, ranging up to 15-fold when analyzing the full dataset.

“WNN analysis also exhibited significant improvements in speed, ranging up to 15-fold when analyzing the full dataset (Figure 2E).”

ATAC-seq data were more capable of separating naive CD8+ and CD4+ T cell states due to reliable detection of cell-type-specific open chromatin regions.

“Interestingly, similar to our CITE-seq analyses, we found that ATAC-seq data were more capable of separating naive CD8+ and CD4+ T cell states due to reliable detection of cell-type-specific open chromatin regions (Figure S3).”

Heterogeneous expression of inflammatory genes in monocyte populations was exhibited in multiple, but not all, volunteers.

“Heterogeneous expression of these genes is exhibited in multiple, but not all, volunteers.”

The analysis identified heterogeneous subsets of myeloid cells that were fully concordant with recent high-resolution scRNA-seq analyses of sorted populations.

“our analysis clearly identified heterogeneous subsets of myeloid cells that were fully concordant with recent high-resolution scRNA-seq analyses of sorted populations”

• cites See et al., 2017Not checked yet

  Not checked in the free run.

• cites Villani et al., 2017Not checked yet

  Not checked in the free run.

The heterogeneity in inflammatory gene expression in monocytes was not related to the vaccination time course.

“This heterogeneity was not related to the vaccination time course, as shown in (E).”

The identified biomarkers were invariant across human volunteers and vaccination time points.

“Moreover, these identified biomarkers were invariant across human volunteers and vaccination time points.”

The separation of naive CD8+ and CD4+ T cell clusters upon UMAP was consistent with the number of incorrect edges (RNA KNN: 984, ATAC KNN: 373, WNN: 322).

“The separation of these clusters upon UMAP visualization (Figure S3) was consistent with the number of incorrect naive CD8+/CD4+ edges identified in each representation (RNA KNN: 984, ATAC KNN: 373, WNN: 322).”

In a dataset of 11,351 paired PBMC profiles from 10x Multiome ATAC+RNA, the combination of modalities exhibited maximal power to separate immune subsets.

“We found that the combination of modalities exhibited maximal power to separate immune subsets (Figure S3).”

T cell groups-and in particular, populations masked in scRNA-seq analyses-all received higher protein modality weights.

“Indeed, when examining the cell-specific modality weights, we found that T cell groups—and in particular, populations that were masked in scRNA-seq analyses—all received higher protein modality weights (Figure 2B).”

With rare exceptions for infrequent cell types, each cluster included cells from all 24 samples.

“With rare exceptions for infrequent cell types, each cluster included cells from all 24 samples.”

For each of the 25 proteins, as well as the RNA transcriptome, WNN analysis exhibited superior or equivalent performance to alternative approaches (MOFA+ and totalVI).

“We found that for each of the 25 proteins (Figure 2D), as well as the RNA transcriptome (Figure S2), WNN analysis exhibited superior or equivalent performance to alternative approaches.”

Cell populations with the highest RNA weights represented hematopoietic progenitor populations.

“Conversely, each of the cell populations with the highest RNA weights represented hematopoietic progenitor populations.”

Markers do not vary across different PBMC samples.

“Heatmap displays pseudobulk averages where cells are grouped by cell type, donor, and vaccination time point and demonstrates that markers do not vary across different PBMC samples.”

The analysis identified extremely rare populations (0.02%) of dendritic cells defined by expression of AXL and SIGLEC6 (ASDC).

“including extremely rare populations (0.02%) of dendritic cells defined by the expression of AXL and SIGLEC6 (See et al., 2017; Villani et al., 2017) (ASDC; Figure S4).”

• cites See et al., 2017Not checked yet

  Not checked in the free run.

• cites Villani et al., 2017Not checked yet

  Not checked in the free run.

Statistical power in unsupervised analysis of either RNA or protein modality separately was insufficient to identify hematopoietic cell populations defined by both RNA and protein markers.

“However, although these cell types are defined by both RNA and protein markers, the statistical power in unsupervised analysis of either modality separately was insufficient to identify these populations, demonstrating the importance of joint analysis.”

Multimodal analysis was able to identify diverse populations of hematopoietic stem cells, LMPP, and progenitors of erythroid, platelet, monocyte, B, and conventional/plasmacytoid DC lineages that could be recovered in scRNA-seq data, even though these groups lacked immunophenotypic markers.

“As a result, our multimodal analysis was able to identify diverse populations of hematopoietic stem cells, lymphoid-primed multipotent progenitors (LMPP), and progenitors of erythroid, platelet, monocyte, B, and conventional/plasmacytoid DC lineages that could be recovered in scRNA-seq data, even though these groups lacked immunophenotypic markers in our CITE-seq experiment.”

The difference in performance was particularly striking for markers of regulatory (CD25) and effector (CD57) T cells.

“The difference in performance was particularly striking for markers of regulatory (CD25) and effector (CD57) T cells.”

The Basal_4 population was specifically characterized by increased chromatin accessibility at CTCF and p53 motifs.

“We found that the Basal_4 population was specifically characterized by increased chromatin accessibility at CTCF and p53 motifs (Demirkan et al., 2000) (Figure S3).”

• cites Demirkan et al., 2000Not checked yet

  Not checked in the free run.

Consistent results were observed in an independent CITE-seq dataset of human PBMCs (Kotliarov et al., 2020).

“We also observed consistent results (Figure S5) in an independent CITE-seq dataset of human PBMCs (Kotliarov et al., 2020).”

• cites Kotliarov et al., 2020Not checked yet

  Not checked in the free run.

Within B cells, a continuous trajectory connecting naive to memory cells was identified, defined by canonical protein markers IgD and CD27, along with a module of correlated genes.

“For example, within B cells, we identified a continuous trajectory connecting naive to memory cells defined by the canonical protein markers immunoglobulin D (IgD) and CD27, along with a module of correlated genes (Figure S5).”

Expanded clonal populations were observed.

“However, we also observed the presence of expanded clonal populations.”

Siglec-1 (CD169) was identified as a protein response biomarker that was robustly induced only in day 3 samples.

“In addition, we identified Siglec-1 (CD169) as a protein response biomarker that was robustly induced only in day 3 samples (Figure 6C).”

Five positive and three negative volunteers were identified for CMV status by stimulating PBMCs with a CMV peptide pool and performing intracellular cytokine staining.

“We assessed the CMV status of each volunteer by stimulating PBMCs with a CMV peptide pool and performing intracellular cytokine staining to determine responses in CD8+ T cells (Table S3; STAR Methods), identifying five positive and three negative volunteers.”

Within CD8+ memory T cells, distinct subpopulations were identified defined by bimodal and mutually exclusive expression of integrin proteins CD49a and CD103.

“Within CD8+ memory T cells, we identified distinct subpopulations defined by bimodal and mutually exclusive expression of the integrin proteins CD49a and CD103 (Figure 5A).”

No significant changes were observed during the time course in overall abundance of broad immune classes.

“We did not observe significant changes during the time course in overall abundance of broad immune classes (Figure 6E;); thus, we focused on identifying more subtle compositional changes.”

Supervised mapping procedure dramatically improved the ability to analyze and interpret query scRNA-seq datasets compared to unsupervised analysis.

“We found that this supervised mapping procedure dramatically improved our ability to analyze and interpret query scRNA-seq datasets compared to unsupervised analysis.”

A panel with three markers was sufficient to achieve 10-fold enrichment for 55 clusters.

“whereas a panel with three markers was sufficient to achieve 10-fold enrichment for 55 clusters.”

A second gradient defined by CD38 expression was observed that, to our knowledge, has not been previously described.

“We also observed a second gradient defined by CD38 expression that, to our knowledge, has not been previously described.”

NK cells were subdivided into five clusters, representing variation across a continuous landscape.

“Similarly, NK cells were subdivided into five clusters, representing variation across a continuous landscape.”

A population of CD14+ Monocytes that has severe transcriptional responses to COVID-19 is highlighted in the analysis.

“Additionally, a population of CD14+ Monocytes that has severe transcriptional responses to COVID-19 is also highlighted in this analysis.”

Mapping a dataset of human PBMCs prior to flu vaccination onto the reference using only transcriptome data and transferring level 2 annotations revealed the presence of multiple high-resolution lymphoid subsets.

“We mapped this dataset onto our reference using only the transcriptome data and transferred our level 2 annotations, revealing the presence of multiple high-resolution lymphoid subsets (Figure S7).”

Overall clonal diversity was consistent across vaccination time points.

“Overall clonal diversity was consistent across vaccination time points, consistent with an expected lack of a lymphoid response to vaccination within 7 days”

Pathway enrichment of top DE genes between day 0 and day 3 myeloid cells exhibits a clear enrichment for components of the interferon response.

“Pathway enrichment (enrichR) of the top DE genes between day 0 and day 3 myeloid cells exhibits a clear enrichment for components of the interferon response.”

Heterogeneity in NK cells is defined by two gradients correlating with CD16 and CD38 protein expression.

“Heterogeneity in NK cells is defined by two gradients correlating with CD16 and CD38 protein expression.”

Cells within a clone exhibit similar molecular profiles.

“Cells within a clone exhibit similar molecular profiles.”

The marker discovery procedure identifies effective panels for well-characterized populations (pDC: CD123+; MAIT cells: CD3+ CD161+ TCRvɑ7.2+; CD4 naive: CD4+ CD45RA+ CD45RB+).

“We confirmed that this marker discovery procedure identifies effective panels for well-characterized populations (plasmacytoid DC [pDC]: CD123+; MAIT cells: CD3+ CD161+ TCRvɑ7.2+; CD4 naive: CD4+ CD45RA+ CD45RB+).”

Predicted marker panels for CD4 CTL cells were validated by bulk RNA-seq.

“Same as in (D) but for CD4 CTL cells.”

For five cell types where a high rate of discordant predictions between Seurat and scArches was observed, Seurat annotations provided the most concordant results.

“For five cell types where we observed a high rate of discordant predictions between Seurat and scArches, we visualize the protein expression of key markers in the reference dataset (left), Seurat-transferred annotations (middle), and scArches-transferred annotations (right). In each case, the Seurat annotations provide the most concordant results.”

Mutually exclusive expression of CD103 and CD49a within CD8+ T memory cells was measured by CITE-seq.

“Mutually exclusive expression of the integrin proteins CD103 and CD49a within CD8+ T memory cells, as measured by CITE-seq.”

Predicted marker panels for the CD8_TEM_5 cluster were validated by bulk RNA-seq.

“Validation of predicted marker panels for the CD8_TEM_5 cluster. We sorted cells based on the marker panels identified in (C), and performed bulk RNA-seq. Each column represents a replicate bulk RNA-seq profile. Heatmap is ordered by genes expected to be DE based on our CITE-seq dataset and are validated by bulk RNA-seq.”

The presence of CD103+ and CD49a+ populations was validated in independent healthy PBMC samples by flow cytometry.

“We validated the presence of the populations in independent healthy PBMC samples by performing flow cytometry for the same markers (Figures 5C and 5D).”

The analysis identified CD43 as a marker with high enrichment power for cytotoxic populations of CD4+ lymphocytes that has not been previously reported.

“In other instances (e.g., cytotoxic populations of CD4+ lymphocytes), our analysis identified CD43 as a marker with high enrichment power that has not been previously reported.”

Modules of differentially expressed genes between CD103+ CD49a− and CD103− CD49+ groups were enriched for T cell activation, differentiation, signaling response, and chemotaxis modules.

“In addition, we identified modules of differentially expressed genes between these two groups (Figure 5E), which were enriched for T cell activation, differentiation, signaling response, and chemotaxis modules (Figure 5F).”

Concordant cell types were identified between query and reference data with three exceptions.

“Concordant cell types are identified between query and reference data with three exceptions, denoted with dashed rectangles.”

Differentially expressed genes and enriched gene ontology terms were identified between CD103+ CD49a− and CD103− CD49+ populations.

“Differentially expressed genes, and enriched gene ontology terms, between CD103+ CD49a− and CD103− CD49+ populations.”

Enrichment panels for cytotoxic CD4+ lymphocytes and CD8_TEM_5 cells were successfully validated in an independent set of PBMCs from healthy donors by conventional flow cytometry followed by bulk RNA-seq.

“For this population, as well as a subgroup of highly cytotoxic CD8+ T cells (CD8_TEM_5), we successfully validated our enrichment panels in an independent set of PBMCs from healthy donors by conventional flow cytometry followed by bulk RNA-seq (STAR Methods).”

Developing and differentiated neutrophils, which are not present in the reference, remain distinct after reference-based mapping.

“These results demonstrate that developing and differentiated neutrophils, which are not present in the reference, remain distinct after reference-based mapping.”

Minimal correlation between CD38 and CD16 expression was observed.

“Finally, we observed minimal correlation between CD38 and CD16 expression (Figure 5K)”

Robust responses were consistently observed only in CD14+ monocytes, CD16+ monocytes, and cDC2 DC.

“In (A)–(D) we consistently observe robust responses only in CD14+ monocytes, CD16+ monocytes, and cDC2 DC.”

Differential expression of integrin-7 between CD103+ CD49a− and CD103+ CD49+ populations was measured by CITE-seq.

“Differential expression of integrin-7 between CD103+ CD49a− and CD103+ CD49+ populations as measured by CITE-seq.”

For all 17,480 (32.9%) of query cells where Seurat and scArches returned different annotations based on the transcriptome, in 73.8% of cases there was stronger support for the Seurat annotation.

“For all 17,480 (32.9%) of query cells where Seurat and scArches returned different annotations based on the transcriptome, we calculated protein-based classification metrics to determine the support for each result (STAR Methods). In 73.8% of cases, we observe stronger support for the Seurat annotation.”

A single marker was capable of achieving effective enrichment of at least 10-fold for 45 clusters.

“We found that a single marker was capable of achieving effective enrichment of at least 10-fold for 45 clusters”

97% of clones consisted only of a single cell.

“and 97% of clones consisted only of a single cell.”

Flow cytometry validates the presence of CD103+ and CD49a+ populations.

“Flow cytometry validates the presence of these populations.”

Naive, intermediate and memory B cells are ordered by their quantitative level of CD27 protein expression.

“Naive, intermediate and memory B cells are ordered by their quantitative level of CD27 protein expression.”

These findings are consistent across volunteers, vaccination time points, and CMV status.

“As expected, and in contrast to the clonotypes reported in Figure 5G, these findings are consistent across volunteers, vaccination time points, and CMV status.”

Consistent trends were observed when restricting the analysis only to individuals with either positive or negative CMV T cell responses.

“Notably, we observed consistent trends when restricting our analysis only to individuals with either positive or negative CMV T cell responses (Figure S5).”

The five CMV-positive samples accounted for 91% of cells within expanded clones.

“We found that the five positive samples accounted for 91% of cells within expanded clones.”

The frequency of broad groups (level 1 annotations) is stable across the vaccination time course.

“Bar plot showing that the frequency of broad groups (level 1 annotations) is stable across the vaccination time course.”

Rare populations of invariant NKT cells defined by use of TRAV10.TRAJ18 were detected.

“We also detect rare populations of invariant NKT cells (defined by the use of TRAV10.TRAJ18).”

The abundance of plasmablasts increases during COVID-19 response.

“The abundance of plamablasts increases during COVID-19 response. p value is computed using an unpaired Wilcoxon test. Annotations were derived from reference-based mapping, and confirm the result reported in Wilk et al. (2020).”

• cites Wilk et al., 2020Not checked yet

  Not checked in the free run.

Excluding MAIT and invariant NKT populations, 31 additional expanded clones consisting of at least 10 cells were identified.

“Excluding these populations, we identified 31 additional expanded clones consisting of at least 10 cells (Figure 5G).”

Both CD14+ classical and CD16+ non-classical monocytes upregulated a shared module of 62 genes highly enriched for transcripts responsive to type I interferon.

“We observed the strongest changes in both CD14+ classical and CD16+ non-classical monocytes, as both cell types upregulated a shared module of 62 genes highly enriched for transcripts responsive to type I interferon (Figures 6A, 6B, and S6; visualization in Figure 6B from Alquicira-Hernandez and Powell [2020]).”

A robust dendritic cell response was observed only within CD1C+ cDC2 cells.

“When we examined dendritic cell populations, we observed a similarly robust response only within CD1C+ cDC2 cells.”

The approach yielded substantial improvements in accuracy and performance compared to scArches.

“We benchmarked our method against scArches, a recently developed method for mapping scRNA-seq queries to reference datasets (Lotfollahi et al., 2020) and observed that our approach yielded substantial improvements in accuracy and performance (Figures 7A, 7B, and S7).”

• cites Lotfollahi et al., 2020Not checked yet

  Not checked in the free run.

There was clear agreement between sorted bulk profiles and CITE-seq clusters for genes expected to be DE based on CITE-seq data.

“In both cases, we examined the expression level of genes that we expected to be DE-based on our CITE-seq data, and we observed clear agreement between the sorted bulk profiles and CITE-seq clusters (Figures 4D and 4E).”

For CD25, WNN analysis achieved the highest correlation between predicted and measured values.

“For CD25, WNN analysis achieved the highest correlation, as cells that are CD25+ are correctly identified as neighbors with other cells that are CD25+ in the dataset.”

The supervised mapping workflow robustly integrates samples from multiple donors and disease states.

“we show how this workflow can improve cell type identification and robustly integrate samples from multiple donors and disease states.”

• AssumptionThe authors state this as an assumption: WNN assumes that modalities do not define conflicting sets of cell states.

The circulating immune atlas was constructed from PBMCs and therefore contains few cells with no nuclei (erythrocytes) or multi-lobed nuclei (granulocytes).

“In addition, our circulating immune atlas was constructed from PBMCs and therefore contains few cells with no nuclei (erythrocytes) or multi-lobed nuclei (granulocytes).”

The dataset of human PBMCs features paired transcriptomes and measurements of 228 surface proteins, representing a multimodal atlas of the immune system.

“We apply our approach to analyze a dataset of human PBMCs featuring paired transcriptomes and measurements of 228 surface proteins, representing a multimodal atlas of the immune system.”

In human bone marrow, progenitor cells and differentiated cells exhibited divergent modality weights.

“For example, our analyses of human bone marrow demonstrated that progenitor cells and differentiated cells exhibited divergent modality weights.”

• AssumptionThe authors state this as an assumption: WNN assumes that modalities do not define conflicting sets of cell states.

A depletion of MAIT cells in COVID-19 samples compared to healthy controls was observed.

“In particular, we observed a depletion of MAIT cells in COVID-19 samples compared to healthy controls.”

The authors have not observed modalities defining conflicting sets of cell states when using molecular data such as chromatin state, gene expression, and surface protein abundance.

“Although we have not observed this when using molecular data such as chromatin state, gene expression, and surface protein abundance, this assumption may be problematic when integrating morphological, functional, and molecular data.”

A dataset of 50,000 cells can be fully processed and mapped in less than 5 min using this approach.

“Using this approach, a dataset of 50,000 cells can be fully processed and mapped in less than 5 min.”

For all protein features, WNN analysis consistently achieved the highest correlation.

“We repeated this analysis for all protein features, and found that, WNN analysis consistently achieved the highest correlation.”

The Pearson correlation between MAIT cell abundance predicted by scRNA-seq reference mapping and measured by CyTOF is 0.911.

“The Pearson correlation between these two methods is 0.911.”

The authors created a freely available web application (Azimuth) that enables users to rapidly map their own datasets online, automating visualization and annotation.

“we have created a web application, freely available at https://azimuth.hubmapconsortium.org/, which enables users to rapidly map their own datasets online, automating the process of visualization and annotation.”

CyTOF analysis of a larger sample set identified a depletion of MAIT cells in COVID-19 samples.

“Moreover, CyTOF analysis of the larger sample set identified a depletion of MAIT cells in COVID-19 samples (Figures 7F and S7).”

The signal for each of these 12 markers was reduced by addition of a proportional amount of unlabeled antibodies.

“The signal for each of these markers was reduced by addition of a proportional amount of unlabeled antibodies.”

The bulk-sorted populations exhibited highly concordant relative expression patterns for DE genes as observed in CITE-seq data.

“The bulk-sorted populations exhibited highly concordant relative expression patterns for DE genes as we observed in CITE-seq data.”

The supervised mapping workflow improves cell type identification.

“we show how this workflow can improve cell type identification and robustly integrate samples from multiple donors and disease states.”

• AssumptionThe authors state this as an assumption: WNN assumes that modalities do not define conflicting sets of cell states.

The protein data withheld from mapping displays the same patterns as in the reference CITE-seq dataset.

“The protein data were withheld from the mapping but displays the same patterns as in (A).”

In supervised analysis, T cells were successfully divided into 12 groups after transferring level 2 annotations.

“In our supervised analysis, we transferred our level 2 annotations, successfully dividing T cells into the 12 groups (Figure 7C, D).”

For all transcriptomic features, similar performance was observed for all methods.

“We repeated the analysis for all transcriptomic features as well (Figure S2) and observed similar performance for all methods.”

Genes whose expression level is correlated with a cell's position along a molecular gradient defined by a single protein were identified.

“In Figures 5G, 5H, and S5, we identify genes whose expression level is correlated with a cell’s position along a molecular gradient defined by a single protein.”

Similar results were obtained when computing excess variance and when repeating the analyses using progenitor subpopulations with high RNA modality weights.

“We obtained similar results when computing a measure of ‘excess variance’, defined as the amount of residual variance observed for each gene after controlling for the mean-variance relationship inherent in single-cell data, and also when repeating these analyses using progenitor subpopulations with high RNA modality weights (Figure S2).”

• AssumptionRests on an assumption: If the WNN procedure is performing well, gene expression levels within the examined neighborhoods should represent a homogeneous population with low levels of variability.

WNN-derived neighborhoods were most reflective of a homogeneous population, as evidenced by the number of genes with variable expression within each neighbor set.

“We found that WNN-derived neighborhoods were most reflective of a homogeneous population, as evidenced by the number of genes that exhibited variable levels of expression within each neighbor set (Figure S2).”

• AssumptionRests on an assumption: If the WNN procedure is performing well, gene expression levels within the examined neighborhoods should represent a homogeneous population with low levels of variability.

Differentially expressed genes between cell states exhibited nearly identical fold-changes in either WNN-derived or RNA-derived clusters.

“Moreover, we found that differentially expressed genes between cell states exhibited nearly identical fold-changes in either WNN-derived or RNA-derived clusters (Figure S2).”

Populations of developing and differentiated neutrophils, which were identified by the original manuscript as uniquely present in COVID-19 samples but absent from the reference, could not be successfully mapped.

“Notably, populations of developing and differentiated neutrophils, which were identified by the original manuscript as being uniquely present in COVID-19 samples but were absent from our reference, could not be successfully mapped (Figure S7).”

Strong quantitative agreement (R = 0.911) was observed in the fraction of MAIT cells predicted by scRNA-seq and measured by CyTOF in the original cohort.

“We observed strong quantitative agreement (R = 0.911) in the fraction of MAIT cells predicted by scRNA-seq and measured by CyTOF in the original cohort (Figure 7E).”

The authors' supervised mapping approach can map scRNA-seq data onto the multimodal reference dataset.

“In addition to constructing a multimodal reference, we demonstrate the ability to map scRNA-seq data onto this dataset.”

In the de-novo visualization, all three populations (neutrophils, activated granulocytes, and perturbed monocyte subpopulations) remain distinct from reference cells even after integration.

“In the de-novo visualization, all three of these populations remain distinct from reference cells even after integration (Figure S7).”

In 79.4% of cases where Seurat and scArches disagreed, there was stronger support for the Seurat annotation based on protein levels.

“In 79.4% of cases, we observe stronger support for the Seurat annotation (Figure S7E).”

In the reference-based visualization, cells that cannot be effectively represented in the reference are often placed at the boundary of cell clusters.

“In the reference-based visualization, the umap_transform functions aims to embed these cells adjacent to their closest neighbors in the reference, which often places these cells at the boundary of cell clusters.”

For 35,619 cells (67.1%), Seurat and scArches returned the same annotation.

“For 35,619 cells (67.1%), Seurat and scArches returned the same annotation.”

The perturbation score helps prioritize cell types that exhibit robust responses, particularly when comparing populations with vastly different abundances.

“We find that this approach helps to prioritize cell types that exhibit robust responses, particularly when comparing populations with vastly different abundances.”

For the remaining 17,480 query cells, Seurat and scArches returned divergent annotations.

“For the remaining 17,480 query cells, the two methods returned two divergent annotations (for example, suppose that Seurat annotated the cell as CD4 Treg, and scArches annotated as NK).”

Transcriptomics alone is often incapable of separating molecularly similar but functionally distinct categories of immune cells.

“Although established technologies like single-cell RNA-seq (scRNA-seq) are capable of discovering new cell types and states in heterogeneous tissues, transcriptomics alone is often incapable of separating molecularly similar, but functionally distinct, categories of immune cells.”

• cites Ding et al., 2020Not checked yet

  Not checked in the free run.

• cites Mereu et al., 2020Not checked yet

  Not checked in the free run.

WNN analysis is an unsupervised framework to learn the relative utility of each data type in each cell, enabling integrative analysis of multiple modalities.

“Here, we introduce “weighted-nearest neighbor” analysis, an unsupervised framework to learn the relative utility of each data type in each cell, enabling an integrative analysis of multiple modalities.”

Distinct populations of T cells such as effector, regulatory, γδ, and MAIT often cannot be effectively separated by scRNA-seq alone, even with sensitive technologies.

“Despite tremendous functional diversity, distinct populations of T cells such as effector, regulatory, γδ, and mucosal associated invariant T (MAIT), often cannot be effectively separated by scRNA-seq alone, even when using the most sensitive and cutting-edge technologies (Ding et al., 2020; Mereu et al., 2020).”

• cites Ding et al., 2020Not checked yet

  Not checked in the free run.

• cites Mereu et al., 2020Not checked yet

  Not checked in the free run.

Defining cell states based on transcriptome alone is challenging because important sources of cellular heterogeneity may not correlate strongly with transcriptomic features despite being identifiable in other modalities.

“More broadly, this exhibits the challenge of defining cell states based on the transcriptome alone, because important sources of cellular heterogeneity may not correlate strongly with transcriptomic features despite being identifiable in other modalities.”

Increased copy number of protein molecules compared to RNA molecules typically leads to more robust detection of protein features in CITE-seq.

“As previously described for CITE-seq (Mimitou et al., 2019; Stoeckius et al., 2017), the increased copy number of protein molecules compared to RNA molecules typically leads to more robust detection of protein features.”

• cites Mimitou et al., 2019Not checked yet

  Not checked in the free run.

• cites Stoeckius et al., 2017Not checked yet

  Not checked in the free run.

Multimodal integration is essential for the unsupervised discovery and annotation of immune cell states.

“We conclude that multimodal integration is essential for the unsupervised discovery and annotation of immune cell states”

The WNN strategy is applicable to diverse multimodal technologies.

“Although we primarily demonstrate our approach on CITE-seq datasets, our strategy is applicable to diverse multimodal technologies.”

Once immune cell states are enumerated, supervised differential analyses are capable of sensitively describing markers that define their molecular state.

“however, once these states are enumerated, supervised differential analyses are capable of sensitively describing markers that define their molecular state.”

The ability of WNN to identify subpopulations that are masked by scRNA-seq alone is not limited to immune or CITE-seq datasets.

“Taken together, these findings demonstrate that the ability of WNN to identify subpopulations that are masked by scRNA-seq alone is not limited to immune or CITE-seq datasets.”

Even within the same dataset, the relative utility of each modality to define cell states may vary across individual cells.

“Even within the same dataset, the relative utility of each modality to define cell states may vary across individual cells.”

WNN results were robust to a range of values for k.

“We confirmed that our results were robust to a range of values for k (Figure S2)”

The anchor-based workflow enables cells to cluster together based on shared biological state, as opposed to sample-of-origin.

“We applied our “anchor-based” workflow (Stuart et al., 2019) to first integrate the samples together, enabling cells to cluster together based on their shared biological state, as opposed to sample-of-origin (STAR Methods).”

• cites Stuart et al., 2019Not checked yet

  Not checked in the free run.

The signal-to-noise for each antibody can be optimized for any individual marker as a function of antibody concentration.

“In addition, we have recently demonstrated that the signal-to-noise for each antibody can be optimized for any individual marker as a function of antibody concentration”

• cites Stoeckius et al., 2018Not checked yet

  Not checked in the free run.

The cell-type-specific differences across modalities could be predicted by the composition of the CITE-seq panel, which contained anti-CD4 and anti-CD8 antibodies but lacked markers to discriminate cDCs.

“With biological foresight, the cell-type-specific differences across modalities could be predicted by the composition of the CITE-seq panel, which contained anti-CD4 and anti-CD8 antibodies but lacked any immunophenotypic markers to discriminate cDCs.”

Protein data have particular utility when defining the state of CD8+ T cells.

“This reflects the particular utility of protein data when defining the state of these cells.”

Incorporation of protein information in the WNN graph does not come at the expense of identifying transcriptomically congruent neighborhoods.

“and the incorporation of protein information in the WNN graph does not come at the expense of identifying transcriptomically congruent neighborhoods (Figure S2; STAR Methods).”

CITE-seq data quality does not deteriorate with greater amounts of total antibody.

“and we have shown that CITE-seq data quality does not deteriorate with greater amounts of total antibody (Stoeckius et al., 2018).”

• cites Stoeckius et al., 2018Not checked yet

  Not checked in the free run.

The protein data in CITE-seq may represent the most informative modality when the antibody panel comprehensively represents all cell subsets with high specificity.

“The protein data in CITE-seq may therefore represent the most informative modality, particularly in cases where the antibody panel comprehensively represents all cell subsets with high specificity.”

The inability to separate T cell populations by scRNA-seq reflects technical challenges driven by minimal RNA content of T cells coupled with high RNase expression.

“This reflects technical challenges driven by the minimal RNA content of T cells coupled with high RNase expression (Andreeff et al., 1978; Lu et al., 2018; Sercan Alp et al., 2015), which hampers scRNA-seq data quality.”

• cites Andreeff et al., 1978Not checked yet

  Not checked in the free run.

• cites Lu et al., 2018Not checked yet

  Not checked in the free run.

• cites Sercan Alp et al., 2015Not checked yet

  Not checked in the free run.

Because results are consistent in both pre- and post-vaccination time points, they likely describe general characteristics of the healthy immune system.

“Because these results are consistent in both pre- and post-vaccination time points, they likely describe general characteristics of the healthy immune system.”

• AssumptionRests on an assumption: The 24 human volunteers sampled are representative of the healthy immune system.

Even optimally sorted groups may contain a minority of contaminating cells from other states.

“We note that although these panels can achieve high levels of enrichment, even optimally sorted groups may contain a minority of contaminating cells from other states.”

NK cells fall along a two-dimensional gradient defined by CD38 and CD16.

“demonstrating that NK cells fall along a two-dimensional gradient defined by these markers.”

The CD103+ and CD49a+ populations are not TRMs that have temporarily detached and re-entered circulation.

“suggesting that they are not TRMs that have temporarily detached and re-entered circulation.”

Within DC subgroups, cDC2s may perform an important role in the downstream priming and activation of the adaptive immune system during this vaccine response.

“This indicates that within DC subgroups, cDC2s may perform an important role in the downstream priming and activation of the adaptive immune system during this vaccine response.”

This expression pattern is consistent with the development of "adaptive" or "memory-like" NK cells observed in CMV seropositive individuals.

“This expression pattern is consistent with the development of “adaptive” or “memory-like” NK cells observed in CMV seropositive individuals (Lee et al., 2015; Schlums et al., 2015).”

• cites Lee et al., 2015Not checked yet

  Not checked in the free run.

• cites Schlums et al., 2015Not checked yet

  Not checked in the free run.

CITE-seq data can be easily mined to identify the best immunophenotypic marker panels for any subpopulation of interest.

“Importantly, we demonstrate that CITE-seq data can be easily mined to identify the best immunophenotypic marker panels for any subpopulation of interest.”

The sPCA procedure can identify the set of principal components that can transform the data in a single modality to best capture the structure in a multimodal dataset.

“Therefore, the sPCA procedure can identify the set of principal components that can transform the data in a single modality to best capture the structure in a multimodal dataset.”

KNN-graphs are prone to the formation of spurious edges, which represent links between cells that share some similarity in molecular profiles but are not in a matched molecular state.

“Recent work has clearly demonstrated that KNN-graphs are prone to the formation of spurious edges, which represent links between cells that share some similarity molecular profiles, but are not in a matched molecular state (Levine et al., 2015).”

• cites Levine et al., 2015Not checked yet

  Not checked in the free run.

The modality weights learned in the WNN procedure may also reflect the biological importance of each modality in determining cellular identity.

“Last, we note that the modality weights learned in our procedure serve not only as a proxy for the technical quality of a measurement type, but may also reflect the biological importance of each modality in determining cellular identity.”

• AssumptionThe authors state this as an assumption: WNN assumes that modalities do not define conflicting sets of cell states.

The modality weights learned in the WNN procedure serve as a proxy for the technical quality of a measurement type.

“Last, we note that the modality weights learned in our procedure serve not only as a proxy for the technical quality of a measurement type, but may also reflect the biological importance of each modality in determining cellular identity.”

• AssumptionThe authors state this as an assumption: WNN assumes that modalities do not define conflicting sets of cell states.

Flow cytometry measurements were used to validate changes in cell type abundance detected by scRNA-seq.

“We used these measurements in Figures 6G and 6H to validate changes in cell type abundance that were detected by scRNA-seq.”

Performing downstream analyses on a weighted combination of data types improves the ability to characterize cellular diversity.

“We demonstrate throughout this manuscript that performing downstream analyses on a weighted combination of data types improves our ability to characterize cellular diversity.”

Spurious edges in KNN-graphs can be identified through the use of the Jaccard metric, which identifies the number of shared nearest neighbors between two cells.

“However, it is possible to identify these spurious edges through the use of the Jaccard metric. This identifies the number of shared nearest neighbors between two cells, thereby exploiting the local density of each data point to separate well-supported from spurious edges.”

The sPCA components should represent the optimal transcriptomic modules that can be used to separate the cell types defined in the multimodal dataset.

“These components should represent the optimal transcriptomic modules that can be used to separate the cell types defined in our multimodal dataset.”

The WNN method can be generalized to analyze single-cell datasets with three or more simultaneously measured modalities.

“The WNN method can be generalized to analyze single-cell datasets with three or more simultaneously measured modalities.”

The sPCA transformation should be more capable of separating cell types in the query dataset than unsupervised PCA.

“This transformation should therefore be more capable of separating cell types in the query dataset.”

Projecting the sPCA transformation onto a query dataset places the query in the same low-dimensional space as the reference.

“As a secondary benefit, projecting the sPCA transformation onto a query dataset places the query in the same low-dimensional space as the reference.”

sPCA can capture both discrete and continuous sources of variation in a multimodal dataset.

“Therefore, it can capture both discrete and continuous sources of variation in a multimodal dataset.”

The sPCA transformation computed on the reference dataset can subsequently be projected onto any query dataset consisting of PBMC.

“We compute the sPCA transformation described above for our reference dataset, and can subsequently project this transformation onto any query dataset consisting of PBMC.”

One advantage of the de-novo UMAP approach is that it can help visualize populations in the query that cannot be effectively represented in the reference.

“One advantage of the de-novo approach is that it can help to visualize populations in the query that cannot be effectively represented in the reference.”

Withholding protein data during mapping allows protein measurements to be used as an independent assessment of prediction quality.

“By withholding the protein data from consideration during the mapping process, we can use the protein measurements as an independent assessment of prediction quality.”

If a cell's protein levels correlate highly with the CD4 Treg centroid but poorly with the NK centroid, this suggests that the Treg annotation is correct.

“If the cell’s protein levels exhibit a high correlation with the centroid of CD4 Treg, but a low correlation with the centroid of NK, this suggests that the Treg annotation is correct.”

Projecting the sPCA transformation enables supervised analysis of query datasets.

“This enables us to perform supervised analysis of the query datasets.”

Together, the methods enable a fully automated pipeline to leverage a multimodal single-cell reference to annotate and visualize new single-cell query datasets, even if only the transcriptome was measured.

“Together, these methods enable a fully automated pipeline to leverage a multimodal single-cell reference to annotate and visualize new single-cell query datasets, even if only the transcriptome was measured.”

sPCA can help pinpoint sources of variation that are of the greatest interest.

“While PCA will identify the directions that explain maximal variance in the source data, sPCA can help pinpoint sources of variation that are of the greatest interest.”

Because the sPCA was computed on a reference defined by a large number of cells and antibodies, the transformation will likely be more informative than an unsupervised PCA computed on a new scRNA-seq query.

“Since our sPCA was computed based on a reference defined by a large number of cells and antibodies, this transformation will likely be more informative than an unsupervised PCA computed on a new scRNA-seq query.”

At the conclusion of the low-dimensional correction procedure, the query dataset is placed in the same low-dimensional representation as the reference.

“Moreover, at the conclusion of this procedure, we have placed the query dataset in the same low-dimensional representation (defined by the sPCA transformation) as the reference.”