Integrated analysis of multimodal single-cell data

Key results

1. ShownUnsupervised per-cell modality weights indicate which data type best resolves each cell, so the WNN graph keeps protein-, RNA- or ATAC-defined distinctions that single-modality graphs blur, across CITE-seq, 10x Multiome and SHARE-seq data.

   Wrong CD8+/CD4+ T edges, human cord blood CITE-seq (8,617 cells): RNA 944, protein 12, WNN 20; naive CD8+/CD4+ edges in 11,351 human Multiome PBMCs: RNA 984, ATAC 373, WNN 322. CD8+ T cells got higher protein and DCs higher RNA weights. SHARE-seq mouse skin (34,774 cells): 23 published populations recovered plus a Basal_4 subgroup separating in ATAC but not RNA space. · Quantifying the relative utility of each…; WNN analysis is a robust and flexible approach…; SHARE-seq ATAC+Gene expression dataset of mouse…; Figure legends

2. ShownIn human bone marrow CITE-seq data, WNN neighborhoods predicted measured protein levels as well as or better than MOFA+ and totalVI latent spaces for every antibody, ran faster, and down-weighted protein automatically as simulated antibody noise rose.

   30,672 human bone marrow mononuclear cells, 25 antibodies: correlation (Pearson; Spearman in supplement) of neighbor-averaged vs measured values was highest or tied for WNN across all 25 proteins, most strikingly CD25 and CD57; RNA reconstruction was similar across methods; up to 15-fold faster on full dataset; Gaussian noise added to protein data drove protein weights to 0. · WNN analysis is a robust and flexible approach…; Simulated addition of protein noise; Comparisons with MOFA+ and totalVI; Figure legends

3. ShownA CITE-seq atlas of PBMCs from eight HIV-vaccine-trial volunteers resolved 57 WNN clusters, several of which merged when either modality was analyzed alone; in silico marker selection yielded compact sorting panels, two of them validated by flow sorting plus bulk RNA-seq.

   210,911 cells across 24 samples (161,764 10x 3′ cells with 228 antibodies; 49,147 5′ ECCITE-seq cells with 54 antibodies), annotated in tiers of 8, 30 and 57 categories. One marker gave at least 10-fold in silico enrichment for 45 clusters, three markers for 55; CD43 emerged for cytotoxic CD4+ T cells; bulk RNA-seq of sorted CD4 CTL and CD8_TEM_5 cells matched CITE-seq differentially expressed genes. · A multimodal atlas of the human PBMCs; Antibodies used (all at 5uL per condition unless…; PBMC CITE-seq datasets of HIV Vaccine Trials…; Identifying targeted immunophenotype panels

4. ShownSupervised principal component analysis (sPCA), guided by the WNN graph, plus anchor-based transfer maps RNA-only queries onto the reference, outperforming scArches on protein-checked labels and revealing MAIT depletion in hospitalized COVID-19 patients, confirmed by mass cytometry (CyTOF).

   Kotliarov human CITE-seq query (53,099 cells; 82 proteins withheld): Seurat and scArches agreed on 35,619 cells (67.1%); among 17,480 discordant cells, protein evidence favored Seurat in 79.4% (73.8% in a figure legend). COVID-19 data: T cells split into 12 groups vs six originally; MAIT fraction by mapping vs CyTOF R = 0.911; CyTOF including 16 more samples confirmed depletion. · Mapping query datasets to multimodal references; Benchmarking Seurat reference-mapping with…; Figure legends

5. ShownThe atlas exposed circulating lymphoid substructure, including CD8+ memory T subsets with mutually exclusive CD49a or CD103, cytomegalovirus (CMV)-linked expanded clones and two NK cell gradients, while vaccination elicited day 3 innate responses concentrated in monocytes and cDC2 dendritic cells.

   CD49a/CD103 subsets confirmed by flow cytometry in independent PBMCs; 97% of clones were single cells, 31 expanded clones had at least 10 cells, and five CMV-responsive volunteers held 91% of expanded-clone cells; NK CD16 and CD38 gradients were minimally correlated. CD14+ and CD16+ monocytes shared a 62-gene type I interferon module; Siglec-1 (CD169) protein rose only at day 3. · Multimodal heterogeneity within lymphoid…; Characterizing the initial innate response to…