baSeq docs › Multiome (RNA + ATAC)

Multiome (RNA + ATAC)

A multiome analysis takes the RNA and the chromatin of the same cells (10x Multiome), checks every cell on both, clusters them on both together, and finds what only the two together can say: which peaks regulate which genes, which transcription factors drive each cell type, and the network they make.

Start a multiome analysis

In a project, click New session and choose Multiome (RNA + ATAC). Use it when both measurements come from the same cells (10x Chromium Single Cell Multiome ATAC + Gene Expression, SHARE-seq, SNARE-seq). For chromatin alone choose scATAC-seq; for RNA alone, single-cell RNA-seq.

What to upload

Per sample, what Cell Ranger ARC leaves in its outs folder, any one of:

The matrix with the fragments, metrics and peaks (best): filtered_feature_bc_matrix.h5 (genes and peaks; its barcodes are ARC's cell calls), atac_fragments.tsv.gz, per_barcode_metrics.csv and atac_peaks.bed. Every cell's insertions are counted in the peaks and its ATAC quality measured from its fragments, which are kept for the coverage tracks.
The filtered_feature_bc_matrix folder (matrix.mtx, barcodes.tsv, features.tsv), with the fragments for the ATAC's quality.
A MuData file (.h5mu, from muon or scverse) with rna and atac modalities, or an .h5ad of both side by side (a feature_types column of GEX and ATAC, raw counts in layers['counts'], as the NeurIPS 2021 data is shared); its cells' metadata and samples are kept.

Several samples: drop each sample's outs folder, or start each file's name with its sample (PBMC1_atac_fragments.tsv.gz), or upload a .zip. The cells kept are those with both modalities; the upload says how many had one only. A Multiome run's two libraries share their barcodes, so a run processed with Cell Ranger and Cell Ranger ATAC apart can be uploaded too (the gene expression .h5 with the ATAC's files).

One graph from two measurements

The RNA goes through the single-cell RNA steps (normalization, variable genes, PCA), the ATAC through scATAC-seq's TF-IDF and LSI, and Harmony corrects each for batches on its own. Multimodal Neighbors then joins them with WNN (Hao et al. 2021): for every cell, how well each modality's neighbours predict it decides its weight for each, so a cell type whose chromatin is more distinct than its RNA (naive T cell subsets) leans on its ATAC, and one whose RNA is (cycling cells, monocytes) on its RNA. The weights are drawn on the joint UMAP. MultiVI, a deep generative model of both, is the choice for many batches, on the GPU. Clustering and every RNA step after it read the joint graph.

Linking peaks to genes

A correlation between a peak's accessibility and a gene's expression across cells is the usual evidence that the peak regulates the gene (ArchR's peak-to-gene links, Signac's LinkPeaks). It is weak evidence alone: genes and peaks of one cell type correlate with each other wherever they are in the genome. Link Peaks to Genes scores a pair by its correlation across metacells of ten similar cells and within cell types, discounted by the ABC model's contact with distance. Chosen on the genetics: of the peaks in 10x's PBMC multiome runs holding a fine-mapped eQTL variant (OneK1K, eQTL Catalogue), this score picked the variant's gene with an average precision of 0.61 (0.07 by chance, 0.55 for the nearest gene, 0.26 for the correlation alone, 0.21 for Signac's z-score against background peaks), and the links it calls name the variant's gene for 67–68% of the peaks they link. The step also says how many links co-variation alone would make, from the same score of genes paired with peaks of other chromosomes.

What only both together can say

Regulatory Domains: genes with ten or more linked peaks (DORCs, Ma et al. 2020), the cell types' identity genes, and each cell's chromatin potential: an arrow to the cells whose RNA its chromatin looks like. Where chromatin opens before the genes are expressed the arrows point forward along a path; where it stays open after them, back: on 10x's embryonic mouse cortex they point from the intermediate progenitors back to the radial glia, whose Notch domains stay open after their genes turn off. Read them with the order of the states known.
TF Drivers: chromVAR says a motif is open, but a family shares a motif; the TF expressed where its motif opens is the one at work (ArchR's positive TF regulators).
Regulatory Network: eRegulons. A TF's co-expressed genes follow cell types too; the network keeps a TF's module only when its motif is enriched in those genes' own linked peaks against other genes' (GC-matched), and its targets are the genes with the motif there: TF, peak, gene. Each eRegulon has an activity per cell from its genes and from its peaks.

Not yet

FASTQ files (Cell Ranger ARC has to be run first), scRNA-seq and scATAC-seq of different cells integrated (label transfer), and footprinting are not part of the multiome analysis yet.

Step reference

Step	Stage	What it does	You get	Key inputs
Initial Plot	Setup	Reads Cell Ranger ARC's output sample by sample: the RNA from the matrix, the ATAC from the fragments (each cell's insertions counted in the peaks and its quality measured, as scATAC-seq does); keeps the cells with both, and their fragments beside the data.	Cells per sample, each cell's RNA UMIs against its ATAC fragments and its TSS enrichment, the fragment sizes and TSS profile.	The files.
Cell Quality (RNA + ATAC)	QC	Keeps a cell when both its RNA (UMIs, genes, mitochondrial share) and its chromatin (fragments, TSS enrichment, fraction in peaks, nucleosome signal, blacklist ratio) pass; upper depth limits five MADs above each sample's median.	Every cell's RNA against its ATAC depth with the cut-offs (kept, failing on one, on both), each measure per sample, the cells kept.	Ten cut-offs.
Remove Doublets (RNA + ATAC)	QC	Scores every cell on its RNA (Scrublet) and its ATAC (ArchR's simulated doublets), combines their ranks, and removes as many of the highest as 10x's rule of thumb expects.	Doublets per sample, the two scores against each other, the doublets' depth.	Doublets from.
Remove Ambient RNA, Normalize Data, Variable Genes, PCA	QC	The single-cell RNA steps, on the RNA. Remove Ambient RNA takes each sample's background RNA out of the counts with SoupX's method, its make-up from the cells' pooled reads (an ARC upload brings no empty droplets), or DecontX's; the peak-gene links and networks correlate the raw counts.	As in single-cell RNA-seq.	As there.
TF-IDF and LSI (ATAC)	QC	Signac's latent semantic indexing of the ATAC, the components that follow depth left out.	Each component's correlation with depth, the singular values, the first two used.	Peaks used, Components.
Peak Sequences and Motifs	QC	scATAC-seq's step on the ATAC: each peak's sequence, GC and JASPAR motifs.	GC content, GC bias per sample, the most common motifs.	None.
Perform Integration (RNA + ATAC)	QC	Harmony on the RNA's PCA and on the ATAC's LSI by batch; each modality's own graph and UMAP.	Each modality's UMAP before and after, and how well the batches mix.	Correct for, Dimensions.
Multimodal Neighbors	QC	One graph from both modalities: WNN (Seurat's weighted nearest neighbours: each cell's weight for its RNA and its ATAC), or MultiVI on the GPU.	The same cells on the RNA, ATAC and joint UMAPs, which modality defines each cell, samples on the joint map.	Method, Dimensions.
Explore UMAPs, Perform Clustering	Analysis	The joint UMAP coloured by both modalities' quality; Leiden on the joint graph at three resolutions.	The clusters, which modality defines each, their samples.	None.
Annotation, DGE, pseudobulk, gene plots, GSEA, trajectories	Analysis	The single-cell RNA steps, on the RNA, with the joint graph's clusters.	As in single-cell RNA-seq.	As there.
Marker Peaks, Differential Accessibility, TF Activity, Motif Enrichment	Analysis	The scATAC-seq steps, on the ATAC, with the same clusters.	As in scATAC-seq.	As there.
Link Peaks to Genes	Analysis	Which peaks regulate which gene: how a peak's accessibility follows the gene's expression across metacells and within cell types, weighed by the ABC model's contact with distance; benchmarked against eQTLs.	The links' accessibility and their genes' expression across groups, the genes with most links, the distances.	Clusters or cell types, Window, Least score.
Coverage, Links and Expression at Genes	Analysis	Signac's CoveragePlot with links and expression: each group's accessibility around a gene, its links as arcs, the gene's expression per group beside them.	One block per gene.	Clusters or cell types, Window; genes in your own words.
Regulatory Domains and Chromatin Potential	Analysis	The genes with ten or more linked peaks (DORCs, Ma et al. 2020), and each cell's chromatin potential: which state's RNA its chromatin looks like.	Genes by their links, DORC accessibility and expression per group, arrows on the UMAP, how far each group's chromatin is from its RNA.	Clusters or cell types, Linked peaks that make a DORC.
TF Drivers	Analysis	The TFs whose expression follows their motif's chromVAR activity (activators) or its closing (repressors), as ArchR finds positive TF regulators.	Each TF by its correlation and its motif's spread, the drivers' activity and expression per group, the strongest on the UMAP.	Clusters or cell types, Correlation.
Regulatory Network	Analysis	eRegulons: each TF's co-expressed genes, kept when its motif is enriched in their linked peaks; its targets those with the motif there (SCENIC's pruning on SCENIC+'s regions). An activity per cell from the genes and from the regions.	The eRegulons, their activity per group from genes and regions, the network, the strongest on the UMAP.	Correlation.

Checked on simulated cells and public data

The steps are tested on simulated cells with a known answer, written as Cell Ranger ARC writes them: four samples of four cell types, each type opening its own sites (holding its own motif) and expressing its marker genes, whose expression follows the same cell's insertions in its type's sites inside them; a TF gene for each motif, expressed in its type; a decoy TF expressed in one type whose motif is in none of its sites; a gene of one type far from any of its sites; planted doublets and low-quality cells; treated cells of one type opening 80 more sites. Cell Quality and Remove Doublets leave out every doublet and low-quality cell, the joint graph's clusters are the types (adjusted Rand index 1.0), Link Peaks to Genes finds all 38 planted links and nothing strong for the far gene, TF Drivers names the four motifs' TFs and no other as activators, each TF's eRegulon holds its type's marker genes and the decoy has none, the regulatory domains are the marker genes, and Differential Accessibility finds the treated sites. Every step runs with the figures of the plot kit and with the older ones.

Every step was run in real kernels on four public datasets and every figure was read: 10x Genomics' PBMC 3k and 10k multiome runs, 10x's embryonic (E18) mouse brain, and the NeurIPS 2021 multiome bone marrow (69,249 cells of 13 donors at 4 sites, with the cell types its organisers annotated). On the bone marrow, clusters of the joint graph matched the experts' 22 cell types better than either modality's alone at resolutions 1.0 and 1.5 (adjusted Rand index 0.69 against 0.59 for the RNA and 0.57 for the ATAC at 1.0; at 0.5 the RNA alone did as well, 0.71 against 0.69), and Harmony raised the share of each cell's neighbours from another site from 26% to 58% (RNA) and from 30% to 60% (ATAC), of 75% were they mixed, while 81% and 77% stayed of the same cell type. The doublets the two modalities agree on: 659 cells in the top 5% of both, where chance puts 151.

Peak-gene links were chosen against fine-mapped eQTLs of blood cell types (OneK1K, eQTL Catalogue): see Linking peaks to genes above. TF Drivers and the Regulatory Network found each tissue's regulators without being told them: on the PBMCs SPI1 and the CEBPs over monocyte genes (CSF1R, LYZ, MNDA), TBX21 and EOMES over NK cells' (KLRD1, GZMH, PRF1), BCL11B over T cells' (CD3D, ITK, LEF1), TCF3 over B cells' (BLK, MS4A1, CD22) and IKZF1 repressing the myeloid genes; on the bone marrow GATA1 and TAL1 over the erythroid program (SPTA1, SLC4A1, ALAS2, KEL; their motifs 4-fold enriched, q below 1e-120), EBF1 over B cells' (CD79B, BANK1, BLK), IRF8 over plasmacytoid DCs' (CLEC4C) and TCF7 over naive T cells' (CCR7, CD28); on the mouse brain Sox2, Eomes, Neurod1, Tbr1, Mef2c, Lhx2 and the Nfi factors, with the E2Fs over the replication genes (Mcm3, Mcm10, Cdt1) and Mef2c over the synaptic ones (Syt1, Scn2a). Its regulatory domains were the Notch pathway (Notch1, Notch2, Lfng) and Neurod2. How many links co-variation alone could make was 30% to 48% of the real ones on blood and marrow, and 63% on the developing brain, whose cell states change continuously.

Bone marrow shared already filtered may lose cells to the default floor of 500 UMIs (5,755 of the 69,249 here): lower it in Cell Quality. MultiVI ran on a CPU on the PBMC 3k: its clusters agreed with the RNA's alone at an adjusted Rand index of 0.56 and the ATAC's at 0.46, where WNN's agreed at 0.92 and 0.78, so WNN is the default; MultiVI is meant for many cells and batches, and was not checked on those, nor on the GPU. Conditions with a known biological difference have not been compared on real data yet.