henhaobio
Sign in

Nature BiotechnologyGenomicsPeer-reviewed journal article

Dictionary learning for integrative, multimodal, and scalable single-cell analysis

Yuhan HaoiD, Tim StuartiD, Madeline H. KowalskiiD, Saket ChoudharyiD, Paul HoffmaniD, Austin HartmaniD, Avi SrivastavaiD, Gesmira MollaiD, Shaista MadadiD, Carlos Fernandez‐GrandaiD, Rahul SatijaiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Mapping single-cell sequencing profiles to comprehensive reference datasets represents a powerful alternative to unsupervised analysis. Reference datasets, however, are predominantly constructed from single-cell RNA-seq data, and cannot be used to annotate datasets that do not measure gene expression. Here we introduce bridge integration, a method to harmonize singlecell datasets across modalities by leveraging a multi-omic dataset as a molecular bridge. Each cell in the multi-omic dataset comprises an element in a dictionary, which can be used to reconstruct unimodal datasets and transform them into a shared space. We demonstrate that our procedure can accurately harmonize transcriptomic data with independent single cell measurements of chromatin accessibility, histone modifications, DNA methylation, and protein levels. Moreover, we demonstrate how dictionary learning can be combined with sketching techniques to substantially improve computational scalability, and harmonize 8.6 million human immune cell profiles from sequencing and mass cytometry experiments. Our approach aims to broaden the utility of single-cell reference datasets and facilitate comparisons across diverse molecular modalities. AvailabilityInstallation instructions, documentations, and vignettes are available at http://www.satijalab.org/seurat

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 1 (26 Feb 2022) · CC BY-NC-ND 4.0

Dictionary learning enables assumption-free cross-modality single-cell mapping via multi-omic bridge datasets and scales single-modality integration to millions of cells in under an hour via atomic sketching.

Existing cross-modality single-cell integration tools rely on simplifying biological assumptions such as equating gene-body chromatin accessibility with transcription, while atlas-scale integration struggles when full datasets cannot fit in memory. Treating multi-omic cells or leverage-score-sampled cells as dictionary atoms aligns disparate modalities without feature-conversion assumptions and scales integration to millions of cells.

Key results

  1. ShownBridge integration mapped human bone marrow scATAC-seq profiles onto an scRNA-seq reference, resolving rare populations and 236 chromatin-primed lagging loci along myeloid differentiation.

    Using a 32,368-cell 10x Multiome bridge, 26,159 scATAC-seq profiles (16,266 whole bone marrow, 9,893 CD34+ enriched) mapped onto a 297,627-cell Azimuth scRNA-seq reference (35 states), identifying innate lymphoid cells (0.15%) and AXL+SIGLEC6+ dendritic cells (0.10%) and 236 cell-cycle/DNA-replication-enriched loci where accessibility precedes expression. · Mapping scATAC-seq data onto scRNA-seq…; Figure legends

  2. ShownBridge integration outperformed variational autoencoder methods in cross-modality cell matching accuracy and runtime across chromatin accessibility and histone modification benchmarks.

    On split 10x Multiome bone marrow data and mouse brain Paired-Tag (scRNA-seq with H3K27ac, H3K27me3, or H3K4me1 scCUT&Tag) replicates, bridge integration achieved higher ground-truth Jaccard similarity than multiVI and Cobolt and ran in 0.8 hours (plus 1.2 hours preprocessing) versus 3.3 hours for Cobolt and 15.7 hours for multiVI, requiring >=50 bridge cells per cell type for robust mapping. · Robustness and benchmarking analysis; Figure legends

  3. ShownAtomic sketch integration harmonized 1,525,710 human lung scRNA-seq profiles across 19 studies on a single core in 55 minutes, uncovering ultra-rare pulmonary ionocytes in 17 studies.

    Using 5,000 leverage-score-sketched atom cells per dataset (95,000 total atoms), atomic sketch integration integrated 1,525,710 lung and upper airway cells in 55 minutes on 1 core, identifying FOXI1+ pulmonary ionocytes (0.047% abundance) in 17 of 19 studies (versus 6 of 19 originally annotated) and 116 positive ionocyte markers via pseudobulk differential expression. · Community-scale integration for human lung…; Figure legends

  4. ShownCombining atomic sketch integration and CITE-seq bridge integration linked 3,461,171 human blood scRNA-seq profiles with 5,170,249 mass cytometry (CyTOF) cells.

    Atomic sketch integration harmonized 3,461,171 peripheral blood scRNA-seq profiles (14 studies, 639 donors) into 30 states, and bridge integration via a 161,764-cell CITE-seq bridge (30 shared surface proteins) mapped 5,170,249 CyTOF cells (119 donors), resolving innate lymphoid cells (0.024%) and inferring 17 unshared proteins including FOXP3 in Tregs. · Community-scale integration of scRNA-seq and…; Figure legends

Methods, data and reproducibility

Projects unimodal single-cell datasets into a shared space via linear reconstructions over multi-omic cell atoms compressed by Laplacian eigendecomposition and aligned with mnnCorrect (bridge integration). For unimodal compendia, samples 5,000 leverage-score atoms per dataset, learns reconstructions independently, integrates the atoms, and reconstructs full datasets (atomic sketch integration).

  • Human bone marrow mononuclear cell (BMMC) Azimuth scRNA-seq reference, scATAC-seq query, and NeurIPS 2021 10x Multiome bridge · human · n = 297,627 scRNA-seq; 26,159 scATAC-seq; 32,368 Multiome cells
  • Mouse brain Paired-Tag joint single-cell RNA and histone modification (H3K27ac, H3K27me3, H3K4me1 scCUT&Tag) replicates · mouse
  • Human motor cortex snmC-seq DNA methylation query, snmC2T-seq bridge, and Allen Brain Atlas scRNA-seq reference · human
  • Human lung and upper airway scRNA-seq compendium from 19 studies · human · n = 1,525,710 cells (95,000 sketched atoms, 5,000 per study)
  • Human peripheral blood scRNA-seq compendium (14 studies), healthy donor PBMC CITE-seq bridge, and CyTOF datasets · human · n = 3,461,171 scRNA-seq (639 donors); 161,764 CITE-seq; 5,170,249 CyTOF (119 donors)
Methods & tools
Seurat, Signac, Azimuth, mnnCorrect, multiVI, Cobolt
Assay conditions & compute
Bridge integration of BMMC data took 0.8 hours (plus 1.2 hours preprocessing); atomic sketch integration of 1,525,710 lung cells ran in 55 minutes on a single core [S6, S8].
Reproducibility gaps
Download URLs and accession IDs for the individual public datasets are deferred to Supplementary Methods and Supplementary Table 1, which are omitted from the text [S12].
Limitations (2)
  • Annotation accuracy for rare cell types degrades when fewer than 50 cells of that population are present in the multi-omic bridge dataset, with failure modes observed below 10 cells [S6]. · stated by the authors
  • Constructing atomic dictionaries via uniform random downsampling instead of leverage-score sketching fails to capture and integrate ultra-rare cell populations such as pulmonary ionocytes (0.047% frequency) [S8]. · stated by the authors

Read this if you map single-cell ATAC, methylation, CUT&Tag, or cytometry data onto scRNA-seq references or integrate multi-million-cell transcriptomic atlases.

Citations
5,614
FWCI
499.38
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISZhou '26Patikas '26Wang '25Xiong '22Stuart '19Niu '25
66% semantic matchBioinformatics
Jiahao Zhou, Jing Wang, Shuofeng Hu+10 more·Nature Computational Science·Jul 31, 2026