Comprehensive integration of single cell data
Tim StuartiD, Andrew ButleriD, Paul HoffmaniD, Christoph HafemeisteriD, Efthymia PapalexiiD, William M. MauckiD, Yuhan HaoiD, Marlon StoeckiusiD, Peter SmibertiD, Rahul SatijaiD
Abstract
Single cell transcriptomics (scRNA-seq) has transformed our ability to discover and annotate cell types and states, but deep biological understanding requires more than a taxonomic listing of clusters. As new methods arise to measure distinct cellular modalities, including high-dimensional immunophenotypes, chromatin accessibility, and spatial positioning, a key analytical challenge is to integrate these datasets into a harmonized atlas that can be used to better understand cellular identity and function. Here, we develop a computational strategy to \"anchor\" diverse datasets together, enabling us to integrate and compare single cell measurements not only across scRNA-seq technologies, but different modalities as well. After demonstrating substantial improvement over existing methods for data integration, we anchor scRNA-seq experiments with scATAC-seq datasets to explore chromatin differences in closely related interneuron subsets, and project single cell protein measurements onto a human bone marrow atlas to annotate and characterize lymphocyte populations. Lastly, we demonstrate how anchoring can harmonize in-situ gene expression and scRNA-seq datasets, allowing for the transcriptome-wide imputation of spatial gene expression patterns, and the identification of spatial relationships between mapped cell types in the visual cortex. Our work presents a strategy for comprehensive integration of single cell data, including the assembly of harmonized references, and the transfer of information across datasets.\n\nAvailability: Installation instructions, documentation, and tutorials are available at: https://www.satijalab.org/seurat
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 15,454
- FWCI
- 514.80
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (2 Nov 2018) · CC BY 4.0
Seurat v3 finds scored mutual-nearest-neighbor anchors in canonical correlation space to integrate scRNA-seq batches and transfer labels, protein levels and transcriptome-wide expression across modalities.
Prior integration approaches based on canonical correlation or mutual nearest neighbors targeted scRNA-seq only, could struggle when datasets shared just some cell types or technical effects masked biology, and could not transfer information between datasets. Seurat v3 provides one anchor framework for atlas assembly and reference-to-query transfer of labels and continuous features, letting scRNA-seq complement scATAC-seq and gene-limited in situ methods such as STARmap.
Key results
ShownAnchor-based integration merged scRNA-seq batches across donors, technologies and labs while keeping non-shared populations distinct, beating Seurat v2, mnnCorrect and Scanorama on silhouette score and scaling to an 18-tissue mouse atlas exposing rare cell types.
Eight human islet datasets (27 donors, five technologies, four labs), one cell type withheld per dataset: Seurat v3 gave the highest silhouette scores vs Seurat v2, mnnCorrect and Scanorama, likewise on six mouse bipolar batches; erroneous anchors scored lower. Tabula Muris (97,029 cells, 18 tissues) yielded plasmacytoid dendritic (0.07%) and mesothelial (0.05%) cells. · Constructing integrated atlases at the scale of…; Anchor scoring; Identification of rare subtypes in the Tabula…; Integration with simulated cell type holdouts
ShownWeighted anchor voting labeled query cells more accurately than scmap-cluster and scmap-cell across benchmark splits, and its prediction scores flagged reference-absent cells and misclassifications, as illustrated in an alpha-cell holdout example.
166 evaluation cases from human islet and mouse bipolar data, each withholding one cell type from the reference (20% of a query capped at 100 cells per type): Seurat v3 had the highest accuracy vs both scmap variants; with alpha cells withheld, query alpha cells got the lowest scores (labeled Unassigned, bottom 20%) and errors scored lower. · Leveraging anchor correspondences to classify…; Transferring cell type labels onto scRNA-seq…; Figure legends
ShownAnchors transferred 25 surface-protein levels from a new bone marrow CITE-seq dataset onto the Human Cell Atlas (HCA) marrow resource, which lacks protein data, revealing a CD8+ CD69+ memory T cell subset with an inflammatory, IFN-γ-linked program supported by sorted-cell bulk RNA-seq.
Within-CITE-seq cross-validation (33,454 cells, 25 proteins): 22/25 proteins strongly correlated, median R=0.852, saturating near 250-500 RNA features. Transfer to 274,599 HCA cells (eight donors): CD69-linked gene module in all eight donors, IFN-γ response enrichment P < 10^−11; top markers differentially expressed in sorted-cell bulk RNA-seq (four replicates each). · Predicting protein expression in human bone…; CITE-seq cross-validation; Protein expression transfer to the HCA; Validation of CD69+ T cell population
ShownAnchoring Allen SMART-Seq4 scRNA-seq to STARmap imputed transcriptome-wide spatial expression in mouse visual cortex and mapped neuronal subtypes in situ, with intratelencephalic (IT), pyramidal tract (PT) and corticothalamic (CT) neurons differing in laminar position within a layer.
45,768 transcripts from 14,249 cells were imputed onto STARmap replicates (1,539 and 890 cells, 1,020 genes); leave-one-out predictions matched measured Lamp5, Cux2, Sst and Bsg, and osmFISH supported Rorb and Syt6; 1,915 (79%) cells scored > 0.5 and 1,210 (50%) were kept; IT vs PT (layer 5) and IT vs CT (layer 6) laminar differences recurred in the second replicate. · Spatial mapping of single cell sequencing data…; Preprocessing STARmap data; Data transfer; Analysis of integrated STARmap data
SuggestedTransferring scRNA-seq labels onto mouse prefrontal cortex scATAC-seq via gene-activity scores split interneurons into Sst, Pvalb (PV), Vip and Lamp5 groups; peaks more accessible in PV than SST cells were most enriched for Mef-family motifs (including Mef2c), then Rora.
Anchors used the gene-activity matrix (dims 1:20, CCA) against 14,249 visual cortex SMART-Seq4 cells; 2,548 nuclei scored > 0.5 into 17 clusters (eight excitatory, four inhibitory); 96% of previously annotated inhibitory cells stayed inhibitory; top 1,000 PV-biased peaks gave Mef2c motif p < 10^-21, Rora p < 10^-7 (Homer). · Projecting cellular states across modalities; Transferring cell type labels onto scATAC-seq…; Identification of differentially accessible…; Figure legends
Methods, data and reproducibility
Diagonalized canonical correlation analysis (CCA) with L2 normalization embeds two datasets jointly; mutual nearest neighbors (anchors) are filtered against original expression, scored by neighborhood overlap and distance-weighted per query cell, yielding correction vectors for integration (merged along a guide tree) or weighted label/feature transfer.
- Eight published human pancreatic islet scRNA-seq datasets spanning CelSeq, CelSeq2, Fluidigm C1, SMART-Seq2 and inDrops · human · n = 27 donors; eight datasets; five technologies; four labs · GSE81076; GSE85241; GSE86469; E-MTAB-5061; GSE84133
- Mouse retinal bipolar cell scRNA-seq UMI matrix (previously annotated cells kept), used for integration and classification benchmarks · mouse · n = six batches · GSE81904
- Tabula Muris plate-based (SMART-Seq2) and droplet (10x Genomics) atlases from FigShare (Version 1 release) · mouse · n = 97,029 cells; 18 tissues (12 shared, 6 SMART-Seq2 only)
- Human Cell Atlas (HCA) healthy bone marrow scRNA-seq from the HCA Data Portal preview site; doublets removed with Scrublet · human · n = 274,599 cells; 8 donors
- New CITE-seq of human bone marrow mononuclear cells (single donor, AllCells; 25 antibodies, 10 hashtags) plus bulk RNA-seq of FACS-sorted CD8+ CD69+/- T cells · human · n = 33,454 CITE-seq cells; 4 bulk replicates per sorted population
- Allen Brain Atlas SMART-Seq4 scRNA-seq of mouse visual cortex (Low Quality and No Class cells removed), used as transfer reference · mouse · n = 14,249 cells; 45,768 transcripts
- Mouse prefrontal cortex scATAC-seq gene-activity and binarized peak matrices from the SOMA Data Portal, part of a 13-tissue resource · mouse · n = 3,482 cells reported; 2,548 classified (score > 0.5)
- STARmap 1,020-gene in situ datasets of mouse visual cortex (two replicates) plus osmFISH of mouse somatosensory cortex as validation · mouse · n = 1,539 and 890 cells; 1,020 genes
- Methods & tools
- Seurat v3 (FindIntegrationAnchors, IntegrateData, FindTransferAnchors, TransferData), Diagonalized CCA approximated via irlba (implicitly restarted Lanczos bidiagonalization), RANN approximate nearest-neighbor search, Seurat v2, mnnCorrect (scran) and Scanorama as integration baselines, scmap-cluster and scmap-cell as classification baselines, CITE-seq with BioLegend TotalSeq antibodies and cell hashing, 10x Genomics 3' scRNA-seq, CellRanger v2.1.0, CITE-seq-Count and HTOdemux, Scrublet doublet removal, DESeq2, mixtools normalmixEM and GOstats, Homer findMotifsGenome.pl (mm9), pysam and deepTools bamCoverage (RPKM) for pseudo-bulk ATAC tracks, Moran's I (ape), silhouette (cluster R package), UMAP
- Code
- https://satijalab.org/seurathttps://github.com/Hoohm/CITE-seq-Counthttps://github.com/pysam-developers/pysamhttps://github.com/weallen/STARmap
- Data & accessions
- https://www.starmapresources.com/data/
- Software & environment
- R/Seurat v3 with irlba, RANN, scran, scmap, mixtools, GOstats, ape; CellRanger v2.1.0 (GRCh38); Homer (mm9); AllCells bone marrow (ABM007F, lot 3008803); BioLegend TotalSeq antibodies, hashtags, FcX (422302); FACS clones HIT3a, RPA-T4, RPA-T8, FN50
- Assay conditions & compute
- No CPU/GPU, runtime or memory figures reported; wet-lab: 10x Genomics 3' scRNA-seq, SONY SH800 sorter, antibody pool (1µg each) for 30 min at 4°C, hashing 20 min, FcX block 10 min, 40µm filtering
- Reproducibility gaps
- No accession or URL is given for the newly generated CITE-seq and sorted-cell bulk RNA-seq data; custom pysam code for splitting scATAC-seq BAMs is not shared; HCA, Allen, SOMA and osmFISH data are cited by portal name only, HCA via a preview site.
Limitations (6)
- Transfer onto scATAC-seq requires first converting accessibility into a predicted gene-activity matrix that assumes accessibility tracks expression; the authors expect anchors to fail where chromatin changes precede transcription, as in developing systems. · stated by the authors
- Three of 25 proteins were poorly imputed in CITE-seq cross-validation, attributed by the authors to weak antibody specificity or no transcriptomic markers tracking those immunophenotypes; residual error elsewhere includes background binding, stochastic expression and noise. · stated by the authors
- Dimensionality (number of canonical or principal components; 20 to 100 here) was chosen per data type rather than finely tuned, and the authors acknowledge that a robust, fully unsupervised way to estimate it remains a fundamental challenge. · stated by the authors
- For weakly detected STARmap genes (e.g., Rorb, Syt6), imputed spatial patterns were stronger than measured ones; the authors list imputation error, alongside stochastic expression and STARmap technical noise, as a possible explanation. · stated by the authors
- The scATAC-seq query was prefrontal cortex while the scRNA-seq reference was visual cortex, risking forced assignment of region-specific neuron types; finer subclass calls rest on marker-locus accessibility and coarse-label agreement rather than orthogonal ground truth. · noted in this summary
- Protein imputation was cross-validated only by randomly splitting one donor's CITE-seq run, so accuracy of cross-donor transfer onto the HCA atlas is not directly quantified; the CITE-seq cell count also differs between text (33,454) and Figure 4 legend (35,543). · noted in this summary
Who should read:Read this if you need the algorithmic basis of Seurat v3 anchors for integrating scRNA-seq batches or transferring labels, protein levels or expression onto scATAC-seq, atlas or in situ data.
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
- Integrated analysis of single cell transcriptomic data across conditions, technologies, and species
Andrew Butler, Paul Hoffman, … Rahul Satija·Nature BiotechnologyPeer-reviewed·Apr 2, 2018
- scJoint: transfer learning for data integration of atlas-scale single-cell RNA-seq and ATAC-seq
Yingxin Lin, Tung-Yu Wu, … Y. X. Rachel Wang·Nature BiotechnologyPeer-reviewed·Jan 20, 2022
- Benchmarking algorithms for joint integration of unpaired and paired single-cell RNA-seq and ATAC-seq data
Michelle Y. Y. Lee, Klaus H. Kaestner, Mingyao Li·Genome biologyPeer-reviewed·Oct 24, 2023
- Atlas-Level Single-Cell and Spatial Transcriptomics Data Integration via PRIME
Xinchao Wu, Xuefei Wang, … Shibiao Wan·bioRxiv·Jun 4, 2026
- Online single-cell data integration through projecting heterogeneous datasets into a common cell-embedding space
Lei Xiong, Kang Tian, … Qiangfeng Cliff Zhang·Nature CommunicationsPeer-reviewed·Oct 17, 2022
- Panoramic stitching of heterogeneous single-cell transcriptomic data
Brian Hie, Bryan D. Bryson, Bonnie Berger·Nature BiotechnologyPeer-reviewed·May 6, 2019