henhaobio
Sign in

Nature BiotechnologyGenomicsPeer-reviewed journal article

Integrated analysis of single cell transcriptomic data across conditions, technologies, and species

Andrew ButleriD, Paul HoffmaniD, Peter SmibertiD, Efthymia PapalexiiD, Rahul SatijaiD

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

Abstract

Single cell RNA-seq (scRNA-seq) has emerged as a transformative tool to discover and define cellular phenotypes. While computational scRNA-seq methods are currently well suited for experiments representing a single condition, technology, or species, analyzing multiple datasets simultaneously raises new challenges. In particular, traditional analytical workflows struggle to align subpopulations that are present across datasets, limiting the possibility for integrated or comparative analysis. Here, we introduce a new computational strategy for scRNA-seq alignment, utilizing common sources of variation to identify shared subpopulations between datasets as part of our R toolkit Seurat. We demonstrate our approach by aligning scRNA-seq datasets of PBMCs under resting and stimulated conditions, hematopoietic progenitors sequenced across two profiling technologies, and pancreatic cell atlases generated from human and mouse islets. In each case, we learn distinct or transitional cell states jointly across datasets, and can identify subpopulations that could not be detected by analyzing datasets independently. We anticipate that these methods will serve not only to correct for batch or technology-dependent effects, but also to facilitate general comparisons of scRNA-seq datasets, potentially deepening our understanding of how distinct cell states respond to perturbation, disease, and evolution.\n\nAvailabilityInstallation instructions, documentation, and tutorials are available at http://www.satijalab.org/seurat

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (18 Jul 2017) · CC BY-NC-ND 4.0

Seurat combines canonical correlation analysis and dynamic time warping to co-embed scRNA-seq datasets from different conditions, platforms, or species, outscoring ComBat and limma in alignment score.

Pooling heterogeneous scRNA-seq samples conflates shifts in cell-type composition with within-type expression changes, and batch correction tools assume a confounder affects all cells uniformly, an assumption broken when responses differ by cell type. Seurat instead learns gene-gene correlation structure shared by two datasets without marker supervision, tolerating density shifts and feature-scale differences so matched subpopulations can be compared.

Key results

  1. ShownAligning control and interferon-beta (IFN-β) stimulated peripheral blood mononuclear cells (PBMCs) grouped cells by type rather than condition; joint clustering resolved more populations than the published marker-based supervised classification.

    14,038 human PBMCs from 8 patients: 13 cell types vs eight supervised clusters, newly separating naive from memory T cells and plasmacytoid from conventional dendritic cells (DCs), plus a rare (0.4%) erythroblast population; cluster proportions matched across conditions (R=0.997), with no dataset-specific cells. · Comparative analysis of stimulated and resting…; Figure legends

  2. ShownAligning deep full-length SMART-Seq2 with shallow 3' MARS-Seq mouse bone marrow data revealed progenitors committed to all eight lineages in both sets, including SMART-Seq2 subsets its original analysis had not resolved, and excluded contaminating natural killer cells as dataset-specific.

    SMART-Seq2 (774 progenitors, 6,558 genes/cell) vs MARS-Seq (2,730 cells, 1,453 genes/cell): SMART-Seq2 megakaryocyte-erythrocyte progenitors mapped only to C1-7, granulocyte-macrophage progenitors to C11-18, early lymphoid progenitors only to C9-10; a joint erythroid diffusion map kept each dataset's pseudotemporal order; JUN/FOS stress response was technology-specific. · Integrated analysis of multiple scRNA-seq…; Figure legends

  3. ShownAlthough human and mouse islet cell-type transcriptomes are poorly conserved, aligned inDrop atlases yielded 10 joint clusters and a rare beta-cell subgroup in both species up-regulating endoplasmic reticulum (ER) stress genes, previously hinted at only by a semi-supervised analysis of human beta cells.

    10,322 islet cells (human 8,533; mouse 1,789); prior mean human-mouse cell-type correlation R=0.42. Human mast cells and murine B cells were flagged dataset-specific; 4.3% of calls disagreed with per-species labels, especially among low-UMI-count cells. Stressed beta cells kept INS1 but raised HERPUD1, GADD45A, ATF3 and ATF4, with shared genes enriched for the unfolded-protein response. · Joint learning of cell types across species; Figure legends

  4. ShownOn an author-designed alignment score of local dataset mixing, Seurat outperformed ComBat and limma on all three dataset pairs, most clearly across species, whereas datasets from different tissues aligned poorly, serving as a negative control.

    Score spans 0 to 1, reflecting how evenly each cell's neighborhood draws from both datasets; bar plots (with an uncorrected baseline) show Seurat above ComBat and limma for PBMC, bone marrow, and islet pairs, but per-method values are not printed in the text. Different-tissue controls scored low even when cells were not flagged as dataset-specific. · Benchmarking alignment and batch correction…; Figure legends

  5. ShownPer-cluster differential expression split the IFN-β program into genes induced in every cell type and cell-type-restricted components, with plasmacytoid DCs (pDCs) showing a response distinct from the correlated myeloid and lymphoid responses.

    ISG15 and IFIT1 were induced in all cells, CXCL10 mainly in myeloid cells; the monocyte marker CD14 dropped 1.98-fold on stimulation; naive and memory CD4+ T cells responded nearly identically, whereas pDC versus conventional DC differences reproduced across patients and pDCs clustered apart in the response-correlation heatmap. · Comparative analysis of stimulated and resting…; Figure legends

Methods, data and reproducibility

Diagonal canonical correlation analysis (CCA) on genes highly variable in either dataset finds correlated basis vectors; cells whose CCA variance explained drops by the 50% cutoff versus per-dataset principal component analysis are discarded, and per-vector metagenes are rescaled to their 95% reference range, then aligned by dynamic time warping.

  • Control and IFN-β-stimulated peripheral blood mononuclear cells (PBMCs) from a published interferon stimulation study · human · n = 14,038 cells from 8 patients
  • Bone marrow hematopoietic progenitors profiled with deep full-length SMART-Seq2 (Nestorowa et al.) · mouse · n = 774 progenitors; 765 shown in Figure 3 tSNE
  • Myeloid-enriched bone marrow hematopoietic progenitors profiled with shallow 3' MARS-Seq (Paul et al.); 18 clusters originally · mouse · n = 2,730 cells; 2,689 shown in Figure 3 tSNE
  • Human pancreatic islet cells profiled with inDrop in a published human-mouse islet study · human · n = 8,533 cells
  • Mouse pancreatic islet cells profiled with inDrop in the same human-mouse islet study · mouse · n = 1,789 cells
  • PBMC datasets from 10X Chromium and SeqWell aligned in an R Markdown tutorial (Supplementary Note 1) · human
  • Dataset pairs from different tissues used as alignment negative controls (Supplementary Figure 11)
Methods & tools
Seurat (open-source R toolkit implementing the alignment workflow), Diagonal canonical correlation analysis (CCA), Principal component analysis (per-dataset baseline for flagging unalignable cells), Dynamic time warping (metagene alignment), Graph-based clustering, t-distributed stochastic neighbor embedding (t-SNE), Diffusion maps (joint erythroid trajectory), Single-cell differential expression tests, merged across species by Fisher's criterion, ComBat (batch correction baseline), limma (batch correction baseline), Alignment score (author-designed local neighborhood mixing metric), GOplot R package (Gene Ontology enrichment plots)
Software & environment
Implemented in Seurat, an open-source R toolkit; an R Markdown tutorial (Supplementary Note 1) is mentioned; no package versions or containers are stated.
Reproducibility gaps
Input datasets are cited by study only, without accession IDs or links; Seurat version is unstated; diagonal CCA, metagene, mapping, and alignment-score details are deferred to Supplementary Methods absent from the text.
Limitations (6)
  • Only pairwise alignment of sequencing-based datasets is implemented; extension to more datasets (repeated alignment to a reference or multi-set CCA) and to spatial or in situ data is proposed as future work. · stated by the authors
  • Alignment depends on shared populations: different-tissue negative controls aligned poorly with low scores, and their unshared cells were not necessarily caught by the dataset-specific filter. · stated by the authors
  • In the human-mouse islet integration, 4.3% of cluster calls were discordant with the original per-species annotations, especially among cells with low UMI counts. · stated by the authors
  • Aligned clusters can carry technical signals: stress-signature T- and B-cell subclusters were judged likely artifacts of culturing, and a JUN/FOS stress response surfaced as a technology-specific effect. · stated by the authors
  • The alignment score only measures how evenly datasets mix within neighborhoods, not whether distinct cell types stay separated, so over-correction would not be penalized; per-method values appear only in bar plots. · noted in this summary
  • One 50% variance-explained cutoff for discarding unalignable cells was used in every example, and no sensitivity analysis of this user-defined threshold is reported in the text. · noted in this summary

Read this if you need to jointly cluster or compare scRNA-seq datasets across conditions, platforms, or species and want the details and validation of Seurat's CCA plus dynamic time warping alignment.

Citations
13,175
FWCI
381.57
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
THISStuart '19Johansen '19Andreatta '20Hie '19Lee '23Specque '25
76% semantic matchGenomics
Tim Stuart, Andrew Butler, Paul Hoffman+7 more·Cell·Jun 1, 2019· 15,454 citations