henhaobio
Sign in

Nature CommunicationsBioinformaticsPeer-reviewed journal article

Inference and analysis of cell-cell communication using CellChat

Suoqin JiniD, Christian Fernando Guerrero‐JuareziD, Lihua ZhangiD, Ivan Chang, Raúl RamosiD, Chen‐Hsiang KuaniD, Peggy S. MyungiD, Maksim V. PlikusiD, Qing Hua NieiD

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

Abstract

Understanding global communications among cells requires accurate representation of cell-cell signaling links and effective systems-level analyses of those links. We constructed a database of interactions among ligands, receptors and their cofactors that accurately represents known heteromeric molecular complexes. Based on mass action models, we then developed CellChat, a tool that is able to quantitively infer and analyze intercellular communication networks from single-cell RNA-sequencing (scRNA-seq) data. CellChat predicts major signaling inputs and outputs for cells and how those cells and signals coordinate for functions using network analysis and pattern recognition approaches. Through manifold learning and quantitative contrasts, CellChat classifies signaling pathways and delineates conserved and context-specific pathways across different datasets. Applications of CellChat to several mouse skin scRNA-seq datasets for embryonic development and adult wound healing shows its ability to extract complex signaling patterns, both previously known as well as novel. Our versatile and easy-to-use toolkit CellChat and a web-based Explorer (http://www.cellchat.org/) will help discover novel intercellular communications and build a cell-cell communication atlas in diverse tissues.

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (21 Jul 2020) · CC BY 4.0

CellChat infers communication probabilities between cell groups using a mass-action model and 2,021 curated ligand-receptor interactions (multimeric complexes, cofactors), then compares signaling networks across scRNA-seq datasets.

Prior tools (SingleCellSignalR, iTALK, NicheNet) typically score single ligand-receptor gene pairs, missing multi-subunit receptors, and CellPhoneDB v2.0 models heteromeric complexes but not soluble agonists/antagonists or stimulatory/inhibitory co-receptors. CellChat adds these cofactors, curated pathway classes, support for continuous pseudotime states, and network, pattern and manifold analyses to compare datasets.

Key results

  1. ShownCellChatDB, a manually curated mouse and human ligand-receptor resource, encodes multimeric ligands and receptors plus soluble agonists, antagonists and stimulatory or inhibitory co-receptors, assigning every interaction to a signaling pathway.

    2,021 validated interactions: 60% paracrine/autocrine, 21% extracellular matrix-receptor, 19% cell-cell contact; 48% involve heteromeric complexes and 25% were newly curated from recent literature; mostly built on KEGG maps; grouped into 229 pathways. · Overview of CellChat; Database construction for ligand-receptor…

  2. SuggestedCellChat returned fewer ligand-receptor (L-R) pairs than CellPhoneDB but, like CellPhoneDB, significantly separated spatially adjacent from distant cell populations. Single-pair tools partly missed multi-subunit complexes and did not reach significance on the spatial test.

    Mean false-positive rates from partial multi-subunit hits: 10.6% (SingleCellSignalR), 14.3% (iTALK). Mean L-R pairs per E14.5 subpopulation pair: CellChat 12, CellPhoneDB 37, iTALK 14, SingleCellSignalR 12 (top 10% kept for the latter two). Only CellChat and CellPhoneDB significantly separated adjacent from distant cells, and they shared ~50% of predictions. · Comparison with other cell–cell communication…; Method comparisons

  3. SuggestedAmong E14.5 placode and dermal condensate (DC) states, CellChat ranked placode as the main signaling hub. It also predicted autocrine Pros1-Axl signaling in DC and Edn3-Ednrb signaling from DC to melanocytes, both consistent with RNAscope transcript patterns.

    Placode sent signals via 44 and received via 19 L-R pairs; 21 placode-to-DC pairs spanned WNT, TGFβ, SEMA3, PTN, PDGF, MK and FGF. RNAscope (n = 4 independent experiments): Axl co-localized with Edn3 and Thy1, concentrated in DC; Dct+ melanocytes in and around placode expressed Ednrb, while Edn3 was enriched in DC cells. · CellChat predicts key signaling events between…; RNAscope in situ assay; Figure legends

  4. SuggestedJoint manifold learning plus ranking by information flow separated shared from context-specific pathways in two comparisons: mouse embryonic versus wound skin, and human atopic dermatitis lesional versus nonlesional skin. It recovered known wound and inflammatory signals.

    Mouse E13.5 vs day 12 wound: 19 out of 34 pathways active in both, nine embryo-only (incl. FGF, WNT), four wound-only (incl. SPP1, MIF, IL1). Human atopic dermatitis (7563 cells): 11 out of 16 pathways higher in lesions; CSF, IL4, CCL, TNF lesion-only; CCL19-CCR7 top CCL pair from inflammatory fibroblasts to inflammatory dendritic cells, matching prior immunofluorescence. · Joint learning of conserved and context-specific…; Joint learning of normal and diseased human skin…; Single-cell RNA-seq datasets, data…

  5. SuggestedIn mouse skin, CellChat recovered signaling supported by prior literature. In day 12 wounds it found myeloid-to-fibroblast TGFβ and a non-canonical WNT (ncWNT) network driven by Wnt5a from one fibroblast population. At E14.5 it found epidermis-derived canonical WNT reaching dermal cells.

    Wound (25 cell groups): 60 significant L-R pairs in 25 pathways; TGFβ dominated by Tgfb1 with Tgfbr1/Tgfbr2, myeloid MYL-A as main mediator; ncWNT driven only by Wnt5a from FIB-D; five outgoing and five incoming communication patterns. E14.5 trajectories: 88 interactions in 22 pathways, with Wnt6-Fzd10/Lrp6 and Wnt6-Fzd2/Lrp6 dominating WNT. · CellChat identifies communication patterns and…; CellChat reveals continuous cell…

Methods, data and reproducibility

For each cell group, a trimean summarizes differentially expressed signaling genes, which enter a mass-action/Hill model (geometric-mean subunits, cofactor terms, group size). Label permutation tests call significant links; graph centrality, non-negative matrix factorization and UMAP manifolds of network similarity then summarize them.

  • Mouse skin wound dermis at day 12 post-wounding, unsorted cells, 10X Genomics scRNA-seq; red blood cells excluded and abundant types subclustered into 25 groups · mouse · n = 21,819 cells (21,557 after excluding red blood cells) · GSE113854
  • Embryonic mouse whole-skin scRNA-seq (unsorted, 10X Genomics), two E13.5 and two E14.5 biological replicates; 11 and 13 annotated populations, respectively · mouse · n = 12,951 cells at E13.5; 12,197 cells at E14.5 after filtering · GSE122043 (GSM3453535, GSM3453536, GSM3453537, GSM3453538)
  • E14.5 subsets: dermal and epidermal diffusion-map trajectories, plus placode, pre-DC, DC1, DC2 and three melanocyte subpopulations (MELA-A, MELA-B, MELA-C) · mouse
  • Processed scRNA-seq of lesional and non-lesional skin from atopic dermatitis patients; CellChat run on 7563 fibroblast, dendritic cell and T cell profiles in twelve groups · human · n = 17,349 cells from four lesional and four non-lesional samples · GSE147424
  • RNAscope in situ hybridization on frozen E14.5 mouse skin sections for Dct, Edn3, Ednrb, Axl and Thy1 · mouse · n = 4 independent experiments
  • CellChatDB ligand-receptor database curated mainly from KEGG plus recent literature, with human-specific additions after ortholog mapping · mouse and human · n = 2,021 interactions in 229 signaling pathways
Methods & tools
CellChat R package with CellChatDB, KEGG Pathway database, STRINGdb protein-protein interaction network (random-walk propagation), NMF R package (Cophenetic and Silhouette metrics), sna, igraph and ggalluvial R packages, UMAP with k-means clustering, Louvain community detection; diffusion map pseudotime, Seurat V3, CellPhoneDB v2.0, SingleCellSignalR and iTALK (benchmarked tools), CellTalkDB, Ramilowski2015, NicheNet and ICELLNET (database comparison), Geometric sketching subsampling, RNAscope kit v2 (Advanced Cell Diagnostics)
Software & environment
CellChat is an R package. Human samples were integrated with Seurat V3 following its v3.2 tutorial. Centrality, pattern and plotting steps use the NMF, sna, igraph and ggalluvial R packages; no container or version pins are described.
Reproducibility gaps
No CellChat release or commit is pinned. Settings for the tool comparison and definitions of the subsampling metrics are left to Supplementary Note 3. The human atopic dermatitis data were reused as processed GEO matrices.
Limitations (6)
  • Binding and mass action happen between proteins, but CellChat infers them from mRNA levels and ignores tissue position. The authors note that single-cell protein data could improve modeling, and that adapting CellChat to spatial transcriptomics is still untested. · stated by the authors
  • Hill-function parameters (dissociation constant Kh, cooperativity n) lack explicit biological grounding and often differ between ligand-receptor pairs. Inferred interactions were robust only within certain parameter ranges on the four tested datasets. · stated by the authors
  • With no ground-truth communication data, evaluation relies on proxies. Overlap with other tools can reward similar methods regardless of accuracy and is confounded by differing ligand-receptor databases; subsampling measured consistency rather than accuracy. · stated by the authors
  • Inference depends on upstream clustering (the number of clusters can change inferred pairs) and on the averaging statistic. The default trimean dropped about 15% more ligand-receptor pairs than a 10% truncated mean, giving fewer interactions than CellPhoneDB. · stated by the authors
  • False-positive rates for SingleCellSignalR and iTALK use CellChatDB's own multi-subunit complex definitions as ground truth. The comparison therefore assumes CellChat's curation is correct rather than independently testing which interactions are real. · noted in this summary
  • Experimental support is RNAscope transcript co-localization at E14.5, and the Pros1 ligand itself was not probed. No perturbation or protein-level assay tested whether Pros1-Axl or Edn3-Ednrb signaling drives migration of dermal condensate cells or melanocytes. · noted in this summary

Read this if you need to infer ligand-receptor signaling from clustered or pseudotime-ordered scRNA-seq data while modeling multi-subunit receptors and cofactors, or to compare communication networks across conditions.

Citations
8,399
FWCI
336.00
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
THISJin '24Efremova '20Hou '25Bocci '26Cabello‐Agu '20Raredon '22
80% semantic matchBioinformatics
Suoqin Jin, Maksim V. Plikus, Qing Hua Nie·Nature Protocols·Sep 16, 2024· 1,450 citations