henhao
Sign in
BioinformaticsbioRxiv
Preprint

MultiNicheNet: a flexible framework for differential cell-cell communication analysis from multi-sample multi-condition single-cell transcriptomics data

Robin BrowaeysiD, Jeroen GilisiD, Chananchida Sang-aramiD, Pieter De BleseriD, Levi HosteiD, Simon Jan TavernieriD, Diether LambrechtsiD, Ruth SeurinckiD, Yvan SaeysiD

Abstract

Dysregulated cell-cell communication is a hallmark of many disease phenotypes. Due to recent advances in single-cell transcriptomics and computational approaches, it is now possible to study intercellular communication on a genome- and tissue-wide scale. However, most current cell-cell communication inference tools have limitations when analyzing data from multiple samples and conditions. Their main limitation is that they do not address inter-sample heterogeneity adequately, which could lead to false inference. This issue is crucial for analyzing human cohort scRNA-seq datasets, complicating the comparison between healthy and diseased subjects. Therefore, we developed MultiNicheNet (https://github.com/saeyslab/multinichenetr), a novel framework to better analyze cell-cell communication from multi-sample multi-condition single-cell transcriptomics data. The main goals of MultiNicheNet are inferring the differentially expressed and active ligand-receptor pairs between conditions of interest and predicting the putative downstream target genes of these pairs. To achieve this goal, MultiNicheNet applies the principles of state-of-the-art differential expression algorithms for multi-sample scRNA-seq data. As a result, users can analyze differential cell-cell communication while adequately addressing inter-sample heterogeneity, handling complex multifactorial experimental designs, and correcting for batch effects and covariates. Moreover, MultiNicheNet uses NicheNet-v2, our new and substantially improved version of NicheNets ligand-receptor network and ligand-target prior knowledge model. We applied MultiNicheNet to patient cohort data of several diseases (breast cancer, squamous cell carcinoma, multisystem inflammatory syndrome in children, and lung fibrosis). For these diseases, MultiNicheNet uncovered known and novel aberrant cell-cell signaling processes. We also demonstrated MultiNicheNets potential to perform non-trivial analysis tasks, such as studying between- and within-group differences in cell-cell communication dynamics in response to therapy. As a final example, we used MulitNicheNet to elucidate dysregulated intercellular signaling in idiopathic pulmonary fibrosis while correcting batch effects in integrated atlas data. Given the anticipated increase in multi-sample scRNA-seq datasets due to technological advancements and extensive atlas-building integration efforts, we expect that MultiNicheNet will be a valuable tool to uncover differences in cell-cell communication between healthy and diseased states.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
137
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (14 Jun 2023) · CC BY-NC-ND 4.0

Combining pseudobulk differential expression, cell-type specificity, sample prevalence and NicheNet-v2 ligand activity, MultiNicheNet ranks condition-specific ligand-receptor-target signaling in multi-sample scRNA-seq with batches or multifactor designs.

Most cell-cell communication tools pool cells across samples, ignoring inter-sample heterogeneity and confounders, and rank ligand-receptor pairs by expression or downstream activity alone, which can yield false inferences in patient cohorts. MultiNicheNet treats samples as units in edgeR models supporting complex contrasts, batches and covariates, and couples this with joint expression- and activity-based ranking, a combination the authors find lacking in existing tools.

Key results

  1. ShownNicheNet-v2 rebuilds the prior around Omnipath ligand-receptor pairs, added signaling and regulatory sources, and in vitro targets for 119 ligands; it predicts ligand target genes substantially better than NicheNet-v1 while ligand activity accuracy stays similar.

    Leakage-controlled fivefold cross-validation with NSGA-II tuning on 111 NicheNet and 52 CytoSig ligand-treatment gold-standard datasets gave the same conclusion; v1/v2 metric values appear only in Supplementary Figure 1/Table 2. Scoring ligand activity by AUPRC instead of Pearson correlation gave ligand AUROC 0.884 vs 0.863 (NicheNet), 0.860 vs 0.827 (CytoSig). · MultiNicheNet uses NicheNet-v2: an update of…; Construction of the NicheNet-v2 model; Evaluation of the performance of NicheNet-v2 in…; Parameter optimization via NSGA-II

  2. SuggestedPredictions agreed with orthogonal data: cutaneous squamous cell carcinoma (cSCC) tumor-specific pairs linked mostly co-localized cell types in Visium tumor data, and top multisystem inflammatory syndrome in children (MIS-C) signals match serum protein rises in independent cohorts.

    cSCC: top 500 tumor-specific pairs concentrated in cell-type pairs with high RCTD-derived spot co-localization (same with cell2location); random, mid-ranked and normal-skin-specific sets were not enriched. MIS-C: IFNG, CCL3, CCL4, RETN, S100A8 and S100A9 were MIS-C-specific; five studies report higher serum (or monocyte flow-cytometry) protein for these. · MultiNicheNet infers tumor-specific cell-cell…; MultiNicheNet compares cell-cell communication…; MultiNicheNet analysis on data from Li et al…; Discussion

  3. ShownAmong macrophage-T cell interactions in anti-PD1-treated breast cancer, pre-therapy macrophage PD-L1 (CD274) to PD-1 (PDCD1) on CD4 and CD8 T cells were the two top-ranked pairs for T cell clonotype expanders (an expected true positive); TGFB1, GDF11 and TNF led in non-expanders.

    PreE vs PreNE contrast, Bassez et al. cohort A; both PD-L1/PD-1 pairs combined strong differential expression (DE) with high ligand activity. Top 50 expander-specific macrophage-T cell pairs also included PDCD1LG2-PDCD1, NECTIN2-TIGIT, LGALS3-LAG3, CD86-CTLA4 and CXCL9-CXCR3; T cell IFNG, CSF1 and CSF2 toward macrophages were expander-specific. · MultiNicheNet highlights critical pre-therapy…; MultiNicheNet analysis on data from Bassez et…

  4. SuggestedA difference-in-differences contrast (on- vs pre-therapy, expanders vs non-expanders) showed stronger therapy-induced chemotactic signaling in expanders, whereas non-expander gains in (lymph)angiogenesis-related signals appeared to stem from a patient subset, revealing within-group heterogeneity.

    Contrast [OnE - PreE] - [OnNE - PreNE] (empirical-null p-values). Expanders: CCL19-CCR7 (fibroblasts to B cells), CXCL12-CXCR4 (endothelial to B cells), CXCL9-CXCR3 (to CD4 T cells). Non-expander gains, often toward endothelial cells, involved VEGFC, FLT4, semaphorin-plexin and IGF2; subgroups did not align with cancer type, hot/cold state, expansion level or cell abundance. · MultiNicheNet elucidates differences in…; MultiNicheNet analysis on data from Bassez et…

  5. ShownCorrecting for source dataset across four atlas studies recovered known idiopathic pulmonary fibrosis (IPF) processes (TGF-beta, extracellular matrix remodeling, myeloid recruitment) and shifted rankings, adding correction-only pairs and down-weighting a dataset-driven artifact.

    Dataset of origin modeled as batch (edgeR, Combat-seq). CCL13-CXCR3 (to CD8 T cells) and GAS6-AXL (to dendritic cells) ranked outside the top 1000 without correction; LGALS3-ANXA2 and MMP7-SDC1 were DE only after correction. Uncorrected, PTPRM-PTPRM in bronchial vessel cells looked IPF-specific, driven by one dataset with higher PTPRM and more qualifying IPF samples. · MultiNicheNet corrects for batch effects in…; MultiNicheNet analysis on integrated lung atlas…; Step 2: Calculating normalized pseudobulk…

Methods, data and reproducibility

Counts are summed per sample and cell type into pseudobulk profiles tested with edgeR quasi-likelihood generalized linear models (batch/covariate terms allowed); ligand and receptor DE, condition- and cell-type-specific expression, sample prevalence and NicheNet-v2 ligand activity are scaled to 0-1 and combined by weighted aggregation.

  • Bassez et al. breast cancer biopsies taken before and on anti-PD1 therapy (cohort A); scRNA-seq with T cell receptor sequencing-defined expander vs non-expander labels · human
  • Ji et al. cutaneous squamous cell carcinoma scRNA-seq of tumor and matched normal skin, patient-paired (dataset attributed to Li et al. in Methods) · human · GSE144240
  • 10x Visium spatial transcriptomics of cSCC tumor tissue, deconvolved with RCTD and cell2location for cell-type co-localization · human · n = two replicate tumor samples from patient 4 · GSE144240
  • Hoste et al. peripheral blood mononuclear cell scRNA-seq from MIS-C patients, healthy siblings, and adults with severe COVID-19 · human
  • Healthy and IPF lungs from four studies (Adams 2020, Reyfman 2019, Morse 2019, Habermann 2020) within the Azimuth-integrated lung atlas, with harmonized annotations · human · n = four of ten atlas source datasets · cellxgene collection 2f75d249-1bec-459b-bf2b-b86221097ced
  • NicheNet ligand-treatment gold standard: in vitro transcriptomes before and after ligand stimulation, with DE genes as true targets · n = 111 ligand treatment datasets
  • CytoSig cytokine signatures: median log fold change (logFC) per gene after ligand treatment; targets at |z-scored logFC| >= 2.5 · human · n = 52 datasets
  • NicheNet-v2 ligand-receptor network and ligand-target matrix built and released in this work · human and mouse versions · Zenodo record 7074291
Methods & tools
MultiNicheNet (multinichenetr R package), nichenetr with NicheNet-v2 ligand-receptor and ligand-target networks, edgeR quasi-likelihood negative binomial GLM with filterByExpr, Combat-seq for pseudobulk batch correction, Empirical-null p-values (Efron; satuRn implementation), Omnipath intercellular communication database, Verschueren et al. immunoglobulin-superfamily ligand-receptor set, CytoSig ligand-response signatures, KnockTF, Dorothea, HuRi, HumanNet, PathwayCommons (version 12), ReMap (2022), NSGA-II multi-objective optimization (nsga2r), RCTD and cell2location via the Spotless benchmark pipeline, Cytoscape
Code
https://github.com/saeyslab/multinichenetrhttps://github.com/saeyslab/nichenetrhttps://zenodo.org/record/8016880https://github.com/saeyslab/spotless-benchmark
Data & accessions
https://zenodo.org/record/7074291https://zenodo.org/record/8016880https://lambrechtslab.sites.vib.be/en/single-cellhttps://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE144240https://cellxgene.cziscience.com/collections/2f75d249-1bec-459b-bf2b-b86221097ced
Software & environment
Open-source R package (multinichenetr) building on nichenetr and edgeR; NSGA-II via nsga2r; prior networks use PathwayCommons version 12 and ReMap 2022; R and package versions are not stated.
Reproducibility gaps
MIS-C (Hoste et al.) data are cited without accession or download link; per-cohort patient numbers are absent from the main text; NicheNet-v1 vs v2 benchmark values are given only in supplementary material.
Limitations (6)
  • MultiNicheNet lacks a comprehensive quantitative benchmark and was compared with other tools only qualitatively, owing to missing in vivo ground truth; spatial co-localization and serum protein agreement neither prove that interactions occur nor that they matter. · stated by the authors
  • Cell types must exist in all compared conditions; samples under min_cells (default 10) are dropped and cell types without two samples per condition or one per condition-batch pair are excluded (myofibroblasts in IPF), so coarse labels that can mask subtype DE are often needed. · stated by the authors
  • Pseudobulk aggregation collapses each sample-cell type pair to one estimate, and with few samples DE testing is underpowered; adjusted-p-value filtering may leave too few DE genes for reliable ligand activity, so the authors suggest logFC-based filtering. · stated by the authors
  • Ligand activity cannot tell downregulation in one group from upregulation in the other (e.g., TNF in non-expanders), and intercellular regulatory networks rest on expression correlation plus prior knowledge, not causal evidence; perturbation or spatiotemporal data are needed. · stated by the authors
  • Breast cancer, cSCC and MIS-C analyses used p-values without multiple-testing correction (and cSCC and IPF lowered min_cells to 5), so the DE gene sets that drive ligand activity scores may contain more false positives than default adjusted-p-value settings would allow. · noted in this summary
  • Default aggregation weights (2 for ligand activity and ligand/receptor specificity, 1 for DE and sample-fraction terms) are recommended without a reported sensitivity analysis, so ranking stability under other weightings is not shown. · noted in this summary

Who should read:Read this if you compare ligand-receptor signaling between patient groups in multi-sample scRNA-seq and need sample-level statistics, batch or covariate correction, or multifactor contrasts such as therapy-response dynamics.

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
THISSolovey '24Bocci '26Jin '24Maryam '26Raredon '22Wilk '22
68% semantic matchBioinformatics
Maria Solovey, Muhammet A. Celik, Felix R. Salcher+5 more·bioRxiv·Mar 1, 2024· 3 citations