Skip the cell-type labels: NetMap reads regulatory networks and cell identity together from mixed single-cell data

Cell-type partitioning and gene regulatory network inference now happen in one step, with networks produced per cell rather than one network per cluster. The prior route clustered highly variable genes, identified cell types from marker genes, and then ran network inference separately on each homogeneous cluster, so each cluster yielded a single network and cell-level changes in regulatory activity were flattened out. On synthetic data generated with scMultiSim, where each simulated cluster carries its own characteristic network, NetMap exceeded the conceptually closest scGeneRAI and a GRNBoost2 baseline given ground-truth labels in on-target edge recall, while holding off-target edges near random.

The network comes from a shallow bottleneck autoencoder that reconstructs expression counts, followed by an explainability method that traces each target gene back to input genes. The autoencoder fits raw counts under a negative binomial or zero-inflated negative binomial distribution to accommodate the overdispersion and excess zeros typical of single-cell data, and several models are averaged into a robust cell-level network. On 10X Genomics PBMC multiome data, NetMap received only the unclustered mixed expression data and reached clustering scores of 1.0, 0.99 and 0.99 across three evaluation cases; against ChIP-seq binding evidence it reached a median AUC of 0.65, above GRNBoost2 and scPRINT, which also used pre-clustered input, while LINGER, which used chromatin accessibility and regulatory priors, achieved the highest median AUC of 0.71.

The network structure can in turn label cells, by running personalized PageRank on the cell-level graph to propagate marker genes. Conventional workflows revert to gene set analysis soon after building the network and stop using its structure; here the structure itself drives PageRank, yielding a probabilistic assignment of each cell to cell types. Across four PBMC technical replicates, NetMap mapped over 95% of cells to the same category as the gene expression clustering, consistently across repeated runs and sequencing platforms; on a mouse CD8+ T-cell dataset spanning naive, activated at 24 hours, and exhausted at 72 and 96 hours, clustering matched up to 95% of cells, with most mismatches between the two adjacent exhausted states.

The demonstrated scope covers synthetic two- and three-cluster data generated with scMultiSim, 10X Genomics PBMC multiome data, and mouse P14 CD8+ T-cell single-cell data at four stages: naive, activated at 24 hours, and exhausted at 72 and 96 hours. The intended reader is an analyst already familiar with the scverse ecosystem who wants cell-level regulatory networks without pre-annotating cell types. The method is released as a Python package and is compatible with scverse.

Explainability methods assume variables are independent, while gene expression features are highly correlated, so attribution is split across correlated variables, signal sharpness drops, and precision is generally low on real data. PageRank assignment is sensitive to marker set size, since largely unequal sets dilute probabilities; the authors recommend keeping sets similarly sized and note that ambiguous markers dilute labels. Literature-derived cell-type definitions disagree, and the same gene can appear in marker sets for several cell types, which produces outliers when overlapping de novo regulons with marker sets.