Uncovering cellular programs and regulatory circuits governing cell fate bifurcations using interpretable machine learning

Introduction/Rationale: Understanding immune cell fate bifurcations requires deciphering how transcriptional and regulatory programs coordinate dynamic transitions. While deep learning models can capture these differences, they often lack interpretability. To address this, we developed a framework that integrates SLIDE—an interpretable machine learning (ML) method that extracts latent factors (LFs) representing cellular programs—with static and dynamic gene regulatory network (GRN) inference to reveal mechanisms driving immune differentiation.

Methods: We applied this framework to human B cells differentiating into plasmablasts (PB) or germinal center (GC) cells, and to T cells undergoing terminal (Texterm) or KLR⁺ cytotoxic (TexKLR) exhaustion. SLIDE identified regulon-level LFs distinguishing each state using scRNA-seq datasets. Static and dynamic state-specific GRNs were reconstructed using CellOracle, and Dictys, benchmarked against SCENIC+. Perturb-seq experiments targeting key TFs (PRDM1, IRF4, IRF8, SPIB, BATF, IKZF1, ETS1) enabled rollback analyses to test SLIDE’s ability to predict early fate bias.

Results: Using SLIDE, we uncovered strikingly specific and transferable LFs defining both B and T cell states. Cross-referencing these LFs with state-resolved GRN linkages revealed precise, TF-centric regulons that orchestrate lineage bifurcation with higher specificity and biological coherence than SCENIC+. Dynamic GRNs exposed distinct TF waves driving state transitions, while rollback analyses showed that SLIDE—without relying on GRNs—accurately predicted early cell-fate predisposition before transitions occurred, outperforming other methods.

Conclusion: By coupling interpretable ML with GRNs, this framework reveals mechanistic regulatory circuits governing immune cell differentiation. Moreover, when applied independently of GRNs to TF perturb-seq data, SLIDE could predict cell fate before bifurcation occurs, highlighting its power to identify transcriptional programs that predefine lineage commitment.