BEACON: Bayesian Contrastive Learning for Single-Cell Gene Regulatory Inference

Gene regulatory networks (GRNs) govern cellular processes by encoding interactions between transcription factors (TFs) and their target genes. While computational GRN inference has traditionally relied on bulk RNA sequencing data, such approaches fail to capture cellular heterogeneity. Single-cell RNA sequencing (scRNA-seq) offers higher resolution but introduces additional challenges, including sparsity, technical noise, and dropout effects. Existing GRN inference methods struggle with these limitations, often yielding near-random predictions in benchmarking studies. To address this, we introduce BEACON (BayEsiAn COntrastive learNing for regulatory inference) for robust GRN edge discovery from static scRNA-seq data. Specifically, BEACON learns gene representations via contrastive loss on known positive and sampled negative edges, then refines edge likelihoods through a Bayesian prior that encodes partial network knowledge. We evaluate BEACON on both simulated and real-world scRNA-seq datasets with experimentally established TF–target regulatory relationships, demonstrating that BEACON outperforms state-of-the-art GRN inference methods on experimental datasets. An ablation study shows that both contrastive embedding and Bayesian priors are essential to achieving this performance gain. These findings suggest that BEACON mitigates key scRNA-seq limitations, providing a scalable and biologically grounded solution for GRN inference.