Fig. 2 |. Overview of the BayesCNet.

a. BayesCNet takes either paired or unpaired single-cell gene expression and chromatin accessibility or chromatin accessibility alone as input. Single cells within each cell type are clustered using a KNN graph and aggregated into metacells. For each gene, BayesCNet models gene expression (or gene activity) as the response variable and chromatin accessibility at peaks within ±250 kb of the transcription starts site (TSS) as covariates. b. Cell type–specific enhancer–gene (EG) linkages are inferred using a Bayesian regression framework with structured priors. One prior captures the cell type hierarchy, reflecting developmental lineage or hierarchical cell type relationships, while the other prior encodes genomic proximity and peak co-accessibility. Both priors are represented as covariance matrices. c. Predicted cell type–specific EG linkages can be integrated with TF motifs and TF expression to construct cell type–specific TF–gene regulatory networks (TF-GRNs). d. Predicted EG linkages can be integrated with GWAS summary statistics to perform partitioned heritability analysis.