STED: flexible cross-modal topic modeling infers cell-type-specific regulatory landscapes from bulk epigenomics

Deciphering cell-type-specific epigenetic landscapes within heterogeneous tissues is restricted by the sparsity, high cost, and limited throughput of current single-cell epigenomic technologies. To bridge the gap between massive legacy bulk epigenomic data and cellular resolution, we present STED (Single-cell Topic modeling and Epigenetic deconvolution), a flexible computational framework that reconstructs cell-type-specific regulatory signals by leveraging single-cell transcriptomic references. Methodologically, STED introduces a versatile topic modeling architecture: users can employ the information-theoretic correlation explanation algorithm to robustly mitigate transcriptomic sparsity, or innovatively adapt BERTopic-a large language model-based framework-to capture high-dimensional semantic cellular states. STED couples these latent regulatory topics with a physics-based "gene activity score" transformation, which acts as a cross-modal bridge to deconvolve bulk chromatin accessibility (ATAC-seq) and histone modification (ChIP-seq/CUT&Tag) profiles. Extensive benchmarking across diverse datasets-including human peripheral blood mononuclear cells (PBMCs), mouse brain, and zebrafish inner ear-demonstrates that STED significantly outperforms existing rigid-reference tools in accuracy and robustness against cross-platform batch effects. In a hematopoietic stem cell differentiation model, STED successfully identifies cell-type-specific transcription factor binding motifs and differential peaks associated with lineage commitment. We further demonstrate STED's biological utility by uncovering tumor-specific super-enhancers in colorectal cancer-a mechanism obscured in bulk signals. STED thus provides a scalable, foundation-model-empowered solution for dissecting gene regulatory networks in complex tissues.