EpiBinder: A Multimodal Framework for Cell-Type-Specific Prediction and Interpretation of Transcription Factor Binding
TL;DR
EpiBinder is presented, a multimodal deep-learning framework for cell-type-specific prediction of TF binding that jointly models DNA sequence with base-resolution epigenetic information, including cytosine methylation from whole-genome bisulfite sequencing and chromatin accessibility from DNase I hypersensitivity data.
Abstract
Transcription factor (TF) occupancy in vivo depends not only on the underlying DNA sequence but also on the local epigenetic environment, which varies across cell types and strongly influences whether sequence-encoded binding potential becomes functional. Here we present EpiBinder, a multimodal deep-learning framework for cell-type-specific prediction of TF binding that jointly models DNA sequence with base-resolution epigenetic information, including cytosine methylation from whole-genome bisulfite sequencing and chromatin accessibility from DNase I hypersensitivity data. Across multiple human cell lines, EpiBinder consistently outperforms strong sequence-only baselines, improving TF-binding prediction by up to 10% in area under the precision-recall curve. Beyond predictive performance, EpiBinder provides base-level attribution maps that enable systematic interrogation of regulatory context, including candidate methylation-sensitive loci, contextual motif dependencies, and putative TF-TF interactions. These results position EpiBinder as a practical framework for modeling and exploring the local regulatory grammar underlying cell-type-specific TF occupancy.