TRAPT: A multi-stage fused deep learning framework for predicting transcriptional regulators based on large-scale epigenomic data

It is challenging to identify regulatory transcriptional regulators (TRs), which control the expression of gene sets via regulatory elements and epigenomic signals, in context-specific studies on the onset and progression of diseases. The use of large-scale multi-omics epigenomic data enables the representation of the complex epigenomic patterns of control of the regulatory elements and the regulators. In this study, we propose Transcription Regulator Activity Prediction Tool (TRAPT), a multi-modality deep learning framework that can predict regulatory TRs in a queried gene set by using large-scale multi-omics epigenomic data that include information on histone modifications, ATAC-seq, and TR-ChIP-seq. We designed a two-stage knowledge-based model of distillation to learn the embedded nonlinear representations of upstream and downstream regulatory potentials, and merged multi-modality epigenomic features from the TRs and the queried gene sets to infer regulator activity. The results of experiments on 570 TR-related datasets showed that TRAPT outperformed state-of-the-art methods in predicting the TRs, especially in terms of forecasting transcription co-factors and chromatin regulators. Moreover, we successfully identified key TRs associated with diseases, genetic variations, cell-fate decisions, and tissues. Our method provided an innovative perspective on identifying TRs by using epigenomic data, and can help researchers better understand the mechanisms of regulation of gene expression.