Deep learning the TF regulatory code for gene expression

Gene transcription is activated through the interaction between cis-regulatory elements (CREs) and transcription factors (TFs). CREs serve as templates to provide binding sites, whereas TFs provide functions to directly initiate transcription. Current research mainly focuses on deciphering cis-regulatory code, while neglecting TF regulatory code. However, CREs alone are not sufficient to determine the binding of TFs, which makes the interpretation of cis-regulatory code ambiguous. In this study, we systematically analyze 13 TF binding profiles associated with transcription initiation and encode them as TF sequence to explore the TF regulatory code. Furthermore, we propose a deep learning model named DeepTF to predict gene expression from TF sequence. Results show that TF binding exhibits conserved positional preferences and combinatorial patterns in promoters and DeepTF is able to predict gene expression with high accuracy (AUROC = 0.97). Meanwhile, cross-cell-line validation (AUROC > 0.90) further confirms the model's transferability. Model interpretation reveals DeepTF successfully captures the TF regulatory grammar associated with gene expression. Compared with the cis-regulatory code, our proposed TF regulatory code is better suited for investigating the relationship between TF binding positions or combinatorial patterns and gene expression. Collectively, DeepTF simplifies gene expression prediction and provides clear biological insights of transcriptional regulation.