Sequence-based modeling of low-affinity transcription factor–DNA binding through deep learning

Multiple layers of molecular determinants and mechanisms affect binding specificity between transcription factors (TFs) and DNA. DNA sequence-based deep learning models using convolutional neural networks (CNNs) and self-attention (SA) transformers have improved modeling accuracy and advanced our understanding of TF–DNA binding specificity through network interpretation. However, the systematic evaluation of various strategies for handling DNA sequence orientations in deep learning models—and their interpretation—remains underexplored, especially in the context of learning low-affinity binding site specificity. Using SELEX-seq data for eight Exd-Hox heterodimers in Drosophila, we compared canonical models with data augmentation and reverse-complement weight-sharing models. We found that reverse-complement weight-sharing CNN models and SA models trained with augmented data with reverse complements outperformed other approaches in modeling binding specificity. In this work, we evaluated several interpretation methods, including Gradient*input, DeconvNet, DeepLIFT, and in silico mutagenesis (ISM). Compared to other interpretation methods, ISM was less sensitive to model hyperparameter settings. In this work, we identified Exd-Ubx binding at low-affinity sites and suggested possible biophysical mechanisms. The findings of this study will be relevant for studying the functional role of low-affinity TF binding in gene regulatory mechanisms with possible implications on TF–DNA binding specificity guided protein design.