Enhancing the interpretability of transcription factor binding site prediction using attention mechanism

Transcription factors (TFs) regulate the gene expression of their target genes by binding to the regulatory sequences of target genes (e.g., promoters and enhancers). To fully understand gene regulatory mechanisms, it is crucial to decipher the relationships between TFs and DNA sequences. Moreover, studies such as GWAS and eQTL have verified that most disease-related variants exist in non-coding regions, and highlighted the necessity to identify such variants that cause diseases by interrupting TF binding mechanisms. To do this, it is necessary to build a prediction model that precisely predicts the binding relationships between TFs and DNA sequences. Recently, deep learning based models have been proposed and have shown competitive results on a transcription factor binding site prediction task. However, it is difficult to interpret the prediction results obtained from the previous models. In addition, the previous models assumed all the sequence regions in the input DNA sequence have the same importance for predicting TF-binding, although sequence regions containing TF-binding-associated signals such as TF-binding motifs should be captured more than other regions. To address these challenges, we propose TBiNet, an attention based interpretable deep neural network for predicting transcription factor binding sites. Using the attention mechanism, our method is able to assign more importance on the actual TF binding sites in the input DNA sequence. TBiNet outperforms the current state-of-the-art methods (DeepSea and DanQ) quantitatively in the TF-DNA binding prediction task. Moreover, TBiNet is more effective than the previous models in discovering known TF-binding motifs.