BTFBS: Binding Prediction of Bacterial Transcription Factors and Binding Sites Based on Deep Learning

Abstract

The binding of transcription factors (TFs) to TF binding sites plays a vital role in the process of regulating gene expression and evolution. With the development of machine learning and deep learning, some successes have been achieved in predicting transcription factors and binding sites. In this paper, we develop a model, BTFBS, which predicts whether the bacterial transcription factors and binding sites combine or not. The model takes both the amino acid sequences of bacterial transcription factors and the nucleotide sequences of binding sites as inputs, and extracts features through convolutional neural network and MultiheadAttention. For the model inputs, we use two negative sample sampling methods: RS and EE. On the test dataset of RS, the accuracy, sensitivity, specificity, F1-score, and MCC of BTFBS are 0.91446, 0.89746, 0.93134, 0.91264, and 0.82946, respectively. Furthermore, on the test dataset of EE, the accuracy, sensitivity, specificity, F1-score and MCC of BTFBS are 0.87868, 0.89354, 0.86394, 0.87996, and 0.75796, respectively. Meanwhile, our findings indicate that the optimal approach for obtaining negative samples in the context of bacterial research is to utilize the whole genome sequences of the corresponding bacteria, as opposed to the shuffling method. The above results on the test dataset have shown that the proposed BTFBS model has a good performance and it can provide an experimental guide.

Keywords

• transcription factors
• binding sites
• deep learning
• bacteria