Predicting transcription factor binding sites by dual-stream multiple instance learning network

The discovery of transcription factor binding sites(TFBSs) is important for modeling potential binding mechanisms and subsequent cellular functions. In recent years, there have been many deep learning methods that have achieved good results in predicting transcription factor binding sites. However, these methods usually follow the fully supervised approach and ignore the weakly supervised information in DNA sequences. In contrast, the currently proposed multiple instance learing(MIL) methods based on weakly supervised learning usually divide the DNA sequence into multiple overlapping subsequences and model each instance separately. These methods do not take into account the connections between overlapping subsequences, and these methods destroy the global information of the sequences in the process of dividing them into overlapping subsequences. In addition, deep learning methods generally perform poorly when there is less training data. We, therefore, propose a new deep learning method, DS-SSB. More specifically, DS-SSB combines the dual-stream multiple instance network with multiple features. First, we combine sequence features and shape features after feature extraction at the instance level to enhance the feature representation of instances. Then, the instance embeddings are aggregated into bag embedding through the dual-stream multiple instance network, and the relationships between the instances are considered in the aggregation process. Finally, the instance features fused into bag features are fused together with the BERT features of the whole sequence at the bag level for the final prediction. Experiments conducted on 690 ChIP-seq datasets showed that DS-SSB achieved good performance in predicting TFBSs. Also, experiments on four datasets further show that our method has an advantage on small datasets as well