SubBERT: Sub-Structure Bidirectional Transformer for Drug–Protein Interaction Prediction
BiTransDPI: Bidirectional Transformer-Based for Drug–Protein Interaction Prediction
## Abstract ##
Considering the dependence between protein properties and sub-structure, we devise an unsupervised task to learn the relationship between sub-structures. Furthermore, for learning from the vivo process, we adopt a single stream to feed the information into the model and encode it together. Specifically, we name the above framework SubBERT, which chooses the bidirectional Transformers as the backbone. Comprehensive experiment results show that SubBERT achieves the best performance in Pearson correlation coefficient r and comparable performance in RMSE on three of five datasets. Furthermore, the ablation experiment demonstrates that the unsupervised pre-training and the sub-structure encoding can improve the performance of SubBERT, which could also be applied to other deep learning models. To the best of our knowledge, SubBERT is first method to unified encode protein and drug.
we propose a deep learning model using the Bidirectional Transformers as the backbone, BiTransDPI, to predict the Drug-Protein binding affinity.
Unlike other methods, we encode the sequences only based on the conserved fragments, and encode the protein and drug into a unified vector.
Moreover, we adopt a novel two-step training strategy to build BiTransDPI. The pre-training step is to learn the interaction between different fragments with unsupervised learning, and the fine-tuning step is for predicting the binding affinity with supervised learning. The comprehensive experiment results have illustrated the superiority of our proposed BiTransDPI.
#### The details about the project to be continuted... ####
