Advancing Drug-Target Interaction prediction with BERT and subsequence embedding.

Area of Science:

• Bioinformatics
• Computational Chemistry
• Drug Discovery

Background:

• Drug-target interaction (DTI) prediction is crucial for discovering new synthetic drugs.
• Traditional methods often encode proteins by amino acids, which may not fully simulate biological processes.

Purpose of the Study:

• To propose a novel deep learning framework for DTI prediction.
• To improve the accuracy and efficiency of DTI prediction by simulating biological processes.

Main Methods:

• A Bidirectional Encoder Representation from Transformers (BERT)-based framework was developed.
• The framework integrates high-frequency subsequence embedding and transfer learning.
• Multi-head self-attention mechanisms were used to learn internal sequence and interaction features.

Main Results:

• The BERT-based model achieved higher average prediction metrics than most baseline methods on three benchmark datasets.
• An ablation study confirmed the superiority of transfer learning.
• The model demonstrated acceptable scalability on datasets with unseen drugs and proteins.

Conclusions:

• The proposed BERT-based framework effectively predicts drug-target interactions.
• Subsequence embedding and transfer learning are key components for improving DTI prediction.
• The model shows promise for accelerating drug discovery and development.