Multi-scaled self-attention for drug-target interaction prediction based on multi-granularity representation.

Area of Science:

• Bioinformatics
• Computational Chemistry
• Drug Discovery

Background:

• Drug-target interaction (DTI) prediction is vital for drug discovery.
• Current deep learning methods for DTI prediction have limitations in encoding chemical information and capturing complex patterns.
• Existing character encoding methods overlook crucial chemical details in drug and protein sequences.

Purpose of the Study:

• To propose a novel deep learning model for improved DTI prediction.
• To address limitations in existing encoding methods and deep model architectures for DTI prediction.

Main Methods:

• Developed a multi-granularity encoding method using sequence segmentation for drugs and proteins.
• Implemented a multi-scaled self-attention network (SAN) to extract diverse local patterns from multi-granularity representations.
• Fused deep representations of drugs and targets for final DTI prediction.

Main Results:

• The proposed model demonstrated superior prediction accuracy compared to strong baseline models on KIBA and Davis datasets.
• The multi-granularity encoding effectively captures chemical textual information.
• The multi-scaled SAN successfully extracts various local patterns in drug and target representations.

Conclusions:

• The novel multi-granularity encoding and multi-scaled SAN model significantly enhance DTI prediction.
• The approach improves the encoding of chemical textual information and extraction of local patterns.
• This work offers a promising advancement for computational drug discovery and development.