AttentionSiteDTI: an interpretable graph-based model for drug-target interaction prediction using NLP sentence-level...

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Drug-target interaction (DTI) prediction is crucial for drug discovery.
• Existing DTI models often lack interpretability and generalizability.
• Identifying key protein binding sites is essential for accurate DTI prediction.

Purpose of the Study:

• To introduce AttentionSiteDTI, an interpretable graph-based deep learning model for DTI prediction.
• To leverage protein binding sites and a self-attention mechanism for enhanced prediction accuracy.
• To improve model generalizability to new, unseen proteins.

Main Methods:

• Developed a graph-based deep learning model, AttentionSiteDTI, inspired by Natural Language Processing sentence classification.
• Utilized protein binding sites and a self-attention mechanism to model drug-target complexes.
• Evaluated model performance on three benchmark datasets and tested generalizability on new proteins.

Main Results:

• AttentionSiteDTI achieved superior performance compared to state-of-the-art models on benchmark datasets.
• The model demonstrated high generalizability, performing well on previously unseen proteins.
• Computational predictions showed high agreement with experimental validation, highlighting practical potential.

Conclusions:

• AttentionSiteDTI offers an interpretable and highly generalizable approach to drug-target interaction prediction.
• The model's ability to identify critical binding sites enhances understanding of drug-target interactions.
• AttentionSiteDTI shows significant potential as an effective pre-screening tool for drug repurposing applications.