EnsDTI: Predicting Drug-Target Interaction With Mixture-of-Experts and Confidence Assessment.

Area of Science:

• Computational chemistry
• Drug discovery informatics
• Bioinformatics

Background:

• Accurate drug-target interaction (DTI) identification is crucial for efficient drug discovery.
• Structure-based methods are accurate but computationally expensive for large chemical spaces.
• Ligand-based methods lack consistency and reliability on unseen data.

Purpose of the Study:

• To develop a computational framework that balances speed and accuracy for DTI prediction.
• To improve the reliability and applicability of DTI prediction tools in drug discovery.
• To provide confidence scores for DTI predictions to aid in candidate filtering.

Main Methods:

• Proposed EnsDTI, a novel framework integrating structure-based and ligand-based DTI prediction approaches.
• Utilized a mixture-of-experts architecture to leverage existing deep learning models for enhanced DTI predictions.
• Incorporated an inductive conformal predictor to provide reliable confidence scores for predictions.

Main Results:

• EnsDTI demonstrated high performance in prediction accuracy across four benchmark datasets.
• The framework achieved excellent confidence estimation capabilities.
• Candidate rankings generated by EnsDTI showed strong correlation with actual docking affinities.

Conclusions:

• EnsDTI effectively bridges the gap between structure-based and ligand-based DTI prediction methods.
• The framework offers a practical and reliable tool for ranking and filtering potential drug candidates in drug discovery.
• EnsDTI's ability to provide confidence scores enhances its utility for efficient lead identification.