Improved Prediction of Drug-Protein Interactions through Physics-Based Few-Shot Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of drug-protein interactions is vital for efficient drug discovery.
• Traditional scoring functions and existing machine learning scoring functions (MLSFs) face limitations, including small data issues and poor interpretability.
• Structure-based drug screening requires robust predictive models.

Purpose of the Study:

• To develop a novel physics-based small data machine learning framework for interpretable and generalizable prediction of drug-protein interactions.
• To address the challenges of limited data and poor interpretability in existing MLSFs.
• To provide a tool for drug screening on targets with scarce positive data.

Main Methods:

• Proposed DrugBaiter, a physics-based small data machine learning framework.
• Employed a strategy of three training phases with three loss functions (score, weight, and ranking).
• Evaluated DrugBaiter on DUD-E (102 targets) and DEKOIS 2.0 (81 targets), comparing it with 14 other MLSFs.

Main Results:

• DrugBaiter significantly improved drug screening performance, especially for targets with few known active compounds.
• The model demonstrated interpretability in describing drug-protein interactions at the atomic level.
• Successfully applied DrugBaiter to screen drugs for the SARS-CoV-2 main protease target.

Conclusions:

• DrugBaiter offers a powerful solution for interpretable and generalizable drug-protein interaction prediction, particularly in low-data scenarios.
• The framework enhances drug screening efficiency and provides atomic-level insights into interactions.
• DrugBaiter is poised to become a valuable tool for screening novel drugs against new targets with limited active data.