SCORCH: Improving structure-based virtual screening with machine learning classifiers, data augmentation, and...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Drug discovery is expensive and time-consuming.
• Computational prediction of protein-ligand binding can accelerate drug discovery.
• Existing machine-learning scoring functions have limitations due to training data biases.

Purpose of the Study:

• To improve the predictive performance of structure-based virtual screening.
• To develop a novel machine-learning scoring function that overcomes limitations of previous methods.

Main Methods:

• Developed SCORCH (Scoring COnsensus for RMSD-based Classification of Hits), a novel machine-learning scoring function.
• Augmented training data by considering multiple ligand poses and labeling based on RMSD.
• Addressed decoy bias by generating property-matched decoys and using consistent preparation/docking methods.
• Employed a consensus of three machine learning approaches for improved performance.

Main Results:

• SCORCH demonstrated improved docking and screening power on independent benchmark datasets.
• Achieved a higher enrichment factor (13.78% vs. 10.86%) and better native pose rank (5.9 vs. 30.4) compared to single-pose methods.
• Outperformed widely used scoring functions in virtual screening and pose prediction tasks.

Conclusions:

• SCORCH rationally addresses key limitations, significantly improving virtual screening performance.
• Provides an uncertainty estimate, aiding in cost and time reduction for drug discovery.
• Enhances the efficiency and accuracy of identifying potential drug candidates.