Improving structure-based virtual screening performance via learning from scoring function components.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Classical scoring functions (SFs) have limitations in drug discovery.
• Machine learning (ML) based SFs show promise but often focus on new representations or algorithms.
• Existing ML SFs often neglect valuable information from decomposed energy components of classical SFs.

Purpose of the Study:

• To propose a novel method, energy auxiliary terms learning (EATL), for developing improved ML-based SFs.
• To leverage energy components from existing SFs as input for new ML models.
• To enhance virtual screening (VS) performance using ML-derived SFs.

Main Methods:

• Developed the energy auxiliary terms learning (EATL) approach.
• Extracted scoring components from existing SFs to serve as input features.
• Created three levels of ML SFs: EATL SFs, docking-EATL SFs, and comprehensive SFs.
• Evaluated SF performance using metrics like ROC and BEDROC on benchmark datasets (DUD-E, AD).

Main Results:

• EATL-based SFs significantly outperformed classical SFs in virtual screening.
• EATL achieved performance comparable to advanced ML-based methods on the DUD-E dataset.
• Effectiveness was validated on the actives as decoys (AD) dataset.
• The approach demonstrated superior ROC and BEDROC scores.

Conclusions:

• The EATL method offers a powerful strategy for developing high-performing ML-based SFs.
• Learning from SF components is an effective way to improve virtual screening power.
• The EATL concept is adaptable and can be extended to other docking programs and SFs.