Improved Method of Structure-Based Virtual Screening via Interaction-Energy-Based Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Virtual screening is crucial for identifying novel drug candidates from large compound libraries.
• Current virtual screening methods lack sufficient accuracy, hindering efficient drug discovery.
• There is a need for improved computational methods to enhance the precision of hit compound identification.

Purpose of the Study:

• To develop and validate a novel virtual screening method, SIEVE-Score, for more accurate drug discovery.
• To utilize protein-ligand interaction energies within a machine learning framework to improve docking pose representation.
• To provide human-interpretable results for better understanding of active vs. inactive compound discrimination.

Main Methods:

• Developed SIEVE-Score, a machine learning-based scoring function utilizing protein-ligand interaction energies.
• Represented docking poses using extracted interaction energy vectors for machine learning models.
• Evaluated SIEVE-Score performance against state-of-the-art methods using the Directory of Useful Decoys, Enhanced (DUD-E) dataset.

Main Results:

• SIEVE-Score demonstrated substantial improvements in enrichment factor at 1% across DUD-E benchmarks.
• The method outperformed existing machine learning scoring functions, interaction fingerprints, and docking software.
• Identified key interactions contributing to the distinction between active and inactive compounds, offering interpretability.

Conclusions:

• SIEVE-Score represents a significant advancement in virtual screening accuracy for drug discovery.
• The method's reliance on interpretable interaction energies enhances its practical applicability.
• The developed approach offers a more reliable tool for enriching potentially active compounds in early-stage drug development.