A generalized protein-ligand scoring framework with balanced scoring, docking, ranking and screening powers.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Machine learning (ML) enhances protein-ligand scoring functions, offering high accuracy and efficiency.
• Current ML scoring functions often excel at specific tasks (e.g., binding affinity, pose prediction, virtual screening) but lack generalizability.
• Developing a single scoring function with balanced performance across critical tasks remains a significant challenge in computational drug design.

Purpose of the Study:

• To develop a novel parameterization strategy for ML-based scoring functions.
• To create a scoring function that achieves balanced performance across diverse tasks like scoring, ranking, docking, and screening.
• To enhance the accuracy and applicability of ML scoring functions in structure-based drug design.

Main Methods:

• Proposed a novel parameterization strategy incorporating an adjustable binding affinity term into mixture density network training.
• Introduced a residue-atom distance likelihood potential.
• Investigated the impact of key elements on prediction accuracy and task preference.

Main Results:

• The developed residue-atom distance likelihood potential demonstrated superior docking and screening power compared to state-of-the-art methods.
• Achieved remarkable improvements in scoring and ranking performance.
• Showcased that model performance for scoring/ranking and docking/screening tasks can be effectively balanced through the proposed parameterization strategy.

Conclusions:

• The innovative parameterization strategy and resulting scoring framework show significant potential for structure-based drug design.
• The study provides a method to balance performance across different tasks for ML scoring functions.
• Highlights the utility of the new approach for improving the efficiency and accuracy of drug discovery pipelines.