Optimization of molecular docking scores with support vector rank regression

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Molecular docking is crucial for identifying potential drug candidates.
• Existing scoring functions often struggle with accurate prediction of binding poses and ligand ranking.
• There is a need for improved methods to integrate and optimize docking scores.

Purpose of the Study:

• To introduce and evaluate the Support Vector Rank Regression (SVRR) algorithm for optimizing molecular docking scores.
• To assess the performance of SVRR in improving binding conformation prediction and ligand screening.
• To compare SVRR with traditional machine learning approaches for docking score optimization.

Main Methods:

• Integration of seven original docking scores from two software using the SVRR algorithm.
• Evaluation of SVRR performance in binding conformation prediction tests.
• Assessment of SVRR in large-scale compound library screening (LS) tests.
• Comparison of SVRR accuracy with Support Vector Classification and Regression on identical datasets.

Main Results:

• SVRR scores achieved an average 12.1% improvement in correctly ranking binding conformations.
• An average RMSD improvement of 16.7% was observed for top-ranked conformations.
• SVRR demonstrated a 46.3% improvement in ranking the correct ligand first in LS tests.
• SVRR exhibited high robustness and generalizability across different training datasets and strategies.
• Traditional SVR algorithms showed significantly lower accuracy improvements compared to SVRR.

Conclusions:

• The SVRR algorithm offers a robust and generalizable approach to optimize molecular docking scores.
• SVRR significantly enhances the accuracy of both binding conformation prediction and ligand screening.
• SVRR holds potential for developing a new class of highly accurate integrative docking scoring functions.