A Consistent Scheme for Gradient-Based Optimization of Protein-Ligand Poses.

Area of Science:

• Computational chemistry
• Structural biology
• Bioinformatics

Background:

• Protein-ligand pose scoring and optimization are critical for molecular docking.
• Existing scoring functions often lack continuous differentiability and are not optimized for numerical stability.
• Analysis of numerical optimization behavior of scoring functions is rarely performed.

Purpose of the Study:

• To present a consistent scheme for protein-ligand pose scoring and gradient-based pose optimization.
• To introduce a novel scoring function (JAMDA) and an optimized BFGS algorithm (LSL-BFGS).
• To evaluate the performance and numerical properties of the proposed scheme.

Main Methods:

• Developed a novel variant of the BFGS algorithm with step-length control: LSL-BFGS (limited step length BFGS).
• Created the empirical JAMDA scoring function, designed for pose prediction and numerical optimizability.
• Validated the JAMDA scoring function using the CASF-2016 benchmark for docking power.

Main Results:

• The JAMDA scoring function achieved high pose prediction performance, ranking top poses within 2 Å RMSD in approximately 89% of cases.
• The combination of JAMDA scoring with LSL-BFGS demonstrated superior optimization locality compared to the classical BFGS algorithm.
• The LSL-BFGS algorithm maintained a low number of scoring function evaluations, similar to the classical BFGS.

Conclusions:

• The presented JAMDA scoring and LSL-BFGS optimization scheme offers a consistent and effective approach for protein-ligand pose prediction.
• The novel scheme enhances optimization locality and maintains high prediction accuracy, outperforming traditional methods.
• The JAMDA scoring and optimization tools are available for non-commercial and academic use, facilitating further research in drug discovery.