Equivariant Scalar Fields for Molecular Docking with Fast Fourier Transforms

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Molecular docking is essential for structure-based virtual screening.
• Current docking algorithms are limited by computationally expensive scoring function optimization.

Purpose of the Study:

• To accelerate molecular docking workflows using machine learning.
• To develop a novel, rapidly optimizable scoring function.

Main Methods:

• Utilized equivariant graph neural networks to parameterize a scoring function based on ligand and protein scalar fields.
• Employed fast Fourier transforms for rapid optimization over rigid-body degrees of freedom.
• Benchmarked the method on decoy pose scoring and rigid conformer docking tasks.

Main Results:

• Achieved performance comparable to Vina and Gnina on crystal structures, but with significantly faster runtimes.
• Demonstrated increased robustness on computationally predicted structures.
• The method's runtime is amortizable, especially for virtual screening with common binding pockets.

Conclusions:

• Machine learning can effectively accelerate molecular docking by enabling rapid scoring function optimization.
• The proposed method offers a promising alternative for high-throughput virtual screening.
• The approach shows potential for improved accuracy with predicted structures.