Equivariant Scalar Fields for Molecular Docking with Fast Fourier Transforms

Explore an innovative approach to molecular docking in this 38-minute conference talk by Bowen Jing from Valence Labs. Dive into the world of AI-driven drug discovery as Jing presents a novel method using equivariant scalar fields and fast Fourier transforms to accelerate structure-based virtual screening. Learn how machine learning can enhance the optimization of scoring functions in docking algorithms, potentially revolutionizing the throughput of drug discovery workflows. Discover the unique functional form of the proposed scoring function, which utilizes multi-channel ligand and protein scalar fields parameterized by equivariant graph neural networks. Understand the advantages of this approach in virtual screening settings, particularly for scenarios with a common binding pocket. Follow along as Jing discusses the method's performance in decoy pose scoring and rigid conformer docking, comparing it to established scoring functions like Vina and Gnina. Gain insights into the robustness of this technique when applied to computationally predicted structures and its potential impact on the field of drug discovery.

- Intro + Background
- Scalar Fields
- Cross-Correlation
- Training
- Inference
- Experiments
- Conclusions