Reproduce “Ligand Pose Optimization with Atomic Grid-Based Convolutional Neural Networks” on OpenResearch

About this paper

Matthew Ragoza, Lillian Turner, David Ryan Koes

This research investigates using gradients from an atomic grid-based convolutional neural network (CNN) for direct ligand pose optimization. An iterative training procedure enables a second CNN (CNN2) to slightly improve average RMSD for random poses and generate more physically realistic structures compared to initial CNNs or AutoDock Vina.

Key insights

• A convolutional neural network trained for pose discrimination can be effectively used for gradient-based pose optimization, provided its input representation (e.g., 3D atomic grid) is differentiable.
• An iterative training procedure, where a subsequent CNN is trained on a dataset augmented with poses optimized by a previous CNN, is crucial for improving optimization performance and generating more physically realistic structures.
• The initial training data distribution significantly influences the CNN's optimization behavior, with iteratively trained models learning to navigate conformational space more effectively and avoid non-physical outcomes.

Summary by alphaXiv.