Molecular docking is an important tool for computational drug discovery that aims to predict the binding pose of a ligand (drug) to a target protein. Identifying a correctly oriented pose requires a scoring function that has a global optimum close to the experimentally observed pose. Additionally, it should also be differentiable with respect to atomic positions so that it can be used for gradient-based pose optimization. We'll describe a differentiable grid-based convolutional neural network scoring function and explore its application in an end-to-end GPU-optimized molecular docking workflow. We'll show that convolutional neural networks trained on experimental data can successfully identify correct binding modes and meaningfully rank and score compounds. We'll also describe several visualization approaches that map the CNN score back to the atomic inputs to help guide medicinal chemistry optimization and provide insight into the functioning of the neural network. We will explore the challenges and limitations of this approach, provide both retrospective and prospective evaluation results, and discuss future directions. The entirety of our approach is available under an open-source license as part of our gnina package (https://github.com/gnina).

GNINA: Deep learning for molecular docking

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.