Protein-ligand scoring functions rank and score molecular structures with the goals of accurately predicting binding affinity, selecting the correct binding pose, and distinguishing between binders and non-binders. Creating a function that can accurately do this has many positive implications pertaining to computational drug discovery. Convolutional neural networks are a specific style of neural networks that leverage translationally equivariant sets of weights that define filters. We developed a convolutional neural network model for protein-ligand scoring that takes a 4-D representation of a protein-ligand structure. Here we describe methods for visualizing the decision process of a trained network by projecting heatmaps onto the input structure using masking, layerwise-relevance propagation, and gradient-based approaches. Additionally, we provide direct visualizations of the first convolutional layer of the network, which reveal the low-level features that maximally activate this this initial layer.

Visualizing convolutional neural network scoring of protein-ligand binding

255th National Meeting (2018)

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