Virtual screening is an important part of the drug discovery process, as it can quickly and cheaply select compounds for evaluation. Improving virtual screening performance would reduce the cost of drug discovery and, by identifying better compounds earlier, potentially reduce failure rates. Machine learning has been shown to be a useful tool for predicting either a compound’s potential binding pose or the affinity of the binding. Our previous work, gnina, utilized convolutional neural networks and a 3D grid based representation of atomic densities to score protein-ligand binding poses. As protein-ligand binding affinity is pose dependent, we reason that there is potential benefit to a joint training approach that simultaneously scores the protein-ligand pose and predicts the binding affinity of that pose. Additionally, recent work has shown the danger of utilizing machine learning on improperly curated datasets, leading to overly optimistic estimates of generalization error, as well as a lack of available training data. In order to address these issues we present a new dataset that aims to expand the available training data by docking ligands into different protein structures with similar binding pockets as defined by the Pocketome database. This expands the available training data by several orders of magnitude while also better mimicking a prospective evaluation. We also evaluate different training/test splits and observe significant differences in estimated performance depending on how the test set is constructed. Overall, we find that grid-based 3D CNNs can achieve comparable affinity prediction performance (R=~0.8) to other machine learning methods, but this is a significant overestimate. More robust evaluations yield significantly lower correlations (R=~0.5).

Predicting protein-ligand binding affinity with gnina

2019 Fall National Meeting

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