The <i>in-silico </i>comparison of protein binding sites has several applications in structure-based drug design; for understanding off-target side effects, the prediction of protein function and the repurposing of known drugs. However, traditional binding-site comparison tools suffer from many limitations due to their optimization and validation processes; benchmarking datasets tend to be small and focussed, resulting in intrinsic biases. In this work, we develop a scoring function based on data, using deep geometric learning in the form of rotation-invariant, graph neural networks (GNN). GNNs coupled with a contrastive objective function generate binding site 'fingerprints' that can be compared rapidly using simple distance metrics. We further extend our GNN architecture with a cross-graph messaging layer, improving effectiveness at the cost of efficiency. We demonstrate that our method is competitive with traditional binding-site similarity tools and a more recent CNN based approach, across a range of recently compiled, large scale benchmark datasets, consistently ranking among the top-scoring tools. Finally, we present case studies showing the utility of the trained model and its ability to generalize to unseen data.

Machine learning models for prediction of protein binding site similarity

ACS Spring 2022

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