<br/>Kinase selectivity is emerging as an important area in drug discovery and clinical development. Finding kinase inhibitors with a particular selectivity profile is also a notoriously challenging problem for both structure-based methods and ligand-based methods given the similarity of the commonly targeted ATP binding site across multiple kinases. Here we apply structure-based convolutional neural networks and ligand-based methods to predict the selectivity profile of kinase inhibitors drawn from public sources of kinase binding data. We compare their performance using both a traditional cross-fold validation approach and an independent dataset for kinase selectivity. We introduce a performance metric termed compound AUC to summarize the degree of similarity between the predicted and observed kinase selectivity profiles for a particular compound. Ligand-based methods and structure-based methods have individual merits, and we examine the strengths and weaknesses of each approach. Our approach to predicting binding differences for similar proteins and compounds has natural extensions into other areas of drug discovery where it is important to bind to a specific target within a highly conserved protein family.

Combining structure-based convolutional neural networks and ligand-based methods to address kinase selectivity

2019 Fall National Meeting

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