Crowdsourced mapping of unexplored target space of kinase inhibitors

Despite decades of intensive search for compounds that modulate the activity of particular proteins, there are currently small-molecule probes available only for a small proportion of the human proteome. Effective approaches are therefore required to map the massive space of unexplored compound-target interactions for novel and potent activities. Here, we carried out a crowdsourced benchmarking of the accuracy of machine learning (ML) algorithms at predicting kinase inhibitor potencies across multiple kinase families. A total of 268 ML predictions were scored in unpublished bioactivity data sets. Top-performing algorithms used kernel learning, gradient boosting and deep learning, with predictive accuracy exceeding that of target activity assays. Subsequent experiments carried out based on the the top-performing model predictions demonstrated that these models and their ensemble can improve the accuracy of experimental mapping efforts, especially for so far under-studied kinases. The open-source ML algorithms together with the novel dose-response data for 905 bioactivities between 95 compounds and 295 kinases provide a unique resource for extending the druggable kinome.