Protein kinases have been a focus of drug discovery efforts due to their central roles in regulating cell cycle, metabolism, and proliferation. Most small molecule kinase inhibitors target the highly conserved ATP-binding pocket, which may lead to selectivity and safety challenges. <i>In silico</i> prediction of kinome-wide activity can aid the evaluation of potential kinase off-targets and selectivity with a much higher throughput than experimental profiling. Here, a multitask Message-Passing Neural Network (MPNN) model for a panel of 300 kinases was developed using public and proprietary datasets of over 20,000 compounds. This model achieves favorable performance on a structurally distinct test set with an auROC of 0.93. Compared to single-task random forest models, the MPNN model yields similarly good performance for large and balanced datasets, but significantly outperforms on a majority of kinases where the datasets are small or imbalanced. Moreover, using a composite selectivity score derived from the activity predictions of individual kinases, the MPNN model was able to differentiate between type I and type III kinase inhibitors based on selectivity profile. Lastly, the MPNN model can provide simple molecular rationales for the observed activity on a chosen kinase.

Kinome-wide activity profile prediction using multitask message-passing neural networks

ACS Spring 2022

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