Machine-learning-guided discovery of selective CK2 inhibitors with reduced CDK off-target activity

A key challenge in drug development is identification of druggable targets, the modulation of which attenuates disease progression, while avoiding inhibition of proteins that lead to dose-limiting toxicities. Here, we investigate a drug target, casein kinase 2 (CK2) - a serine/threonine kinase implicated in multiple cancers - for which existing inhibitors have so far failed in clinical trials. Using molecular and pharmacoepidemiology approaches, we show that small molecules targeting cyclin-dependent kinase (CDK) family members CDK1/2/7/9, including the existing CK2 inhibitors, have a higher risk to induce adverse effects or failing in clinical trials. We then establish a machine learning (ML)-assisted discovery pipeline to redesign more specific and allosteric lead compounds against CK2, with more selective on-target binding and favourable off-target profile. Importantly, we show that such selective design is possible when standard molecular docking, CDK-liability filtering and ML algorithms are combined with an error prediction model. Taken together, we report a simple yet efficient ML-powered drug discovery pipeline and novel submicromolar inhibitors targeting the clinically relevant CK2 kinase with no clinically approved antagonists currently available. Our prediction pipeline was able to achieve a 90% hit-rate, significantly reducing the need for subsequent wet-lab validation.