Leveraging Machine Learning and AlphaFold2 Steering to Discover State-Specific Inhibitors Across the Kinome

Protein kinases are structurally dynamic proteins that control downstream signaling cascades by phosphorylating their substrates. Protein kinases regulate their function by adopting several conformational states in their active site determined by the movements of several motifs such as the αC-Helix, DFG residues and the activation loop. Each conformational state represents a distinct physicochemical environment that accepts or precludes ligand binding. However, most of the kinome have not been crystalized across these possible conformational states. It has been shown that shallow Multiple Sequence Alignments (MSA) can enable AlphaFold2 (AF2) to model kinases in alternative conformations. However, it is unclear if these models can be leveraged for structure-based drug discovery. Additionally, there are several machine learning tools to predict protein-ligand interactions based on ligand chemotype and binding pocket properties, but these models cannot be used to identify ligands with clear state specificity. Here, we first present an approach called AlphaFold2 Steering (AF2-Steering), a systematic methodology to direct AF2 to sample kinases in the active and inactive conformations. We use our approach to model the protein kinome in precise conformational states. We demonstrate the utility of these AF2-steered kinase models by employing them in a prospective virtual screening study that integrates machine learning with docking to find state specific inhibitors for well-studied and dark kinases that lack structures in the active conformational state. We then experimentally validate the hits, an essential step often overlooked, and later experimentally confirm the conformation-specificity of the ligands identified for FLT3, a protein kinase that currently lacks an active state crystal structure. Against a strict binding criterion of at least 1μM Kd, our modelled structures achieved an overall hit rate of 53%. We also confirm the conformation-specificity of 4/7 FLT3 ligands, thus demonstrating the value of MSA-steered AF2 modelled kinase structures combined with machine learning and docking to guide conformation-specific kinase drug discovery.