How far are we in the rapid prediction of drug resistance caused by kinase mutations?

Abstract

Protein kinases regulate various cell signaling events in a diverse range of species through phosphorylation. The phosphorylation occurs upon transferring the terminal phosphate of an ATP molecule to a designated target residue. Due to the central role of protein kinases in proliferative pathways, point mutations occurring within or in the vicinity of ATP binding pocket can render the enzyme overactive, leading to cancer. Combatting such mutation-induced effects with the available drugs has been a challenge, since these mutations usually happen to be drug resistant. Therefore, the functional study of naturally and/or artificially occurring kinase mutations have been at the center of attention in diverse biology-related disciplines. Unfortunately, rapid experimental exploration of the impact of such mutations remains to be a challenge due to technical and economical limitations. Therefore, the availability of kinase-ligand binding affinity prediction tools is of great importance. Within this context, we have tested six state-of-the-art web-based affinity predictors (DSX-ONLINE, KDEEP, HADDOCK2.2, PDBePISA, Pose&Rank, and PRODIGY-LIG) in assessing the impact of kinase mutations with their ligand interactions. This assessment is performed on our structure-based protein kinase mutation benchmark, BINDKIN. BINDKIN contains 23 wild type-mutant pairs of kinase-small molecule complexes, together with their corresponding binding affinity data (in the form of IC50, Kd, and Ki). The web-server performances over BINDKIN show that the raw server predictions fail to produce good correlations with the experimental data. However, when we start looking in to the direction of change (whether a mutation improves/worsens the binding), we observe that over Ki data, DSX-ONLINE achieves a Pearsons R correlation coefficient of 0.97. When we used homology models instead of crystal structures, this correlation drops to 0.45. These results highlight that there is still room to improve the available web-based predictors to estimate the impact of protein kinase point mutations. We present our BINDKIN benchmark and all the related results online for the sake of aiding such improvement efforts. Our files can be reached at https://github.com/CSB-KaracaLab/BINDKIN