A novel framework for the discovery of MAPK-activated protein kinase 2 (MAPKAPK2) inhibitors using a multi-feature deep learning ensemble.

MAPKAPK2 is a promising therapeutic target in numerous diseases. However, many MAPKAPK2 inhibitors are plagued by low solubility and permeability, and none have advanced through clinical trials. New computational methods utilizing deep learning can speed up inhibitor identification. This study aims to develop and validate a novel framework for MAPKAPK2 inhibitor discovery utilizing an ensemble of ten individual models trained on various feature sets. We trained DNN models using 21 molecular featurizers and 28 layer-size settings, and selected ten high-performing feature-architecture combinations to establish the ensemble. We explored various voting methods in conjunction with the ensemble and used the ensemble to generate a ranking of potential MAPKAPK2 inhibitors from an in-house compound set. Potential inhibitors satisfying Lipinski and Veber Rules not containing PAINS structures were selected for enzyme assay testing, and a molecular docking simulation was performed to investigate interactions. The individual model with the highest evaluation metrics was trained on functional-class fingerprints. Meanwhile, the ten-model voting ensemble reported an accuracy of 0.969 on a testing set. One novel MAPKAPK2 inhibitor, S021-0180, was identified out of seven tested with enzyme assays. The molecular docking simulation revealed critical ligand-residue interactions within the binding site. The novel computational framework was successful in identifying a novel MAPKAPK2 inhibitor as a promising inhibitor for further optimization in future studies. The established ensemble can be used to evaluate more compound sets for novel MAPKAPK2 inhibitors. Moreover, we anticipate that this new framework can be applied to all protein kinases for rapid compound screening.