Identification of Novel Fourth-Generation Allosteric Inhibitors Targeting Inactive State of EGFR T790M/L858R/C797S and T790M/L858R Mutations: A Combined Machine Learning and Molecular Dynamics Approach

Targeted therapy with an allosteric inhibitor (AIs) is an important area of research in patients with epidermal growth factor receptor (EGFR) mutations. Current treatment of nonsmall cell lung cancer patients with EGFR mutations using orthosteric inhibitors faces challenges like resistance and stopping over phosphorylation. Notably AIs have been introduced to overcome this resistance and increase inhibitory potency that binds to pockets other than the ATP-binding site (orthosteric site). Recently, fourth-generation AIs, EAI045, have been discovered to potently and selectively inhibit various EGFR mutations but limited antiproliferative effects in the absence of the antibody cetuximab. The purpose of this work is to identify nontoxic, potent small AIs through various screening pipelines and explore their molecular mechanism. In the discovery of AIs, structural similarity search, high-throughput virtual screening, and machine learning-guided QSAR modeling, several candidates were identified. Machine learning was employed to guide the QSAR model based on 2D descriptors and DFT-derived quantum chemical descriptors followed by a PCA reduction technique, which enabled the prediction of the biological activity (IC 50 ) of screened drugs against various EGFR mutations such as T790M/L858R/C797S and T790M/L858R. In addition, multinanosecond (ns) and microsecond (μs) classical molecular dynamics (MD) simulations run on protein–ligand binding complex to check the stability of binding dynamics for T790M/L858R/C797S and T790M/L858R mutations with lower IC 50 and higher docking score compounds. The molecular mechanics generalized Boltzmann surface area (MM/GBSA) calculation revealed that the five hit allosteric molecules for T790M/C797S/L858R and two for T790M/L858R mutations had a high binding affinity. The results were corroborated further by MM/GBSA employing the normal-mode analysis entropy method to perform additional screening. Furthermore, the compounds’ efficacy was confirmed using path-dependent ligand unbinding free energy techniques such as Jarzynski averaged free energy profiles obtained from adaptive steered MD, relative residence time, and umbrella sampling simulations, which were compared to a reference inhibitor. However, path-independent alchemical approaches like streamlined alchemical free energy perturbation and binding free energy estimator 2 (BFEE2) were employed to validate the results and identify potent compounds. These findings pave the way to identification of novel potential fourth-generation AIs, which require further experimental validation.