Machine learning and structure-based modelling for the identification of allosteric AKT1 inhibitors in ovarian cancer: an in silico approach

Abstract

Aberrant activation of the PI3K/AKT/mTOR signalling pathway plays a critical role in ovarian cancer progression and chemoresistance, with AKT1 representing a key therapeutic target. In this study, an integrated artificial intelligence-driven and structure-based computational workflow was developed to identify new allosteric AKT1 inhibitors. A curated dataset of 3,919 experimentally validated AKT1 inhibitors from ChEMBL was used to train regression models based on Random Forest and Extreme Gradient Boosting algorithms, which were combined through a stacking ensemble approach to enhance predictive performance. A transfer learning-optimized recurrent neural network was subsequently fine-tuned on highly potent inhibitors (pIC 50 ≥ 8.0) to generate 8,995 new candidate molecules. Multi-parameter optimization incorporating predicted potency, Quantitative Estimate of Drug-likeness (QED), Synthetic Accessibility (SA) and Pan-Assay Interference Compounds (PAINS) filtering reduced the library to 72 candidates. These compounds were evaluated by Glide Extra Precision docking against the validated AKT1 allosteric binding pocket (PDB ID: 7NH5). TIL 50 and TIL 72 demonstrated docking scores (-10.268 kcal/mol and -9.928 kcal/mol, respectively) comparable to reference inhibitor MK-2206 (-10.287 kcal/mol) and preserved critical interactions with ASN54, GLU85, and TRP80, which are essential for stabilization of the inactive PH-in conformation of AKT1. Collectively, this multi-step computational strategy identified a range of chemically diverse, synthetically accessible AKT1 allosteric inhibitor candidates suitable for further experimental validation.