Integrative machine learning-guided in silico and in vitro approach reveals selective small molecule inhibitors targeting mutant IDH1

Abstract

Mutations in Isocitrate Dehydrogenase 1 (IDH1) are found in over 80% of WHO grade II/III gliomas. These mutations, specifically R132H, confer a neomorphic activity that converts α-ketoglutarate (α-KG) into the oncometabolite 2-hydroxyglutarate (2HG), a key driver in glioma pathogenesis. While the therapeutic potential of targeting mutant IDH1 (MT-IDH1) is established, the discovery of novel and selective inhibitors remains a priority. Leveraging the growing availability of pharmacological data, we developed regression-based ML models to predict pIC₅₀ values and identify potent inhibitors of MT-IDH1. We trained these models using a dataset of 1631 compounds from ChEMBL, utilising 208 molecular descriptors derived from RDKit. Among the three algorithms evaluated, the Random Forest model demonstrated superior performance, achieving high predictive accuracy on the training set and robust generalisability on the test set. Feature importance analysis provided critical insights related to lipophilicity, halogen, and electronic factors as the key molecular determinants of inhibitor activity. We subsequently deployed this model to screen drug databases, identifying five promising hits. These candidates were further validated through in silico molecular docking, molecular dynamics simulation, and MM/PBSA free energy calculations. Experimental in vitro enzymatic assays confirmed that these compounds selectively inhibit MT-IDH1 with IC₅₀ values in the micromolar range, while exhibiting no significant activity against the WT-IDH1. While the mechanism of action of these compounds as IDH inhibitors is yet to be established, our results support these compounds as potent and selective hits. They offer a promising foundation for structural optimisation and the development of next-generation therapeutics against MT-IDH1 malignancies.