Machine-Learning-Guided Identification of Potent Osimertinib Analogues Targeting the EGFR Triple Mutant

Background: Epidermal growth factor receptor (EGFR) mutations, including exon 19 deletions and L858R, are major drivers of advanced non-small cell lung cancer (NSCLC). Although osimertinib is an established EGFR tyrosine kinase inhibitor (TKI), the emergence of T790M and C797S mutations can reduce its effectiveness. This study aimed to identify and computationally prioritize osimertinib analogues with potential activity against the EGFR L858R/T790M/C797S triple mutant using an integrated machine-learning-assisted cheminformatics and structure-based approach. Methods: A library of 297,420 osimertinib analogues retrieved from PubChem was represented using physicochemical descriptors and 2048-bit Morgan fingerprints. Principal component analysis, t-SNE, UMAP, K-means clustering, and HDBSCAN were used to characterize chemical space, while Tanimoto similarity and Isolation Forest analysis were applied to assess structural similarity and scaffold novelty. Shortlisted compounds were subsequently evaluated using Glide extra-precision (XP) docking, induced-fit docking (IFD), molecular dynamics (MD) simulations, and density functional theory (DFT) calculations. Results: K-means analysis identified an osimertinib-containing cluster comprising 17,149 compounds, providing a focused chemical space for analogue prioritization. A clear inverse relationship was observed between Tanimoto similarity to osimertinib and quantitative estimate of drug-likeness (QED) among the top 30 compounds (Pearson’s r = −0.86). Three compounds satisfied the predefined combination of Tanimoto similarity ≥ 0.4 and QED ≥ 0.5. Subsequent molecular modelling highlighted compounds 155515815 and 122474632, together with osimertinib, as compounds exhibiting favorable and stable interactions with key EGFR residues. MD simulations further supported the stability of the protein–ligand complexes, while DFT analysis provided additional insight into their electronic and molecular properties. Conclusions: The integrated machine-learning-assisted and structure-based workflow enabled systematic exploration of a large osimertinib analogue library and identified computationally prioritized candidates for further investigation against the EGFR L858R/T790M/C797S triple mutant. Compounds 155515815 and 122474632 showed favorable computational profiles across multiple analyses and may serve as starting points for further scaffold optimization. Experimental studies are required to determine their biological activity and potential to address osimertinib resistance.