AI-Assisted Screening of Erlotinib-like Molecules for Mutant EGFR

Poster presented at the 2026 VCU VIP Showcase and Poster Symposium — 30 October 2026, Commons Ballroom, University Student Commons, Virginia Commonwealth University, Richmond, Virginia — part of Transformative Learning Week. The work is from the Vertically Integrated Project AI in Pharmaceutical Sciences: Bench to Bedside, VCU School of Pharmacy. Mutations in the epidermal growth factor receptor (EGFR) kinase domain drive a subset of non-small cell lung cancers. The T790M “gatekeeper” mutation causes resistance to first-generation reversible inhibitors such as erlotinib by increasing the kinase’s affinity for ATP; third-generation osimertinib binds irreversibly to cysteine 797 and overcomes T790M, but the C797S mutation replaces that cysteine with serine and confers resistance. We asked whether molecules built on erlotinib’s 4-anilinoquinazoline scaffold, intended as noncovalent binders, could be prioritized computationally against the T790M/C797S EGFR crystal structure 8H7X, with erlotinib as both scaffold and docking reference. Methods. A generator combined substituents at three scaffold positions (844 combinations); property filters and de-duplication left 529 molecules, ranked by Morgan-fingerprint Tanimoto similarity to erlotinib. A ring-label bug in the generator’s original text-based assembly had built all ten initially selected liberal-library structures incorrectly; the corrected builder joins molecular graphs and was cross-checked on all 844 combinations. The docking panel held erlotinib plus 29 candidates. AutoDock Vina 1.2.7 docked RDKit/Meeko-prepared ligands into one rigid 8H7X receptor and box (exhaustiveness 32; three seeds; 90 runs). Follow-up checks covered Vina score components, pose contacts, neutral and +1 charge states docked with Vina and Vinardo, and GNINA 1.3.2 rescoring. Findings. Similarity to erlotinib and docking rank were not interchangeable, and the leading candidates’ advantages changed with scoring method, pose and charge state. Twenty-six of 29 candidates scored more favorably than erlotinib under Vina, but much of the similarity trend reflected differences between library groups rather than a within-group relationship. For the three Vina leaders, fewer rotatable bonds — and so a smaller flexibility correction — outweighed weaker modeled protein-contact terms. GNINA and Vina rankings barely agreed; Vinardo erased most of the Vina leaders’ advantage; and alternative poses and charge states changed which placement looked best. In an exploratory follow-up, LibFix_07 kept a favorable mean score across the methods and states tested, but its binding orientation is unresolved. Limitations. Rigid single-ligand docking into a 3.40 Å structure cannot capture the ATP competition through which T790M acts, and all scores reported are predictions, not measured potency. Redocking the co-crystallized 8H7X ligand reproduced its pose only at the third-ranked mode, which is stated on the poster and bounds how much weight any top-ranked pose should carry. Educational context. The first four authors are undergraduate students, and they came to this problem from four different disciplines — chemistry and biochemistry, computer science, data science, and communication arts. The work is from a Vertically Integrated Project, a multi-year research team spanning academic levels, in which students use AI tools as inquiry partners under a curriculum whose central practice is verifying machine-generated claims rather than accepting them. The disciplinary mix was load-bearing rather than incidental: the errors the models did not catch were caught by different students drawing on different coursework. Catching the ring-label bug before docking, and reporting the redocking control and the scoring-function disagreements rather than the single most favorable number, are the substantive results of that practice. Analysis code, archived docking runs and checksums are held in a private team repository and are available from the corresponding author on request. Files in this record: the poster as displayed (PDF), a print-resolution PDF at final artwork size (63 × 42 in), and the editable PowerPoint source.