Generative AI for Drug Discovery: A GPT-2 and LSTM Based Models for Designing EGFR Inhibitors

Abstract

The design of novel EGFR inhibitors is a critical focus in cancer drug discovery. This study leverages generative AI models, specifically fine-tuned GPT-2 and LSTM architectures, to generate new chemical structures targeting the Epidermal Growth Factor Receptor (EGFR). The models were trained on a curated dataset of approximately 500,000 bioactive ligands from the ChEMBL database, with SMILES strings as input. After generating a batch of 1,000 molecules, post-generation filtering was applied based on Lipinskis rule of five, Quantitative Estimate of Drug-likeness (QED), and Synthetic Accessibility Scores (SAS) to ensure drug-like properties. Molecular docking studies were performed using PyRx with AutoDock Vina, focusing on the crystal structure of EGFR (PDB ID: 1M17). The LSTM model outperformed GPT-2, achieving a higher validity rate (90.98% vs. 52.27%) and similar uniqueness rates, while also producing molecules with stronger binding affinities, ranging from -9.4 to -10.4 kcal/mol. The results indicate that the LSTM model is more effective for generating chemically valid EGFR inhibitors, offering promising candidates for further experimental validation. This study demonstrates the potential of generative AI to accelerate the identification of effective cancer therapeutics.