Abstract
Lung cancer remains the leading cause of cancer-related mortality worldwide, necessitating innovative strategies for accelerating drug discovery. This study proposes a deep learning-based molecular generation framework that integrates fragment-based drug design, Relational Graph Convolutional Networks (RGCN), and a Wasserstein Generative Adversarial Network (WGAN) to discover novel compounds targeting the P2X7 receptor (P2X7R), a validated oncogenic driver in lung cancer progression. The framework begins with the fragmentation of known P2X7R-targeting drugs to create a fragment library, which is then used to construct new candidate molecules. These were encoded as molecular graphs to enable RGCN-based feature learning and subsequently passed to a WGAN generator for de novo molecule creation. Evaluation of generated compounds included assessments of chemical validity, drug-likeness (via Lipinski’s Rule of Five), quantitative estimate of drug-likeness (QED), lipophilicity (LogP), and structural similarity using the Tanimoto coefficient and molecular docking.
The model generated 4,498 valid molecules, of which 614 were unique and 100% novel. Approximately 64% of these satisfied drug-likeness criteria, with QED values predominantly above 0.6 and LogP values within the optimal pharmacokinetic range. The Tanimoto similarity between the novel molecule and the seed molecule yielded a maximum score of 59%. However, the molecular docking results showed that the novel molecule achieved a better docking score against the P2X7R than the seed molecule. Despite limited computational resources restricting training data to 5,000 SMILES samples, the framework demonstrated robust molecular diversity, high validity, and strong physicochemical alignment with drug-like characteristics. The study validates the feasibility of integrating fragment-based molecular generation, WGAN, and RGCN, providing a scalable, target-specific pipeline for AI-driven lung cancer drug discovery.