MolGraphGAN: a graph transformer-adversarial framework for target-specific molecule generation

The synthesis of small molecule target-specific compounds remains challenging due to the combinatorial complexity of chemical space and the limited interpretability of existing deep generative models. This work introduces MolGraphGAN, a graph transformer-adversarial architecture designed to generate chemically valid, synthetically accessible and target-conditioned drug-like compounds. Compared to standard adversarial approaches that start from noisy random inputs, MolGraphGAN is trained directly on real molecular graphs that are derived from canonical SMILES strings modelled from benchmark datasets such as ZINC-250K and ChEMBL-27 to enable the model to learn realistic atomic connectivity and chemical combinations. The proposed system integrates an Edge-Aware Attention (EAA) module, a novel enhancement inspired by Graph Attention Networks (GATs), which explicitly incorporates bond-type embeddings to improve atom-bond interaction learning and long-range dependency modelling within molecular structures. MolGraphGAN demonstrates scalability across multiple therapeutic classes, showing robust performance on diverse datasets and protein targets. The synthesized molecules are intensively evaluated by benchmarking metrics, which include validity, novelty, uniqueness, diversity, QED and SA, alongside target-specific in silico protocols encompassing molecular docking and drug-target interaction predictions. Additionally, MolGraphGAN embeds explainable AI protocols through the presentation of attention maps highlighting substructure having maximum influence on predicted binding such that it provides interpretability alongside generative effectiveness. By integrating adversarial learning, transformer-constructed graph models, edge-aware attention and explainability, MolGraphGAN provides a generalizable and experimentally usable pipeline for accelerating target-centric drug discovery in a variety of therapeutics.