De Novo Antibiotic-Like Molecule Design via Descriptor-Guided Patch-Based GANs

Abstract

The rise in antimicrobial resistance necessitates rapid discovery of novel antibiotics, but the majority of generative pipelines are unable to produce clinically viable candidates. The structural and topological complexity of actual antibiotics is not captured by current models, which lack pharmacological depth and are frequently trained on datasets such as QM9 and ZINC. We present a modular, descriptor-guided framework for antibiotic-like molecule design that combines a property-aligned β-VAE for interpretable encoding, a descriptor-to-latent conditioner for controllable sampling, and a patch-based graph generator for fragment-wise synthesis. Trained on a curated subset of ChEMBL containing clinically validated antibiotics, our framework supports end-to-end generation from RDKit descriptors to final molecules, with adversarial refinement for topological realism. Beyond favorable ADMET profiles, our method establishes joint satisfaction reporting of drug-likeness thresholds as a benchmark standard, resists mode collapse under 50k-sample stress tests, and surfaces ligands that outperform ciprofloxacin and co-crystal references in docking assays. These results highlight chemically meaningful, pharmacologically informed generation − overcoming limitations of black-box pipelines and general-purpose datasets.