Deep Generative Models for Drug Design

Abstract

Deep generative models have transformed small-molecule drug design by replacing brute-force virtual screening of commercial libraries (10^7-10^9 compounds) with targeted generation of novel molecules satisfying multiple design constraints. These models learn the distribution of drug-like chemical space from training data (ChEMBL, ZINC, DrugBank) and sample from it to propose new molecules with desired target affinity, physicochemical properties, synthesisability, and safety profile. Variational autoencoders, generative adversarial networks, reinforcement learning agents, diffusion models, and autoregressive transformers each bring different strengths to this multi-objective generative design problem. We present the Generative Drug Design Framework (GDDF), evaluating five deep generative approaches -- SMILES-based VAE, graph-based GAN, RL-guided property optimisation, 3D diffusion for structure-based design, and chemical language transformers -- across four drug design tasks (target-specific lead generation, multi-property optimisation, scaffold hopping, and fragment-based design). Our Generative Drug Design Score (GDDS) measures chemical validity, novelty versus training data, target activity prediction, drug-likeness, and synthesisability. 3D diffusion models achieve the highest GDDS (0.924) through structure-based generation that directly accounts for protein pocket geometry, while chemical language transformers achieve the highest novelty and scaffold diversity (0.960) through learned chemical grammar that extrapolates beyond training distribution