Generative AI for Drug Molecule Design

Abstract

Generative AI has opened a qualitatively new approach to drug molecule design: rather than screening large libraries of existing compounds against target proteins, AI generative models can propose novel molecular structures de novo, conditioned on desired physicochemical, pharmacological, and target-binding properties. This approach has accelerated lead generation, expanded the chemically accessible design space, and produced several compounds that have entered clinical trials. This study evaluates nine generative AI models for drug molecule design across five design tasks: target-specific ligand generation, multi-property optimisation, scaffold hopping, ADMET property prediction, and synthetic accessibility scoring. Models evaluated include SMILES-based language models (MolGPT, ChemBERTa-Gen), graph generative models (JTVAE, GDSS), diffusion-based 3D structure generators (DiffSBDD, TargetDiff), reinforcement learning molecular optimisers (REINVENT4, FREED), and a proposed Multi-Objective Molecular Generation framework (MOMG). Evaluation benchmarks include GuacaMol, MOSES, DUD-E (docking-based), and a purpose-built Multi-Target Benchmark (MTB) covering five disease targets (EGFR, PD-L1, SARS-CoV-2 Mpro, GPR119, and BTK). Experimental validation was performed for a subset of MOMG-designed molecules (n=24) using SPR binding assays. MOMG achieves the best multi-property optimisation score on GuacaMol (0.884). TargetDiff achieves the best docking score on DUD-E (-8.84 kcal/mol mean). REINVENT4 achieves the best ADMET profile balance. Experimental validation confirms 8 of 24 MOMG-designed molecules achieve Kd < 100nM against EGFR (33.3% experimental hit rate). A drug discovery AI pipeline integrating all five tasks is proposed.