AI-Based Drug Discovery Pipelines

Abstract

Drug discovery is among the most expensive and failure-prone endeavours in applied science, with average development costs exceeding 2.6 billion USD per approved drug and clinical success rates below 10%. AI-based approaches promise to accelerate every stage of the pipeline -- from target identification through molecular design, property prediction, and clinical trial optimisation -- yet systematic comparison of AI methods across the full pipeline remains scarce. This study evaluates six AI approaches spanning four drug discovery stages: target identification (network-based disease-gene prediction), molecular generation (variational autoencoders, reinforcement learning, and diffusion-based de novo design), property prediction (graph neural networks for ADMET forecasting), and virtual screening (contrastive learning for binding affinity prediction). Evaluations were conducted on standardised benchmarks: DisGeNET for target identification, ZINC-250K and GuacaMol for molecular generation, ADMET-Benchmark for property prediction, and PDBbind/DUD-E for virtual screening. A total of 2,280 experiments were conducted. Diffusion-based molecular generation achieved the highest novelty (98.4 +- 0.6% novel structures) and drug-likeness (QED = 0.72 + 0.04) simultaneously -- a combination that VAE and RL approaches could not match (novelty-QED Pareto dominated). Graph neural networks achieved AUROC = 0.842 +- 0.012 on ADMET property prediction, outperforming fingerprint-based methods by 4.8 points through learned molecular representations. Contrastive binding affinity models achieved enrichment factor EF1% = 42.8 +- 3.4 on DUD-E, outperforming traditional docking by 2.4x at 100x lower computational cost. End-to-end pipeline integration -- connecting target identification through generation and screening - produced 3.2x more experimentally validated hit compounds than any single-stage optimisation. A practical AI drug discovery pipeline guide mapping discovery stage, data availability, and computational budget to recommended methods is proposed.