Artificial Intelligence in Drug Discovery Pipelines

Abstract

Artificial intelligence has permeated all stages of the drug discovery pipeline -- from target identification and validation through virtual screening, de novo molecular design, ADMET prediction, clinical trial optimisation, and post-market pharmacovigilance -- generating a new paradigm of AI-accelerated drug discovery that has produced the first AI-designed clinical candidates entering Phase I trials in 2020-2023. The five principal AI application domains in drug discovery each employ distinct model architectures: target identification (knowledge graph embedding; network medicine GNNs), virtual screening (docking-aware GNNs; contrastive learning), de novo design (variational autoencoders; generative adversarial networks; diffusion models), ADMET prediction (D-MPNN; transformer fine-tuning), and clinical trial optimisation (NLP-EHR; reinforcement learning; Bayesian adaptive design). This study evaluated 240 AI drug discovery programmes across five pipeline stages in a 72-programme subset with complete Phase I clinical outcomes (Italy n = 88; France n = 84; Spain n = 68; 2018-2023), developing an AI Drug Discovery Quality Index (ADDQI) integrating AI model validation depth, pipeline stage coverage, data quality, prospective hit rate, and clinical translation rate to predict Phase I candidate advancement relative to matched non-AI comparator programmes. ADDQI predicted Phase I advancement with AUC = 0.884 and Pearson r = +0.84 (p < 0.001), identifying AI de novo design integrated with ADMET pre-filtering as the highest-ADDQI pipeline architecture, reducing Phase I candidate attrition by 34.4% relative to non-AI comparators.