From Algorithmic Prediction to Therapeutic Evidence: Artificial Intelligence in Drug Discovery, Translation, and Responsible Governance
TL;DR
This structured critical narrative review examines AI as part of an iterative discovery system linking data, algorithms, medicinal chemistry, experimental biology, pharmacology, manufacturing, and clinical development and proposes five author-defined evidence levels and an author-synthesized stage-gated governance framework.
Abstract
Artificial intelligence (AI) now supports target identification, molecular representation, virtual screening, structure prediction, de novo design, synthesis planning, drug repurposing, and prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET). These capabilities can expand the chemical and biological hypotheses evaluated before costly experiments, but computational novelty is not equivalent to a safe or effective medicine. This structured critical narrative review examines AI as part of an iterative discovery system linking data, algorithms, medicinal chemistry, experimental biology, pharmacology, manufacturing, and clinical development. The synthesis emphasizes evidence quality, applicability domain, prospective validation, reproducibility, and regulatory credibility. A detailed case analysis of halicin shows both the legitimate achievement and the boundary of AI-prioritized discovery: a graph neural network selected a structurally unusual antibacterial candidate, followed by laboratory and animal validation, yet the finding did not establish clinical efficacy or approval. Additional analysis covers AlphaFold-enabled structural hypothesis generation, AI-designed molecules, multi-objective optimization, ADMET modeling, automated laboratories, clinical-development support, and AI-enabled target discovery. The article proposes five author-defined evidence levels and an author-synthesized stage-gated governance framework. Major limitations include biased or noisy data, distribution shift, shortcut learning, inadequate uncertainty estimation, proprietary opacity, environmental cost, intellectual-property ambiguity, and unequal access to data and computation. AI can accelerate learning and improve prioritization, but therapeutic value requires progressively stronger experimental, translational, clinical, and, where applicable, regulatory evidence.