Artificial Intelligence-Driven Natural Product Drug Discovery: From Computational Genome Mining to Clinical Translation.

Abstract

Natural products (NPs) have historically yielded numerous therapeutic agents, yet their integration into modern drug discovery has been constrained by chemical complexity, low abundance, laborious dereplication, and limited target annotation. Convergence of multi-omics technologies with high-resolution structural and biological data has created unprecedented opportunities for artificial intelligence (AI) to accelerate NP-based therapeutics development. This review provides an operational, end-to-end workflow that explicitly connects computational predictions to medicinal chemistry decision points, addressing a critical gap between computational prediction and clinical translation. We trace the complete discovery pipeline: computational mining of biosynthetic gene clusters (BGCs) and metabolomes, deep learning (DL)-assisted structural elucidation and dereplication, network-based target identification using protein-ligand prediction, and generative molecular design inspired by NP scaffolds (including large language models, diffusion models, and genetic algorithms). Critical evaluation of current limitations (data scarcity, lack of standardized ontologies, model interpretability) is complemented by discussion of emergent strategies (foundation models trained on multi-modal data, graph neural networks, autonomous closed-loop laboratories). Representative case studies, including the synthetic AI-designed clinical benchmark rentosertib, illustrate the current evidence spectrum from discovery-level validation to early clinical benchmarking, while also highlighting that most AI-enabled NP discovery workflows remain at the preclinical or proof-of-concept stage, with limited quantitative evidence of improved clinical productivity. We conclude with an Outlook proposing feasible developments for 2025-2030: self-driving laboratories with reported acceleration in specific experimental contexts, foundation models enabling hypothesis-free chemical space exploration, and sustainability-aware AI frameworks embedding biodiversity impact assessments. This operational focus fills a critical gap between algorithmic capability and clinically actionable NP-derived leads. Importantly, while AI has demonstrably accelerated several early discovery steps, quantitative comparisons with classical NP workflows remain limited, and most reported advances are supported by preclinical or proof-of-concept studies rather than systematic evidence of improved time-to-lead, cost reduction, or clinical success rates.