Artificial intelligence for Traditional Chinese Medicine classic formula development: a technology readiness assessment using metabolic syndrome as a case study

Abstract

Traditional Chinese medicine (TCM) formulas, developed through millennia of clinical practice, represent a rich legacy of approximately 100,000 documented prescriptions (99,584 named formula entries in the second edition of the Dictionary of Traditional Chinese Medicine Formulas), yet on average fewer than ten new TCM products were approved annually by China’s NMPA from 2013 to 2024 (77 approvals in total). The experience-based screening paradigm is misaligned with the precision medicine framework demanded by metabolic syndrome (MetS), a multi-target disorder affecting an estimated 1.54 billion adults (95% CrI 1.35–1.76) in 2023. Although artificial intelligence (AI)—encompassing knowledge graphs (KGs), large language models (LLMs), and high-throughput experimentation—has been proposed as a solution, its actual readiness for TCM development remains unclear. This review critically examines KG, LLM, and high-throughput validation technologies enabling precision phytomedicine for MetS, providing a realistic roadmap for AI-driven TCM research and development. The three technology streams exhibit asymmetric readiness. At the stream level, knowledge-graph approaches span Technology Readiness Level (TRL) 3–6, large language model approaches span TRL 1–5, and high-throughput validation platforms span TRL 3–5. LLM development is limited by three TCM-specific hallucination types that lack evaluation benchmarks. KG entity recognition is the most mature component, yet reasoning is hindered by the absence of negative data, missing dose–time dimensions, and unencoded monarch–minister–assistant–envoy hierarchies. High-throughput platforms are technically feasible but remain mismatched to MetS functional endpoints—tumor-derived organoid paradigms show limited direct transferability and would require substantial adaptation. Precision phytomedicine must follow a “foundations before algorithms” sequence: institutionalize negative data, build TCM-Hallu benchmarks, establish functional-endpoint evaluation, and develop AI-assisted regulatory review. Algorithmic advances will remain in an incomplete information space until knowledge infrastructure gaps are closed.