Global Research Trends in AI-Assisted Natural Product Drug Discovery: A Scientometric and Keyword Co-Occurrence Network Analysis (2016-2025)

Abstract

Objectives: Artificial Intelligence (AI) methods are increasingly applied to natural-product chemistry and drug discovery, but the resulting literature has grown rapidly and unevenly across subthemes. This study maps the structure and evolution of this interdisciplinary research front using bibliometric network analysis, addressing the gap left by narrative reviews that describe individual AI applications without quantifying how the broader field is organized. Materials and Methods: A total of 4277 English-language articles and reviews published between 2016 and 2025 were retrieved from Scopus using a Boolean search string combining natural-product/ phytochemical terminology with AI/machine-learning terminology. A Python-based co-occurrence network analysis (NetworkX, Louvain community detection, scikit-learn TF-IDF labeling) was used to build a keyword co-occurrence network, detect clusters, and identify keyword bursts-conceptually equivalent to CiteSpace/VOSviewer-style mapping. Generic MEDLINE/Emtree indexing terms (e.g, “article,” “human,” “nonhuman”) were removed from the Index Keywords field before network construction to prevent indexing artifacts from dominating the map. Results: Annual output rose from 90 documents in 2016 to 1282 in 2025, with two especially sharp expansion years (+ 65.66% in 2020% and + 61.66% in 2025). The keyword co-occurrence network (n=120, E=6903, density=0.967) resolved into five clusters (Q=0.1514, weighted mean silhouette S=0.5326): network pharmacology/herbal-medicine mechanisms (n=30), natural-product genome mining and biosynthesis informatics (n=26), machine learning/ AI predictive methods (n=24), molecular docking and structure-based virtual screening (n=22), and phytochemical profiling/analytic chemistry (n=18). All five clusters carry near-identical mean publication years (2022.5-2022.8), indicating the field’s major subthemes matured together rather than sequentially. Keyword bursts in 2024-25 cluster around drug-delivery and nanocarrier engineering (nanocarrier, encapsulation, targeted drug delivery) and oncology applications (tumor microenvironment, antineoplastic agents, phytogenic). The relatively low network modularity, despite high publication volume, signals a field whose vocabulary is still highly interconnected across subthemes rather than splintering into separate specialities-useful information for funders and journal editors deciding whether to treat this as one integrated field or several emerging sub-disciplines. Unlike previous narrative reviews that describe individual AI applications in isolation, this study provides the first corpus-level, quantitative map of how 4277 documents collectively self-organize into thematic clusters, revealing a structurally compressed field (all five subtheme clusters share mean publication years within a 0.3-y band) whose vocabulary remains too densely interconnected to be treated as separate specialities. Conclusion: A finding that no narrative review can produce and that has direct implications for funding allocation, journal scope decisions, and future research design. The study also demonstrates that output leadership (China) and citation-impact leadership (United States, Canada, Switzerland, Germany) are empirically dissociated, providing actionable intelligence for international collaboration strategy.