Machine Learning For Specialized-Metabolite Prediction In Medicinal Plants: A Validation-Focused Review Of Five Ethnomedicinal Species

Abstract

Machine learning (ML) is widely used to quantify, annotate and prioritize plant specialized metabolites, but reported performance is rarely comparable across studies because validation designs differ. We critically reviewed ML applications published between 2021 and 2026, using five ethnomedicinal species with contrasting chemistry as case studies: Curcuma longa, Withania somnifera, Zingiber officinale, Piper nigrum and Mentha × piperita. Each study was coded by task, data type, algorithm, validation design and reported metric, and graded on a five-tier validation scale. Species-level evidence was concentrated in spectroscopic quantification, authentication and extraction optimization, almost always with internal validation only; no ML-based biosynthetic-gene or pathway prediction was found for any of the five species. The most transferable methods, namely multi-omics gene predictors, pathway classifiers and mass-spectrometry class annotators, were developed on model or reference taxa. Their scores declined under stricter evaluation: gene-prediction AUC-ROC fell from 0.924 within species to 0.584 in an external taxon. We propose minimum validation and reporting practices and research priorities for transferring these methods to medicinal species.