Integrating Biosynthetic, Genomic and Ecological Open Data for Medicinal Plant Research: A Leakage-Aware Evidence-Prioritization Framework

Medicinal plant research increasingly combines heterogeneous public data, but data leakage and unsupported biological inference remain major risks. We developed a leakage-aware framework separating taxon–compound evidence ranking from environmental niche characterization. Six taxa and ten molecules or broad classes formed 60 taxon–compound pairs (27 supported and 33 below-threshold background); primary modeling used 36 specific-molecule pairs (8 supported and 28 unlabeled background) and five compound-matched pathway/chemical predictors. Under leave-one-taxon-out validation, the prespecified balanced random forest achieved a balanced accuracy of 0.621 (95% fold interval: 0.500–0.800; accuracy: 0.694; precision: 0.250; recall: 0.500; permutation: p = 0.154). Matched-pathway-only and molecular-weight-only benchmarks achieved 0.662 and 0.358, respectively, and a post hoc logistic comparator achieved 0.646. The cross-molecule balanced accuracy was 0.746 (fold interval: 0.516–0.975). Evidence scores correlated moderately with out-of-fold probabilities (Spearman: ρ = 0.39, p = 0.017). Environmental analyses used 248 SoilGrids and 324 NASA POWER taxon × exact-cell rows. The spatially restricted PERMANOVA was non-significant for soil (R2 = 0.257, p = 0.067) and climate (R2 = 0.233, p = 0.075), whereas grouped taxon classifiers achieved balanced accuracies of 0.479 and 0.547 (permutation: p = 0.005 for both). Environmental-only compound controls were non-significant. The outputs provide an auditable exploratory ranking workflow, but predictive validity for taxon–compound prioritization was not established.