MechVal v1.0: a confidence-graded reference set of 31 gut-microbial drug biotransformations for scoring model explanations

MechVal is a confidence-graded reference set of 31 relationships between a named drug and a gut-microbial enzyme, assembled so that the explanations produced by machine-learning models of microbiome-mediated drug metabolism can be scored rather than only interpreted. Each row records the substrate, the reaction, the enzyme or gene, the organism, the direction of effect (inactivation, activation, toxification or altered exposure), the toxicological relevance, a confidence grade, and the source with its DOI. Eight rows are grade A, meaning the enzyme was identified by gene knockout, heterologous expression, purified-enzyme assay or causal inhibitor rescue in a cited primary report; these eight form the strict subset for primary reporting. The set is the ground truth behind the mechanistic precision@k protocol described in the accompanying review article. Grading criteria, limitations and the rows deliberately excluded from the set are documented in the README. Grades were assigned by one author; no inter-annotator agreement is reported, and regrading is invited.