Hierarchical Information Leakage in Machine-Learning Classification of Marine Microbial Enzymes for Microplastic Research

Abstract

This record contains the curated dataset, executable analysis code, and complete computational outputs supporting the study “Auditing Hierarchical Information Leakage in Marine Microbial Enzyme Classification for Microplastic Research.” The study evaluates classification of the recorded top-level Enzyme Commission class using three prespecified feature sets. Model A includes optimum temperature, optimum pH, and EC levels 2–4. It serves as a positive leakage control because EC levels 2–4 are subordinate components of the EC hierarchy defining the target. Model B includes optimum temperature and optimum pH only. Model C includes optimum temperature, optimum pH, and genus derived from the normalized organism-name field as an EC-independent taxonomic predictor. The original study dataset contained 950 records. After excluding three records with missing required values and removing 144 exact normalized duplicate enzyme records, the analysis-ready dataset contained 803 observations representing 376 accessions. Model performance was evaluated using five repeats of five-fold stratified cross-validation grouped by accession, ensuring that records sharing an accession were assigned to the same validation fold. The deposited materials include the processed dataset, encoded-class mapping, dataset audit, complete fold-level metrics, class-specific sensitivities, out-of-fold predictions, paired model comparisons, confusion matrices, figures, software requirements, and executable analysis code. The response variable is the recorded top-level enzyme-function class. The reference to microplastic research describes the study context and intended area of application. The deposited data contain no direct measurements of microplastic biodegradation, polymer-specific catalytic activity, substrate loss, reaction products, mineralization, or toxicity. Accordingly, the models and deposited outputs should not be interpreted as predicting or demonstrating any of these outcomes.