Optimizing Feature Selection for Unbalanced Metabolomics Data with A Background Factor: A Comparative Study in Parkinson’s Disease

Exploring associations between molecular features and categorical factors, like disease status, is a crucial task in metabolomics data analysis, demanding exceptional rigor. However, data imbalance and crossed factorial design complicate the selection of appropriate analysis settings. This paper presents a comparative study of candidate analysis settings for unbalanced metabolomics data, focusing on Parkinson’s disease with gender as a background factor. The study evaluates two statistical test methods, pairwise t-tests and 2-way ANOVA, combined with two multiple hypothesis correction strategies: global correction and two-stage correction. Additionally, an unconditional analysis setting is examined, which can be easily misapplied due to its disregard for data imbalance and factor interactions. We profile the characteristics of each setting through experiments on real-world datasets from human samples and provide practical guidance for selecting appropriate settings to enhance the reliability of metabolomics studies.