Epsilon Archive for Student Projects

Abstract

This literature study investigates the challenges and opportunities of using machine learning for biomarker discovery in metabolomics, with emphasis on how data characteristics influence model performance. In general, simple machine learning methods are well suited for small datasets and offer higher interpretability, but they may fail to capture complex biological relationships between metabolites, while more advanced methods can model such complexity but require larger datasets, careful tuning and often provide limited interpretability. Importantly, good statistical performance does not necessarily reflect true biological relevance, as evaluation metrics may indicate strong predictive ability even when underlying biological relationships are not properly captured, particularly in high-dimensional and imbalanced datasets. For this reason, model results must be interpreted with caution and supported by robust validation strategies, although commonly used validation approaches can still produce overly optimistic results if not correctly applied and more rigorous methods are often constrained by data availability and computational demands. At the same time, metabolomics data itself presents major challenges due to high dimensionality, missing values and technical variation, which makes data preprocessing a critical step in ensuring reliable results. Overall, the findings suggest that in metabolomics-based biomarker discovery, data quality, preprocessing and study design have a greater impact on outcomes than the choice of machine learning model alone.