Biomarker selection in metabolomics

Biomarker identification is of fundamental importance in many fields of biology and medicine. Modern analytical techniques provide a wealth of information, often leading to a holistic view of the state of biological systems. Out of this huge mass of data those variables that are indicative of class differences, the biomarkers, need to be identified in order to obtain either better predictive models, to obtain a better understanding of the complexity of living systems and to prioritize experimental follow-up work. Many statistical approaches focus on creating predictive models, and inspecting the model coefficients to see which variables could be termed putative biomarkers. In practice, however, the applicability of this approach is hindered by the unfavourable sample-to-variable ratio, making it virtually impossible to both build and validate predictive models. We will discuss two principled approaches to obtain good biomarker estimates anyway, not relying on accurate error estimations, and show using experimental spike-in data that also unpredictive models may be used to good effect in the quest for biomarkers.