On the interpretation of high throughput MS based metabolomics fingerprints with Random Forest

We discuss application of a machine learning method, Random Forest (RF), for the extraction of relevant biological knowledge from metabolomics fingerprinting experiments. The importance of RF margins and variable significance as well as prediction accuracy is discussed to provide insight into model generalisability and explanatory power. A method is described for detection of relevant features while conserving the redundant structure of the fingerprint data. The methodology is illustrated using two datasets from electrospray ionisation mass spectrometry from 27 Arabidopsis genotypes and a set of transgenic potato lines