Abstract
Analytical drift is a major source of bias in mass spectrometry based metabolomics confounding interpretation and biomarker detection. So far, standard protocols for sample and data analysis have not been able to fully resolve this. We present a combined approach for minimizing the influence of analytical drift on multivariate comparisons of matched or dependent samples in mass spectrometry based metabolomics studies. The approach is building on a randomization procedure for sample run order, constrained to independent randomizations between and within dependent sample pairs (e.g. pre/post intervention). This is followed by a novel multivariate statistical analysis strategy allowing paired or dependent analyses of individual effects named OPLS-effect projections (OPLS-EP). We show, using simulated data that OPLS-EP gives improved interpretation over existing methods and that constrained randomization of sample run order in combination with an appropriate dependent statistical test increase the accuracy and sensitivity and decrease the false omission rate in biomarker detection. We verify these findings and prove the strength of the suggested approach in a clinical data set consisting of LC/MS data of blood plasma samples from patients before and after radical prostatectomy. Here OPLS-EP compared to traditional (independent) OPLS-discriminant analysis (OPLS-DA) on constrained randomized data gives a less complex model (3 versus 5 components) as well a higher predictive ability (Q2 = 0.80 versus Q2 = 0.55). We explain this by showing that paired statistical analysis detects 37 unique significant metabolites that were masked for the independent test due to bias, including analytical drift and inter-individual variation.