We demonstrate how different normalization techniques in GC-MS analysis impart unique properties to the data, influencing any biological inference. Using simulations, and empirical data, we compare the most commonly used techniques (Total Sum Normalization 'TSN'; Median Normalization 'MN'; Probabilistic Quotient Normalization 'PQN'; Internal Standard Normalization 'ISN'; External Standard Normalization 'ESN'; and a compositional data approach 'CODA'). When differences between biological classes are pronounced, ESN and ISN provides good results, but are less reliable for more subtly differentiated groups. MN, TSN, and CODA approaches produced variable results dependent on the structure of the data, and are prone to false positive biomarker identification. In contrast, PQN exhibits the lowest false positive rate, though with occasionally poor model performance. Because ESN requires extensive pre-planning, and offers only mixed reliability, and ISN, TSN, MN, and CODA approaches are prone to introducing artefactual differences, we recommend the use of PQN in GC-MS research.
Keywords: GC-MS; biomarker identification; log-ratio transformations; olfactory communication; pheromones; pre-processing; size effects.
© 2018 WILEY Periodicals, Inc.