Analytical challenges of untargeted GC-MS-based metabolomics and the critical issues in selecting the data processing strategy

Background: A challenge of metabolomics is data processing the enormous amount of information generated by sophisticated analytical techniques. The raw data of an untargeted metabolomic experiment are composited with unwanted biological and technical variations that confound the biological variations of interest. The art of data normalisation to offset these variations and/or eliminate experimental or biological biases has made significant progress recently. However, published comparative studies are often biased or have omissions. Methods: We investigated the issues with our own data set, using five different representative methods of internal standard-based, model-based, and pooled quality control-based approaches, and examined the performance of these methods against each other in an epidemiological study of gestational diabetes using plasma. Results: Our results demonstrated that the quality control-based approaches gave the highest data precision in all methods tested, and would be the method of choice for controlled experimental conditions. But for our epidemiological study, the model-based approaches were able to classify the clinical groups more effectively than the quality control-based approaches because of their ability to minimise not only technical variations, but also biological biases from the raw data. Conclusions: We suggest that metabolomic researchers should optimise and justify the method they have chosen for their experimental condition in order to obtain an optimal biological outcome.