Evaluation of a New Approach for Principal Component Analysis Application in Metabolomics Studies

Principal component analysis (PCA) has been widely used in metabolomics studies, and loading plots are frequently used for data interpretation to test the performance of an experiment design. PCA is an unsupervised machine-learning approach that describes high-throughput data and has advantages such as lower overfitting risk compared to supervised approaches. However, the PCA is not well applied in metabolomics studies, especially when partial separation is observed. While methods were developed for estimating variations among sample classes, there is no easy approach to determine if the PCA loadings could be used to identify important metabolites when the PCA score plot shows partial separation in metabolomics studies. In this study, we developed a novel approach that can help researchers assess if PCA showed partial separation and whether the PCA loadings can be used for metabolite identification. We applied the score projection values to the groups’ separation direction and used the t-test of the projected values to determine the score plot separation conditions. The new approach was demonstrated with two different metabolomics datasets under various PCA score plot separation conditions. Our study offers an easy and efficient way to evaluate partial separation in PCA score plots, enabling loadings to be more effectively applied in the discovery of important metabolites, providing a robust tool for advancing metabolomics research.