Application of fuzzy c-means clustering in data analysis of metabolomics.

Fuzzy c-means (FCM) clustering is an unsupervised method derived from fuzzy logic that is suitable for solving multiclass and ambiguous clustering problems. In this study, FCM clustering is applied to cluster metabolomics data. FCM is performed directly on the data matrix to generate a membership matrix which represents the degree of association the samples have with each cluster. The method is parametrized with the number of clusters (C) and the fuzziness coefficient (m), which denotes the degree of fuzziness in the algorithm. Both have been optimized by combining FCM with partial least-squares (PLS) using the membership matrix as the Y matrix in the PLS model. The quality parameters R(2)Y and Q(2) of the PLS model have been used to monitor and optimize C and m. Data of metabolic profiles from three gene types of Escherichia coli were used to demonstrate the method above. Different multivariable analysis methods have been compared. Principal component analysis failed to model the metabolite data, while partial least-squares discriminant analysis yielded results with overfitting. On the basis of the optimized parameters, the FCM was able to reveal main phenotype changes and individual characters of three gene types of E. coli. Coupled with PLS, FCM provides a powerful research tool for metabolomics with improved visualization, accurate classification, and outlier estimation.