Use of Non-Negative Matrix Factorization to Understand Exercise Effects on Metabolites

• Activity Number: 377
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Douglas A. Marsteller*+ and S. Stanley Young and K. Eric Milgram and John V. St. Peter and Mark A. Pirner
• Companies: PepsiCo and National Institute of Statistical Sciences and PepsiCo and PepsiCo and PepsiCo
• Keywords: matrix factorization ; metabolomics ; high-dimensional data

Abstract

Non-negative matrix factorization (NMF) was used in parallel to supervised methods to uncover signals and patterns in a metabolomic dataset where exercise was the main effect. A cross-over design of 27 adult males age 20-30 was executed with plasma sampled prior to and following moderate exercise on two occasions resulting in 213 observations. The plasma samples were separated using gas and liquid chromatography followed by mass spectrometry. There were 638 identified plasma metabolites available for statistical analysis of which 347 were named chemicals. Treatment of missing data was studied and the data was blocked by occasion to create a holdout set. Where appropriate, any differences between the software suites: Orange, SAS BASE 9.3, SAS JMPpro 10, and R (v 2.15.2), were outlined with a focus on unsupervised matrix factorization such as NMF, rSVD/PCA, and other high-D prediction techniques. For supervised analysis; Random Forest, a mixed effect model, PLS-DA, and others were employed. Overall, the potential utility of a complement of algorithms and their contribution to information in the understanding of biological networks influence in human exercise is discussed.