Optimal Algorithm for Metabolomics Classification and Feature Selection varies by Dataset

Metabolomics, the systematic identification and quantification of all metabolites in a biological system, is increasingly applied towards identification of biomarkers for disease diagnosis, prognosis and risk prediction. Applications of metabolomics extend across the health spectrum including Alzheimer's, cancer, diabetes, and trauma. Despite the continued interest in metabolomics there are numerous techniques for analyzing metabolomics datasets with the intent to classify group membership (e.g. Control or Treated). These include Partial Least Squares Discriminant Analysis, Support Vector Machines, Random Forest, Regularized Generalized Linear Models, and Prediction Analysis for Microarrays. Each classification algorithm is dependent upon different assumptions and can potentially lead to alternate conclusions. This project seeks to conduct an in depth comparison of algorithm performance on both simulated and real datasets to determine which algorithms perform best given alternate dataset structures. Three simulated datasets were generated to validate algorithm performance and mimic 'real' metabolomics data: (Han et al., 2011) independent null dataset (no correlation, no discriminatory variables), (Davis, Schiller, Eurich, & Sawyer, 2012) correlated null (no discriminating variables), (Guan et al., 2009) correlated discriminatory. This comparison is also applied to 3 open-access datasets including two Nuclear Magnetic Resonance (NMR) and one Mass Spectrometry (MS) dataset. Performance was evaluated based on th