Target-Decoy MineR for determining the biological relevance of variables in noisy datasets.

Abstract

Machine learning algorithms excavate important variables from big data. However, deciding on the relevance of identified variables is challenging. The addition of artificial noise, 'decoy' variables, to raw data, 'target' variables, enables calculating a false-positive rate and a biological relevance probability for each variable rank. These scores allow the setting of a cut-off for informative variables, depending on the required sensitivity/specificity of a scientific question. We tested the function of the Target-Decoy MineR (TDM) using synthetic data with different degrees of perturbation. Following, we applied the TDM to experimental Omics (metabolomics, transcriptomics and proteomics) results. The TDM graphs indicate the degree of difference between sample groups. Further, the TDM reports the contribution of each variable to correct classification, i.e. its biological relevance. An implementation of the algorithm in R is freely available from https://bitbucket.org/cesaremov/targetdecoy_mining/. The Target-Decoy MineR is applicable to different types of quantitative data in tabular format. Supplementary data are available at Bioinformatics online.