Using expert driven machine learning to enhance dynamic metabolomics data analysis.

Abstract

Data analysis for metabolomics is undergoing rapid progress thanks to the proliferation of novel tools and the standardization of existing workflows. However, as datasets and experiments continue to increase in size and complexity, standardized workflows are often not sufficient. In addition, as the ground truth for metabolomics experiments is intrinsically unknown, there is no way to critically evaluate the performance of tools. Here, we investigate the problem of dynamic multi-class metabolomics experiments using a simulated dataset with a known ground truth and evaluate the performance of tinderesting, a new and intuitive tool based on gathering expert knowledge to be used in machine learning, and compare it to EDGE, a statistical method for sequence data. This paper presents three novel outcomes. First we present a way to simulate dynamic metabolomics data with a known ground truth based on ordinary differential equations. This method is made available through the MetaboLouise R package. Second, we show that the EDGE tool, originally developed for genomics data analysis, is highly performant in analyzing dynamic case vs control metabolomics data. Last, we introduce the tinderesting method to analyse more complex dynamic metabolomics experiments that performs on par with statistical methods. This tool consists of a Shiny app for collecting expert knowledge, which in turn is used to train a machine learning model to emulate the decision process of the expert. This approach does not replace traditional data analysis workflows for metabolomics, but can provide additional information, improved performance or easier interpretation of results. The advantage is that the tool is agnostic to the complexity of the experiment, and thus is easier to use in advanced setups. All code for the presented analysis, MetaboLouise and tinderesting are freely available.