Extensive benchmark of machine learning methods for quantitative microbiome data
Year: 2021
Type: conference-paper
Abstract: 1. IntroductionCharacterization of microbial communities with omics technologies shed to light powerfulbiomarkers for diagnosis and prognosis in human health [1]. In particular, shotgun metagenomicsallows a highly precise microbiome profiling. Indeed, prediction of phenotypic features, such asclinical status or disease states can help to stratify patients which is the first step toward precisionmedicine. Many machine learning (ML) methods have been developed to tackle classification andregression problems yet statistical specificities of metagenomic data make difficult the learning task[2]. In the present work, we compare the commonly used ML methods on a quantitativemetagenomics dataset.2. MethodsWe developed a workflow in R to browse and compare ML methods for classification or regressionimplemented in the caret package [3]. A table where microbial features (species, functions,metabolites) are quantified across a set labelled samples (e.g: control/disease) is taken as input. Then,the selected models are trained and evaluated with repeated 10 fold cross validation. Each model istrained 100 times with different random splits. The Activeon Proactive workflow engine was used toefficiently distribute the computing load on multiple servers. 3. ResultsWe applied our workflow on a dataset where the fecal microbiota of patients with cardiovasculardiseases is compared to healthy controls using shotgun metagenomics. Each model was tested onregression or classification problems, expected to be easy or difficult to predict. We compared themodels with several indicators including predictive performance (F-score, Rð), stability acrossiterations and computational resources consumption. We also explored the impact of commonpreprocessing steps to remove non informative variables (near zero variance features, linear combo,etc.). We observed that a wide range of common methods show similar predictive performance (svm,pls, glm, rf, etc.) although some can be very slow (spls). Finally, the choice of the best model may beguided by other criteria like the interpretability that can give insights in the underlying biologicalhypotheses in order to provide insightful medical decisions.References1. Laura Judith Marcos-Zambrano et al. Applications of machine learning in human microbiome studies: areview on feature selection, biomarker identification, disease prediction and treatment. FURQWieUV iQPicURbiRlRg\, vol. 12, p. 313, 2021.2. Isabel Moreno-Indias et al. Statistical and machine learning techniques in human microbiome studies:contemporary challenges and solutions. FURQWieUV iQ MicURbiRlRg\, vol. 12, p. 277, 2021.3. Max Kuhn. Caret: classification and regression training. AVWURSh\VicV SRXUce CRde LibUaU\, ascl-1505,2015. https://topepo.github.io/caret/
Source: HAL
Language: English
Topic: Gut microbiota and health
Subfield: Molecular Biology
Domain: Life Sciences
Keywords: machine learning benchmark, shotgun metagenomics, microbial community classification, regression analysis, cardiovascular disease, fecal microbiota, model stability, feature preprocessing, precision medicine, biomarkers, R workflow, computational resource consumption, interpretability, model selection criteria, metabolites
Open Access status: green
OpenAlex ID: W7223422028
