Machine Learning algorithms reveals country-specific metagenomic taxa from American Gut Project data
This repository includes code based on article published in MIE congress to be held on 29th-31st May, 2021.
Liñares-Blanco J, Fernandez-Lozano C, Seoane JA, Lopez-Campos G. Machine Learning Algorithms Reveals Country-Specific Metagenomic Taxa from American Gut Project Data. Studies in Health Technology and Informatics. 2021 May;281:382-386. DOI: 10.3233/shti210185.
DOI: 10.3233/shti210185
The data used in this work was downloaded from the AGP (ftp://ftp.microbio.me/AmericanGut/). Three files were downloaded: otu table, otu tree and clinical data. Phyloseq package was used to manage this data and filtered by feces samples.
In recent years, microbiota has become an increasingly relevant factor for the understanding and potential treatment of diseases. A major problem has been the variability that exists between different countries of origin when designing a global treatment. For this reason, in this work, based on the data reported by the largest study of microbioma in the world, a classification model has been developed based on Machine Learning (ML) capable of predicting the country of origin (United Kingdom vs United States) according to metagenomic data. The data were used for the training of a glmnet algorithm and a Random Forest algorithm. Both algorithms obtained similar results (0.698 and 0.672 in AUC, respectively). Furthermore, thanks to the application of a multivariate feature selection algorithm, eleven metagenomic genres highly correlated with the country of origin were obtained. An in-depth study of the variables used in each model is shown in the present work.
There are two main files in this repository: preprocessing analyis and machine learning code. In addition, utils functions are included in a utils.r file. To run this code, make sure you have the following:
The packages we´ve used:
install.packages(c('mlr', 'h2o', 'FCBF', 'parallelMap', 'BiocManager'))
BiocManager::install("phyloseq")Jose Liñares Blanco1 Carlos Fernandez Lozano1 Jose A. Seoane2 Guillermo Lopez Campos3
1 Department of Computer Science and Information Technologies, Faculty of Computer Science, University of A Coruña, CITIC, Campus Elviña s/n, A Coruña, 15071, Spain
2 Stanford Cancer Institute, Stanford University School of Medicine, Stanford CA, USA
3 Wellcome-Wolfson Institute for Experimental Medicine, Queen´s University Belfast, Belfast, UK
If you have any questions, please feel fre to contact ([email protected])