mGPS

A microbiome fine-scale biogeography tool maps the transfer of antimicrobial resistance genes

Yali Zhang, Leo McCarthy, S. Emil Ruff, and Eran Elhaik

This repository contains all code used for data cleaning, modelling and results of the Microbiome Geographic Population Structure (mGPS). mGPS is a biogeographical tool that employs multi-output model for hierarchical location predictions (i.e., continent/country/city) and location coordinates (latitude + longitude) using bacterial abundance data as input.
For the mGPS interface - the freely available public tool with a friendly user interface developed with the shiny package in R (version 4.0.3), please visit: mGPS_interface

Data

All data and geographical cooridnates needed to reproduce the analyses are in the Data.zip.Unzip the file before running the analyses. Description of folders:

• Metasub
  • complete_metadata.csv - MetaSUB environmental and geo data
  • metasub_taxa_abundance.csv - MetaSUB taxa abundance data
  • MetaSUB City Metadata - Sheet1 new.csv - MetaSUB city information data
  • After_process_megares_amr_class_rpkmg.csv - MetaSUB AMR data
• Soil
  • Dataset_01_22_2018_enviro.csv - Soil origin metadata
  • Dataset_01_22_2018_taxa.csv - Soil OTU taxa read data
• Marine
  • marine_taxa.csv - Marine taxa and metadata
• Geo: Geodata required for produce the analysis and figures

Usage

The interface of mGPS and its detailed description can be found inthe folder mGPS_interface。

The script mGPS.r, has two parts:

• species_select() - function using random forest for selecting optimal Geographically Informative Taxa (GIT) that used to built the mGPS prediction model:
  • x -- prediction variables
  • y -- target variable
  • remove_correlated -- should correlated predictor variables be removed (>98% correlation).
  • subsets -- The variable subset sizes to try
  • cores -- number of cores to utilize

Example

species_select(x = metasub_data[, taxa],
               y = metasub_data$city,
               remove_correlated = F,
               c(50,100,200,300,500),
               cores = 8)

• mGPS() - a machine-learning-based function that utilizes microbial relative sequence abundances to yield a fine-scale source site for microorganisms
  • training -- Taxa data of samples used to train the model.
  • testing -- taxa abundance data of samples for which predictions are to be generated
  • classTarget -- granularity for geographic class prediction either country,city or transit station etc.
  • variables -- a vector containing names of species or taxa to be used as variables for prediction. This needs defining even if all taxa/species in the training data frame are to be used, so that geographic information are not mistakenly used as predictors.
  • hierarchy -- The geographic hierarchy for predictions i.e. continent –> city -> latitude -> longitude
  • nthread -- number of threads to utilize
  • coast -- (optional) data.frame of co-ordinates for predictions to be bound by

If no test set is given then a trained model is returned that takes a test set as the input.

For this implementation of mGPS, hyperparameter tuning is carried out at every level of the chained model using a small grid search applied to the training set provided, the same validation fold splits are used at every level. Predictions are then generated using the provided test data set .

Example

mGPS(training = train, 
     testing = test, 
     classTarget = "city",
     variables = optVars,
     nthread = 8,
     hierarchy = c('continent','city','latitude','longitude'), 
     coast=coastlines)

Results and figures

Each dataset used in the analysis has its own folder. Each dataset has a {dataset}_make.R file for finding Geographically Informative Taxa and generating predictions using mGPS and cross validation the results are saved to the corresponding outputs folder. There is also a {dataset}_plots.Rmd file for each dataset in the analysis which when knit produces plots and tables found in the manuscript.

Each make and plots file for the corresponding datasets should be run in turn to reproduce the full analysis. After cloning the repo the plots files can be run without first running the corresponding make file as the predictions are stored in this repo.

AMR analysis part is described in the Metasub/Scripts/AMR part.

Dependencies

The required packages for mGPS can be found in packages.r. The packages used for mGPS interface an be found in mGPS_interface/packages.r