MB-GAN

Microbiome Simulation via Generative Adversarial Network

Introduction

MB-GAN is a deep learning simulation framework for simulating realistic microbiome data

Manuscript is avaialble at bioRxiv https://doi.org/10.1101/863977

Information provided by this repository

Data folder

The folder “data” contains:

• The real microbiome data used to generate MB-GAN samples analyzed in the manuscript

• The csv files of the MB-GAN samples

• The simulated data by alternative methods: Normal-To-Anything and metaSPARSim

• The processed data used to generate figures and tables in the manuscript

R codes

The folder “Rcode” includes the following R scripts:

• generate_NorTA.R: Simulate data by Normal-To-Anything (NorTA)

• generate_metaSPARSim.R: Simulate data by metaSPARSim

• process_MBGAN.R: Converts the csv files of MB-GAN samples to Rdata files

• calculate_unifrac_dist.R: Calculate UniFrac distance for real microbiome samples and samples generated by three simulation methods (MBGAN, NorTA, metaSPARSim)

• check_first_level_property.R: Compare sample sparsity, abundance, alpha diversity (by Shannon index), and beta diversity (by non-metric multidimensional scaling (nMDS) analysis)

• check_second_level_property.R: Compare taxa-taxa associations using correlation structure and proportionality

• mbgan_application_supp15.R: Compare differential abundance analysis results based on the real microbiome and MBGAN simulated samples

• MiRKAT_simulation.R and MiRKAT_result_summary.R: Implement MiRKAT-based MWAS using MBGAN-simulated samples

• generate_abundance_heatmap.R: Compare the abundances of the top 60 most abundant taxa across different datasets

The folder “Rcode_smalldataset” includes the similar R codes as above for a small microbiome dataset dicussed in the supplement.

Python codes

The root folder includes the following Python scripts:

• mbgan_train_demo.py: codes to train a MB-GAN network
• mbgan_inference_casectrl.py: codes to simulate new microbiome abundances using trained model parameters

The folder "code_check_convergence" includes Python scripts to validate the model convergence.

The folder "models" includes the trained model weights.

The folder "outputs" includes simulation outputs.

The folder "models_smaller_dataset" includes the following Python scripts (similar to above):

• inference_case_ctrl.py: codes to simulate new microbiome abundances for a smaller microbiome dataset using trained model parameters

Contact

Xiaowei Zhan xiaowei.zhan@utsouthwestern.edu

Quantitative Biomedical Research Center

Center for the Genetics of Host Defence

Department of Population and Data Sciences

UT Southwestern Medical Center

Dallas, TX 75390-8821