[CODE] Low-dimensional representations of genome-scale metabolism

This library implements a variational autoencoder (VAE), specifically altered to interface and load flux vector sampled from the null space of GSMs. The COBRApy library is used to interface with these GSMs. Dependencies This project was origionally created using the python venv utility, for setup information see this page. Once venv is set up run: python -m pip install -r requirements.txt
Organization of the code The main files implementing the VAE and dataset can be found in the root folder. fluxDataset.py contains the implementation of the FluxDataset class which manages loading flux samples generated the fluxSampler.py. It cannot be run independently. fluxFile contains the implementation of FluxFile class which acts as a wrapper around the flux samples files and manager renaming columns etc. fluxModel/py contains the implementation of FluxModel class which acts as a wrapper around a GSM model (which are assumed to be stored in some ./gems/ folder.) fluxSampler.py is a python script which acts as a wrapper around the COBRApy implementation of OptGpSampler and ACHR. reproducibility.py contains some functionality related to reproducibility with random states. trainer.py is a python script that is used to train VAEs. vae.py constraints the implementation of the FluxVAE class implementing the main VAE functionality along side some extra functionality specific to loading samples from different GSM sources consistently. vaeTrainer.py contains the implementation of the VAETrainer class which trainer.py is running. ./tools/ contains some additional python scripts used to experiment with the trained VAEs. ari_score.py implements a script which automatically runs K-Means then compares its results some original clustering. extract_fluxes.py implements a script used to extract specific fluxes from a flux sample file. flux_processor.py implements a script allowing a specific flux dataset to be loaded, passed to different stages of some trained VAE and then additionally applies some transformation e.g. t-SNE or PCA. gmm_classifier.py allows data to be loaded then fits a GMM model to predict cell type, the classifier is then applied to some test data. kmeans_cluster.py allows fluxes to be loaded and possibly passed to some VAE stage, it then applies K-means clustering saving the result. loss_scoring.py implements a script allowing multiple losses.csv (in the format output by trainer.py) to be combined and averaged. nearest_centroid.py allows fluxes to be loaded to some VAE stage and then trains a nearest_centroid classifier to predict some original label, the classifier is then applied to some test data. prepare_test_train.py implements a script which automatically splits a collection of samples from fluxSampler.py into a test set and many train sets. ./misc/ contains common functionality used in ./tools/ and ./ (root). It is kept here to make code organization easier. ari.py contains functionality related to ari_score.py classifier.py contains functionality related to both classification scripts. constants.py contains constants used throughout the codebase. fluxDataset.py contains functionality related to then FluxDataset class. io.py contains some general I/O functionality used in different scripts. kmeans.py contains some functionality related to kmeans_cluster.py parsing.py contains many different argparse parsers used in many different scripts. vae.py contains functionality related to loading a FluxVAE and interacting with it. Setting up a dataset The scripts implemented here should work with any GSMs which can be loaded by COBRApy. To set upt a new dataset called new_dataset create a folder new_dataset with a subfolder gems. The gems subfolder should contain any GSM files. Passing new_dataset folder to fluxSampler.py will popular the main folder with flux samples. A FluxDataset can not be set up loading form this folder. It will crease .pkl files and some other .json files to cache some information in order to speed up interaction. Recommended: The ./data/ folder was used during the project to store all data not to be saved to the repo and is already included in the .gitignore Training a VAE Once a dataset has been setup trainer.py can be run. Using the -h tag will reveal all arguments. python trainer.py -h
The only required arguments are -d (dataset folder) and a final main_folder (where the VAE and a losses.csv file will be saved to) are required, but generally other arguments are required. As an example: python trainer.py -d ./data/samples/new_dataset/ -e 32 --lr 0.0001 ./data/models/test/
will train a new VAE for 32 epochs on a dataset stored at ./data/samples/new_dataset/ using a learning rate of 0.0001. The VAE will be saved along with its losses and some files logging which parameter were used to ./data/models/test

Cain, S., Merzbacher, C., & Oyarzún, D. (2024). Low-dimensional representations of genome-scale metabolism. Zenodo. https://doi.org/10.5281/zenodo.11471995