This repository contains a python implementation (tested on version 3.7.6) for the DeepGenome model described in:
Dudek, N.K. & Precup, D. Towards AI-designed genomes using machine learning. In prep for submission (2023).
Here, we introduce a framework for training a machine learning algorithm to learn the basic genetic principles underlying the gene composition of bacterial genomes. Given a set of desired pathways (e.g., glycolysis, TCA cycle, and butane degradation), our variational autoencoder (VAE) model strives to generate synthetic bacterial "genome vectors" - vectors denoting the full complement of genes that would need to be encoded to support a viable cell that supports the user-defined input functions.
pip install requirements.txt
The core module is main.ipynb. This juptyer notebook is used to load the data, process and filter it, perform data exploration, train a single VAE model or load a trained model (i.e. the best model resulting from hyperparameter optimization), and then analyze the results and generate the figures seen in the manuscript.
To perform hyperparameter tuning, use the hpo.py module. Note that it is recommend to do this on an HPC cluster due to significant compute requirements.
Data was sourced from the KEGG database. A substantial amount of processing was performed on the data prior to input into the model. You can access processed training data here and test data here.