This package is the accompanying code to our paper "Deep in the Bowel: Highly Interpretable Neural Encoder-Decoder Networks Predict Gut Metabolites from Gut Microbiome" - https://www.biorxiv.org/content/10.1101/686394v1.abstract

This code runs on python 3.6 or newer on Anaconda 3 environment.

Main dependencies:

1. pytorch 1.1.0
2. scikit-learn 0.20.1
3. graphviz 0.10.1
4. matplotlib 3.0.2

The results reported in the paper can be reproduced by running script main_cv_1dir.py For example, to reproduce the Nonneg-Sparse-NED performance reported in table 3, run:

python main_cv_1dir.py --model BiomeAESnip --sparse 0.06 --learning_rate 0.01 --batch_size 20 --latent_size 70 --activation "tanh_tanh" --data_type "clr" --nonneg_weight --normalize_input --draw_graph

For other experimental configuration change the options accordingly.

This script will also generate the computational graphs shown in the paper.