Receptor2Odorant

Implementation of paper "Matching receptor to odorant with protein language and graph neural networks"

@inproceedings{
hladi{\v{s}}2023matching,
title={Matching receptor to odorant with protein language and graph neural networks},
author={Matej Hladi{\v{s}} and Maxence Lalis and Sebastien Fiorucci and J{\'e}r{\'e}mie Topin},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=q9VherQJd8_}
}

Data

Our aim is to keep the information about olfactory receptor experiments up-to-date, organized and easily accessible. We're going to keep all the data in one place at https://github.com/chemosim-lab/M2OR. We hope that it will be a useful resource for both biology practitioners as well as machine learning experts.

Currently, we are working on a website for quick analyses of the data that everyone could use as a base resource. In the meantime the up-to-date data are available as a csv in the M2OR repo.

To reporoduce the results in the paper, see Original data section below.

Usage

For easy usage we created a Singularity container and included definition file in _container/Receptor2Odorant_singularity.def. You can find instructions how to bild the container here. In further sections we assume the container was sucessfully built.

Start the container in an interactive session:

singularity exec --containall --nv -B <bind_dir>:/mnt,<HOME>/.cache/huggingface/transformers Receptor2Odorant_singularity.sif bash

Source conda and activate the environment:

source /opt/miniconda3/etc/profile.d/conda.sh
conda activate base

Run scripts with a config file:

(Assuming that Receptor2Odorant dir is binded as /mnt)

python /mnt/Receptor2Odorant/scripts/main.py --config /mnt/configs/config_train.yml

or

python /mnt/Receptor2Odorant/scripts/main_odor_embedding.py --config /mnt/configs/config_odor_embedding_train.yml

There are several types of config files available: config_precompute.yml,config_train.yml,config_eval.yml,config_predict.yml,config_predict_single.yml,config_odor_embedding_train.yml,config_odor_embedding_predict.yml

config_precompute:

This config file is giving instructions to precompute protein representation and saving it to .h5 file using pyTables library.

config_train:

The training config for the main OR-molecule model.

config_eval, config_predict, config_predict_single:

Evaluation and prediction configs. Difference between predict and predict_single is that the former expects a precomputed protein representation for a bulk of pairs and the second one is running protein representation model on the fly.

config_odor_embedding_train:

The config to train GNN-based odor prediction model that was used in section 6 Agreement with odor perception to create abstract odor families (i.e. clusters) for the molecules.

config_odor_embedding_predict:

Prediction config for odor prediction model which outputs molecule's embedding alongside its odor labels.

Original data

The link below leads to the original dataset used in the paper.

https://drive.google.com/drive/folders/1Sb6nPaSgeX66Wo5uG8NYRmcMfvGr76_K?usp=sharing

Download data file: Data_figures.zip and Dataset.zip unzip them and place in the giving folder (create the folders if necessary):

• Content of Data from Data_figures.zip to Receptor2Odorant/Figures/Data
• Content of Dataset from Dataset.zip to Receptor2Odorant/RawData/m2or
• Content of _seq_dist_matrices from _seq_dist_matrices.zip to Receptor2Odorant/RawData/m2or/_seq_dist_matrices

Origianl data used in experiments:

• Folders from mixDiscard_20220608-144844.zip in the Google drive are original data splits used in the experiments. To create new splits see code in Receptor2Odorant/datasets, in particular Receptor2Odorant/datasets/m2or.py for olfactory receptor data.