Mol-PECO

This is the workspace for odor chemical recognition by laplacian graph neural network. Mol-PECO (Molecular Representation by Positional Encoding of Coulomb Matrix) is the deep learning framework of extacting molecular representation for odor perception prediction. Mol-PECO's strucutre is as follows:

Text

• Overleaf document

Dataset

Data source

All comes from pyrfume-data (https://github.com/pyrfume/pyrfume-data). After data cleaning, there exists 8503 molecule-odor descriptors pairs, which is 1.5 folds higher than Google's (https://arxiv.org/abs/1910.10685) with ~5000 pairs. The folder of pyrfume_sois_canon mainly includes the preprocessed datasets with:

• Adjacent matrix
• Coulomb matrix
• Atom array
• Eigenvectors and eigenvalues of adjacent matrix
• Eigenvectors and eigenvalues of Coulomb matrix

Data split

The dataset was split by second order iterative stratification, same as Google's split method (https://arxiv.org/abs/1910.10685), to handle the imbalanced distribution of odor descriptors with train/val/test of 8/1/1.

Requirements

• python-3.7+
• deepchem
• matplotlib
• openpyxl
• pyTorch
• pandas
• rdkit
• scipy
• seaborn
• sklearn
• tensorboard
• tqdm
• imbalanced-learn

Quick Start

% python -m venv venv                    # create python virtual env
% source ./venv/bin/activate             # activate venv
% pip install -r requirements.txt        # install requirements (confirmed to work with CUDA 11.1 on ubuntu 18.04)

% python preprocess_atom2id.py           # calculate to adjacent matrix, Coulomb matrix, etc.
...
100%|█████████████████████████████████████| 12336/12336 [07:28<00:00, 27.49it/s]

% python train_pyrfume_lpe.py            # train and evaluate the LGNN model with pyrfume dataset.
...
min loss at 12: 0.62    max auc at 12: 0.58:   2%|▊                                    | 14/600 [21:44<14:48:30, 90.97s/it]]
...

Train the pure GNN models

Command line python train_pyrfume.py --out_dir ../GNN-Coulomb --gnn_matrix coulomb

Parameters:

• out_dir: the output folder during training; ${out_name}$, please assgin the output folder in ${out_name}$
• gnn_matrix: the matrix that we want to model; please choose from adjacent and coulomb

Train the conventional classifier with fingerprints

The scripts are located in folder baselines. Command line python run_fgs.py --fp mordreds --model knn

Parameters:

• fp: the fingerprints that we want to model; please choose from mordreds, bfps, and cfps
• model: the classifier that we want to train; please choose from knn, rf, gb, smote-knn, smote-rf, and smote-gb; knn refers to K- Nearest Neighbor, rf refers to Random Forest, gb refers to Gradient Boosting, and smote refers to Synthetic Minority Oversampling Technique.