thermo_GNN

This is the repository containing the code associated with the paper "Graph-based deep learning models for thermodynamic property prediction: The interplay between target definition, data distribution, featurization, and model architecture". Code is provided "as-is". Minor edits may be required to tailor the scripts for different computational systems.

Table of Contents

• Features
• Requirements
• Installation
• Updates
• Quick Start
  • Folder Structure
  • Train
• Script
• Dataset
• Citation
• Acknowledgement

Features

• Atom-fingerprint:
• Mol-feature:
• Ringcount feature:
• MLP_Trigonometric:
• KAN_Trigonometric:

More information on the meaning of the individual features, we refer to the associated manuscript.

Requirements

To use CPUs, Suitable for x86 and ARM platforms. To use GPUs, you will need: * cuda >= 8.0 * cuDNN

Installation

To download the code git clone https://github.com/chimie-paristech-CTM/tree/main cd thermo_GNN To set up the thermoGNN conda environment: conda env create -f environment.yml To install the thermoGNN package, activate the thermoGNN environment and run the following command within the thermoGNN directory: conda activate thermo_GNN pip install -e .

Updates

• ☑ add Atom fingerprint.
• ☑ add Mol-feature.
• ☑ add Ringcount feature.
• ☑ add Mlp_Trigonometric.
• ☑ add KAN_Trigonometric.

Quick Start

Folder Structure

. ├── README.md ├── dataset/smalldataset │ ├── data/ │ │ ├── *data*.csv │ ... └── chemprop/ . ├── *data*.csv column |smiles | target_label| data: │smiles1| target_label1| │smiles2| target_label2| │smiles3| target_label3| ... └──

Train

To train a model, run:

python train.py --data_path <path> --dataset_type <type> --save_dir <dir> --epochs <epoch> --input_features_type <input_features_type> --aggregation <norm> --output_fingerprint <output_fingerprint> --model <model> where: 1. <path> is the csv file path not the dir path. 2. <aggregation> containing [sum, mean, norm] controlling the output type of outputhead. 3. `<inputfeaturestype>` containing [chemprop, jpca, moleculelevelfeature] controling the type of input feature. 4. `<outputfingerprint>containing [atom, mol] controlling the type of output fingerprint. 5.` containing [dpmnn, kantrigonometric, mlptrigonometric] controlling the type of output fingerprint.

For example: python train.py --data_path ./dataset/singledata/lipo_train.csv --dataset_type regression --output_fingerprint atom --save_dir ./lipo/checkpoint --epochs 2 --input_features_type molecule_level_feature --aggregation norm

A full list of available command-line arguments can be found in chemprop/args.py.

If installed from source, python train.py can be replaced with chemprop_train.

Notes: * The default metric for classification is AUC and the default metric for regression is RMSE. Other metrics may be specified with --metric <metric>. * --save_dir may be left out if you don't want to save model checkpoints. * --quiet can be added to reduce the amount of debugging information printed to the console. Both a quiet and verbose version of the logs are saved in the save_dir.

Script

the folder 'dataset_preparation' contains all scripts to process the original datasets.

Dataset

This link datasets contains qm9, paton, qmugs, pc9, and qmugs1.1 datasets.

Citation

If (parts of) this work are used as part of a publication, please cite the paper: @article{***, title={Graph-based deep learning models for thermodynamic property prediction: The interplay between target definition, data distribution, featurization, and model architecture}, author={Bowen Deng ,Thijs Stuyver}, journal={ChemRxiv}, year={2024} } Furthermore, since the work is based on chemprop, please also cite the paper in which this code was originally presented: @article{***, title={Chemprop: A Machine Learning Package for Chemical Property Prediction}, author={Esther Heid ,Kevin P. Greenman ,Yunsie Chung ,Shih-Cheng Li ,David E. Graff ,Florence H. Vermeire ,Haoyang Wu ,William H. Green ,Charles J. McGill}, journal={ChemRxiv}, year={2023} }

Acknowledgement

• PyTorch implementation of chemprop: https://github.com/chemprop/chemprop