Path Complex Neural Network (PCNN)

This repository contains the official implementation of our paper
Path Complex Neural Network for Molecular Property Prediction,
presented at the ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling.

Path-Complex-Neural-Network

As a generalization of simplicial complexes, path complexes provide a powerful framework for modeling the connections and interactions within a set, such as the atoms in a molecule. By integrating with molecular mechanics models, path complexes can enhance the accuracy of predictions related to molecular properties.

PCNN (Path Complex Neural Network) is a model that utilizes path complexes to represent and predict molecular data. Our findings highlight the potential of path-based approaches in the molecular sciences, providing new insights into the relationships between molecular structure and function.

The Path Complex Message-passing Architecture:

The Path Complex Neural Network Module:

Table of Contents

1. Environment Requirements
2. Installation Steps
3. Data Download and Configuration
4. Running the Project
5. Experimental Results

Environment Requirements

This project requires:

• Python Version: 3.11
  • Use Python 3.11, as this is the version used for development and testing of the code.
• CUDA Version: 11.7
  • To fully utilize GPU acceleration, ensure that your environment supports CUDA 11.7.

Installation Steps

1. Create a Virtual Environment

Recommended to use conda:

conda create -n myenv python=3.11
conda activate myenv

2. Install Dependencies

Install the necessary Python libraries from requirements.txt:

pip install -r requirements.txt

3. Verify CUDA Installation

Check that CUDA 11.7 is correctly installed on your system:

nvcc --version

Data Download and Configuration

Download Data

Download the datasets from MoleculeNet. Place the datasets into the data directory in your project folder.

Configure Dataset Usage

To use different datasets, modify the path.yaml file in the config directory:

select_dataset: "qm7"  # Replace "qm8" with "qm7" or "qm9" as needed

Running the Project

Execute the project with the configured dataset by running:

python main_c.py

or

python main_r.py

Experimental Results

Comparison with Non-Pretrained Models

The following table presents the comparison of PCNN with various GNN architectures. The best performance values are highlighted in bold, and standard deviation values are indicated in subscripts.

Method	QM7	QM9	Tox21	HIV	MUV
GNN
GIN	110.3 (7.2)	0.00886 (0.00005)	*	*	*
GAT	103.0 (4.4)	0.01117 (0.00018)	*	*	*
GCN	100.0 (3.8)	0.00923 (0.00019)	*	*	*
D-MPNN	103.5 (8.6)	0.00812 (0.00009)	0.759 (0.007)	0.771 (0.005)	0.786 (0.014)
Attentive FP	72.0 (2.7)	0.00812 (0.00001)	0.761 (0.005)	0.757 (0.014)	0.766 (0.015)
GTransformer	161.3 (7.1)	0.00923 (0.00019)	*	*	*
SGCN	131.3 (11.6)	0.01459 (0.00055)	*	*	*
DimNet	95.6 (4.1)	0.01031 (0.00076)	*	*	*
HMGNN	101.6 (3.2)	0.01239 (0.00001)	*	*	*
Mol-GDL	62.2 (0.4)	0.00952 (0.00013)	0.791 (0.005)	0.808 (0.007)	0.675 (0.014)
Pretrain_GNN
N-Gram_RF	92.8 (4.0)	0.01037 (0.00016)	0.758 (0.009)	0.787 (0.004)	0.748 (0.002)
N-Gram_XGB	81.9 (1.9)	0.00964 (0.00031)	0.758 (0.009)	0.787 (0.004)	0.748 (0.002)
PretrainGNN	113.2 (0.6)	0.00922 (0.00004)	0.781 (0.006)	0.799 (0.007)	0.813 (0.021)
GROVER_base	94.5 (3.8)	0.00986 (0.00055)	0.743 (0.001)	0.625 (0.009)	0.673 (0.018)
GROVER_large	92.0 (0.9)	0.00986 (0.00025)	0.735 (0.001)	0.682 (0.011)	0.673 (0.018)
MolCLR	*	*	*	0.750 (0.002)	0.796 (0.019)
GEM	58.9 (0.8)	0.00746 (0.00001)	0.781 (0.001)	0.806 (0.009)	0.817 (0.005)
DMP	74.4 (1.2)	*	0.791 (0.004)	0.814 (0.004)	*
SMPT	*	*	0.797 (0.001)	0.812 (0.001)	0.822 (0.008)
PCNN	53.6 (2.1)	0.00683 (0.00005)	0.801 (0.002)	0.823 (0.004)	0.827 (0.015)

Citation

If you find this code useful, please cite:

@inproceedings{
li2024path,
title={Path Complex Neural Network for Molecular Property Prediction},
author={Longlong Li and Xiang LIU and Guanghui Wang and Yu Guang Wang and KELIN XIA},
booktitle={ICML 2024 Workshop on Geometry-grounded Representation Learning and Generative Modeling},
year={2024},
url={https://openreview.net/forum?id=FlnGcMp6FL}
}