# PremuNet

**Repository Path**: Tomhappy/PremuNet

## Basic Information

- **Project Name**: PremuNet
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-04-08
- **Last Updated**: 2026-04-08

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# A Pretrained Multi-Representation Network for Molecular Property Prediction
### [Paper](https://www.sciencedirect.com/science/article/pii/S1566253523004086) | [Github](https://github.com/A-Gentle-Cat/PremuNet) 

## OverView

PremuNet can extract information from multiple views and combine them interactively through pre-training and fine-tuning. The network consists of two branches: a Transformer-GNN branch that extracts SMILES and graph information, and a Fusion Net branch that extracts graph and conformation information, called PermuNet-L and PermuNet-H respectively. We employ masked self-supervised methods to enable the model to learn information fusion and achieve improved performance in downstream tasks.
<div align="center">
<img width="2500" src="DrawGraph/totalmodel.png" alt="totalmodel" style="zoom:14%;" />
</div>

## Dataset
### Dataset for pretraining.
Our pre-training process employs both the ChEMBL and PCQM4M datasets. The ChEMBL dataset can be accessed through [this link](https://chembl.gitbook.io/chembl-interface-documentation/downloads), 
while the PCQM4M dataset is available for download [here](https://dgl-data.s3-accelerate.amazonaws.com/dataset/OGB-LSC/pcqm4m-v2.zip).
### Dataset for finetune
We use MoleculeNet as our benchmark test. The datasets and splits of MoleculeNet in our experiment are available at [this link](http://snap.stanford.edu/ogb/data/graphproppred/csv_mol_download/). Download and unzip it into your dataset directory.

We have added 3D coordinates to the molecules in MoleculeNet to further improve prediction performance. The source for these 3D coordinates can be accessed via [here](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/JNGTDF). 
You should download the `molecule_net.tar.gz` file. Unzip it and place it in your dataset directory
## Setup Environment
```shell
pip install -r requirements.txt
```
## Pretrain
The pretraining weights can be found at [this link](https://drive.google.com/drive/folders/16ANfFYW4uKYrBzn56v94hkAGu4yGg7tM?usp=share_link).
Please download them to the project root directory and unzip them.
If you want to train your own pre-training weights, you can refer to the following steps.
### SMILES-Transformer
#### For Atom-Level Pretrain
**Step 1: Download PCQM4M Dataset**

**Step 2:**
```shell
cd Transformer1D
python pretrain_trfm.py --data /pcqm4m-v2/raw/data.csv.gz
```
Note that you should replace parameter with your own data path.
#### For Graph-Level Pretrain
The pre-training of Graph-Level Transformer can follow the introduction in [this link](https://github.com/DSPsleeporg/smiles-transformer).
### PremuNet-H
PremuNet-H module, we used the code from [this link](https://github.com/teslacool/UnifiedMolPretrain).
You can consider using their code and steps for pre-training, or use our code with the following steps:
```shell
cd UnifiedMolPretrain
python train.py --num-layers 12 --batch-size 128 --enable-tb --node-attn --use-bn --pred-pos-residual --mask-prob 0.25
```
## Finetuning
The configuration files for each dataset are located in the /configs directory. These parameters can be utilized directly as provided, or they can be modified as needed. The step for finetuning on MoleculeNet is as follows
### Step 1: Download Dataset
You should unzip the downloaded dataset and place them separately under the `/dataset/` directory.
### Step 2: Get Molecular Feature
```shell
python pretrans_data.py --dataset_dir ${your dataset directory}
```
### Step 3: Start finetuning
```shell
bash evaluate_all.sh
```

## Cite
Zhang H, Wu J, Liu S, et al. A pre-trained multi-representation fusion network for molecular property prediction[J]. Information Fusion, 2024, 103: 102092.
## Reference
This project utilizes the code from the [UnifiedMolPretrain](https://github.com/teslacool/UnifiedMolPretrain) and [Smiles-Transformer](https://github.com/DSPsleeporg/smiles-transformer) projects. The successful implementation of this project is attributed to the outstanding research papers and well-structured code provided by these initiatives.