MMPolymer: A Multimodal Multitask Pretraining Framework for Polymer Property Prediction

[Paper] [Slide] [Website]

MMPolymer is a multimodal multitask pretraining framework that incorporates both 1D sequential and 3D structural information into polymer property prediction

The overview of our proposed MMPolymer

Website

You can try MMPolymer online by clicking on this link

Dependency

• The code has been tested in the following environment

  Package	Version
  Python	3.8.13
  PyTorch	1.11.0
  CUDA	11.3.1
  RDKit	2022.9.5

• Please install via the yaml file

  conda env create -f env.yml
  conda activate MMPolymer

Data Preparation

The origin data have beed placed in the fold ./dataset/data, and please further process these data as follows

cd dataset
python pretrain_data_process.py
python finetune_data_process.py

Training

Please download the checkpoint and place it to the fold ./ckpt

bash train.sh

Inference

bash inference.sh

Application

After training, you can use following scripts for actual application

• Take psmiles (e.g., *CC(*)C) as input and predict all properties

  python get_prediction_results.py --input_data '*CC(*)C'

• Take a csv file as input and predict all properties

  python get_prediction_results.py --input_data $CSV_FILE_PATH 

• If you just want to predict a specific property (e.g., Eat)

  python get_prediction_results.py --input_data '*CC(*)C' --property Eat
  python get_prediction_results.py --input_data $CSV_FILE_PATH --property Eat

Citation

If this work can help you, please cite it

@inproceedings{wang2024mmpolymer,
  title={MMPolymer: A Multimodal Multitask Pretraining Framework for Polymer Property Prediction},
  author={Wang, Fanmeng and Guo, Wentao and Cheng, Minjie and Yuan, Shen and Xu, Hongteng and Gao, Zhifeng},
  booktitle={Proceedings of the 33rd ACM International Conference on Information and Knowledge Management},
  pages={2336--2346},
  year={2024}
}

Acknowledgment

This code is built upon Uni-Mol and Uni-Core. Thanks for their contribution.