ProCeSa

This is the source code of "ProCeSa: Contrast-Enhanced Structure-Aware Network for Thermostability Prediction with Protein Language Models" (https://doi.org/10.1021/acs.jcim.4c01752).

conda environment

/procesa/environment.yml

esm model

Download pretrained esm1b, esm1v, esm2 model from original esm github. Put these models in /dwnl_ckpts.

data generation

Use scripts in /procesa/FLIP/baselines/scripts/ to generate dgl graph pkl files. Generated data will be saved in /datasets.

(Due to file size limit, the hotprotein-S dataset is in https://drive.google.com/file/d/1VvAXKw01hMrKBMsOMDB5OKcQlvSzN0TN/view?usp=drive_link.)

Train and evaluate

Run scripts in /procesa/scripts to train and evaluate models. Results will be saved in /procesa/results/. The correspondence between results shown in paper and running scripts are shown in figure below.

For hotprotein-s2c2 and hotprotein-s2c5, you can change DATANAME and EXPNAME to run other experiments, like s2c2_1 and model31. For hotprotein-S, you can change EXPNAME to run other experiments, like model116.

Results

Results are saved in /procesa/results folder.

Citations

If you make use of this code or the ProCeSa algorithm in your work, please cite the following paper:

@article{zhou2025procesa,
  title={ProCeSa: Contrast-Enhanced Structure-Aware Network for Thermostability Prediction with Protein Language Models},
  author={Zhou,  Feixiang and Zhang,  Shuo and Zhang,  Huifeng and Liu,  Jian K.},
  journal={Journal of Chemical Information and Modeling},
  year={2025},
  publisher={American Chemical Society (ACS)}
}