This is the official implementation of the paper: Protein Design with Dynamic Protein Vocabulary.

We refactored the codebase using the DVAGen framework (🎉recently accepted to EMNLP 2025 Demo).

Updates

• [2025/09/19] 🎉🎉ProDVa was accepted to NeurIPS 2025 Spotlight!

Quick Start

Setup

Download the repository and install ProDVa:

git clone https://github.com/sornkL/ProDVa.git
cd DVAGen
pip install -e .

Download the model weights from 🤗HuggingFace ProDVa Collection. All datasets and model weights are provided in the Collection, with details listed in the tables below.

Datasets used in our paper (see Appendix C.2 for more details):

Datasets	Task Descriptions	Links
CAMEO	Protein Design from Function Keywords (Section 4.2)	Download
Molinst-SwissProtCLAP	Protein Design from Textual Descriptions (Section 4.3)	Download

Each dataset includes training/validation/test sets, a fragment mapping file, and a FAISS index directory constructed from the training set.

Model weights used in our paper:

Models	Descriptions	Links
ProDVa-CAMEO	ProDVa trained on CAMEO (Section 4.2)	Download
ProDVa-Molinst-SwissProtCLAP	ProDVa trained on Molinst-SwissProtCLAP (Section 4.3)	Download

Designing Proteins with ProDVa

Simply use the following command to launch a CLI chat tool for designing proteins.

dvagen chat --config_path examples/chat.yaml

Evaluation

The evaluation pipeline typically involves two stages: protein design and assessment using certain metrics. Run the following command to launch the pipeline:

dvagen eval --config_path examples/eval.yaml

This command will prompt ProDVa to design proteins using the given test set. Note that for evaluation (or benchmarking the results), please refer to our latest work: PDFBench. We are working on integrating the evaluation metrics into our codebase.

Training

By default, we use deepspeed to launch the training script. To train ProDVa, use the following command:

dvagen train [deepspeed_args] --config_path examples/train.yaml

Details of the configuration files are available in the examples/README.md file. We also include some file (e.g., training set, mapping file) examples in the README. Most of the arguments are directly inherited from DVAGen while we ignore the unused ones. Some newly added arguments specific to ProDVa are also documented.

Citation

@article{liu2025protein,
  title={Protein Design with Dynamic Protein Vocabulary},
  author={Liu, Nuowei and Kuang, Jiahao and Liu, Yanting and Sun, Changzhi and Ji, Tao and Wu, Yuanbin and Lan, Man},
  journal={arXiv preprint arXiv:2505.18966},
  year={2025}
}

If you find the codebase helpful, please also cite the DVAGen framework:

@inproceedings{du2025dvagen,
  title={DVAGEN: Dynamic Vocabulary Augmented Generation},
  author={Du, Wei and Liu, Nuowei and Wang, Jie and Kuang, Jiahao and Ji, Tao and Wang, Xiaoling and Wu, Yuanbin},
  booktitle={Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
  pages={365--372},
  year={2025}
}