🌍 Pre-requisite:

• Python3, numpy, pandas, pytorch(2.6.0+cu118),peft(0.14.0),bitsandbytes
• esm3(https://github.com/evolutionaryscale/esm)
• evo2(https://github.com/arcinstitute/evo2)

🏛️ Installation:

• 1.Download the source code in this repository.
• 2.Download the weights of EiRA at https://huggingface.co/zengwenwu/EiRA, and make sure they locate in the same folder.
• 3.Unzip all .zip packages.

CheckPoint Name	Description	Size
DNAbinder_lora_ft32_DNAlen50_cross_att_DNAtransformer_DNAbinder.pth	DNA transformer and Cross attention weights of DNA-informed EiRA	2.23G
EiRA_checkpoint_DNAbinder_lora_ft32_DNAlen50.zip	LoRA weighht of DNA-informed EiRA	23.9M
EiRA_checkpoint_vanilla_lora_ft32_repeat_penalty.zip	LoRA wight of EiRA without DPO	10.1M
DPO_checkpoint_VanillaLora_part_data_no_repeat.zip	LoRA wight of EiRAD with DPO	12.1M

🔍 Protein Generation like ESM3

EiRA inherits the flexible multi-modal editing feature of ESM3 and supports the free combination of multiple tracks of prompts.

$ python run_EiRA.py \
          --weight_dir "The local path of the downloaded weight file"
          --SRC_PDB_path "The path of your template"
          --designed_seq_save_path "Result path"
          --design_num "Number of designed sequences"
          --inform_position "The constant residue indices in the template, like: 0,1,2,3,5,6,7,8,9"
          --device cuda:0
          --chain Template chain (like "A")

🧬 DNA-conditioned DBP Generation

Protein editing under DNA conditions (EVO2 embedding).

$ cd DBPdesign
$ python run_EiRA_withDNA.py \
          --weight_dir "The local path of the downloaded weight file"
          --SRC_PDB_path "The path of your template"
          --designed_seq_save_path "Result path"
          --design_num "Number of designed sequences"
          --inform_position "The constant residue indices in the template, like: 0,1,2,3,5,6,7,8,9"
          --device cuda:0
          --chain Template chain (like "A")
          --DNA "Target DNA sequence, like AGCTCGC"

⚔️ Domain adaptive post-training

• 1.Before starting the training, you should first read the Tokenization Tutorial and prepare the tokens.
• 2.Run the following command depend on the number of GPUs available to you.

$ torchrun --nproc_per_node=8 --master_port=29512 pretrain.py \
          --gpu 0,1,2,3,4,5,6,7
          --token_dir "The path contains all token in .npz format"
          --save_path "Checkpoint path to be saved"
          --batch_size 20
          --fine_tuning_num 16
          --epochs 5
          --prefetch_factor 30
          --num_workers 16

🛠️ Binding site-informed preference optimization

• 1.Before starting the DPO, you should first read the DPO Tokenization Tutorial and prepare the preference pair tokens.
• 2.Run the following command depend on the number of GPUs available to you.

$ torchrun --nproc_per_node=8 --master_port=29512 DPO_train.py \
          --gpu 0,1,2,3,4,5,6,7
          --feature_dir "The path contains all preference pair token in .npz format"
          --lora_checkpoint_path "Path to the LoRA checkpoint of EiRA"
          --save_path "Checkpoint path to be saved"

📄 Citation

@article {Zeng2025.09.02.673615,
   author = {Zeng, Wenwu and Zou, Haitao and Li, Xiaoyu and Wang, Xiaoqi and Peng, Shaoliang},
   title = {Improved multimodal protein language model-driven universal biomolecules-binding protein design with EiRA},
   elocation-id = {2025.09.02.673615},
   year = {2025},
   doi = {10.1101/2025.09.02.673615},
   publisher = {Cold Spring Harbor Laboratory},
   abstract = {The interactions between proteins and other biomolecules, such as nucleic acids, form a complex system that supports life activities. Designing proteins capable of targeted biomolecular binding is therefore critical for protein engineering and gene therapy. In this study, we propose a new generative model, EiRA, specifically designed for universal biomolecular binding protein design, which undergo two-stage post-training, i.e., domain-adaptive masking training and binding site-informed preference optimization, based on a general multimodal protein language model. A multidimensional evaluation reveals the SOTA performance of EiRA, including structural confidence, diversity, novelty, and designability on eight test sets across six biomolecule types. Meanwhile, EiRA provides a better characterization of biomolecular binding proteins than generic models, thereby improving the predictive performance of various downstream tasks. We also mitigate severe repetition generation in the original language model by optimizing training strategies and loss. Additionally, we introduced DNA information into EiRA to support DNA-conditioned binder design, further expanding the boundaries of the design paradigm.Competing Interest StatementThe authors have declared no competing interest.National Key R\&D Program of China, 2022YFC3400400NSFC Grants, U19A2067Key R\&D Program of Hunan Province, 2023GK2004, 2023SK2059, 2023SK2060Top 10 Technical Key Project in Hunan Province, 2023GK1010Key Technologies R\&D Program of Guangdong Province, 2023B1111030004},
   URL = {https://www.biorxiv.org/content/early/2025/09/05/2025.09.02.673615},
   eprint = {https://www.biorxiv.org/content/early/2025/09/05/2025.09.02.673615.full.pdf},
   journal = {bioRxiv}
}

☎ Contact

Please feel free to contact us at wwz_cs@hnu.edu.cn, if you have any question!