# [A trimodal protein language model enables advanced protein searches](https://www.nature.com/articles/s41587-025-02836-0) You can download the final accepted version [here](https://drive.google.com/file/d/11rNOUlSr8CSpBc5QsJH8QXa1oje_qyXT/view?usp=drive_link) in Nature Biotechnology, provided for open access purposes. **Try our online server** [here](http://search-protrek.com/). **Finetuning ProTrek for diverse tasks** [here](https://colab.research.google.com/github/westlake-repl/SaprotHub/blob/main/colab/ColabSeprot.ipynb?hl=en#scrollTo=paX3gluumu7J). **Download function embeddings and descriptions** [here](https://huggingface.co/datasets/westlake-repl/faiss_index/tree/main/SwissProt/ProTrek_650M_UniRef50/text/subsections/). **Download billion-scale protein embeddings** [here](https://protrek.westlake.edu.cn/) More information about ProTrek, see [FAQ](https://github.com/westlake-repl/ProTrek/wiki/FAQs) If you have any question about the paper or the code, feel free to raise an issue! > Every year, we have 2 PhD positions for international students at Westlake University! Join us [here](https://x.com/duguyuan/status/1897101692665258245). Contact yuanfajie@westlake.edu.cn
Table of contents - [News](#News) - [Overview](#Overview) - [Environment installation](#Environment-installation) - [Download model weights](#Download-model-weights) - [Download Foldseek binary file](#Download-Foldseek-binary-file) - [Obtain embeddings and calculate similarity score](#Obtain-embeddings-and-calculate-similarity-score) - [Deploy your server locally](#Deploy-your-server-locally) - [Add custom database](#Add-custom-database) - [Citation](#Citation)
## News - **2025/10/02:** ProTrek is now online in Nature Biotechnology, see [here](https://www.nature.com/articles/s41587-025-02836-0). - **2025/08/20:** We have shared all database embeddings generated by ProTrek, see [here](https://protrek.westlake.edu.cn/). - **2025/08/15:** We added the the full OMG and MGnify databases, which contains over 3 billion protein sequences from metagenomic sequencing. - **2025/02/20:** ProTrek search results have been experimentally validated, see [here](https://x.com/duguyuan/status/1892419416924836164). - **2025/02/09:** We added the NCBI database, which contains 700 million protein sequences - **2025/01/01:** We added the GOPC database, which contains 2 bilion protein sequences from global ocean microbiome. - **2024/09/27:** We added the OMG_prot50 database, which contains 200 million protein sequences from metagenomic sequencing. - **2024/09/04:** We built [ColabProTrek](https://colab.research.google.com/github/westlake-repl/SaprotHub/blob/main/colab/ColabProTrek.ipynb?hl=en). ColabProTrek has joined [OPMC](https://theopmc.github.io/). ## Overview ProTrek is a tri-modal protein language model that jointly models protein sequence, structure and function (SSF). It employs contrastive learning with three core alignment strategies: (1) using structure as the supervision signal for AA sequences and vice versa, (2) mutual supervision between sequences and functions, and (3) mutual supervision between structures and functions. This tri-modal alignment training enables ProTrek to tightly associate SSF by bringing genuine sample pairs (sequence-structure, sequence-function, and structure-function) closer together while pushing negative samples farther apart in the latent space. ProTrek achieves over 30x and 60x improvements in sequence-function and function-sequence retrieval, is 100x faster than Foldseek and MMseqs2 in protein-protein search, and outperforms ESM-2 in 9 of 11 downstream prediction tasks. ## Environment installation ### Create a virtual environment ``` conda create -n protrek python=3.10 --yes conda activate protrek ``` ### Clone the repo and install packages ``` bash environment.sh ``` ## Download model weights ProTrek provides pre-trained models with different sizes (35M and 650M), as shown below. For each pre-trained model, Please download all files and put them in the `weights` directory, e.g. `weights/ProTrek_35M/...`. | **Name** | **Size （protein sequence encoder）** | **Size （protein structure encoder）** | **Size （text encoder）** | Dataset | | ------------------------------------------------------------ | ------------------------------------- | -------------------------------------- |-------------------------| --------------------- | | [ProTrek_35M](https://huggingface.co/westlake-repl/ProTrek_35M) | 35M parameters | 35M parameters | 130M parameters | Swiss-Prot + TrEMBL50 | | [ProTrek_650M](https://huggingface.co/westlake-repl/ProTrek_650M) | 650M parameters | 150M parameters | 130M parameters | Swiss-Prot + TrEMBL50 | We provide an example to download the pre-trained model weights. ``` huggingface-cli download westlake-repl/ProTrek_650M \ --repo-type model \ --local-dir weights/ProTrek_650M ``` > Note: if you cannot access the huggingface website, you can try to connect to the mirror site through "export > HF_ENDPOINT=https://hf-mirror.com" ## Download Foldseek binary file To run examples correctly and deploy your demo locally, please at first download the Foldseek binary file from [here](https://drive.google.com/file/d/1B_9t3n_nlj8Y3Kpc_mMjtMdY0OPYa7Re/view?usp=sharing) and place it into the `bin` folder. Then add the execute permission to the binary file. ``` chmod +x bin/foldseek ``` ## Obtain embeddings and calculate similarity score Below is an example of how to obtain embeddings and calculate similarity score using the pre-trained ProTrek model. ```python import torch from model.ProTrek.protrek_trimodal_model import ProTrekTrimodalModel from utils.foldseek_util import get_struc_seq # Load model config = { "protein_config": "weights/ProTrek_650M/esm2_t33_650M_UR50D", "text_config": "weights/ProTrek_650M/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext", "structure_config": "weights/ProTrek_650M/foldseek_t30_150M", "from_checkpoint": "weights/ProTrek_650M/ProTrek_650M.pt" } device = "cuda" model = ProTrekTrimodalModel(**config).eval().to(device) # Load protein and text pdb_path = "example/8ac8.cif" seqs = get_struc_seq("bin/foldseek", pdb_path, ["A"])["A"] aa_seq = seqs[0] foldseek_seq = seqs[1].lower() text = "Replication initiator in the monomeric form, and autogenous repressor in the dimeric form." with torch.no_grad(): # Obtain protein sequence embedding seq_embedding = model.get_protein_repr([aa_seq]) print("Protein sequence embedding shape:", seq_embedding.shape) # Obtain protein structure embedding struc_embedding = model.get_structure_repr([foldseek_seq]) print("Protein structure embedding shape:", struc_embedding.shape) # Obtain text embedding text_embedding = model.get_text_repr([text]) print("Text embedding shape:", text_embedding.shape) # Calculate similarity score between protein sequence and structure seq_struc_score = seq_embedding @ struc_embedding.T / model.temperature print("Similarity score between protein sequence and structure:", seq_struc_score.item()) # Calculate similarity score between protein sequence and text seq_text_score = seq_embedding @ text_embedding.T / model.temperature print("Similarity score between protein sequence and text:", seq_text_score.item()) # Calculate similarity score between protein structure and text struc_text_score = struc_embedding @ text_embedding.T / model.temperature print("Similarity score between protein structure and text:", struc_text_score.item()) """ Protein sequence embedding shape: torch.Size([1, 1024]) Protein structure embedding shape: torch.Size([1, 1024]) Text embedding shape: torch.Size([1, 1024]) Similarity score between protein sequence and structure: 28.506675720214844 Similarity score between protein sequence and text: 17.842409133911133 Similarity score between protein structure and text: 11.866174697875977 """ ``` ## Deploy your server locally We strongly recommend deploying it on your university server and have your IT department (or a computer science student) do it, it's easy for them. Please follow the steps below: ### Step 1: Install the environment Please follow the instructions in the [Environment installation](#Environment-installation) section. ### Step 2: Download the Foldseek binary file Please follow the instructions in the [Download Foldseek binary file](#Download-Foldseek-binary-file) section. ### Step 3: Download the pre-trained model weights Please download the weights of [ProTrek_650M](https://huggingface.co/westlake-repl/ProTrek_650M) and put them into the `weights` directory, i.e. `weights/ProTrek_650M/...`. Please follow the instructions in the [Download model weights](#Download-model-weights) section. ### Step 4: Download pre-computed faiss index We have built faiss index for fast retrieval using the embedding computed by [ProTrek_650M](https://huggingface.co/westlake-repl/ProTrek_650M). Please download the faiss index from [here](https://huggingface.co/datasets/westlake-repl/faiss_index_ProTrek_650M_UniRef50/tree/main) and put it into the `faiss_index` directory, i.e. `faiss_index/SwissProt/...`. You can follow the below command to download the faiss index. ``` huggingface-cli download westlake-repl/faiss_index --repo-type dataset --local-dir faiss_index/ ``` ### Step 5: Run the server After all data and files are prepared, you can run the server by executing the following command. Once you see the prompt ``All servers are active! You can now visit http://127.0.0.1:7860/ to start to use.``, you can visit the specified URL to use the server. ``` # Important: The server will occupy the ports 7860 to 7863, please make sure these ports are available! python demo/run_pipeline.py ``` ### Step 6(optional): SSH port forwarding If the software is deployed on your remote server, you can use SSH port forwarding to connect to the server. Specifically, you can run the code below on your local computer: ``` ssh -NL 7860:localhost:7860 user@remote_server ``` Then you can visit http://127.0.0.1:7860/ on your local computer. ## Add custom database You can add your custom database to the server for retrieval. Please follow the instructions below: ### Step 1: Build the faiss index You can build the faiss index through a ``.fasta`` file: ``` python scripts/generate_database.py --fasta example/custom_db.fasta --save_dir faiss_index/Custom/ProTrek_650M/sequence ``` ### Step 2: Add the index to the config file You need to add the index to the ``demo/config.yaml``: ``` ... sequence_index_dir: - name: Swiss-Prot index_dir: faiss_index/SwissProt/ProTrek_650M/sequence # Add your custom database here - name: Custom index_dir: faiss_index/Custom/ProTrek_650M/sequence ... frontend: sequence: [ 'Swiss-Prot', # Add your custom database here 'Custom', ] ... ``` Finally, you can run the server to use the custom database. ## Citation If you find ProTrek useful for your research, please consider citing the following paper: ``` @article{su2025trimodal, title={A trimodal protein language model enables advanced protein searches}, author={Su, Jin and He, Yan and You, Shiyang and Jiang, Shiyu and Zhou, Xibin and Zhang, Xuting and Wang, Yuxuan and Su, Xining and Tolstoy, Igor and Chang, Xing and others}, journal={Nature Biotechnology}, pages={1--7}, year={2025}, publisher={Nature Publishing Group US New York} } ``` ### Other resources - [Evolla](https://www.biorxiv.org/content/10.1101/2025.01.05.630192v1) and its [online server](http://www.chat-protein.com/) - [Pinal](https://www.biorxiv.org/content/10.1101/2024.08.01.606258v2) and its [online server](http://www.denovo-pinal.com/) - [SaprotHub](https://www.biorxiv.org/content/10.1101/2024.05.24.595648v5) and its [online server](https://colab.research.google.com/github/westlake-repl/SaprotHub/blob/main/colab/SaprotHub_v2.ipynb?hl=en)