This server is a protein structure prediction tool that processes prediction requests from users and capable of returning various scores for protein sequences.

To install the environment, follow these steps:

git clone https://github.com/Oaklight/protein-score-server.git
cd protein-score-server
conda env create -f env/environment.yaml
conda activate esm
pip install -r env/requirements.txt

Configuration File:

• Copy server.yaml.sample to server.yaml :

cp server.yaml.sample server.yaml

• Edit server.yaml with your settings.

The server uses the server.yaml file for configuration. Currently configurable items include:

• api_key: API key for Hugging Face Hub login.
• history_path: History result storage path.
• intermediate_pdb_path: Intermediate PDB file storage path.
• model: Model configuration
  • name: model name, esmfold or protenix (bytedances' alphafold3 implementation)
  • replica: GPU device and replications mapping, should be in <device>: <num_replica> format. For esmfold case, use _: <num_replica> instead.
• task_queue_size: Task queue size, default to 50.
• timeout: Timeout for async prediction result retrieval, default to 15 seconds.
• backbone_pdb:
  • reversed_index: path for reverse index from pdb id to pdb file path
  • parquet_prefix: path prefix for parquet files
  • pdb_prefix: path prefix for pdb files

For example, see server.yaml

After the config are set, run these commands inside the project folder:

conda activate esm
uvicorn main:app --host 0.0.0.0 --port 8000

Users can send POST requests to http://your-host:8000/predict/ to get predictions. The request body comprises of these fields: seq , name , type , seq2 .

• seq: String, representing the protein sequence.
• name: String, representing the name of the reference protein.
• type: String, representing the task type, currently supports "plddt", "tmscore", "sc-tmscore", "pdb".
• seq2: String, representing the sequence of the reference protein. Used only for sc-tmscore task. You may choose to provide either seq2 or name

1. pLDDT

{
    "seq": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST",
    "type": "plddt"
}

2. TMscore

{
    "seq": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST",
    "name": "1a0a.A", # must provide for tasks that require a reference structure
    "type": "tmscore"
}

3. sc-TMscore

{
    "seq": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST",
    "seq2": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST", # choose to provide either seq2 or name
    "type": "sc-tmscore"
}

or

{
    "seq": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST",
    "name": "1a0a.A", # choose to provide either seq2 or name
    "type": "sc-tmscore"
}

4. pdb

{
    "seq": "MKRESHKHAEQARRNRLAVALHELASLIPAEWKQQNVSAAPSKATTVEAACRYIRHLQQNGST",
    "type": "pdb"
}

The server will return a JSON response containing two fields: job_id and prediction .

• job_id: String, representing the task ID.
• prediction: String, currently only indicating the prediction is in processing.

{
    "job_id": "0a98a981748c4b7eacfd5e0957905ced", # this is a uuid4 hex string
    "prediction": ... # not very useful at this moment
}

Users can send GET requests to http://your-host:8000/result/{job_id} to get prediction results. The header of the request should contain Content-Type: application/json .

The server will return a JSON response containing two fields: job_id and prediction .

{
    "job_id": "0a98a981748c4b7eacfd5e0957905ced", # this is a uuid4 hex string
    "prediction": 0.983124
}

• 202: Job is being processed. Please wait.
• 400: Task input information error. Check detailed messages.
• 404: Task ID does not exist in server records.
• 429: Server job queue is currently full. Please wait.

• Recommend to use an exponential backoff strategy with a base of 3 when querying for results.
• Example of querying is available in test.py.

To stop the server, use Ctrl+C in the terminal where the server is running.

This server is licensed under the Apache License 2.0.