gene retrieval via API

procyon/app/main.py

@@ -0,0 +1,100 @@
+import os
+from typing import Optional
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel, Field
+import pandas as pd
+from loguru import logger
+
+# Import the key functions from the existing codebase
+from procyon.inference.retrieval_utils import startup_retrieval, do_retrieval
+
+app = FastAPI()
+
+# Global variables to store model and device
+model = None
+device = None
+data_args = None
+all_protein_embeddings = None
+
+
+class RetrievalRequest(BaseModel):
+    task_desc: str = Field(description="The task description.")
+    disease_desc: str = Field(description="The disease description.")
+    instruction_source_dataset: str = Field(
+        description="Dataset source for instructions - either 'disgenet' or 'omim'"
+    )
+    k: Optional[int] = Field(
+        default=None,
+        description="Number of top results to return. If None, returns all results",
+        ge=1,
+    )
+
+
+@app.on_event("startup")
+async def startup_event():
+    """Initialize the model and required components on startup"""
+    global model, device, data_args, all_protein_embeddings
+
+    if not os.getenv("HF_TOKEN"):
+        raise EnvironmentError("HF_TOKEN environment variable not set")
+    if not os.getenv("CHECKPOINT_PATH"):
+        raise EnvironmentError("CHECKPOINT_PATH environment variable not set")
+    if not os.getenv("HOME_DIR"):
+        raise EnvironmentError("HOME_DIR environment variable not set")
+    if not os.getenv("DATA_DIR"):
+        raise EnvironmentError("DATA_DIR environment variable not set")
+    if not os.getenv("LLAMA3_PATH"):
+        raise EnvironmentError("LLAMA3_PATH environment variable not set")
+
+    # Use the existing startup_retrieval function
+    model, device, data_args, all_protein_embeddings = startup_retrieval(
+        inference_bool=True
+    )
+    logger.info("Model loaded and ready")
+
+
+@app.post("/retrieve")
+async def retrieve_proteins(request: RetrievalRequest):
+    """Endpoint to perform protein retrieval"""
+    global model, device, data_args, all_protein_embeddings
+
+    if not all([model, device, data_args, all_protein_embeddings]):
+        raise HTTPException(status_code=500, detail="Model not initialized")
+
+    # Use the existing do_retrieval function
+    results_df = do_retrieval(
+        model=model,
+        data_args=data_args,
+        device=device,
+        instruction_source_dataset=request.instruction_source_dataset,
+        all_protein_embeddings=all_protein_embeddings,
+        task_desc=request.task_desc,
+        disease_desc=request.disease_desc,
+    )
+
+    results_df = results_df.fillna("")
+
+    # Return all results if k is None, otherwise return top k
+    if request.k is None:
+        return {"results": results_df.to_dict(orient="records")}
+    return {"results": results_df.head(request.k).to_dict(orient="records")}
+
+
+if __name__ == "__main__":
+    """
+    This API endpoint will allow users to perform protein retrieval for a given disease description using the 
+        pre-trained ProCyon model Procyon-Full.
+    This API script can be run directly using the command `python main.py`
+    this script will start the FastAPI server on port 8000
+    The API will be available at http://localhost:8000
+    An example request can be made using curl:
+    curl -X POST "http://localhost:8000/retrieve" \
+     -H "Content-Type: application/json" \
+     -d '{"task_desc": "Find proteins related to this disease", 
+          "disease_desc": "Major depressive disorder",
+          "instruction_source_dataset": "disgenet",
+          "k": 1000}'
+    """
+    import uvicorn
+
+    uvicorn.run(app, host="0.0.0.0", port=8000)

procyon/inference/retrieval_utils.py

@@ -0,0 +1,199 @@
+import os
+from pathlib import Path
+
+from huggingface_hub import login as hf_login
+from loguru import logger
+import pandas as pd
+from typing import Dict, Optional, Tuple, Union
+import torch
+
+from procyon.data.inference_utils import (
+    create_input_retrieval,
+    get_proteins_from_embedding,
+)
+from procyon.evaluate.framework.utils import move_inputs_to_device
+from procyon.model.model_unified import UnifiedProCyon
+from procyon.training.train_utils import DataArgs
+
+CKPT_NAME = os.path.expanduser(os.getenv("CHECKPOINT_PATH"))
+
+
+def startup_retrieval(
+    inference_bool: bool = True,
+) -> Tuple[
+    Union[UnifiedProCyon, None],
+    Union[torch.device, None],
+    Union[DataArgs, None],
+    torch.Tensor,
+]:
+    """
+    This function performs startup functions to initiate protein retrieval:
+    Logs into the huggingface hub and loads the pre-trained ProCyon model.
+    Args:
+        inference_bool (bool): OPTIONAL; choose this if you do not intend to do inference;
+        then the model will not be loaded.
+    Returns:
+        model (UnifiedProCyon): The pre-trained ProCyon model
+        device (torch.device): The compute device (GPU or CPU) on which the model is loaded
+        data_args (DataArgs): The data arguments defined by the pre-trained model
+        all_protein_embeddings (torch.Tensor): The pre-calculated protein target embeddings
+    """
+    logger.info("Now running startup functions for protein retrieval")
+
+    logger.info("Now logging into huggingface hub")
+    hf_login(token=os.getenv("HF_TOKEN"))
+    logger.info("Done logging into huggingface hub")
+
+    if inference_bool:
+        logger.info("Inference is enabled.")
+
+        # load the pre-trained ProCyon model
+        model, device, data_args = load_model_onto_device()
+    else:
+        logger.info("Inference is disabled.")
+        # loading the model takes much time and memory, so we skip it if we don't need it
+        model = None
+        device = None
+        data_args = None
+
+    # Load the pre-calculated protein target embeddings
+    logger.info("Now loading protein target embeddings")
+    all_protein_embeddings, all_protein_ids = torch.load(
+        os.path.join(CKPT_NAME, "protein_target_embeddings.pkl")
+    )
+    all_protein_embeddings = all_protein_embeddings.float()
+    logger.info(
+        f"shape of precalculated embeddings matrix: {all_protein_embeddings.shape}"
+    )
+    logger.info("Done loading protein target embeddings")
+    logger.info("Done running startup functions for protein retrieval")
+
+    return model, device, data_args, all_protein_embeddings
+
+
+def load_model_onto_device() -> Tuple[UnifiedProCyon, torch.device, DataArgs]:
+    """
+    Load the pre-trained ProCyon model and move it to the compute device.
+    Returns:
+        model (UnifiedProCyon): The pre-trained ProCyon model
+        device (torch.device): The compute device (GPU or CPU) on which the model is loaded
+        data_args (DataArgs): The data arguments defined by the pre-trained model
+    """
+    # Load the pre-trained ProCyon model
+    logger.info("Now loading pretrained model")
+    # Replace with the path where you downloaded a pre-trained ProCyon model (e.g. ProCyon-Full)
+    data_args = torch.load(os.path.join(CKPT_NAME, "data_args.pt"))
+    model, _ = UnifiedProCyon.from_pretrained(checkpoint_dir=CKPT_NAME)
+    logger.info("Done loading pretrained model")
+
+    logger.info("Now quantizing the model to a smaller precision")
+    model.bfloat16()  # Quantize the model to a smaller precision
+    logger.info("Done quantizing the model to a smaller precision")
+
+    logger.info("Now setting the model to evaluation mode")
+    model.eval()
+    logger.info("Done setting the model to evaluation mode")
+
+    logger.info("Now applying pretrained model to device")
+    logger.info(f"Total memory allocated by PyTorch: {torch.cuda.memory_allocated()}")
+    # identify available devices on the machine
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model.to(device)
+    logger.info(f"Total memory allocated by PyTorch: {torch.cuda.memory_allocated()}")
+
+    logger.info("Done loading model and applying it to compute device")
+
+    return model, device, data_args
+
+
+def do_retrieval(
+    model: UnifiedProCyon,
+    data_args: DataArgs,
+    device: torch.device,
+    instruction_source_dataset: str,
+    all_protein_embeddings: torch.Tensor,
+    inference_bool: bool = True,
+    task_desc_infile: Path = None,
+    disease_desc_infile: Path = None,
+    task_desc: str = None,
+    disease_desc: str = None,
+) -> Optional[pd.DataFrame]:
+    """
+    This function performs protein retrieval for a given disease using the pre-trained ProCyon model.
+    Args:
+        model (UnifiedProCyon): The pre-trained ProCyon model
+        data_args (DataArgs): The data arguments defined by the pre-trained model
+        device (torch.device): The compute device (GPU or CPU) on which the model is loaded
+        instruction_source_dataset (str): Dataset source for instructions - either "disgenet" or "omim"
+        all_protein_embeddings (torch.Tensor): The pre-calculated protein target embeddings
+        inference_bool (bool): OPTIONAL; choose this if you do not intend to do inference
+        task_desc_infile (Path): The path to the file containing the task description.
+        disease_desc_infile (Path): The path to the file containing the disease description.
+        task_desc (str): The task description.
+        disease_desc (str): The disease description.
+    Returns:
+        df_dep (pd.DataFrame): The DataFrame containing the top protein retrieval results
+    """
+    logger.info("Now performing protein retrieval")
+
+    if instruction_source_dataset not in ["disgenet", "omim"]:
+        raise ValueError(
+            'instruction_source_dataset must be either "disgenet" or "omim"'
+        )
+
+    logger.info("entering task description and prompt")
+    if task_desc_infile is not None:
+        if task_desc is not None:
+            raise ValueError(
+                "Only one of task_desc_infile and task_desc can be provided."
+            )
+        # read the task description from a file
+        with open(task_desc_infile, "r") as f:
+            task_desc = f.read()
+    elif task_desc is None:
+        raise ValueError("Either task_desc_infile or task_desc must be provided.")
+
+    if disease_desc_infile is not None:
+        if disease_desc is not None:
+            raise ValueError(
+                "Only one of disease_desc_infile and disease_desc can be provided."
+            )
+        # read the disease description from a file
+        with open(disease_desc_infile, "r") as f:
+            disease_desc = f.read()
+    elif disease_desc is None:
+        raise ValueError("Either disease_desc_infile or disease_desc must be provided.")
+
+    task_desc = task_desc.replace("\n", " ")
+    disease_desc = disease_desc.replace("\n", " ")
+    logger.info("Done entering task description and prompt")
+
+    if inference_bool:
+        logger.info("Now performing protein retrieval")
+
+        # Create input for retrieval
+        input_simple = create_input_retrieval(
+            input_description=disease_desc,
+            data_args=data_args,
+            task_definition=task_desc,
+            instruction_source_dataset=instruction_source_dataset,
+            instruction_source_relation="all",
+            aaseq_type="protein",
+            icl_example_number=1,  # 0, 1, 2
+        )
+
+        input_simple = move_inputs_to_device(input_simple, device=device)
+        with torch.no_grad():
+            model_out = model(
+                inputs=input_simple,
+                retrieval=True,
+                aaseq_type="protein",
+            )
+        # The script can run up to here without a GPU, but the following line requires a GPU
+        df_dep = get_proteins_from_embedding(
+            all_protein_embeddings, model_out, top_k=None
+        )
+
+        logger.info("Done performing protein retrieval")
+
+        return df_dep

pyproject.toml

@@ -20,6 +20,7 @@ dependencies = [
    "jupyter==1.0.0",
    "kiwisolver==1.4.5",
    "llvmlite==0.42.0",
+   "loguru==0.7.3",
    "matplotlib==3.8.3",
    "multiprocess==0.70.16",
    "networkx==3.2.1",
@@ -50,6 +51,10 @@ dependencies = [
    "virtualenv==20.28.0",
    "wandb==0.16.3",
    "bert-score>=0.3.13",
+   "argparse>=1.4.0",
+   "huggingface-hub==0.23.4",
+   "fastapi>=0.109.0",
+   "uvicorn>=0.27.0",
 ]
 
 [project.optional-dependencies]