TxGNN Restful API - FastAPI with latest CUDA & PyTorch Support and Docker Compose

[l4b4r4b4b4]

This PR provides an out-of-the-box the implementation of a RESTful API that exposes TxGNN's functionality for zero-shot therapeutic predictions and explanations using geometric deep learning (GNN). This API is deployed using FastAPI, with CUDA support for GPU acceleration, and containerized using Docker. Additionally, the repository includes a Makefile, docker-compose.yml, and .env file for convenient management and deployment of the application.

It all worked after pushing the stack to latest tech meaning: CUDA: 12.4 PyTorch: 2.4.0 DGL: 2.4.0+cu124

TL;DR - Quickstart

1. Clone the repository

   git clone https://github.com/healthecosystem/TxGNN
   cd TxGNN
   cp .env.example .env
   

2. Run the application using Make

   make run
   

3. Access the API Once the services are up, the API will be locally accessible at port 8883.

Features

• CUDA-enabled: GPU-accelerated predictions for faster performance.
• FastAPI: A modern, fast web framework for building RESTful APIs.
• Docker: Containerized deployment for easy setup and scalability.

API Endpoints

Health Check

• Endpoint: /healthz
• Method: GET
• Description: Returns the health status of the API.

Predict Drug Replacement

• Endpoint: /predict
• Method: GET
• Parameters: disease (str)
• Description: Predicts a drug replacement for a given disease.

Explain Drug Replacement

• Endpoint: /explain
• Method: GET
• Parameters: disease (str), drug (str)
• Description: Explains why a drug is recommended as a replacement for the specified disease.

txgnn  | [32mINFO[0m:     Will watch for changes in these directories: ['/app']
txgnn  | [33mWARNING[0m:  "workers" flag is ignored when reloading is enabled.
txgnn  | [32mINFO[0m:     Uvicorn running on [1mhttp://0.0.0.0:80[0m (Press CTRL+C to quit)
txgnn  | [32mINFO[0m:     Started reloader process [[36m[1m1[0m] using [36m[1mWatchFiles[0m
txgnn  | DGL backend not selected or invalid.  Assuming PyTorch for now.
txgnn  | Setting the default backend to "pytorch". You can change it in the ~/.dgl/config.json file or export the DGLBACKEND environment variable.  Valid options are: pytorch, mxnet, tensorflow (all lowercase)
txgnn  | Number of available CUDA devices: 1
txgnn  | Device 0: NVIDIA GeForce RTX 3080 Ti
txgnn  | Found local copy...
txgnn  | Found local copy...
txgnn  | Found local copy...
txgnn  | Found saved processed KG... Loading...
txgnn  | Splits detected... Loading splits....
txgnn  | Creating DGL graph....
txgnn  | Done!
txgnn  | Loading pre-trained GNN model ... <txgnn.TxGNN.TxGNN object at 0x7f576fe05e50>
txgnn  | Opening file ... 
txgnn  | Inittialise model with config ...  {'n_hid': 512, 'n_inp': 512, 'n_out': 512, 'proto': True, 'proto_num': 3, 'attention': False, 'sim_measure': 'all_nodes_profile', 'bert_measure': 'disease_name', 'agg_measure': 'rarity', 'num_walks': 200, 'walk_mode': 'bit', 'path_length': 2}
txgnn  | Initialising ... 
txgnn  | Setting G ...  
txgnn  | initialize_node_embedding ...  
txgnn  | evaluate_graph_construct valid ...  
....
txgnn  | Set model ... 
txgnn  | Set best model ... 
txgnn  | Loading pre-trained GNN model successfull! <txgnn.TxGNN.TxGNN object at 0x7f576fe05e50>