GCN Solubility Prediction

This project implements and deploys a Graph Convolutional Network (GCN) model for predicting solubility values of molecular compounds based on their SMILES (Simplified Molecular Input Line Entry System) representations. This GCN architecture builds on graph-based representations of molecules to provide accurate predictions of solubility.

Project Overview

Objective

To create a robust GCN model capable of predicting solubility (e.g., solvation free energy) of chemical compounds by leveraging molecular graph representations. The project includes:

• Training and optimization using Bayesian techniques.
• Deployment for real-time predictions.

Key Features

1. Model Architecture:

   • Implements GCN layers with graph convolutions and pooling operations.
   • Flexible hyperparameter tuning for layers, dropout, and learning rates.
   • Designed for scalable training on molecular datasets.

2. Bayesian Optimization:

   • Utilizes Ax-platform for Bayesian optimization.
   • Finds the optimal hyperparameters (e.g., number of layers, learning rates) to minimize prediction error (RMSE).

3. Prediction Service:

   • Supports predictions using single or ensembles of models.
   • Generates outputs with associated uncertainty (mean and standard deviation for ensemble models).

4. Molecular Data Handling:

   • Converts SMILES strings to graph representations using RDKit.
   • Creates feature and adjacency matrices for GCN input.

---

Folder Structure

• model.py: Contains the implementation of the GCN model, including convolution, pooling layers, and utilities for training and testing.

• optimization.py: Automates hyperparameter tuning and model optimization using Bayesian optimization.

• runner.py: Initializes the prediction service and makes predictions based on input SMILES strings.

• collection.py: Manages the dataset, including molecule-to-graph conversion and data loading utilities.

• model_service.py: Provides a unified service for loading models, running predictions, and managing single or multi-model setups.

• config.py: Defines project settings and configurations using environment variables.

---

Setup Instructions

Dependencies

1. Clone the repository
2. Install Poetry if not already installed:

pip install poetry

3. Install dependencies:

poetry install

Configuration

Modify .env or config.py to set paths and default parameters:

• data_file_name: Path to the dataset CSV.
• model_path: Directory to save/load trained models.
• node_vec_len, max_atoms: Feature lengths for molecular graphs.
• Other hyperparameters such as learning_rate and n_epochs.

---

Workflow

1. Training and Optimization:
   • Use runner.py for automatic model optimization:

     poetry run python runner.py

---

Example Prediction

Input SMILES: CC(=O)N1CCCC1
Expected Output:

CC(=O)N1CCCC1 SFE: -5.14 ± 0.513

---

Notes

• This project is a work in progress. Further refinements to the model architecture and deployment are ongoing.
• Ensure the data/train.csv dataset file is available for training.