LogP Predictor

This repository contains a Graph Neural Network (GNN) model to predict the LogP of small molecules. The model is built with PyTorch and PyTorch Lightning. The following figure shows a parity plot comparing predicted LogP values to true values on the test set.

Installation

1. Clone the repository:

   git clone https://github.com/daandtu/logP-predictor.git
   cd logp-predictor

2. Create a virtual environment and install dependencies:

   python3 -m venv .venv
   source .venv/bin/activate
   pip install -e .

Usage

The command-line interface (CLI) provides several commands to download data from ChEMBL, train a model, make predictions, and evaluate performance.

Download Data

First, download the LogP dataset from the ChEMBL database:

logp-predictor download

This will create a data/logp.csv file.

Train a Model

To train a new model from scratch:

logp-predictor train --epochs 100 --batch-size 64

Model checkpoints will be saved in the checkpoints/ directory, and logs will be stored in logs/.

Make Predictions

You can use a trained model to predict the LogP for a given SMILES string. Make sure to point to a valid checkpoint file.

logp-predictor predict --checkpoint checkpoints/last.ckpt --smiles "CCO"

Evaluate a Model

To evaluate a model on the test set and see the performance metrics:

logp-predictor evaluate --checkpoint checkpoints/last.ckpt

Visualize Results

To generate a parity plot comparing the model's predictions to the true values on the test set:

logp-predictor visualize --checkpoint checkpoints/last.ckpt --output results/parity_plot.png

This will create an image file like the one shown at the top of this README.