SMILES Drug Generation Model

This project implements a transformer-based model for generating drug-like molecules using SMILES (Simplified Molecular Input Line Entry System) strings. The model incorporates molecular properties and scaffolds to guide the generation process.

Project Structure

DrugModel/
│
├── model/
│   ├── __init__.py
│   ├── embeddings.py
│   ├── attention.py
│   ├── transformer_blocks.py
│   ├── transformer.py
│   └── utils.py
│
├── training/
│   ├── __init__.py
│   ├── train.py
│   ├── validation.py
│   ├── tokenizer.py
│   └── model_utils.py
│
├── data/
│   ├── __init__.py
│   ├── dataset.py
│   ├── raw/
│   │   ├── run.sh
│   │   └── sdf_to_smiles.py
│   └── processing/
│       ├── preprocess.py
│       └── run.sh
│
├── utils/
│   ├── __init__.py
│   ├── sascorer.py
│   ├── metrics.py
│   └── loss.py
│
├── config.py
├── requirements.txt
└── main.py

Installation

1. Clone the repository:

git clone https://github.com/Mahmoud-A1i/DrugModel.git
cd DrugModel

2. Create a virtual environment (optional but recommended):

python -m venv venv
source venv/bin/activate  # On Windows, use venv\Scripts\activate

3. Install the required packages:

pip install -r requirements.txt

How to Train with Your Own Data

1. Prepare your data: Place your SMILES strings in a text file (one per line) in the data/processing/ directory.

2. Preprocess the data:

cd data/processing
./run.sh

This will run the preprocess.py script, which filters the SMILES strings based on drug-likeness criteria and calculates molecular properties. The output files will be saved in the data/ directory.

3. Train the model:

python main.py train --train_file data/processed_smiles_train.csv

Usage

To generate new SMILES strings using a trained model:

python main.py generate --input_file path/to/input.csv --output_file path/to/output.csv --weights_path path/to/model_weights.pt

To evaluate the model:

python main.py test --test_file data/processed_smiles_test.csv --weights_path path/to/model_weights.pt

Model Architecture

Our model is based on the transformer architecture with some modifications tailored for SMILES string generation:

1. Encoder:

   • The encoder processes the input partial SMILES string using self-attention mechanisms.
   • It then applies cross-attention to incorporate information from the molecular properties and scaffold.

2. Decoder:

   • The decoder generates the output SMILES string token by token.
   • It uses self-attention on the generated sequence and cross-attention to attend to the encoder's output.

3. Property and Scaffold Integration:

   • Molecular properties are embedded and added to the encoder's output.
   • The scaffold is processed separately and integrated via cross-attention in the encoder.

This architecture allows the model to consider not just the partial SMILES string, but also the desired molecular properties and scaffold information during generation. The self-attention mechanisms help the model understand the relationships between different parts of the SMILES string, while the cross-attention mechanisms allow it to incorporate the additional property and scaffold information effectively.

Acknowledgements

I acknowledge the use of the base transformer architecture from https://github.com/hkproj/pytorch-transformer, which served as a starting point for the model developed in this project. The architecture was substantially modified and adapted to suit the specific requirements of SMILES generation and molecular property prediction.

I also utilized code from the Molecular Sets repository (https://github.com/molecularsets/moses), specifically the methods found in utils/sascorer.py, including calculateScore, novelty, and fraction_valid. Due to the outdated nature of the library, direct usage was not feasible; therefore, I integrated and adapted the necessary components into this project to ensure compatibility and functionality.