Drug Design with Diffusion Models

This project was conducted at École des Ponts ParisTech, under the supervision of Paraskevi Gkeka, Gabriel Stoltz, Tony Levièvre, and Régis Santet.
The goal was to explore how diffusion models can be applied to drug design, starting from fundamentals in image generation to protein-ligand docking.

---

Project Overview

The project is divided into two main parts:

1. Diffusion Models for Images

• Implemented a simple Denoising Diffusion Probabilistic Model (DDPM) on the Fashion-MNIST dataset.
• The implementation was inspired by an open-source notebook from the CNRS.
• Code is available in the image_generation folder.

2. DiffDock for Drug Design

• Studied the DiffDock model from CSAIL, MIT (Gabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay, Tommi Jaakkola).
• Adapted a Colab notebook (original here) to simulate ligand docking on the BCR-ABL protein complex.
• Explored docking procedures, analyzed results, and visualized interactions.
• Performed a comparative study between DiffDock and GNINA, focusing on RMSD (Root Mean Square Deviation) metrics to evaluate prediction accuracy.
• Code and experiments are available in the diff_dock folder.

References

• Ho, J., Jain, A., & Abbeel, P. (2020). Denoising Diffusion Probabilistic Models. NeurIPS.
• Corso, G., Stärk, H., Jing, B., Barzilay, R., & Jaakkola, T. (2023). DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. CSAIL, MIT.