Reproduce “Sesame: Opening the door to protein pockets” on OpenResearch

About this paper

Raúl Miñán, Carles Perez-Lopez, Javier Iglesias, Álvaro Ciudad, Alexis Molina

Nostrum Biodiscovery researchers developed SESAME, a generative model utilizing flow matching to predict ligand-agnostic apo-to-holo protein conformational changes. The model improved molecular docking accuracy and facilitated cryptic pocket identification, outperforming traditional methods and prior deep learning baselines in generating relevant protein structures.

Key insights

• Generative models utilizing flow matching can efficiently and effectively predict ligand-agnostic apo-to-holo conformational changes, offering a scalable alternative to computationally intensive methods.
• The model's generalization capabilities are dependent on the training data distribution; models trained on large conformational changes perform poorly on small ones and vice-versa, suggesting a need for specialized or more diverse training data.
• Generating more accurate holo-like protein backbone conformations significantly enhances downstream applications, improving both cryptic pocket identification and the success rate of molecular docking protocols.

Summary by alphaXiv.