SILVR: Guided Diffusion for Molecule Generation.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Generating novel, synthetically accessible drug compounds with desired properties (high affinity, low toxicity) is a significant challenge in drug design.
• Existing machine learning models for de novo molecule generation often require extensive, target-specific tuning.

Purpose of the Study:

• To introduce a novel method, selective iterative latent variable refinement (SILVR), for conditioning generative models without retraining.
• To enable the generation of new molecules tailored to specific protein binding sites using fragment-based information.

Main Methods:

• SILVR conditions existing diffusion-based equivariant generative models using fragment hits from a reference dataset.
• The SARS-CoV-2 main protease fragment dataset from the COVID Moonshot project was used for conditioning.
• The SILVR rate parameter controls the degree of conditioning, allowing for generation of molecules with shapes similar to input fragments.

Main Results:

• Moderate SILVR rates successfully generated novel molecules that fit protein binding sites without explicit protein structure input.
• The method demonstrated the ability to merge up to three fragments into new molecules without compromising the quality of the underlying generative model.
• Generated molecules exhibited shapes consistent with the original fragments, indicating successful binding site adaptation.

Conclusions:

• SILVR offers a powerful, retraining-free approach to condition generative models for targeted drug design.
• The method facilitates the generation of novel molecules that conform to protein binding pockets using fragment-based conditioning.
• SILVR is a generalizable technique applicable to various protein targets and diffusion-based generative models in drug discovery.