Generating knotted polymer and protein structures by machine learning.

Area of Science:

• Computational chemistry and materials science
• Biophysics and structural biology
• Machine learning and artificial intelligence

Background:

• Knotted molecules, found in nature, possess unique biological and material properties.
• Conventional molecular simulation methods can produce knotted conformations, but direct generation of specific knot types is challenging.
• Diffusion models excel at generative tasks but require specialized steering for complex topological constraints.

Purpose of the Study:

• To develop a machine learning framework for directly generating polymer and protein conformations with user-defined knot types.
• To integrate a knot-type classifier with a diffusion model to guide molecular topology generation.
• To explore the potential for designing novel knotted polymers and proteins.

Main Methods:

• Development of a diffusion-based machine learning framework.
• Integration of a Transformer-based knot-type classifier (achieving >99% accuracy) to guide the diffusion process.
• Application of the framework to generate polymer conformations and subsequently adapt it for protein backbone generation.

Main Results:

• The classifier-guided diffusion model successfully generated polymer conformations with specified knot types.
• Generated polymer structures accurately reproduced key statistical properties like radius of gyration and knot size distributions.
• The adapted model generated structurally plausible knotted protein designs, validated by sequence-structure consistency tests.

Conclusions:

• The developed classifier-guided diffusion framework offers a novel route for the direct generation of molecular topologies, including specific knot types.
• This approach facilitates the design of complex knotted polymers and proteins with tailored properties.
• The findings open new avenues in computational design for advanced materials and biological applications.