StruCloze: A Unified Framework for Backmapping and Inpainting Biomolecule Structures

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Atomistic resolution is crucial for understanding biomolecular function.
• Coarse-grained (CG) models are vital for simulating large biomolecular systems.
• Reconstructing atomistic detail from CG models and handling missing regions are significant challenges.

Purpose of the Study:

• To develop a deep learning framework, StruCloze, for reconstructing atomistic structures from CG models.
• To enable inpainting of missing or disordered regions in biomolecular structures.
• To provide a general solution bridging reduced representations with full atomistic detail.

Main Methods:

• Developed StruCloze, a deep learning framework utilizing a masked learning strategy.
• Pretrained on diverse CG levels and biomolecule types (proteins, nucleic acids).
• Fine-tuning for specific representations to optimize performance.

Main Results:

• Achieved state-of-the-art accuracy in reconstructing protein and nucleic acid structures from CG models.
• Demonstrated superior transferability and speed compared to existing methods.
• Excelled at inpainting missing structural regions.

Conclusions:

• StruCloze offers a general and accurate solution for atomistic structure reconstruction and inpainting.
• The framework facilitates structural refinement and integrative modeling.
• Enables rapid local structure prediction and analysis of biomolecular dynamics.