DiAMoNDBack: Diffusion-Denoising Autoregressive Model for Non-Deterministic Backmapping of Cα Protein Traces

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Artificial Intelligence in Molecular Modeling

Background:

• Coarse-grained protein models offer computational efficiency for large-scale simulations but lack atomistic detail.
• Restoring all-atom resolution from coarse-grained models (backmapping) is crucial for mechanistic insights.
• Existing backmapping methods may struggle with accuracy, bond formation, and generating diverse configurations.

Purpose of the Study:

• To introduce DiAMoNDBack, an autoregressive denoising diffusion probability model for protein backmapping.
• To enable the generation of realistic, diverse all-atom protein structures from Cα-only representations.
• To provide a transferable and accurate method for reconstructing atomistic protein details.

Main Methods:

• Developed DiAMoNDBack, an autoregressive denoising diffusion model.
• Trained the model on over 65,000 Protein Data Bank structures.
• Conditioned residue-by-residue generation on Cα trace and local atomic context.

Main Results:

• Achieved state-of-the-art performance in reconstructing all-atom protein structures.
• Demonstrated accurate bond formation and avoidance of steric clashes.
• Generated diverse ensembles of side-chain configurations consistent with Cα traces.

Conclusions:

• DiAMoNDBack effectively restores atomistic detail to coarse-grained protein models.
• The model's local and autoregressive nature ensures transferability across different proteins.
• Publicly released DiAMoNDBack as an open-source Python package for broader accessibility.