Neural Upscaling from Residue-Level Protein Structure Networks to Atomistic Structures

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Coarse-graining simplifies complex biomolecular systems for efficient simulations.
• Topological coarse-graining, using protein structure networks (PSNs), offers significant computational savings but loses atomistic detail.
• Reconstructing atomistic detail from coarse-grained models is a key challenge.

Purpose of the Study:

• To introduce a novel
• neural upscaling
• method for inferring atomic coordinates from protein structure networks (PSNs).
• To evaluate the effectiveness of neural upscaling in recovering atomistic structural information, particularly for intrinsically disordered proteins.
• To bridge the gap between computationally efficient coarse-grained models and the need for atomistic detail in molecular dynamics.

Main Methods:

• Developed a machine learning approach combined with physically-guided refinement to infer atomic coordinates from PSNs.
• Utilized the constraints and configuration likelihoods inherent in PSNs to guide the upscaling process.
• Applied the method to a 1 μs atomistic molecular dynamics trajectory of Aβ1-40.

Main Results:

• Neural upscaling successfully recapitulated detailed structural information from PSNs.
• The method was particularly effective in recovering transient secondary structure features in intrinsically disordered proteins.
• Demonstrated the ability to impute atomistic coordinates from network representations.

Conclusions:

• Scalable network-based models can be combined with neural upscaling to achieve atomistic detail in protein structure and dynamics.
• Neural upscaling offers a promising strategy for enhancing the utility of coarse-grained models in biophysical simulations.
• This approach facilitates the study of complex biological systems where both computational efficiency and atomistic resolution are required.