Accurate Sampling of Macromolecular Conformations Using Adaptive Deep Learning and Coarse-Grained Representation.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Protein conformational sampling is vital for understanding biochemical functions and predicting thermodynamic properties.
• Traditional methods often rely on sequential sampling, which can be inefficient for large systems.
• Generative deep neural networks offer potential for parallel and statistically independent sampling of molecular configurations.

Purpose of the Study:

• To develop a hierarchical deep learning approach for accurate and efficient conformational sampling of large molecular systems.
• To improve the generation of samples from high-dimensional multimodal equilibrium distribution functions.
• To address the limitations of existing methods in handling complex protein dynamics.

Main Methods:

• Developed a hierarchical approach using expressive normalizing flows with rational quadratic neural splines and coarse-grained representation.
• Incorporated system-specific priors and adaptive, property-based controlled learning to minimize high-energy structure generation.
• Utilized an equivariant transformer model for backmapping from coarse-grained to fully atomistic representations.

Main Results:

• Demonstrated applicability on one-shot configurational sampling of a protein system exceeding one hundred amino acids.
• Achieved enhanced expressivity, mitigating invertibility constraints within the normalizing flow framework.
• Successfully tested the hierarchical model's capacity on the folding/unfolding dynamics of the peptide chignolin.

Conclusions:

• The hierarchical normalizing flow model provides an effective method for protein conformational sampling.
• The approach enhances sampling efficiency and accuracy for large molecular systems.
• This method holds promise for advancing the study of protein dynamics and thermodynamics.