Generalizable Protein Folding Pathway Exploration with DA2-GRASP: Extending Beyond Miniproteins.

Area of Science:

• Computational Biology
• Biophysics
• Artificial Intelligence in Biochemistry

Background:

• Protein dynamics are vital for biological processes and implicated in diseases like Alzheimer's and Parkinson's.
• Predicting static protein structures is advanced by AI, but capturing dynamic folding pathways remains challenging.
• Understanding protein folding is key to deciphering biological functions and disease mechanisms.

Purpose of the Study:

• To present DA2-GRASP, a novel computational framework for efficient and accurate mapping of protein folding pathways.
• To overcome limitations in simulating high-dimensional protein dynamics using conventional methods.
• To enable detailed mechanistic insights into protein folding and its relation to disease.

Main Methods:

• Integration of deep learning (variational autoencoder) with advanced sampling techniques.
• Learning low-dimensional latent representations of protein conformations.
• Multidirectional generative sampling guided by potential energy gradients for efficient pathway reconstruction.

Main Results:

• DA2-GRASP achieves sublinear computational scaling with sequence length, outperforming traditional molecular dynamics.
• The framework accurately quantifies mutation effects on folding thermodynamics, crucial for disease mutation studies.
• Enabled atomistic characterization of folding for medium-sized proteins (e.g., ubiquitin, SUMO) on standard workstations.

Conclusions:

• DA2-GRASP offers a versatile and powerful framework for exploring protein folding dynamics and functional consequences.
• Provides new mechanistic insights into how proteins with similar folds navigate different folding pathways.
• Facilitates research into protein misfolding diseases by enabling tractable and accurate simulations.