Sampling Conformational Ensembles of Highly Dynamic Proteins via Generative Deep Learning.

Area of Science:

• Computational Biology
• Biophysics
• Deep Learning Applications

Background:

• Protein conformational ensembles are crucial for biological function, especially for Intrinsically Disordered Proteins (IDPs).
• Investigating IDP conformational dynamics and disease-related aggregation is computationally and experimentally challenging.
• Understanding protein structure-function relationships requires comprehensive sampling of conformational landscapes.

Purpose of the Study:

• To introduce a novel deep learning model, Internal Coordinate Net (ICoN), for learning and predicting protein conformational changes.
• To generate novel synthetic protein conformations by interpolating in the learned latent space.
• To comprehensively sample and analyze the conformational landscape of the amyloid-beta 42 (Aβ42) monomer.

Main Methods:

• Developed Internal Coordinate Net (ICoN), a deep learning model trained on Molecular Dynamics (MD) simulation data.
• Utilized latent space interpolation to generate novel synthetic protein conformations.
• Applied the ICoN model to sample the conformational landscape of the Aβ42 monomer.

Main Results:

• ICoN successfully learned physical principles of protein conformational changes from MD data.
• Generated novel synthetic conformations with complex sidechain and backbone arrangements.
• Identified distinct conformational clusters for Aβ42 that rationalize experimental findings.
• Discovered novel conformations with atomistic details and distinct sidechain rearrangements, validated by experimental studies.

Conclusions:

• Deep learning, specifically ICoN, offers a powerful approach for comprehensive protein conformation sampling.
• The method can identify functionally relevant conformations and aid in understanding disease mechanisms like Aβ42 aggregation.
• ICoN is a transferable approach applicable to various protein systems and available data.