Sampling Conformational Ensembles of Highly Dynamic Proteins via Generative Deep Learning

Area of Science:

• Biophysics
• Computational Biology
• Structural Biology

Background:

• Protein conformational ensembles are crucial for biological function, particularly for intrinsically disordered proteins (IDPs).
• Understanding and computationally sampling these ensembles, especially for dynamic proteins like amyloid-β42 (Aβ42), presents significant challenges.
• Existing methods struggle to comprehensively explore the vast conformational landscape of proteins.

Purpose of the Study:

• To develop an unsupervised deep learning model, Internal Coordinate Net (ICoN), for learning and sampling protein conformational dynamics.
• To rapidly identify novel synthetic protein conformations with complex structural arrangements.
• To apply ICoN to the amyloid-β42 monomer to comprehensively sample its conformational landscape and rationalize experimental findings.

Main Methods:

• Developed an unsupervised deep learning model, Internal Coordinate Net (ICoN), trained on molecular dynamics (MD) simulation data.
• Utilized latent space interpolation within ICoN to generate novel synthetic conformations.
• Applied the ICoN model to the amyloid-β42 monomer to explore its conformational landscape.

Main Results:

• ICoN successfully learned physical principles of protein conformational changes from MD data.
• The model efficiently generated novel synthetic conformations with sophisticated backbone and sidechain arrangements.
• Comprehensive sampling of the Aβ42 monomer's conformational landscape revealed functionally relevant clusters and rationalized experimental data.
• Identified novel conformations with atomistic details and distinct sidechain rearrangements, validated by EPR and amino acid substitution studies.

Conclusions:

• Internal Coordinate Net (ICoN) provides a powerful and transferable deep learning approach for sampling protein conformational ensembles.
• The method enhances the understanding of protein dynamics, IDPs, and disease-related protein aggregations.
• Deep learning can effectively leverage learned atomistic motions for advanced protein conformation sampling and discovery.