A Phase-Resolved Geometric Deep Learning Framework Maps Structural Determinants of Disease-Associated Protein...

Area of Science:

• Computational biology
• Structural biology
• Neuroscience

Background:

• Protein aggregation is a key driver of neurodegenerative diseases.
• Current computational methods often oversimplify aggregation into static risk scores, failing to distinguish between nucleation and elongation phases.
• Understanding the structural basis of these distinct aggregation phases is crucial for developing targeted therapies.

Purpose of the Study:

• To develop a phase-resolved computational framework (SKALE 2.0) that distinguishes between protein nucleation and elongation.
• To identify mutation-specific structural determinants influencing these aggregation phases.
• To link atomic-level protein topology to phase-specific aggregation kinetics.

Main Methods:

• Developed SKALE 2.0, a geometric deep learning framework representing proteins as multimodal structural graphs.
• Trained the model on three-dimensional protein topology to learn mutation-induced aggregation phenotypes.
• Compared SKALE 2.0 with baseline models including protein language models and AlphaFold-derived features.
• Utilized recombinant SOD1 experiments to validate predicted mutations.

Main Results:

• SKALE 2.0 successfully recovered a conserved transition from nucleation to elongation across multiple proteins (SOD1, TDP-43, MAPT, PRNP).
• The framework resolved distinct mutation-specific phase sensitivities, outperforming non-phase-aware baselines.
• Learned geometric features indicated nucleation is linked to buried hydrophobic changes, while elongation involves solvent-accessible interfaces.
• Experimental validation confirmed predicted mutations' effects on aggregation initiation and propagation.

Conclusions:

• Explicit phase conditioning is essential for accurately modeling protein aggregation.
• SKALE 2.0 provides a mechanistic link between protein structure, phase-specific assembly kinetics, and disease.
• The framework enables the constraint-aware design of aggregation suppressors for neurodegenerative diseases.