A Phase‐Resolved Geometric Deep Learning Framework Maps Structural Determinants of Disease‐Associated Protein Aggregation and Guides Suppressor Design

Protein aggregation drives major neurodegenerative diseases, yet most computational predictors collapse assembly into static risk scores and do not resolve the distinct structural determinants of nucleation and elongation. Here, we present SKALE 2.0, a phase-resolved geometric deep learning framework that represents proteins as multimodal structural graphs and learns mutation-induced aggregation phenotypes directly from three-dimensional topology. Across SOD1, TDP-43, MAPT, and PRNP, SKALE 2.0 recovered a conserved latent transition from nucleation to elongation while resolving distinct mutation-specific phase sensitivities. Representative protein language model, AlphaFold-derived feature, and non-phase-aware structural baselines failed to recover both phase-dependent mutation modulation and phase separability, indicating that explicit phase conditioning is essential. The learned geometry showed that nucleation is preferentially coupled to buried hydrophobic perturbations, whereas elongation is shaped by solvent-accessible interfaces that support fibril propagation. This framework explains how pathogenic variants can remain globally folded yet acquire aggregation competence through localized structural rewiring. Recombinant SOD1 experiments validated predicted suppressor, enhancer, and phase-switch mutations, demonstrating that initiation and propagation can be tuned independently. SKALE 2.0 links atomic topology to phase-specific assembly kinetics and enables a constraint-aware design of aggregation suppressors.