Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Diffusion-map phenotyping from exactly solvable brain-dynamics models via descriptor-family kernel fusion

Julian Kędys1,2, Cezary Mazurek1; 1Poznan Supercomputing and Networking Center, 2Polish Academy of Sciences

Presenter: Julian Kędys

Comparing brain dynamics across heterogeneous subjects requires representations that are cross-subject comparable, mechanistically interpretable, and compact for unsupervised phenotyping. We present a pipeline integrating consensus shared latent coordinates, exactly solvable pairwise maximum-entropy (Ising) models, and multi-view kernel learning to produce a low-dimensional mechanistic phenotyping space. Whole-brain time series are projected into population-universal shared coordinates by aligning and fusing complementary methods (SRM, MCCA, Group PCA/ICA) via Procrustes rotation and stability-aware consensus weighting. With K=12 shared dimensions, the 2^K=4096 state space permits exact partition-function computation, yielding normalised, directly comparable energy landscapes. Energy-landscape and phase-diagram analyses extract heterogeneous per-subject mechanistic descriptors: attractor structure, barrier spectra, kinetic summaries, and near-criticality observables. We integrate these descriptor families via type-specific similarity kernels (Wasserstein for distributions, radial-basis for scalars, spectral for matrices), combined through multiple kernel learning. Diffusion-map embedding of the fused kernel produces interpretable axes along which subjects are positioned by their full dynamical profile. On resting-state functional ultrasound data from heterogeneous autism-model mice, stable subtype clustering emerges along axes reflecting switching kinetics, landscape ruggedness, and criticality proximity, robust under ablation and bootstrap resampling. The framework supports mechanistic phenotyping when large-sample designs are impractical.

Topic Area: Methods, Tools, Theory & Neural Coding