Graph-based embeddings of laminar cortical columns reveal interpretable microstructural patterns associated with Alzheimer’s disease risk

Abstract

Brain aging and Alzheimer’s disease (AD) risk are associated with spatially heterogeneous alterations in cortical microstructure that may be obscured by region-averaged imaging measures. We present a graph-based framework for depth-resolved and interpretable modeling of cortical microstructure using diffusion MRI and quantitative susceptibility mapping (QSM). Cortical profiles from 58 participants were sampled at 21 locations along the gray–white matter-to-pial trajectory within 68 Desikan–Killiany regions and incorporated into subject-specific graphs using joint mean diffusivity (MD) and QSM features. Graphs were embedded into an eight-dimensional latent space using the GAUDI variational graph autoencoder. An APOE ε4-aligned latent direction was derived from the standardized embeddings, and participant scores were subsequently adjusted for age, sex, and body mass index for anatomical attribution. Repeated nested cross-validation showed moderate held-out carrier–noncarrier separation (Cohen’s d=0.53), but the effect was not significant relative to a permutation null. Lasso–SHAP back-projection revealed a distributed cortical pattern with strongest attribution in bilateral pericalcarine cortex and additional contributions from insular, temporal, lingual, and paracentral regions. Depth-resolved attribution was nonuniform: infragranular MD provided the largest global contribution, followed by supragranular QSM, with additional supragranular and granular MD contributions. These findings demonstrate that graph-based modeling of multimodal cortical-column profiles can provide an anatomically interpretable representation of distributed and depth-dependent microstructural organization aligned with APOE genetic risk.