Cortical geometry dominates structure-function coupling in resting-state fMRI: a multi-model held-out comparison on the Dallas Lifespan Brain Study

The relationship between structural and functional brain connectivity remains a central question in network neuroscience. Two contrasting accounts have shaped the field: one emphasizes connectome topology—network motifs, communication paths, and global spectral structure—while the other foregrounds cortical geometry as the dominant organizing principle. To date, no held-out, multi-model comparison has decomposed the contributions of these distinct hypotheses on the same dataset. Using the Dallas Lifespan Brain Study (n = 192) with rs-fMRI and DTI tractography in a Schaefer-100 parcellation, we locked a 70/30 subject-level train/test split (n_train = 135; n_test = 57) and fit four model families on the training cohort: (i) coupling potential, comprising classical link-prediction motifs derived from oscillator dynamics; (ii) connectome spectral regression on the structural Laplacian eigenmodes; (iii) geometric spectral regression on Gaussian-kernel eigenmodes derived from MNI parcel centroids; and (iv) a hybrid combining all three. The hybrid substantially outperformed any single family on held-out data (test R² = 0.123; group-mean r = 0.75; per-subject mean r = 0.38). Ablation analysis revealed that geometric features uniquely explained 50% of the hybrid's R², connectome spectral features contributed 10%, and motif-based predictors uniquely contributed only 3%. Two spin-null analyses—hemispheric centroid rotation and parcel-label permutation, 100 iterations each—confirmed that geometric performance is specific to the actual cortical layout (z = 3.36 and z = 23.20 respectively; p = 0.01 in both cases). These findings identify cortical geometry as the dominant correlate of structure–function coupling at structurally distant pairs in adult brains, with connectome topology contributing measurably but to a smaller extent, and link-prediction motifs as essentially redundant once geometric and spectral features are present. Geometric eigenmodes should serve as a baseline benchmark in future structure–function modeling work.