Predicting functional activity from structural connectivity

Structural connectivity is among the most important constraints on a network since it restricts and defines the sort of information it can process.  The functional responses of a voxel should therefore be strongly influenced by its pattern of connectivity; correspondingly, patterns of connectivity should be highly predictive of function.  We present the use of anatomical connectivity to predict functional activity in each gray matter voxel of the brain, both cortical and subcortical.  Diffusion-weighted images were acquired from 18 healthy subjects (mean age=26.4, 7M:11F) using echo planar imaging (64 slices, voxel size 2x2x2mm, 128x128 base resolution, diffusion weighting isotropically distributed along 30 directions, b-value 700s/mm²).  Automated cortical and subcortical parcellation was performed in each subject’s T1 scan, using the Destrieux atlas from Freesurfer 4.5.  This defined 169 regions, which were then registered to each subject’s diffusion image.  Probabilistic diffusion tractography was carried out using FSL-FDT with 25000 streamline samples, from each region to all other regions.  We also acquired fMRI data while the same subjects passively viewed blocks of various stimulus categories.  Functional images were also registered to each subject’s diffusion image.  For each region, we modeled the fMRI data as a function of connection probability using linear regression in a leave one subject out cross validation (LOOCV) regime.  The model surpassed performance distributions generated from randomly permuted data, indicating that the connectivity data are structured sufficiently for prediction, assuring that we are not modeling high-dimensional noise.  We also performed group analysis though LOOCV, registering the resulting group average to the remaining left-out subject’s brain, as a benchmark standard.  Across subjects, the connectivity model significantly outperformed the benchmark at p=5.7e-15.  Comparing within subject voxelwise residuals between the models revealed that connectivity was a better predictor than the benchmark in all 18 subjects at p<1e-5.  This demonstrates the prospect of using anatomical connectivity to predict function; furthermore, the resulting model coefficients could be potentially revealing of mechanistic principles underlying the functional organization of the brain.