Seed-based correlation analysis

The left and right thalami were individually seeded by extracting the average time-series of all voxels within the region defined by the Harvard-Oxford thalamus probabilistic atlas, thresholded at 90%. The mPFC, PCC, and left dlPFC seeds were created using spherical ROIs with 5 mm radii centered at MNI coordinates (0,47,−2), (0,−49,40), and (−38,52,20), respectively. These were the peak coordinates for regions involved in the default mode and executive control networks^53,54. The resulting time-series from the five ROIs were placed in a first-level GLM to determine the relationship of each seed with every other voxel in the brain. Nuisance regressors were included in the GLM, which consisted of the six rigid body head-motion parameters, CSF, WM, global signal, and their derivatives, resulting in a 18-parameter model. A separate analysis without global signal regression (GSR) was performed to determine if results were influenced by motion artifacts. Frames censored for excessive FD were additionally included in the GLM for motion scrubbing. The first-level GLM analysis was run using FSL FEAT⁵⁵. The resulting parameter estimates for each participant were incorporated in a higher-level GLM, which included any variables with significant group differences as covariates. This higher-level GLM analysis was run using FSL randomise⁵⁶, a permutation-based method that uses threshold-free cluster enhancement to correct for family-wise error rate in order to determine brain areas with group differences in parameter estimates between TD and ASD. Permutation statistical tests were deemed most appropriate for group-level analysis since the distribution of the difference in group means is not known and equal variances were assumed. Finally, for each of these FSL-identified “clusters,” its mean parameter estimate was extracted for each individual using Featquery⁵⁵. The significance threshold for identified clusters was set to p = 0.01 after Bonferroni correction for five seeds (α = 0.05, m = 5).

To validate our observed connectivity findings in the ABIDE cohort, we ran whole-brain connectivity analysis using a similar GLM approach as in our cohort with the five seeds of interest. To account for potential confounding effects from differences in scan acquisition parameters, we included site as a covariate in our ABIDE GLM analysis as one-hot encoded variables.