A cortical hierarchy of localized and distributed processes revealed via dissociation of task activations, connectivity changes, and intrinsic timescales

Abstract

Many studies have identified the role of localized and distributed cognitive functionality by mapping either local task-related activity or distributed functional connectivity (FC). However, few studies have directly explored the relationship between a brain region's localized task activity and its distributed task FC. Here we systematically evaluated the differential contributions of task-related activity and FC changes to identify a relationship between localized and distributed processes across the cortical hierarchy. We found that across multiple tasks, the magnitude of regional task-evoked activity was high in unimodal areas, but low in transmodal areas. In contrast, we found that task-state FC was significantly reduced in unimodal areas relative to transmodal areas. This revealed a strong negative relationship between localized task activity and distributed FC across cortical regions that was associated with the previously reported principal gradient of macroscale organization. Moreover, this dissociation corresponded to hierarchical cortical differences in the intrinsic timescale estimated from resting-state fMRI and region myelin content estimated from structural MRI. Together, our results contribute to a growing literature illustrating the differential contributions of a hierarchical cortical gradient representing localized and distributed cognitive processes.

Keywords: Cortical gradients; Cortical hierarchy; Functional connectivity; Myelin mapping; Task activations; Timescales; fMRI.

Fig. 1.. Data analysis schematic for assessing localized versus distributed processes during multiple task states.

a) Characterizing localized functionality by estimating regional task activation changes. To identify the task activation change of each brain region, we estimated the task activation magnitude of each brain region across 24 task conditions. Localized processes were operationally defined as the magnitude of task activation change relative to baseline (see Methods). b) Characterizing a region’s distributed functionality by estimating the strength of its global FC strength relative to its resting-state FC. To identify the global FC change of a region, we compared the task-state global FC and compared it relative to its resting-state global FC. Thus, a region’s reduced global FC during task states indicated that it reflected more localized processes. c,d) To more simply compare localized and distributed processes across cortical areas, we mapped the previously described functional network assignments (Ji et al., 2019) into transmodal and unimodal networks. Unimodal networks included: primary and secondary visual networks, auditory network, and somatomotor network. Transmodal networks included all other networks.

Fig. 2.. Dissociating localized versus distributed processes across the cortical hierarchy by estimating regional task activations and FC changes.

a) The resting-state principal macroscale gradient (PG1) from Margulies and colleagues, which provides a spatial framework to characterize unimodal to transmodal activity (Margulies et al., 2016). b) Task activation magnitudes relative to baseline (absolute t-values), averaged across 24 task conditions. c) Averaged task FC changes for each region relative to resting-state FC. d) The task activation magnitudes averaged within transmodal and unimodal regions. Unimodal regions had significantly higher task activation magnitudes across multiple tasks relative to transmodal regions. Boxplots indicate the interquartile range of the distribution, dotted black line indicates the mean, grey line indicates the median, and the distribution is visualized using a swarm plot. e) Averaged task FC changes (relative to resting state) for transmodal and unimodal regions. In contrast to the task activation magnitude, unimodal regions significantly decreased their FC relative to transmodal regions. f) We also reproduced a result from our previous study (Ito et al., 2019), demonstrating that regions with higher task-evoked activations decreased their FC more during task states. g) Task activation magnitudes were positively correlated with PG1. h) Task FC changes were negatively correlated with the PG1. All p-values (for correlation analyses) were estimated using a spatial autocorrelation-preserving permutation test to generate random surrogate brain maps (Burt et al., 2020). (*** = p < 0.001, ** = p < 0.01, * = p < 0.05).

Fig. 3.. Hierarchy of intrinsic timescales estimated during resting-state fMRI explains regional differences in task activations and FC.

a) The intrinsic timescale for each cortical region. We estimated the intrinsic timescale of each region by fitting a 3-parameter exponential decay function to the autocorrelation function obtained during resting-state fMRI (Murray et al., 2014). b) The estimated exponential decay functions for two example regions with fast (blue) and slow (red) timescales. Fits were estimated for each subject separately. Error bars denote the 95% confidence interval (across subjects). c) The intrinsic timescale (i.e., the rate of decay) was significantly greater for transmodal regions relative to unimodal regions. Boxplots indicate the interquartile range of the distribution, dotted black line indicates the mean, grey line indicates the median, and the distribution is visualized using a swarm plot. d) Across the cortical hierarchy, the intrinsic timescale was negatively correlated with task activation magnitude across multiple tasks, consistent with the notion that regions with fast timescales respond in a stimulus/task-locked manner. e) In contrast, the intrinsic timescale was positively correlated with the task-state FC change, consistent with the hypothesis that regions with slow timescales have a larger temporal receptive field and can integrate information from lower-order cortical areas. All p-values (for correlation analyses) were estimated using a spatial autocorrelation-preserving permutation test to generate random surrogate brain maps. (*** = p < 0.001, ** = p < 0.01, * = p < 0.05).

Fig. 4.. Intrinsic and task-state differences in hierarchical cortical organization are associated to local myelin density.

a) Cortical myelin content within each parcel estimated from a T1w/T2w contrast map (Burt et al., 2018). b) Across cortical regions, myelin content and the intrinsic timescale are negatively related, suggesting that lower-order brain regions operate at faster intrinsic timescales. c) Across cortical regions, myelin content is positively correlated with the magnitude of task-evoked activations, suggesting that lower-order brain regions tend to have higher task-evoked activations (consistent with stimulus-locked activity). d) Across cortical regions, myelin content is correlated with task-state FC decreases, suggesting that higher-order brain regions change their task FC strength less (consistent with information integration with other brain regions). All p-values were estimated using a spatial autocorrelation-preserving permutation test to generate random surrogate brain maps. (*** = p < 0.001, ** = p < 0.01, * = p < 0.05).

Fig. 5.. Summary of positive and negative associations between the resting-state principal gradient, task activations, task FC change, intrinsic timescales, and myelin content using the standard task GLM and the peak (block) activation approaches (Figures for the replication cohort are in Supplementary Fig. 1.).

a) The standard task GLM approach for the exploratory cohort. We use standard task GLM modeling to estimate activation coefficients for each brain region, and FIR task modeling to remove the mean task-evoked response prior to computing task FC. Note that all measures reported in this study were strongly associated with PG1, which was hypothesized to reflect hierarchical organization in the brain (Margulies et al., 2016). b) The peak activation approach for the exploratory cohort. We estimate the peak activation magnitude at each block (across all blocks) without task regression. Task activations are estimated by averaging peak magnitudes across all blocks for each brain region. Task FC estimates are obtained by correlating block-to-block variance (using peak magnitudes) between all pairs of brain regions. Positive and negative association strengths are typically stronger using the peak activation approach. All correlations were found to be statistically significant using an FDR-corrected p-value of p < 0.01. All p-values were estimated using a spatial autocorrelation-preserving permutation test to generate random surrogate brain maps.

Fig. 6.. Better prediction of task-evoked activations for transmodal regions than unimodal regions via activity flow mapping.

a) The activity flow mapping algorithm, which was originally derived from connectionist principles (Cole et al., 2016; Ito et al., 2020). Briefly, the task-evoked activation of a brain region j can be predicted by summing the task-evoked activations of all other brain regions weighted by their FC weights with region j. A core assumption of this algorithm is that the task-evoked activity of region j is generated from a distributed process, rather than from a local (or internal) process. b) To evaluate whether some brain regions are better predicted via activity flow mapping, we can characterize the mean absolute error of the activity flow predictions (i.e., ‘activity flow MAE’) across task conditions. This evaluates the mean absolute error of the activity flow mapping algorithm for every brain region. c) We find that transmodal regions have significantly lower activity flow MAE relative to unimodal regions. Boxplots indicate the interquartile range of the distribution, dotted black line indicates the mean, grey line indicates the median, and the distribution is visualized using a swarm plot. d) We found a negative association between the intrinsic timescale of regions and activity flow MAE (i.e., slower intrinsic timescales have better activity flow predictions, consistent with the view that a wider temporal receptive field facilitates better information integration from different regions). e) We found a positive association between the myelin content of regions and activity flow MAE. This is consistent with the notion that lower-order cortical regions process information more locally (i.e., reflecting internally generated activity). f,g) Across cortical regions, activity flow MAE was positively/negatively associated with task activation magnitude/FC change. (*** = p < 0.0001, ** = p < 0.01, * = p < 0.05).