The Functional Relevance of Task-State Functional Connectivity

Abstract

Resting-state functional connectivity has provided substantial insight into intrinsic brain network organization, yet the functional importance of task-related change from that intrinsic network organization remains unclear. Indeed, such task-related changes are known to be small, suggesting they may have only minimal functional relevance. Alternatively, despite their small amplitude, these task-related changes may be essential for the ability of the human brain to adaptively alter its functionality via rapid changes in inter-regional relationships. We used activity flow mapping-an approach for building empirically derived network models-to quantify the functional importance of task-state functional connectivity (above and beyond resting-state functional connectivity) in shaping cognitive task activations in the (female and male) human brain. We found that task-state functional connectivity could be used to better predict independent fMRI activations across all 24 task conditions and all 360 cortical regions tested. Further, we found that prediction accuracy was strongly driven by individual-specific functional connectivity patterns, while functional connectivity patterns from other tasks (task-general functional connectivity) still improved predictions beyond resting-state functional connectivity. Additionally, since activity flow models simulate how task-evoked activations (which underlie behavior) are generated, these results may provide mechanistic insight into why prior studies found correlations between task-state functional connectivity and individual differences in behavior. These findings suggest that task-related changes to functional connections play an important role in dynamically reshaping brain network organization, shifting the flow of neural activity during task performance.SIGNIFICANCE STATEMENT Human cognition is highly dynamic, yet the functional network organization of the human brain is highly similar across rest and task states. We hypothesized that, despite this overall network stability, task-related changes from the intrinsic (resting-state) network organization of the brain strongly contribute to brain activations during cognitive task performance. Given that cognitive task activations emerge through network interactions, we leveraged connectivity-based models to predict independent cognitive task activations using resting-state versus task-state functional connectivity. This revealed that task-related changes in functional network organization increased prediction accuracy of cognitive task activations substantially, demonstrating their likely functional relevance for dynamic cognitive processes despite the small size of these task-related network changes.

Keywords: computational model; human connectome project; machine learning; network coding models; network neuroscience; task connectivity.

Figure 1.

Predicted boost to activity flow-based predictions using task-state FC. A, The propagation and activation rules used in neural network modeling provide a framework for modeling the flow of neural activity through networks. The propagation rule can be visualized via the arrows connecting distal nodes (e.g., region $i$) to a given target node, $h$, via a connection with strength $w_{hi}$. The activation rule can be visualized via the summing of incoming activity in the target node $h$, then passing through an activation function, $f$. The equation for computing the activity level $a$ in node $h\mathrm{i}\mathrm{s}\mathrm{a}\mathrm{s}\mathrm{f}\mathrm{o}\mathrm{l}\mathrm{l}\mathrm{o}\mathrm{w}\mathrm{s}$: $a_h=f(\Sigma a_i{w}_{hi}$); adapted from McClelland and Rogers, 2003. B, We recently developed the activity flow mapping framework, applying neural network modeling to empirical connectivity (and activity) estimates. We showed that activity flow mapping can predict independent (held-out) task activations using resting-state FC (Cole et al., 2016; Ito et al., 2017; figure is from the study by Cole et al., 2016). C, An illustration of simplified activity flow prediction of task activity in neural population Y based on task activity in neural population X, based on the resting-state FC between X and Y. D, The hypothesized boost in prediction accuracy by using the FC estimates from the same state as the task activity estimates. Note that the to-be-predicted task activity levels are carefully removed before estimating task-state FC (Cole et al., 2019) to avoid circularity.

Figure 2.

Neural populations are thought to interact via sigmoidal activation functions, helping explain widespread FC decreases from resting to task states. A, Spiking model simulation results from Ito et al. (2020a), showing sigmoidal relationship between inputs and outputs of neural populations. B, Activity increases and decreases from resting state in the spiking model (Ito et al., 2020a) resulted in decreased variance and correlations in simulated excitatory neurons. Empirical results in both spiking populations and fMRI show that variance and correlations decrease as activity levels increase or decrease from resting state (He, 2013; Ito et al., 2020a). C, There is substantial evidence that biological neural populations have a sigmoidal relationship, which could help explain well known neural variability quenching effects from rest to task (He, 2013) as well as reductions in FC from rest to task (Ito et al., 2020a). The relationship between activity levels across neural populations (i.e., FC) changes as a function of overall activity levels. Figure adapted from Ito et al. (2020a).

Figure 3.

The fMRI data-processing workflow for comparing task activation predictions based on task-state versus resting-state FC. Publicly available Human Connectome Project fMRI data (young adult dataset; N = 352) was split into separate discovery and replication datasets (N = 176 each). Circularity is carefully avoided in the predictions by (1) FIR regression to remove cross-block mean task responses before task-state FC estimation, and (2) removal of each to-be-predicted brain region from the set of predictors in the activity flow-mapping step. The goal of the primary analyses is to compare task activation predictions based on resting-state FC to predictions based on task-state FC.

Figure 4.

Resting-state and task-state FC are similar and mostly decrease from rest to task. A, Resting-state FC correlations averaged across N = 176 subjects (discovery set). Network names are listed on the right. B, Mean task-state FC correlations averaged across the 24 task conditions. Plotted on the same scale as in A. The similarity of the connectivity matrices in A and B (top triangle) was r = 0.90. C, Subtraction between data plotted in A and B, plotted on the same scale as in A. Decreased FC was apparent between all networks, with the exception of the frontoparietal network (which had widespread small increases) and between the cingulo-opercular and default-mode networks. D, A cortical surface plot of the networks listed in A (Ji et al., 2019; available at https://github.com/ColeLab/ColeAnticevicNetPartition).

Figure 5.

Task-state FC improves correlation-based activity flow models. A, Activity flow predictions using resting-state correlation FC across all nodes and conditions (r = 0.51 similarity to actual activations, similarity computed for each subject separately before averaging r values). Network colors correspond to those in Figure 4A. B, Actual activations (fMRI GLM β values) across all nodes and conditions. C, Task-state correlation FC-based activity flow predictions across all nodes and conditions (r = 0.66 similarity to actual activations). D, Activity flow prediction accuracies using resting-state correlation FC, calculated condition-wise separately for each node. FWE corrected for multiple comparisons using permutation testing. All nodes were statistically significant above 0. E, Task-state correlation FC-based activity flow prediction accuracies. FWE corrected for multiple comparisons using permutation testing. All nodes were statistically significant above 0. F, Task-state versus resting-state correlation FC-based activity flow prediction accuracy differences. FWE corrected for multiple comparisons using permutation testing; 93% of nodes were statistically significant above 0 (nonsignificant nodes are gray).

Figure 6.

Multiple-regression FC is similar across rest and task, and mostly decreases from rest to task. A, Resting-state multiple-regression FC averaged across N = 176 subjects (discovery set). Network names are listed on the right. B, Mean task-state multiple-regression FC averaged across the 24 task conditions. The similarity between the matrices shown in A and B was r = 0.94. C, Subtraction between data plotted in A and B. Only 2.6% of connections were statistically different from 0. D, A cortical surface plot of the networks listed in A (Ji et al., 2019; available at https://github.com/ColeLab/ColeAnticevicNetPartition).

Figure 7.

Task-state FC improves multiple-regression-based activity flow models. A, Resting-state multiple-regression FC-based activity flow predictions across all nodes and conditions (r = 0.46 similarity to actual activations, similarity computed for each subject separately before averaging r values). Network colors correspond to those in Figure 6A. B, Actual activations (fMRI GLM β values) across all nodes and conditions. C, Task-state multiple-regression FC-based activity flow predictions across all nodes and conditions (r = 0.76 similarity to actual activations). D, Resting-state multiple-regression FC-based activity flow prediction accuracies, calculated condition wise separately for each node. FWE corrected for multiple comparisons using permutation testing; 98% of nodes were statistically significant above 0. E, Task-state multiple-regression FC-based activity flow prediction accuracies. FWE corrected for multiple comparisons using permutation testing. All nodes were statistically significant above 0. F, Task-state versus resting-state multiple-regression FC-based activity flow prediction accuracy differences. FWE corrected for multiple comparisons using permutation testing. All nodes were statistically significant above 0.

Figure 8.

Visualizing three example task conditions across diverse cognitive domains. These values are all present in Figure 7—three of the columns from the matrices shown in Figure 7A–C—but are visualized here on cortical anatomy. R² values are based on the average-then-compare approach, quantifying the similarity of what is being visualized (i.e., group-level rather than individual-level similarity). The average-then-compare approach resulted in higher accuracies than the compare-then-average approach used in Table 2 (and elsewhere; see Table 1 for clarification on average-then-compare vs. compare-then-average approaches). A, Activations for the motor task, left-hand movement condition is shown. Note the right somatomotor activations corresponding to the left hand movement, which is most prominent for the actual and task-state FC-based predictions. Also note the scale difference for resting-state FC predictions, which contribute to R² (not Pearson correlation) values. B, Activations for the language task, story comprehension condition. Note the left-lateralized language network activation pattern in all three maps. C, Activations for the working memory task, two-back face stimuli condition. Note the larger improvement in group-level prediction accuracy in this task condition relative to the motor and language conditions. It will be important for future research to identify factors underlying these task condition differences.

Figure 9.

Intrinsic, individual, task-specific, and task run-specific factors contribute to shaping each individual's task activations. A, Distinct sources of FC were used for activity flow mapping with the same task activations (24 conditions, first run only). The prediction accuracies can be directly compared given that they are predicting the same set of activations (prediction accuracy computed across 360 regions and 24 conditions). B, Prediction accuracies computed as unscaled R² (not r²) values, which can be interpreted (once multiplied by 100) as the percentage of variance in the to-be-predicted activity pattern explained by the prediction. The red lines indicate baselining to quantify the relative effects in C. C, Same as B, but baselined (as indicated in B) to highlight the relative effects and normalized such that all values add up to 1. This facilitates the interpretation of the role of intrinsic, individual, task, and task run-specific factors in producing task activations. These normalized R² values can be interpreted as the proportion of the explained variance (in the IndivTaskRun results) contributed to by each factor. D, Detailed descriptions of the FC sources.

Figure 10.

Activity flow routes are better described by similar task conditions. A, Generalization of the task-state FC for each condition was tested across the whole-cortex task activation pattern for each condition, quantified by Pearson correlation r values. All cells of the matrix were significantly >0; p < 0.05 Bonferroni corrected for multiple comparisons. B, Identical to A, but with resting-state FC used instead of task-state FC. The number of time points going into each FC estimate was matched to the task-state FC estimates. Again, all cells of the matrix were significant; p < 0.05 Bonferroni corrected. Variation along the y-axis reflects the effect of the amount of time contributing to each FC estimate, while variation along the x-axis reflects the predictability of task activation patterns. C, Subtraction of the matrix in B from the matrix in A. D, Thresholded version of the matrix in C; p < 0.05 Bonferroni corrected. E, Similarity of task activation patterns across each pair of task conditions. Pearson correlations using whole-cortex activation patterns. F, Similarity of task-state FC patterns across each pair of task conditions. Pearson correlations across whole-cortex multiple-regression FC matrices.

Figure 11.

Activity flow prediction accuracy increases with the amount of data contributing to FC estimates. A, Prediction accuracies reported as Pearson correlations, collapsed across all 360 nodes and 24 task conditions. Task data were included from all 24 task conditions, similar to the prior task-general analyses but also including the to-be-predicted condition (to include more data). B, The same results reported as unscaled R² (not r²) values, which can be interpreted (once multiplied by 100) as the percentage of variance of the to-be-predicted activity pattern. These values range from negative infinity to 1, with the negative values quantifying how much worse the predictions are than predicting the mean of the data. C, The same results as in B, but with the 5 min results excluded to improve legibility of the other results.