Distinct brain networks for adaptive and stable task control in humans

Dosenbach et al. 10.1073/pnas.0704320104.

Supporting Information

Files in this Data Supplement:

SI Figure 5
SI Figure 6
SI Table 2
SI Figure 7
SI Figure 8
SI Figure 9
SI Figure 10
SI Text
SI Figure 11
SI Table 3

Fig. 5. Histogram of the distribution of all 741 r values (0.01 bins). The interregional r values for our set of 39 ROIs were roughly symmetrically distributed around zero, with the preponderance of values occurring in the range from 0.2 to 0.2. This distribution showed a local minimum at 0.2 and a long right-sided tail of r values >0.2.

Fig. 6. Graph formed by putative task control ROIs differs from random and regular graphs. (A) Graph created by thresholding the 39 ´39 regionwise correlation matrix at r ³ 0.2 contains 61 connections and consists of eight distinct components (separate sets of connected regions). (B) Sample random graph with the same number of nodes (ROIs) and edges (functional connections) as the task-control graph. (C) Regular graph with the same number of nodes and edges as the task-control graph.

Fig. 7. Clustering coefficients vs. characteristic path lengths. Plot of characteristic path length (Lp; x axis) against clustering coefficient (Cp; y axis) for five graph definition thresholds (r ³ 0.2, r ³ 0.175, r ³ 0.15, r ³ 0.125, and r ³ 0.1) comparing task-control graphs (red squares) with both regular (blue triangles) and random (black diamonds) graphs (100 iterations, error bars = SD).

Fig. 8. Task-control graph nonrandom even at lower thresholds. (A) 2D pseudoanatomical rendering of task-control graph created by thresholding correlation matrix at r ³ 0.125. All interregional correlations were significant at P < 5 ´ 10^-4 (two-tailed; Bonferroni corrected; t test). (B) 2D pseudoanatomical rendering of task-control graph created by thresholding correlation matrix at r ³0.1. All interregional correlations were significant at P < 0.005 (two-tailed; Bonferroni corrected; t test).

Fig. 9. Hierarchical cluster tree of all functional connections between putative task control regions. The x axis shows the distance (1-r) used for clustering with an average-linkage algorithm (UPGMA). In this tree, positively correlated clusters are close (distance <1) and negatively correlated clusters are far apart (distance >1). The bootstrap confidence intervals (CIs) are shown in white next to each branch point. The colors show eight clusters generated by cutting the tree at a distance of 0.9, such that all clusters have a CI of >90%.

Fig. 10. Clusters of putative task control regions. (A) The eight clusters displayed on an inflated rendering of the brain. ROIs are color-coded according to clusters (4). (B) The eight cluster colors superimposed on the component structure of the task control graph (r ³ 0.2). The clustering and graph analysis classifications of putative task control regions showed only two differences. The black ovals mark regions that were separate components, but clustered together. The red line demarcates the separation of a component into two clusters.

Fig. 11. Highly connected hubs. (A) Frontoparietal component at the r ³0.2 and r ³ 0.175 thresholds. The diameter of the yellow circles is scaled in accordance with the number of direct connections (unitary path length) to other ROIs (node degree). The numbers next to each ROI provide its node degree. Red ovals indicate putative network hubs. (B) Cinguloopercular component at the r ³0.2 and r ³ 0.175 thresholds. The diameter of the black circles is scaled according to the number of direct connections to other ROIs.

Table 2. Graph metrics at different correlation (r) thresholds

* For set of 100 random graphs, SDs are displayed (±).

Table 3. Graph metrics at different correlation (r) thresholds: Node degrees preserved for controls

* For set of 100 randomized/latticized graphs, SDs are displayed (±).

SI Text

Cluster Validation

Hierarchical clustering and validation measures were calculated using standard algorithms included in the Statistics and Bioinformatics Toolboxes available for Matlab 7.2. Similar to Cordes et al. we used (1 - r) (1) as the distance measure for clustering. Using (1 - r) as the distance measure ensured that strongly positively correlated ROIs would be clustered together, while negatively correlated regions would be separated. Like Salvador et al. (2) we chose the commonly used average linkage method (UPGMA) because it can handle both more chained and more clustered data (3).

Each cluster tree branch point was validated using bootstrapping (3). Bootstrap verification of the cluster tree was based on methods originally developed for phylogenetic and genechip data trees (3) and implemented in Matlab 7.2 similar to (www.mathworks.com/products/demos/bioinfo/phybootdistdemo/phybootdistdemo.html). One thousand bootstrap replicates were created by randomly sampling from the pool of 74 individual subject correlation matrices with replacement. The bootstrap replicates, which consisted of 74 correlation matrices, were converted to a single mean correlation matrix. All 1,000 mean bootstrap matrices were clustered, creating 1,000 bootstrap cluster trees. The bootstrap confidence intervals (CIs) were generated by counting the number of bootstrap trees that contained subtrees consistent of the same ROIs as a subtree in the original cluster tree. Bootstrap subtrees were counted as the same if they contained all of the same ROI as the original subtree, independent of the ordering of regions within the subtree. Using the bootstrap CIs as a guide for cutting the cluster tree generated eight clusters (SI Fig. 9), all of which had CIs >90%, demonstrating statistical robustness.

The eight clusters were strikingly similar to the eight components generated by the (r ³ 0.2) graph analysis (SI Fig. 10). The cinguloopercular, frontoparietal, cerebellar, occipital and temporoparietal junction components from the graph analysis mapped precisely onto clusters derived from the hierarchical tree. There were only two small differences between the component classification scheme generated by the graph analysis (r ³ 0.2) and the clusters (SI Fig. 10B). The eight-ROI default/fusiform component split into two separate clusters (default, fusiform), while the vmPFC and bilateral middle temporal cortex (not connected in the graph) were combined into a cluster.

Both graph-analysis and clustering showed the cerebellar ROIs to be functionally more closely connected to the frontoparietal component/cluster, and the cinguloopercular component/cluster to be functionally connected to visual regions in occipital cortex (SI Fig. 9).

To test for global distortions in the cluster tree, we calculated the correlation between the original correlation matrix and the matrix of distances that represent the tree (cophenetic correlation). A high cophenetic correlation (0.8284) between the original and the clustering matrix indicated that the cluster tree faithfully represents the original distance matrix (3).

Highly Connected Network Hubs

At r ³0.2 and r ³ 0.175, the two regions with the highest number of connections in the frontoparietal component, were the left and right IPS indicating that they might act as central network "hubs" (SI Fig. 11A). The left and right aI/fO were the most highly connected regions within the cinguloopercular component, identifying them as putative hubs (SI Fig. 11B).

Data Acquisition and Processing

All images were obtained using the same Siemens MAGNETOM Vision 1.5 Tesla scanner (Erlangen, Germany). T1-weighted sagittal MPRAGE structural image (TE = 4 ms, MR frame = 9.7 msec, TI = 300 msec, flip angle = 12°, 128 slices with 1.25 ´ 1 ´ 1-mm voxels) was obtained to compute atlas transformation. Functional imaging was performed using a blood oxygenation level-dependent (BOLD) contrast sensitive asymmetric spin-echo echo-planar sequence (volume TR = 2.5 sec, inplane resolution 3.75 ´ 3.75 mm, T2* evolution time = 50 msec, a = 90°). Whole-brain coverage was obtained with 16 contiguous, 8-mm-thick axial slices prescribed parallel to the anterior commissure-posterior commissure plane.

Functional Connectivity Preprocessing

Preprocessing for functional connectivity analyses steps included removal of the linear trend, temporal band-pass filtering (0.009 Hz < f < 0.08 Hz) and spatial smoothing (6 mm full width at half maximum). Several sources of variance unlikely to represent regionally specific neuronal activity were removed by regression from the timeseries at each voxel.

ROI Definition

The 39 ROIs used in the analyses had been previously shown to carry either sustained, start-cue, or error-related activity, or a combination of these signals. Five of the ROIs identified by the cross-studies analysis lacked contralateral homologues. However, previous studies have shown resting state functional connections to be highly symmetric (18, 19). To avoid artifactual laterality effects, we generated five ROIs in the opposite hemisphere by mirroring. Two ROIs in the inferior temporal cortex (coordinates: -33, 2, -29; 34, 2, -29) were excluded because of susceptibility inhomogeneity artifacts.

1. Cordes D, Haughton V, Carew JD, Arfanakis K, Maravilla K (2002) Magn Reson Imaging 20:305-317.

2. Salvador R, Suckling J, Coleman MR, Pickard JD, Menon D, Bullmore E (2005) Cereb Cortex 15:1332-1342.

3. Handl J, Knowles J, Kell DB (2005) Bioinformatics 21:3201-3212.

4. Van Essen DC, Dickson J, Harwell J, Hanlon D, Anderson CH, Drury HA (2001) J Am Med Inform Assoc 411:359-1378.

• Early Edition
• Archives
• Online Submission

• Feature Articles

  (Expanded research articles of exceptional breadth)

• Commentaries

  (Expert commentary on leading research)

• Letters

  (Online-only letters to the editor)

• Inaugural Articles

  (Articles by recently elected NAS members)

• PNAS Profiles

  (Biographical profiles of Academy members)

• Sustainability Science

  (Interactions of human and environmental systems)

• Collected Papers

  (Editorials, Perspectives, Reviews, Colloquia, etc.)

• Special Features

  (Solicited articles on exceptional research topics)

• Sackler Colloquia

  (Scientific reports held under NAS auspices)