Figure 4. Unique contribution of topography versus coupling.

(A) Task basis maps are extracted from group-averaged task contrasts (n=86, 47 unique) using ICA to ensure correspondence of basis maps across subjects. These maps represent the basic building blocks of any activation pattern, and subject task basis maps (obtained by applying the ICA weights to subject task contrast maps) are not influenced by misalignment problems. (B) Dual regression against rfMRI data is performed using either the (potentially misaligned) group task basis maps or the (functionally localised) subject task basis maps. (C) CCA results of group-task-based rfMRI maps and network matrices and of subject-task-based rfMRI maps and netmats. The results show rUV (i.e., the correlation between the first U and V obtained from the CCA analysis describing the strength of association between the rfMRI and behavioural measures). The null line (i.e., p=0.05 based on permutation testing) is shown as a dotted line at 0.68; results below this line do not reach significance. The blue bars show the main CCA results using the complete data, and the red bars show partial CCA results computed after regressing out any variance that can be explained by network matrices from the spatial maps and vice versa prior to running the CCA. The results show a general decrease in rUV for all measures when comparing partial to full CCA results. The strongest partial CCA result (red bars on right) are found when using rfMRI spatial maps, and the associated netmats showed the weakest results (“§”). However, the partial CCA results for the spatial maps (i.e., the red bars on the right) still reach significance. All of the partial CCAs also showed lower rU-Uica compared to the full CCAs (not shown here).

Figure 4—figure supplement 1. Similarities between cross-subject variations estimated from different rfMRI measures.

Subject-by-subject correlation matrices are estimated (A), and vectorised (B; one subject correlation matrix being estimated for each measure type). The first column of the similarities (C; highlighted) shows the relationship (full correlation) between the ICA network matrix and various other measures, such as PFM spatial maps and amplitudes, and ICA spatial maps. These results show that the ICA network matrix is closely related to PFM spatial maps. The first row of the similarities (C; highlighted) shows the same relationship after taking into account all the other elements (i.e. the partial correlation between different measures). This reveals that PFM spatial maps are strongly linked to the ICA network matrix, even after accounting for any variance that can be explained by ICA spatial maps and PFM amplitudes. Similar results are obtained for ICA 200 and 25 dimensionality and for partial and full network matrices (D). These findings are consistent with the simulation results in Table 1, showing that estimated network matrices and spatial topography to a large extent overlap in terms of the interesting cross-subject variability they represent. Additionally, the results show that while dual regression ICA spatial maps are able to capture some of the subject spatial variability, subject maps estimated by PROFUMO capture considerably more spatial variability over and above the dual regression maps.