Article Figures & Data
Figures
- Figure 1.
CCA of spectrally resolved MEG data and a large set of behavioral and other subject measures results in one significant mode (r = 0.94, p < 10−4, corrected for multiple comparisons by permutation testing). For the used parcellation, see Extended Data Figure 1-1, with anatomic labels found in Extended Data Figure 1-2. A visualization of the analysis approach is shown in Extended Data Figure 1-3. A, The canonical mode arranges behavioral variables and subject measures on a positive-negative axis, similar to what has been previously reported for hemodynamic measures of brain connectivity (Smith et al., 2015; for comparison, see also Extended Data Fig. 1-4, showing the fMRI-based brain-behavior mode for the set of subjects used here). At maximum positive correlation, there are mostly subject measures indexing CP such as reading skills and vocabulary knowledge, while on the negative end of the spectrum are subject measures like somatic problems, and tobacco consumption (thresholded at a correlation coefficient of |r| > 0.25). B, Correlation of CCA-derived subject measure scores and connectivity scores for the first canonical variates identified, i.e., the first canonical mode. In color the behavioral score for the working memory test is shown per subject. For a detailed overview of how certain methodological choices impact results see Extended Data Figure 1-5. The first CCA mode as visualized here also explains a significant amount of variance in the data (Extended Data Fig. 1-6). Penn matr. = Penn matrix test; ASR = Achenbach Adult Self Report; DSM = Diagnostic and Statistical Manual; MMSE = mini mental state examination; SSAGA = Semi-Structured Assessment for the Genetics of Alcoholism; unadj = unadjusted for age effects. Please note that some secondary measures (e.g., similar metrics for tobacco consumption) are left out to avoid redundancy.
- Figure 2.
A–E, MCC (i.e., edges whose connectivity on a subject-by-subject level covaries with the mode, and thus behavior, in short MCC) for each frequency band, from δ-band, θ-band, α-band, and β-band to γ-band. MCC values of frequency bands are color coded (two colors per band, positive and negative, encoding the correlation coefficients, in a range from –0.5 to.5 to the observed canonical mode). Edges rendered on the brain templates are thresholded at the 0.5th bottom and 99.5th top percentile (of the permutation-based null distribution) for visualization. Each of the bands shows a preference for either positive or negative relationship to the mode but not a mixture of both. This is also visible in the histograms in the bottom row that depict the distributions of all (i.e., unthresholded) correlation coefficients, with comparison to null distributions generated by a permutation test (n = 10,000, in gray). For additional control analyses demonstrating the robustness of results, see also Extended Data Figure 2-1.
- Figure 3.
A, Connectogram-style visualization (Irimia et al., 2012) of the MCC edges presented in Figure 2. This figure shows the top 50 MCC edges (i.e., the top 50 positive and negative edges, respectively), across all five frequency bands (globally tresholded across all bands pooled). Nodes are represented on the outside of the ring, in an approximately anatomically-faithful anterior-posterior and left-right arrangement. Anatomically labeled parcel identities are listed in Extended Data Figure 1-2. Each parcel’s grayscale value indicates its average relationship to the CCA mode, i.e., its average accumulated MCC values across frequency bands, white indicates a positive relationship to the mode, and black indicates a negative relationship (averaged over bands and all edges, i.e., connections of that parcel). The five-band histograms on the outer ring indicate the band-specific accumulated MCC values for each node (same color coding as for the top edges in Fig. 1A and the maps in Fig. 1B, histograms pointing outwards are positive, inward histograms indicate negative accumulated MCC values). B, Maps of accumulated MCC values for all frequency bands. These represent the average MCC values, over all connections, for each parcel. Thus, these accumulated maps represent the overall involvement of each parcel in predicting behavior and provide complementary information to the MCC edges by showing the presence of connectivity foci, i.e., nodes of particular importance to the MCC network, acting as sinks or hubs. Thresholded at the 10th and 90th percentiles for negative and positive accumulated MCC values, respectively. NB, no suprathreshold accumulated MCC values are observed in the γ-band. L = left; R = right; pos. = positive; neg. = negative; SUBCORT = subcortical.
Extended Data
Extended Data Figure 1-1
Visualization of the used parcellation (n = 100, lateral view). Parcel identities can be found in Extended Data Figure 1-2, with anatomical labels. Download Figure 1-1, TIF file.
Extended Data Figure 1-2
Table showing the parcel identities with anatomical labeling derived from the BNA atlas (Jiang, 2013). Mention of several areas means the parcel has been created by merging these originally separate BNA parcels into a single parcel. m = medial; l = lateral; r = rostral; a = anterior; p = posterior; v = ventral; c = caudal; i = inferior; o = orbital; p = pregenual; ag = agranular; rv = rostroventral; sg = subgenual; cv = caudoventral; cl = caudolateral; dg = dorsal granular; lv = lateroventral; dm = dorsomedial; vm = ventromedial; tla = tongue and larynx area; ms = medial superior; iv = intermediate ventral; rd = rostrodorsal; cd = caudodorsal; rp = rostroposterior; op = opercular; vla/d = ventral a/dysgranular; da = dorsal agranular; ms = medial superior; ulhf = upper limb, head, and face region; op = opercular; pg= pregenual; ms = medial superior; ll = lower limb region; dld = dorsal dysgranular; iv = intermediate ventral; pc = postcentral; ip = intraparietal; vld = ventraldysgranular; vlg = ventral granular; dla = dorsal agranular; cvl = caudal ventral lateral; ta = temporal agranular; lp = lateral posterior. Download Figure 1-2, DOCX file.
Extended Data Figure 1-3
Illustration of analysis approach. FC as defined by envelope correlations within five conventional frequency bands are extracted from n = 89 subjects. For each subject, these spectrally resolved connectivity features are first concatenated (A, forming matrix FC1) and subjected to group PCA, retaining the first 22 principal components per subject (resulting in matrix FC2). The same approach is used for reducing dimensionality of the subject measures of all subjects (concatenated vector B1), retaining the first 22 principal components (resulting matrix B2). These two matrices (FC2 and B2) are then subject to CCA, which identifies within each matrix the optimal linear weighting to maximize correlation of features between the two sets of variables (i.e., brain vs behavior), transforming matrices FC2 and B2 into FC3 and B3. There, each row represents one mode where the newly formed (i.e., linearly recombined) connectivity and behavior canonical variates correlate most strongly (first mode here is indicated by thin black line). Download Figure 1-3, TIF file.
Extended Data Figure 1-4
First CCA mode using fMRI-derived connectivity (n = 89). A, Behavioral ranking (thresholded at |rho| > 0.25) and brain-behavior mode visualization across subjects. B, Correlation of CCA-derived subject measure scores and connectivity scores for the first canonical variates identified, i.e., the first canonical mode. In color, the behavioral score for the top-ranking behavioral variable in this mode, here line orientation discrimination, is shown per subject. Download Figure 1-4, EPS file.
Extended Data Figure 1-5
Comparison of different CCA analyses. A, Correlation of first canonical variates (first mode) of connectivity versus behavior (y-axis show level of significance of correlation in negative logarithmic scale). B, Explained variance (in percent) of the behavioral variables in the first mode, same connectivity models as in A. In the order of appearance: (1) original analysis (five frequency bands, 100 parcels); (2) 20-bin resolved MEG connectivity data; (3) fMRI connectivity used instead of MEG connectivity; (4) spectrally unresolved MEG broadband connectivity (1–40 Hz) used; (5) 50 parcel MEG connectivity data used; (6) as original analysis, without source leakage correction performed. See also Materials and Methods, Validation of CCA results (analyses numbered 4–8). Download Figure 1-5, EPS file.
Extended Data Figure 1-6
The amount of variance (%) explained by the first canonical variate (i.e., first mode) in the full set of behavioral variables. Permutation test n = 1000. Green and red lines indicate differently thresholded top and bottom percentiles of permutation based null model. The first mode explains significantly more variance than permutation-based (pseudo-)results. Download Figure 1-6, EPS file.
Extended Data Figure 2-1
Stability of MCC results for different analyses A, Exemplarily for the α-band, MCCs for the original (main) analysis, and three control analyses. First panel, Results from original analysis, visualizing the α-band MCCs. Second panel, CCA has now been performed separately for each frequency band, here the MCC result for the α-band model is shown. Third panel, For this analysis, data were split into a training and test set, where the training was carried out on the first two resting-state sessions and testing was carried out on the third session. α-Band MCCs are shown for the CCA mode that was learned in the training set and applied to the test set. Fourth panel, In this analysis, data were split into training and test set by a leave-one-subject-out approach. α-Band MCCs are shown for the CCA mode as established for the left-out subjects (and their predicted canonical variates; for details, see Materials and Methods). All α-band MCCs show comparable patterns with the majority of suprathreshold patterns being negative, posterior MCC patterns related to the identified CCA mode. B, Summary of the distribution of suprathreshold MCCs across all frequency bands. Complementary to first row, this shows all frequency bands from δ to γ. Almost all analyses behave the same way, apart from the β-band in the frequency band separate CCA analyses (where β-band CCA did not yield any significant mode). C, Similarity of MCC patterns to the original main analyses (first panel and Fig. 2). Here, correlation was performed over the whole 4950-element MCC vector in each frequency band. High correlations indicate good correspondence with original MCC patterns from the main analysis (as visualized in Fig. 2). See also Materials and Methods, Validation of CCA results (analyses 1–3). Download Figure 2-1, EPS file.
In this issue
