Modulatory interactions between the default mode network and task positive networks in resting-state

Subjects

The resting-state fMRI data was derived from the Beijing_Zang dataset of the 1000 functional connectomes project (http://fcon_1000.projects.nitrc.org/) (Biswal et al., 2010). This dataset originally contained 198 subjects. The first 64 subjects without large head motions were included in the current analysis (40 female/24 male). The mean age of these subjects was 21.1 years (range from 18 to 26 years of age). This study involves analyzing public available dataset, which doesn’t need IRB approval. Further, we didn’t use any patient identification features in this study.

Scanning parameters

The MRI data were acquired using a SIEMENS Trio 3-Tesla scanner from Beijing Normal University. 230 resting-state functional data were acquired for each subject using TR of 2 s. The resolution of the fMRI images was 3.125 × 3.125 × 3 mm³ with 64 × 64 × 36 voxels. T1-weighted three-dimensional magnetization-prepared rapid gradient echo (MP-RAGE) images were acquired for all the subject using the following parameters: 128 slices, TR = 2530 ms, TE = 3.39 ms, slice thickness = 1.33 mm, flip angle = 7°, inversion time = 1100 ms, FOV = 256 × 256 mm².

Functional MRI data analysis

Spatial ICA

Spatial ICA was conducted to define networks for the PPI analysis using Group ICA of fMRI Toolbox (GIFT) (http://icatb.sourceforge.net/) (Calhoun et al., 2001). Twenty components were extracted. Among the 20 ICA maps (see Fig. S1), we identified the DMN and task positive network components by visually comparing the IC maps with previous studies (Biswal et al., 2010; Cole, Smith & Beckmann, 2010). Two components were identified as DMN, with one more anteriorly localized (Fig. 1A) and the other more posteriorly localized (Fig. 1B). We also identified four components that represented different task positive networks, i.e., the salience, dorsal attention, left executive, and right executive networks (Figs. 1C through 1F). Time series associated with these six components were obtained for each subject for following PPI analysis. To aid interpretations of the PPI results, simple correlations among the six networks were calculated for each subject. The correlation values were transformed into Fisher’s z, and statistical significances were tested across subjects using one sample t-test.

PPI analysis

Physiophysiological interaction analysis, along with its variant psychophysiological interaction, were first proposed by Friston and colleagues to characterize modulated connectivity by another region or a psychological manipulation (Friston et al., 1997). The present analysis focused on the modulation of connectivity by other regions or networks. Specifically, time series of two networks were used to define an interaction model using a linear regression framework. ${\displaystyle y={\beta }_{N1}\cdot x_{N1}+{\beta }_{N2}\cdot x_{N2}+{\beta }_{PPI}\cdot x_{N1}\cdot x_{N2}+\epsilon }$ where x_N1 and x_N2 represented the time series of two networks. Critically, we were interested in whether the interaction term of the two time series was correlated with the time series of a given voxel or region y (the effect of β_PPI). A positive interaction effect implies that the connectivity between the resultant region and one of the networks is positively modulated by the other network. While a negative interaction effect implies that the connectivity between the resultant region and one of the networks is negatively modulated by the other network. It should be noted that the PPI analysis is different from partial correlation analysis, which simply examines a linear relationship between two regions by controlling the activity of a third region (Marrelec et al., 2006). A partial correlation is similar to the effects of β_N1 and β_N2 in the current model where the activity of x_N2 or x_N1 is controlled, respectively, which cannot directly examine the interaction of the two variables.

In practice, the time series of the two networks were deconvolved with a hemodynamic response function (hrf), so that the PPI term was calculated in the neuronal level but not hemodynamic level (Gitelman et al., 2003). The deconvolution procedure can in principle minimize noises when calculating PPI terms (Gitelman et al., 2003), and has been shown to provide better statistical results in previous empirical analysis (Di & Biswal, 2013a).

Before PPI analysis, the time series of each network were preprocessed in the following steps. Six rigid-body motion parameters, the first principle component time series of white matter (WM) signal, and the first principle component time series of cerebrospinal fluid (CSF) signal were regressed out from the original time series by using linear regression model. The subject specific WM and CSF masks were derived from their own segmented WM and CSF images, with a threshold of 0.99 to make sure that GM voxels were excluded from the masks. Next, a high-pass filter of 0.01 Hz was applied on the time series to minimize low frequency scanner drift. The preprocessed time series of two networks were first deconvolved with the hrf using a simple empirical Bayes procedure, so that the resulting time course represented an approximation to neural activity (Gitelman et al., 2003). Next, the two neural time series were detrended and point multiplied, so that the resulting time series represented the interaction of neural activity between two networks. And lastly, the interaction time series was convolved with the hrf, resulting in an interaction variable in BOLD level. The PPI terms were calculated for each pair of the six networks.

Network-wise PPI analysis was first conducted to directly examine the relationships among networks, which is similar to von Kriegstein and Giraud (Von Kriegstein & Giraud, 2006). In the network-wise analysis, the dependent variable was the time series of a network rather than the time series of every voxel in the brain. In the PPI linear regression model, the main effects of the two networks, and the PPI effects between them were added as independent variables along with a constant regressor. After model estimation, cross-subject one-sample t-tests were performed on the beta values of PPI effects. The critical p value was set as p < 0.05 after Bonferroni correction (corresponding to a raw p value of 8.33 × 10⁻⁴ after correcting for totally 60 comparisons).

In addition, voxel-wise PPI analysis was also performed to identify regions across the whole brain that were associated with a PPI effect. PPI terms were calculated for each pair of the six networks, resulting in 15 PPI effects. Then separate PPI models were built for each subject using the general linear model (GLM) framework. The GLM model contained two regressors representing the main effects of two networks’ time series, one regressor representing the PPI effect, two regressors representing WM and CSF signals, and six regressors representing head motion effects. An implicit high pass filter of 1/100 Hz was used. For each PPI effect, a 2nd-level one sample t-test was conducted to make group-level inference. Simple t contrast of 1 or −1 was defined to reveal positive or negative PPI effects, respectively. The resulting clusters were first height thresholded at p < 0.001, and cluster-level false discovery rate (FDR) corrected at p < 0.0033 based on random field theory (Chumbley & Friston, 2009). The cluster-level p value was chosen to take into account the total 15 PPI effects. The resulting clusters were labeled according to their peak coordinates using Talairach Daemon (Lancaster et al., 2000), after taking into account the discrepancies between MNI space and Talairach space (Lancaster et al., 2007).

Simple correlations among networks

The mean correlations among the six networks are listed in Table 1. There was a positive correlation between the anterior and the posterior DMNs (M_{Fisher’s z} = 0.359). However, the correlations among the four task positive networks were mixed. The salience network revealed a positive correlation with the dorsal attention network (M_{Fisher’s z} = 0.333), but a negative correlation with the right executive network (M_{Fisher’s z} = −0.142). There was a positive correlation between the left and right executive networks (M_{Fisher’s z} = 0.427). The correlations between DMN components and task positive components were also mixed. The anterior DMN showed negative correlations with the salience network (M_{Fisher’s z} = −0.299) and the dorsal attention network (M_{Fisher’s z} = −0.530), while positive correlations with the left executive network (M_{Fisher’s z} = 0.184) and the right executive network (M_{Fisher’s z} = 0.247). Similarly, the posterior DMN revealed a negative correlation with the salience network (M_{Fisher’s z} = −0.251), while positive correlations with the left executive network (M_{Fisher’s z} = 0.320) and the right executive network (M_{Fisher’s z} = 0.188).

Network-wise PPI analysis

Significant network-wise modulatory interactions are illustrated in Fig. 2 (see also Table S1 for a full list of statistics). First, positive modulatory interactions were observed among the DMNs, the salience network, and the executive networks. Positive modulatory interactions were observed among the anterior DMN, salience, and right executive networks in all of the three ways. The time series of anterior DMN were correlated with the interactions of the salience and right executive networks (M_beta = 0.060; t = 4.77, p = 1.14 × 10⁻⁵). The time series of salience network were correlated with the interactions of the anterior DMN and right executive network (M_beta = 0.054; t = 4.09, p = 1.25 × 10⁻⁴). And the time series of the right executive network were correlated with the interactions of the anterior DMN and salience network (M_beta = 0.109; t = 8.27, p = 1.19 × 10⁻¹¹). The left executive time series were also correlated with the interactions of the anterior DMN and salience network (M_beta = 0.048; t = 3.67, p = 4.98 × 10⁻⁴). In addition, the time series of the right executive network were correlated with the interactions of the posterior DMN and salience network (M_beta = 0.045; t = 3.81, p = 3.17 × 10⁻⁴). Second, a negative modulatory interaction was also observed among the anterior DMN, posterior DMN and right executive network. The time series of the anterior DMN were negatively correlated with the interactions of the posterior DMN and right executive network (M_beta = −0.039; t = −0.404, p = 1.48 × 10⁻⁴). Lastly, positive modulatory interactions were also observed among task positive networks. The left executive network time series were correlated with the interactions of the salience and right executive network (M_beta = 0.046; t = 4.01, p = 1.66 × 10⁻⁴), and the right executive network times series were correlated with the interactions of the salience and left executive network (M_beta = 0.053; t = 3.94, p = 2.06 × 10⁻⁴). The right executive network time series were also correlated with the interaction effects of the dorsal attention and left executive networks (M_beta = 0.058; t = 4.31, p = 5.91 × 10⁻⁵).

Voxel-wise PPI analysis

The voxel-wise PPI results of the anterior DMN with the four task positive networks are illustrated in Fig. 3. A full list of regions that showed significant PPI effects in all the fifteen voxel-wise analyses can also been found in Table S2. The regions that revealed positive modulatory interactions with the anterior DMN and salience network mainly resembled a typical task positive network (Fig. 3A). These regions included the bilateral dorsolateral prefrontal cortex (mainly the middle and superior frontal gyrus, BA 9 and 10), bilateral parietal lobule (mainly the precuneus and inferior parietal lobule, BA 7 and 40), and left middle temporal gyrus (BA 37). Additionally, a small cluster in the posterior cingulate (BA 29) also revealed positive modulatory interactions. In contrast, several regions showed negative modulatory interactions, including the middle portion of cingulate gyrus (BA 24), bilateral putamen, and right insula (BA 13). For the modulatory interactions of the anterior DMN and dorsal attention network, positive effects were observed in the bilateral dorsolateral prefrontal cortex (mainly the middle and superior frontal gyrus, BA 9, and 47), and bilateral parietal lobule (mainly the inferior parietal lobule and supramarginal gyrus, BA 40) (Fig. 3B). No negative effects were observed. Only one region located in the right inferior parietal lobule (BA 40) revealed negative modulatory interactions with the anterior DMN and left executive network (Fig. 3C). No positive effects were observed. For the modulatory interactions of the anterior DMN and right executive network (Fig. 3D), positive effects were observed in the bilateral insula (BA 13), middle portion of cingulate gyrus (BA 24), right inferior parietal lobule (BA 40), and right middle frontal gyrus (BA 10). The bilateral insula and cingulate gyrus resembled the typical salience network. Negative effects were observed in the right middle frontal gyrus (BA 6).

The voxel-wise PPI results of the posterior DMN with the four task positive networks are shown in Fig. 4. Only positive effects were observed in the modulatory interactions of the posterior DMN and salience network, which were localized in the anterior portion of cingulate gyrus (BA 32), posterior portions of cingulate gyrus (BA 31), and left inferior parietal lobule (BA 40) (Fig. 4A). For the modulatory interactions of the posterior DMN and dorsal attention network, only positive effects were observed, which were localized in the right middle occipital gyrus (BA 19), left inferior and middle frontal gyrus (BA 44/47), right cerebellum, and left supramarginal gyrus (BA 40) (Fig. 4B). No significant modulatory interactions were found between the posterior DMN and left or right executive networks.

The PPI results of networks within the DMN and within task positive networks are shown in Fig. 5. Only negative effects were found for the modulatory interactions between the anterior and posterior DMNs, which were localized in the superior frontal gyrus (BA 6), left middle occipital gyrus (BA 19), and right precuneus (BA 7). For the modulatory interactions of the salience network and dorsal attention network (Fig. 5B), positive effects were observed in the medial frontal gyrus (BA 6), subcortical nuclei including right thalamus and left claustrum, and right postcentral gyrus (BA 2). Negative effects were observed in the left inferior frontal gyrus (BA 9). For the modulatory interactions of the salience network and left executive network (Fig. 5C), positive PPI effects were observed in the medial frontal gyrus (BA 8), left superior temporal gyrus (BA 39), and left middle frontal gyrus (BA 6). No negative effects were observed. For the modulatory interactions of the salience network and right executive network (Fig. 5D), positive effects were observed in the superior frontal gyrus (BA 8), right inferior frontal gyrus (BA 47), right superior temporal gyrus (BA 39), and right precentral gyrus (BA 9). No negative PPI effects were observed. For the modulatory interactions of the dorsal attention network and left executive network (Fig. 5E), positive effects were observed in the left inferior parietal lobule (BA 40) and left middle frontal gyrus (BA 6). No negative effects were observed. Only one cluster in the right precuneus (BA 39) revealed positive modulatory interactions with the dorsal attention network and right executive network (Fig. 5F). Lastly, for the modulatory interactions of the left and right executive networks (Fig. 5G), positive PPI effects were observed in the bilateral precuneus (BA 7). No negative effects were observed.