Amplitudes of resting-state functional networks - investigation into their correlates and biophysical properties

Abstract

Resting-state fMRI studies have shown that multiple functional networks, which consist of distributed brain regions that share synchronised spontaneous activity, co-exist in the brain. As these resting-state networks (RSNs) have been thought to reflect the brain's intrinsic functional organization, intersubject variability in the networks' spontaneous fluctuations may be associated with individuals' clinical, physiological, cognitive, and genetic traits. Here, we investigated resting-state fMRI data along with extensive clinical, lifestyle, and genetic data collected from 37,842 UK Biobank participants, with the object of elucidating intersubject variability in the fluctuation amplitudes of RSNs. Functional properties of the RSN amplitudes were first examined by analyzing correlations with the well-established between-network functional connectivity. It was found that a network amplitude is highly correlated with the mean strength of the functional connectivity that the network has with the other networks. Intersubject clustering analysis showed the amplitudes are most strongly correlated with age, cardiovascular factors, body composition, blood cell counts, lung function, and sex, with some differences in the correlation strengths between sensory and cognitive RSNs. Genome-wide association studies (GWASs) of RSN amplitudes identified several significant genetic variants reported in previous GWASs for their implications in sleep duration. We provide insight into key factors determining RSN amplitudes and demonstrate that intersubject variability of the amplitudes primarily originates from differences in temporal synchrony between functionally linked brain regions, rather than differences in the magnitude of raw voxelwise BOLD signal changes. This finding additionally revealed intriguing differences between sensory and cognitive RSNs with respect to sex effects on temporal synchrony and provided evidence suggesting that synchronous coactivations of functionally linked brain regions, and magnitudes of BOLD signal changes, may be related to different genetic mechanisms. These results underscore that intersubject variability of the amplitudes in health and disease need to be interpreted largely as a measure of the sum of within-network temporal synchrony and amplitudes of BOLD signals, with a dominant contribution from the former.

Keywords: Dual regression; GWAS; Network amplitude; Resting-state fMRI; Temporal synchrony; UK Biobank.

Fig. 1

(A) Schematics of temporal concatenation group-ICA. Group-ICA decomposes the temporally concatenated fMRI data from participants into a set of independent spatial maps and a set of corresponding timeseries. The ICA dimension (K) was set to 25 in this study. The spatial maps from group ICA are used in the first stage of dual regression to derive subject-specific network timeseries, which are then subsequently used in the second stage of the dual regression to obtain subject-specific spatial maps. (B) Traditional dual regression stage 1. Network timeseries (D) are obtained using the original fMRI data, where the BOLD signal in each voxel ($x_i$) fluctuates with standard deviation ${\sigma }_i$. Network amplitudes are defined as the standard deviations of the estimated network timeseries. $E$ denotes residuals. (C) New, distinct network timeseries ($\widehat{D}$) are obtained using temporally normalized fMRI data, where the standard deviation of the BOLD signal is set to 1 for every voxel. Temporal synchrony of each network is defined as the standard deviation of each of these new network timeseries.

Fig. 2

Effects of temporal synchrony and amplitudes of BOLD signals on network amplitude computed in dual regression Stage 1. The timeseries ${\mathit{d}}_k$ of network $k$ corresponds to row $k$ of ${\left({\mathit{G}}^T\mathit{G}\right)}^{-1}{\mathit{G}}^T\mathit{X}$. ${\left({\mathit{G}}^T\mathit{G}\right)}^{-1}$ does not affect intersubject differences in the network amplitude, and therefore is greyed out in the figure. A simplified example of ${\mathit{g}}_k^T\mathit{X}$ is presented below using two voxels. For illustration purpose, in the toy example, $x_1\left(t\right)$ and $x_2\left(t\right)$ are described as voxel timeseries of the same frequency that are perfectly aligned with a phase difference of 0. (A) Effect of temporal synchrony of BOLD signals. The network amplitude decreases due to the phase differences ($\theta$) between the two timeseries denoted in green. The network amplitude is also small when the two voxels are expected to be anticorrelated based on the ICA weights (denoted in green) but their timeseries are positively correlated. (B) Effect of amplitudes of BOLD signals. As the BOLD signal amplitudes become half (denoted in green) – assuming the synchrony is unchanged – the network amplitude decreases by half.

Fig. 3

(A) Within-subject correlation. For each participant, the FC matrix is computed from network timeseries using partial correlation. Absolute values of the FC matrix are taken and averaged (across all other networks) for each network to compute the absolute FC. Similarly, positive and negative FC are computed using positive and negative elements in $C$, respectively. A Pearson correlation coefficient, $r$, is computed between network amplitudes and each of the absolute, positive, negative FC. The violin plots on the right-hand side shows the distributions of the correlation coefficients computed for 37,842 participants. (B) Correlation across participants. Top: illustrative scatter plots for network 1 (default mode network) amplitudes and each of the absolute, positive, and negative FC of network 1 computed from 37,842 participants are shown along with their correlation coefficients. Bottom: the bar graphs show the correlations obtained for all 21 networks.

Fig. 4

Ward's clustering shows a clear separation of the sensory networks (green) and cognitive networks (purple). The order of networks and dendrogram are presented on the top along with the correlation matrix ($\in {\mathbb{R}}^{21\times 21}$) computed from the network amplitude matrix ($\in {\mathbb{R}}^{\text{participant}\times 21}$). Each network is labeled with the conventional functional network name on the left (Beckmann et al., 2005; Damoiseaux et al., 2006; Lee et al., 2013; Veer et al., 2010).

Fig. 5

Manhattan plots showing the associations between 4,897 non-imaging variables and sensory/cognitive amplitude. The association strengths are presented as the Pearson correlation P values that have been converted to $-{\log }_{10}P$ (note: subscripts s and c indicate sensory and cognitive, respectively). The horizontal lines indicate $-{\log }_{10}P=20$. The non-imaging variables are categorized into 15 groups, which are denoted with different colours for visualization. (A) Associations between sensory amplitude and non-imaging variables. (B) Associations between cognitive amplitude and non-imaging variables. (C) Differences in the correlation $P$ values between sensory and cognitive amplitudes. Variables with positive $-{\log }_{10}\left(P_s/P_c\right)$ have stronger associations with the sensory amplitude than cognitive amplitude.

Fig. 6

(A) Correlations between network amplitude and temporal synchrony across participants. (B) Correlations between network amplitude and BOLD amplitude across participants. The networks are color-coded such that green and purple colours represent sensory and cognitive networks, respectively, based on the clustering analysis result in Fig. 4. (C) Full (below diagonal) and partial (above diagonal) correlations of the network amplitudes, temporal synchrony, and BOLD amplitudes. The correlation between network amplitude and temporal synchrony is high for every network (diagonal elements in the red box) even after removing all other information (diagonal elements in the blue box). The networks are presented in the same order as in panels (A) and (B).

Fig. 7

(A) Within-subject (across networks) correlations (left) and correlations across participants (right) between temporal synchrony and (absolute/positive/negative) FC. (B) Within-subject correlations (left) and correlations across participants (right) between BOLD amplitudes and (absolute/positive/negative) FC. Detailed descriptions of the plots and procedures to compute the correlation coefficients are provided in Fig. 3. Most of the $P$ values of the correlations in (A) and (B) survived the Bonferroni correction ($P$_corr < 0.001). The few that did not survive correction are denoted as ns.

Fig. 8

Regression coefficients (beta) estimated from the multiple linear regressions to analyze the relationship between a network amplitude/temporal synchrony/BOLD amplitude and non-imaging variables. In total, 63 multiple linear regression analyses were conducted with the same non-imaging variables (systolic blood pressure, body fat %, haemoglobin concentration, sleep duration, age, sex, age $\times$ sex, age², and age²$\times$ sex; variables were normalized except for sex coded as 0 and 1 for female and male) as predictors. Each of the normalized network amplitudes, temporal synchrony, and BOLD amplitudes of the 21 networks was used as the dependent variable. The estimated regression coefficients from the multiple linear regression analyses are arranged according to amplitude types (columns) and predictors (rows). For brevity, only the beta coefficients for (A) age, (B) sex, (C) systolic blood pressure, (D) body fat %, and (E) sleep duration are presented. The networks are colour coded such that green and purple colours represent sensory and cognitive networks, respectively, based on the clustering analysis result in Fig. 4. Bonferroni-corrected $P$ values for the beta coefficients are indicated: *: $P$ < 0.05, **: $P$ < 0.01, ***: $P$ < 0.001.