Background

Brain networks and subnetworks coalesce and dissolve over time, state transitions during resting-state reflect the fluctuation of our brain homeostasis, which might be modulated by various stimulations/brain dysfunctions. Studies in depression have reported diminished fluctuations in dynamic functional connectivity^1,2; it is still obscure, however, how the state interactions are manipulated by stress. In the current study, we examined and compared the state alternation frequencies in pre- and post-stress resting-state runs in two independent datasets, to explore the stress-induced changes of the brain homeostasis. We hypothesized that the brain would transit more frequently across different states under stress, and the post- versus pre-stress state alternations change would differentiate those depression vulnerable under both cohorts.

Methods

Subjects: Seventeen healthy subjects (age = 22.00 ± 0.68 years) in cohort/experiment 1 (Exp.1) and thirty-nine healthy subjects (age = 23.10 ± 2.70 years) in cohort/experiment 2 (Exp.2) were recruited through advertisement in two independent studies to complete similar scanning procedures. The 13-item Beck Depression Inventory (BDI) was used to assess depression symptoms ³. All subjects in Exp.2 were subdivided into low BDI group (BDI ≤ 7, n = 25, none to mild depression) and high BDI group (8 ≤ BDI ≤ 17, n = 14, moderate to severe depression); Subjects in Exp.1 were not divided given the smaller sample size.
Data acquisition: In both cohorts/experiments, a mathematic calculation task (examples in Fig. 1A) and a counting breath task were employed to induce a stressful and relaxing state, respectfully. A resting-state fMRI (6min and 15s in Exp.1 and 15min in Exp.2) was acquired before and after each task (Fig. 1A) with conventional fMRI parameters (TR = 2.5s, voxel size = 3.5 × 3.5 × 3.85 mm³ in Exp.1; TR = 2s, voxel size = 3.5 × 3.5 × 4 mm³ in Exp.2). Subjects’ stressful levels were evaluated pre- and post-tasks. More details regarding the experiments are found in our previous publications ^4,5.
Data analysis: The whole-brain co-activation pattern (CAP) analysis ⁶, which uses k-means clustering to classify time points into multiple groups based on spatial correlations of their instantaneous co-activation patterns, was implemented to define/divide brain states. Specifically, whole-brain time-series from each resting-state run were demeaned and variance-normalized after standard preprocessing, and the mean time-courses for 246 ROIs from a well-established template were extracted ⁷. For each task session (pre- and post-task resting-state runs in Exp.1 or Exp.2), all subjects’ data were concatenated into a 2D matrix [(subjects × time-points × 2 runs), ROIs], and entered into the k-means classification (k = 2,3 … 10 for Exp.1 considering the short scanning period; k = 2, 3 … 15 for Exp.2) with the spatial correlation of the CAPs as the distance measure ⁸. All classifications were repeated 100 times with randomized initial values to find the optimal clustering. CAPs were derived by calculating the centroids of each subgroup. Alternation frequency (AF) was calculated through the transition rate of CAPs, which is the number of times that the brain state shifts from the current CAP to another CAP, divided by the number of time-points minus one ⁹ (Fig. 1B).

Results

Individuals’ stressful levels increased after the math task in both cohorts (Fig. 2), indicating successfully inducing a stressful state.
We here choose k = 2, 7, 10 for Exp.1 and k = 2, 10, 15 for Exp.2 to show the CAPs and AFs under stress conditions (Fig. 3 and Fig. 4); the results of the relaxation task are not shown considering no significant AF changes were observed post-relaxation task.
Some CAPs obtained from the two stress experiments show a high consistency with well-known functional networks regardless of the choice of k, others cover more than one network at a time. When k = 2 (Fig. 3A for Exp.1 and Fig. 4A for Exp.2), CAP1 involves sensory-motor areas and subcortical regions, CAP2 locates in the default mode network (DMN). When k = 10 (Fig. 3C for Exp.1 and Fig. 4B for Exp.2), CAP3 of Exp.1 and CAP2 of Exp.2 cover the DMN areas, CAP4 of Exp.1 and CAP1 of Exp.2 resemble the executive control network (ECN), and CAP6 and CAP7 of the Exp.1 and CAP9 of the Exp.2 include the sensory-motor regions.
The AF significantly increased post-stress tasks for 8 out of 9 k values in Exp.1 and 12 out of 14 k values in Exp. 2 (examples shown in Fig. 3D for Exp.1 and Fig. 4D for Exp.2). Interestingly, the AF change was positively correlated with the BDI scores in Exp.1 except k = 8 (k = 2 in Fig. 3E), and the AF increase post-stress was significantly greater in the high BDI group than the low BDI group in Exp.2 except k = 12 and k = 13 (k = 2, 10, 15 in Fig. 4E).

Conclusions

In the current study, we examined the stress-induced state alternation changes, consistent findings were reported in two independent cohorts. As we hypothesized, the stress task aggravated the state transition rate, indicating the stress-induced disturbance of the brain homeostasis. In addition, the transition frequency increase under stress could differentiate individuals with more severe depression symptoms, implying their incapacity to recover to the pre-stress state.

Acknowledgements

The authors thank Dr. Gary H. Glover for helpful comments on the abstract. This study was supported by the National Natural Science Foundation of China (31271080, 61571258), and the National Key Research and Development Program of China (2017YFA0205904, 2017YFC0108900).