Linked patterns of symptoms and cognitive covariation with functional brain controllability in major depressive disorder

Summary

Background

Controllability analysis is an approach developed for evaluating the ability of a brain region to modulate function in other regions, which has been found to be altered in major depressive disorder (MDD). Both depressive symptoms and cognitive impairments are prominent features of MDD, but the case–control differences of controllability between MDD and controls can not fully interpret the contribution of both clinical symptoms and cognition to brain controllability and linked patterns among them in MDD.

Methods

Sparse canonical correlation analysis was used to investigate the associations between resting-state functional brain controllability at the network level and clinical symptoms and cognition in 99 first-episode medication-naïve patients with MDD.

Findings

Average controllability was significantly correlated with clinical features. The average controllability of the dorsal attention network (DAN) and visual network had the highest correlations with clinical variables. Among clinical variables, depressed mood, suicidal ideation and behaviour, impaired work and activities, and gastrointestinal symptoms were significantly negatively associated with average controllability, and reduced cognitive flexibility was associated with reduced average controllability.

Interpretation

These findings highlight the importance of brain regions in modulating activity across brain networks in MDD, given their associations with symptoms and cognitive impairments observed in our study. Disrupted control of brain reconfiguration of DAN and visual network during their state transitions may represent a core brain mechanism for the behavioural impairments observed in MDD.

Funding

National Natural Science Foundation of China (82001795 and 82027808), National Key R&D Program (2022YFC2009900), and Sichuan Science and Technology Program (2024NSFSC0653).

Research in context.

Evidence before this study

Brain state transitions are crucial for supporting adaptive cognitive, emotional, and behavioural functions by enabling optimal reconfiguration of functional brain networks. Disrupted control of brain reconfiguration required for higher cognitive and behavioural functions may contribute to the functional deficits associated with major depressive disorder (MDD). However, most studies of brain function have focused on static/average measures of regional brain activity or functional connectivity rather than on the organisation of brain dynamics in whole brain networks. Brain network controllability model integrating both metabolism (energy cost) and dynamics is promising for providing a mechanistic explanation of brain reconfiguration. Early work suggested a controllability problem in MDD and the change of controllability was associated with recovery after antidepressant treatment in MDD. However, MDD is a complex syndrome characterised by both clinical symptoms and cognitive impairments. Cumulative evidence has documented alterations in multiple brain measures, including both structure and function that have been associated with depression severity as well as cognitive function in MDD. Thus, the brain alterations can not be exclusively interpreted by clinical symptoms, calling for integrated analyses of brain imaging features as well as symptoms and cognition in MDD.

Added value of this study

We found that brain average controllability especially in dorsal attention network (DAN) and the visual network was correlated with both clinical symptoms and cognitive impairments in MDD.

Implications of all the available evidence

These findings highlight the importance of DAN and visual network in modulating activity across brain networks in MDD, given their associations with symptoms and cognitive impairments observed in our study. Controllability may represent a promising approach for investigating the pathophysiology of MDD.

Introduction

Brain state transitions are crucial for supporting adaptive cognitive, emotional, and behavioural functions by enabling optimal reconfiguration of functional brain networks. Disrupted control of brain reconfiguration required for higher cognitive and behavioural functions may contribute to the functional deficits associated with major depressive disorder (MDD). However, most studies of brain function have focused on static/average measures of regional brain activity or functional connectivity (FC) rather than on the organisation of brain dynamics in whole brain networks.

Brain network controllability model integrating both metabolism (energy cost) and dynamics is promising for providing a mechanistic explanation of brain reconfiguration.1, 2, 3 This model is rooted in the idea that external energy input from brain regions to brain networks drives transitions from the current state to a desired state to meet the current demands,^2,4,5 which had been validated by the deterministic relationship between input energy and target state on a theoretical level.⁶ Regions with higher controllability can drive state transitions with lower input of impulse response energy that reflects the total magnitude and extent of neuronal signal spread.⁷ Early work suggested altered controllability in MDD, represented as decreased average and modal controllability in the default mode network (DMN) and frontoparietal network (FPN) in patients, and the change of controllability in DMN and FPN as well as the global controllability were associated with improvements in symptoms after antidepressant treatment in MDD.⁸ These findings reveal that controllability is correlated with symptom severity in MDD. For healthy populations, controllability measures are related to multidomain cognitive task performance such as executive function,^1,9 and the controllability of dorsal attention network (DAN) contributes to predicting both fluid intelligence and cognitive flexibility more than other brain networks do.¹ Additionally, the variation in regional controllability accounted for approximately half of the interindividual variability in cognition, in which the controllability in DMN had a high contribution.⁹ However, whether and how altered controllability is associated with impaired cognition in MDD have yet to be determined. Therefore, a detailed characterisation of the brain functional substrate from the perspective of controllability and its associated patterns with these symptomatic and cognitive disturbances is warranted.

MDD is a complex syndrome characterised by both clinical symptoms and cognitive impairments.^10,11 Cumulative evidence has documented alterations in multiple brain measurements associated with MDD, including both structure12, 13, 14 and function15, 16, 17 mainly in DMN, DAN, and FPN, which have been associated with depression severity¹² and cognitive deficits.¹⁷ Cognitive processes affect the onset, maintenance, recurrence, and course of MDD.¹⁸ Furthermore, cognitive control deficits and cognitive biases contribute to emotional dysregulation in depression.¹⁹ Altered controllability may impact concentration/attention, memory processes, and behavioural flexibility, which could disrupt adaptive abilities and thereby increase hopelessness and depression.¹⁸ On the other hand, emotion can also modulate cognition.²⁰ Thus, alterations in brain measures cannot be exclusively interpreted by clinical symptoms or cognition, given the intercorrelation of these features, warranting an integrated association analysis between controllability and clinical symptoms and cognition in one model.

Multivariate approaches can leverage patterns in the intercorrelation of features thus helping to better characterise the nature and magnitude of the associations.^21,22 Sparse canonical correlation analysis (sCCA) has been shown to be a powerful machine learning technique for identifying the covariance between brain imaging features and a wide array of nonimaging features, including behavioural, cognitive, biological, and environmental factors, in healthy populations and mental disorders.^21,23, 24, 25 For MDD, three brain-behaviour linked dimensions were revealed by CCA; specifically, the DMN FC was associated with mood/somatic symptoms, the FPN and visual network were linked to anhedonia, and the sensorimotor network was associated with insomnia.²⁶ The symptom-related FC determined by CCA analysis can identify patients with MDD with an accuracy of more than 95%.²⁷ The identified brain-behaviour linked dimensions can help identify the neural subtypes of depression and predict the treatment response to repetitive transcranial magnetic stimulation.^26,28 Compared to patients with MDD but without childhood maltreatment, patients with MDD and childhood maltreatment had a stronger canonical correlation between network FC, which was mainly related to the DMN, and Hamilton Depression Rating Scale (HAMD) scores.²⁹

Therefore, the purpose of the current study was to characterise the pattern of multivariate covariation between clinical features (symptoms and cognition) and brain controllability in MDD and to identify the most prominent features that drive the association by using sCCA. We restricted our analysis to first-episode medication-naïve patients to reduce the potential impact of pharmacological treatments on brain function measures. We hypothesised that brain controllability, especially those in DMN, FPN, and DAN, is related to both cognition and clinical symptoms in first-episode medication-naïve patients with MDD.

Methods

Participants

The diagnosis of MDD was determined by consensus between two experienced psychiatrists using the Structured Clinical Interview for Diagnostic and Statistical Manual of Mental Disorders-IV (SCID; DSM-IV). The inclusion criteria were as follows: 1) aged 18–65 years; 2) experienced a first episode of MDD; 3) had no history of receiving any psychiatric treatments, including psychiatric medications, psychotherapy, and neuromodulation therapy; and 4) had no psychiatric comorbidities (anxiety disorders, etc.). HAMD (17 items) was used to evaluate clinical symptoms.³⁰ The cognitive tests used in our study reflect different aspects of cognitive functions, as follows: 1) three aspects of executive function, specifically, inhibitory cognitive control, assessed by the Stroop Color and Word Test,³¹ cognitive flexibility, assessed by the Wisconsin Card Sorting Test (WSCT),³² and working memory, assessed by Trail Making Test B³³ and Backward Digital Span Test³⁴; 2) verbal memory, assessed by the Logical Memory Test³⁴ and Forward Digital Span Test³⁴; 3) episodic memory, assessed by the Visual Memory Test³⁴; 4) processing speed, assessed by the Digit Symbol Substitution Test³⁵ and Trail Making Test A³³; and 5) semantic fluency, assessed by the Fluency Test.³⁶ HAMD scores were evaluated by psychiatrists and cognitive tests were conducted by trained researchers. Ninety-nine patients with MDD (42 male/57 female, 30 ± 10.3 years old) in West China Hospital of Sichuan University were ultimately recruited in our study (Table 1).

Table 1.

Demographic and clinical characteristics of all patients with major depressive disorder in the present study.

	Abbreviation	All participants	Range	Male	Female	Differences between sexes (P values)
Number of patients	/	99	/	42	57	/
Age, years	/	30.0 ± 10.3	18–63	28.1 ± 9.3	31.4 ± 10.9	0.12
Education, years	/	14.3 ± 2.6	9–16	14.7 ± 2.4	14.1 ± 2.7	0.24
Illness durationa, weeks	/	32.2 ± 27.5	1–105	32.7 ± 32.1	31.9 ± 23.8	0.9
HAMD-total scoreb	/	24.1 ± 6.1	8–39	24.6 ± 6.9	23.6 ± 5.5	0.44
Stroop test
Stroop-C total timec, seconds	STRC_T	66.3 ± 25.0	35–205	67.8 ± 27.0	65.3 ± 23.5	0.63
Stroop-C accurate countd	STRC_C	111.6 ± 1.4	103–112	111.4 ± 1.7	111.7 ± 1.0	0.3
Stroop-CW inhibiting scorec,e	Inhibit	156.2 ± 60.6	61–383.6	167.7 ± 57.8	147.7 ± 61.8	0.11
Logical memory
Immediate scored	LM1	11.1 ± 4.5	2–20	11.4 ± 4.1	10.9 ± 4.8	0.59
Delayed scored	LM2	9.4 ± 4.9	0–20	10.0 ± 4.2	8.9 ± 5.3	0.25
Visual memory
Immediate scored	VM1	4.2 ± 2.2	0–9	4.7 ± 2.2	3.9 ± 2.1	0.05
Delayed scored	VM2	3.9 ± 2.1	0–9	4.4 ± 2.1	3.5 ± 2.0	0.03
Digital span test
Forward-accurate countd	DSTF_C	11.3 ± 2.3	5–14	11.3 ± 2.2	11.3 ± 2.5	0.96
Backward-accurate countd	DSTB_C	8 ± 3.2	2–14	8.4 ± 2.8	7.6 ± 3.4	0.25
Digital symbol substitution test-accurate countd	DSST_C	55.1 ± 15.2	16–90	55.7 ± 13.9	54.7 ± 16.2	0.75
Fluency test-accurate countd	VFT_C	20.4 ± 4.8	11–34	20.4 ± 4.5	20.4 ± 5.1	0.95
Trail making test
A-completing timec, seconds	Trail_A	38 ± 14.5	15–89	37.2 ± 15.0	38.5 ± 14.2	0.65
B-completing timec, seconds	Trail_B	57.5 ± 17.8	31–114	57.2 ± 19.6	57.8 ± 16.6	0.89
Wisconsin card sorting test
Total correctd	WSCT_C	38.1 ± 7.3	14–48	39.2 ± 5.6	37.2 ± 8.2	0.15
Perseverative errorsc	WSCT_PE	2.8 ± 3.3	0–16	2.1 ± 2.1	3.2 ± 3.9	0.07
Categories completedd	WSCT_cat	5.5 ± 1.4	1–8	5.7 ± 1.2	5.4 ± 1.6	0.28

Abbreviations: Stroop-C/CW, Stroop Color/Color-Word Test; HAMD, Hamilton Depression Rating Scale.

^aData were available in 91 of 99 participants (39 male and 52 female).

^bAccording to the HAMD-total score, 9 of 99 patients were mild depression (8–16), 37 of 99 patients were moderate depression (17–23), and 53 of 99 patients were severe depression (≥24).

^cLower value of the variable reflects better cognitive performance.

^dLager value of the variable reflects better cognitive performance.

^eCalculated formula: total time + ([2 × mean time per word] × number of uncorrected errors).

For comparative analyses of controllability, 73 sex-matched (31 male/42 female, 35.5 ± 12.3 years old) healthy controls (HCs) determined by clinicians according to the SCID-Non-Patient Version who did not undergo neuropsychological testing were also included. All HCs were recruited from the local area via internet advertisements, while the patients were seeking treatment in our hospital. This is an observational study, we recruited participants as many as we could, with sample size determination, randomization, and blinding not applicable to the present study. Sex identification was self-reported by all participants and we did not restrict the ratio of male and female participants in our study. Exclusion criteria for all participants included: 1) any history of affective illness comorbidity; 2) significant systemic or neurologic illness; 3) substance abuse or dependency; and 4) pregnancy and other MR contraindications. The HCs group differed in age distribution from the patient group (P = 0.002, two-sample t-test), so age was considered a covariate in between-group comparisons.

Ethics

The study was approved by the local research ethics committee of the West China Hospital of Sichuan University (reference number: 2018(610)), and written informed consent was obtained from all participants.

MRI data acquisition and preprocessing

MRI data were obtained using a 3.0 T magnetic resonance (MR) system (Siemens Trio) with an eight-channel phased-array head coil. Participants were fitted with earplugs and foam pads and instructed to keep their heads still and eyes closed but without falling asleep or having systematic thoughts. Resting-state functional MR imaging (fMRI) data were acquired using a gradient-echo echo-planar imaging sequence (repetition time/echo time [TR/TE] = 2000/30 ms, field of view [FOV] = 240 × 240 mm², matrix size = 64 × 64, flip angle = 90°, voxel size = 3.75 × 3.75 × 5 mm³, slice thickness = 5 mm with no gap). Each functional run contained 175 volumes, and each brain volume comprised 30 slices. High-resolution three-dimensional T1-weighted structural images were obtained using a spoiled gradient recall sequence with the following parameters: TR/TE = 1900/2.26 ms, flip angle = 9°, FOV = 256 × 256 mm², acquisition matrix = 256 × 256, voxel size = 1 × 1 × 1 mm³, and slice thickness = 1 mm. MRI images were inspected by two experienced neuroradiologists to confirm the absence of gross abnormalities and visible movement artifacts.

Imaging preprocessing was conducted based on the pipeline described in Yeo et al.³⁷ This procedure included 1) discarding the first four volumes to ensure blood oxygen level-dependent signal stabilisation; 2) performing slice time correction using SPM12; 3) correcting head motion using the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL) and controlling for spurious variance and its derivatives from head motion, whole brain, ventricular, and white matter signals; and 4) performing bandpass temporal filtering (0.01–0.1 Hz). Whole brain signals were also controlled to reduce nonneural signals and control motion artifacts.³⁸ Censoring of image frames was not conducted, as data scrubbing can cause inflated connectivity estimates in a specific region.³⁹ The mean frame displacement (FD) was required to be smaller than 0.3 mm, and the maximum FD was required to be smaller than 1 mm. T1 images were preprocessed using FreeSurfer (5.3.0).³⁷ Surface mesh representations of the cortex from individual T1 images were constructed and registered to a common spherical coordinate system. Boundary-based registration in the FsFast software package was used to align the structural and functional images. Then, the fMRI images were aligned to the common spherical coordinate space and smoothed with a 6-mm full-width half-maximum kernel. Finally, the fMRI images were downsampled to the fsaverage5 mesh with 10,242 vertices in each hemisphere. A sanity check was conducted as in a previous study³⁷ to control the quality of imaging data by performing a seed-based whole-brain FC analysis with the anterior and posterior cingulate cortex and the hand motor region as seeds and then inspecting whether the networks to which the seeds belonged were activated. Time series were extracted from preprocessed fMRI images in fsaverage5 space using the Schaefer template with 400 cortical parcels that can be assigned to seven networks, including the visual network, somatomotor network, ventral attention network, limbic network, DAN, FPN, and DMN (Figure S1).⁴⁰ The FC between all 400 parcels was calculated using the wavelet coherence approach which is less sensitive to outliers than a Pearson correlation⁴¹ and is less affected by interregional differences in hemodynamic response functions.⁴²

Controllability metrics

The detailed methods used to calculate brain controllability, including average and modal controllability, have been published elsewhere.^1,24 In brief, human brain dynamics (state transitions) were modelled using control theory as the observation of similarities of brain state transitions and engineering control processes. This approach is based on the idea that external energy input can transform a state into other target states to adapt to current brain function demands. Regions with higher controllability can drive state transitions with a lower input of impulse response energy that reflects the total magnitude and extent of neuronal signal spread.⁷

Two commonly used metrics, average and modal controllability, quantify two different aspects of brain control properties.² The average controllability quantifies the average energy cost of a given node to support general state transitions. Brain regions with higher average controllability can more easily drive the transition of functional states with lower input energy, indicating that they can play a crucial role in state transitions. Average controllability showed a positive correlation with nodal degree, indicating that hub brain regions tend to have higher average controllability.² Modal controllability assesses the ease of transition from current states to hard-to-reach states that require higher energy costs to perform complex goal-specific operations. Regions with higher modal controllability can more easily drive the brain state towards hard-to-reach states. Higher modal controllability is preferentially located in weakly connected brain regions or networks such as FPN that serves high-order cognition.² In the present study, we focused on average and modal controllability at a network level by taking the mean controllability of parcels within the same brain network.

The mathematical details of controllability are as follows. The control model was represented by a simplified noise-free linear time-invariant model as follows^1,2:

$$\begin{array}{c}x\left(t+1\right)=A·x\left(t\right)+B·u\left(t\right)\end{array}$$ (1)

where $x\left(t\right)$ is an $N\times 1$ (N = 400, here) vector that describes the brain state representing the blood oxygen level-dependent signal intensity of resting-state functional magnetic resonance imaging data at time t. $A$ is an $N\times N$ system adjacent matrix derived from the FC matrix as follows:

$$\begin{array}{c}A=\frac{FC}{1+{\lambda }_{\mathit{max}}}-I\end{array}$$ (2)

where ${\lambda }_{\mathit{max}}$ is the maximum eigenvalue of the FC matrix and I is an identity matrix. B is the control input matrix that represents the selection of control nodes. Since we set one control node at a time, B is an $N\times 1$ vector with only the selected nodes being one and the others being zero. u is the input control energy applied to the control point and is given by the following equation:

$$\begin{array}{c}u=\frac{1}{2}{d}^T{W}^{-1}d\end{array}$$ (3)

where d is the difference between the target state and the naturally reached state without control input.

The average controllability quantifies the average energy cost of a given node to support state transitions. As a previous study² has shown, the average controllability of a node (i) equals the trace of the controllability Gramian matrix (W) as follows:

$$\begin{array}{c}{W}_i=\sum _{\tau =0}^{\infty }{A}^{\tau }{B}_i{B}_i^T{A}^{\tau }\end{array}$$ (4)

which is proportional to the input energy (u).

The modal controllability of a specific node can be calculated by the following equation:

$$\begin{array}{c}{\varnothing }_i=\sum _{j=1}^N\left(1-{\lambda }_j^2\left(A\right)\right)v_{ij}^2\end{array}$$ (5)

where ${\lambda }_j$ is the jth eigenvalue of A and $V=\left[v_{ij}\right]$ is the eigenvector matrix of A. A detailed derivation of the formula can be found in the study by Duan et al.⁴³

sCCA

To control for potential confounding effects on the sCCA results, age, sex, education, and head motion were regressed out from controllability, while age, sex, and education were regressed out from clinical data (symptom and neuropsychological data) in the full dataset before the data were input into the sCCA model. Before conducting sCCA, all the data were standardised to a mean of zero and a standard deviation of one as in previous studies.^21,23

sCCA evaluates correlations between linear combinations of each multivariate dataset (referred to as pairs of canonical variates or modes) rather than focusing on individual feature-by-feature simple correlations, thus, it can estimate global relationships with a reduced probability of Type I and Type 2 errors.^44,45 In our study, sCCA was used to capture brain controllability patterns related to clinical symptoms and cognitive function in patients with MDD. Elastic net regularisation was used in our analysis to avoid overfitting the sCCA model, as in our prior work.²⁴ Elastic net regularisation combines both the L¹-norm penalty of least absolute shrinkage and selection operator (LASSO) and the L²-norm penalty of ridge method.⁴⁶ The L¹-norm can force weights to zero, thus achieving variable sparsity, and the L²-norm distributes the weights more evenly across all the features, both of which help reduce the complexity of this model, thus minimizing the overfitting of sCCA. Mathematically, sCCA is defined as follows:⁴⁷

$$\begin{array}{c}\underset{u,v}{\max }Corr\left[u^{T}X,v^{T}Y\right]=\underset{u,v}{\max }\frac{Cov\left[u^{T}X,v^{T}Y\right]}{\sqrt{Var\left[u^{T}X\right]Var\left[v^{T}Y\right]}}\end{array}$$ (6)

$${\text{Subject to}\left|\left|u\right|\right|}_2^2\le 1,{\left|\left|v\right|\right|}_2^2\le 1,{\left|\left|u\right|\right|}_1\le c_1,{\left|\left|v\right|\right|}_1\le c_2$$

Abbreviations: corr, correlation; cov, covariance; var, variance.

X and Y are the two groups of multidimensional variables (X denotes brain functional controllability and Y indicates clinical features in this study); u and v are the corresponding weights of variables; ${\Vert ·\Vert }_1$ and ${\Vert ·\Vert }_2$ denote the L¹-norm and L²-norm, respectively, which are used to achieve variable sparsity; and c₁ and c₂ are constants. The L²-norm was fixed at 1 for both u and v, while the L¹-norm (c₁ and c₂) was tuned by using ten-fold cross-validation with a grid search technique.

Specifically, the grid search method with an increment of 0.1 and a range of zero (suggesting the highest sparsity) to one (suggesting the lowest sparsity) was used to determine the combination of sparsity parameters that would achieve the highest average correlation of the first canonical variate across the ten-fold cross-validation. The best sparsity parameters were 0.7 for clinical data (ratings of the HAMD and cognitive tests, 33 in total) and 0.5 for average controllability (one value for each of the seven networks), while they were 0.6 for clinical data and 0.7 for modal controllability (Figure S2). The obtained optimal regularisation parameters were subsequently applied to calculate canonical correlations in the group of patients with MDD.

Statistics

Permutation testing with 1000 iterations was used to assess the significance of the canonical correlations (i.e., controllability was permuted in our case), and Bonferroni correction was used to control the Type I error rate at P < 0.05. As permutations can induce arbitrary axis rotations that lead to an altered order of sCCA modes,⁴⁸ we matched the canonical variates derived from permuted data with those resulting from the original data by comparing the weights of the clinical data. To make the text more understandable, we note here that 1) canonical variates are the latent variables created by the linear combination of variables in each sCCA mode; 2) canonical weights are the beta coefficients in linear combinations of variables in each dataset; 3) canonical cross-loadings are the correlation between each variable and the opposite canonical variate; and 4) canonical correlation indicates the correlation between the canonical variates. We performed sCCA for clinical data and average controllability, and for clinical data and modal controllability separately.

Reliability and reproducibility of sCCA

To assess the reliability and reproducibility of our sCCA analysis and ensure the robustness of our results, we applied a machine learning technique to conduct sCCA mode selection as in previous studies.^21,23 First, we randomly sampled 50% of the original dataset 5000 times, yielding 5000 training sets, and then conducted sCCA analysis on each training sample. Subsequently, the weights obtained from the training set were used to calculate the canonical correlation in the remaining 50% of the original dataset (which was treated as the test set). The redundancy reliability (RR) score, which indicates the similarity of results using the original dataset and the averaged results using resampling datasets, was calculated for each sCCA mode.²³ Mathematically, the RR score is defined as follows:²³

$$\begin{array}{c}RR=\frac{\left|corr\left(r_U^{Ts},r_U\right)\right|+\left|corr\left(r_V^{Ts},r_V\right)\right|}{2}\end{array}$$ (7)

where $r_U^{Ts}$ are the mean correlations between each variable of the clinical data in the test set and the canonical variate of controllability in the training set; $r_V^{Ts}$ represents the mean correlations between each variable of controllability in the test set and the canonical variate of the clinical data in the training set; $r_U$ represents the correlations between each variable of the clinical data and the canonical variate of controllability in the original set; and $r_V$ represents the correlations between each variable of controllability and the canonical variate of the clinical data in the original set.

Only modes that met the following robustness criteria were reported in the present study: 1) being statistically significant (P_Bonferroni<0.05); 2) having a median RR-score not less than 0.8²³; and 3) having an average canonical correlation of test sets that was at least half that of the training sets.

To examine the reliability of the estimated canonical weight of features, a bootstrapping procedure (using random sampling with replacement) was conducted. We only regarded features whose confidence interval of canonical weight did not cross zero and whose canonical cross-loading was statistically significant (P < 0.05) as stable contributing features.

As there are no consistent definitions of specific brain networks yet,⁴⁹ we used the Schaefer template with 400 parcels and assigned them to the most widely used network parcellation (Yeo's seven networks) in our main analysis.^37,40 Given the implicit resolution of the seven-network atlas, we used a finer Yeo's 17-network parcellation³⁷ to repeat our main analysis and validate the main results.

Predictive power of controllability

As an exploratory analysis, we evaluated the predictive power of the average controllability (because we only found a significant correlation between average controllability and clinical scores, see results) of all seven networks and the network with the highest canonical weight (DAN, see results), respectively, for the clinical canonical score (i.e., linear combination score of symptoms and cognition obtained from sCCA) by building ridge regression models. The regularisation parameter of the ridge regression model was optimised by maximizing the Pearson correlation between the actual and predicted clinical canonical variate score with leave-one-out cross-validation as a previous study did.⁵⁰ Then, the optimised model parameter was applied in a ten-fold cross-validation to obtain the predicted clinical canonical score for each participant. The Pearson correlation coefficient between the actual and predicted clinical scores was calculated. The ten-fold cross-validation was repeated five times and an averaged Pearson correlation coefficient value was computed to converge on a true estimate of the test statistic, independent of which participants were randomly selected in each fold. The average coefficient of determination (R²) and root mean squared error (RMSE) of the model were calculated. Finally, permutation testing by randomly shuffling the clinical score was used to estimate the significance of the correlation coefficient with 1000 iterations.

Role of funders

The funders had no role in the study design, data collection, analyses, or interpretation, and were not involved in the writing of the paper or in the decision to submit the paper for publication. None of the authors have been paid to write this article by a pharmaceutical company or other agency.

Results

sCCA

We separately conducted sCCA for average and modal controllability. We found that there were no significant sCCA modes of modal controllability and clinical variables that met all three criteria (Figs. 1b and 2b). Only the first mode for the sCCA analysis of average controllability (seven networks) and clinical data (33 symptom and neuropsychology scores) met the criteria for significance (r_mean = 0.54, P_Bonferroni = 0.03 [Permutation test], median RR score = 0.8, proportion of training/test = 53.8%) and accounted for 31.2% of the covariance (Figs. 1a, 2a and 3a, and Table S5). Thus, subsequent analyses focused only on average controllability.

Fig. 1.

Covariance explained by the canonical variates of average/modal controllability and clinical data. Three sCCA modes for average controllability and clinical data (a) (P_Bonferroni was 0.28 for mode 1, <0.0001 for mode 3, and 0.007 for mode 4) and two modes for modal controllability and clinical data (b) (P_Bonferroni was 0.007 for mode 1, and 0.021 for mode 2) were statistically significant at P_Bonferroni<0.05 (∗) (Permutation test). However, only the first sCCA mode for average controllability met all three robustness criteria.

Fig. 2.

Redundancy Reliability (RR) scores of the sCCA modes for average(a)/modal(b)controllability and clinical data. The horizontal line in the boxplot denotes the median RR score, while the black dot in the boxplot denotes the mean of RR score. The red dashed line represents an RR score of 0.8.

Fig. 3.

Sparse canonical correlation analysis (sCCA) for average controllability and both symptoms and cognition. Panel a shows the mean canonical correlation of sCCA modes, with the mode that met all three criteria highlighted in green. Error bar denotes standard error. For the first sCCA mode of average controllability, panel b shows the canonical correlation of the first mode with the original data, panel c shows the canonical weight of clinical variables, and panel d shows the canonical weight of average controllability. Asterisks (∗) in panels c and d represent the stably contributed features. Abbreviations: FPN, frontoparietal network; DAN, dorsal attention network. For the abbreviations of clinical variables, see Table 1 and Table S1.

With the optimal regularisation parameters, 23 of 33 clinical features and three of seven networks’ average controllability had non-zero canonical weights (Fig. 3c and d, and Tables S1 and S3). The stable contributing features in the first sCCA mode of average controllability included HAMD1 (depressed mood), HAMD3 (suicide), HAMD7 (work and activities), HAMD12 (gastrointestinal symptoms), WSCT (total correct [WSCT_C], perseverative errors [WSCT_PE], and categories completed [WSCT_cat]) (Fig. 3c and Figure S3a, and Table S1), and average controllability of the DAN and the visual network (Fig. 3d and Figure S3b, and Table S3).

Specific clinical symptoms (depressed mood, suicidal ideation and behaviour, impaired work and activities, and gastrointestinal symptoms) had the highest negative canonical cross-loading with brain features, while cognitive flexibility was most strongly and positively correlated with average controllability (Fig. 4a and Table S2). These findings indicate that severe clinical symptoms and poorer cognitive flexibility are related to lower average controllability in MDD. The average controllability of the DAN and the visual network had the highest correlation with the integrated assessment of symptoms and cognition, particularly the DAN, which had the highest weight (Fig. 4b and Table S4). In case–control comparisons, decreased average controllability of the DAN and increased average controllability of the limbic and frontoparietal networks in patients with MDD compared with healthy controls were revealed by permutation testing (10,000 iterations) after controlling for age, sex, and head motion (Figure S4).

Fig. 4.

Canonical cross-loading for clinical variables(a)and average controllability(b)in the first sCCA mode. Canonical cross-loading is the correlation between each variable and the opposite canonical variate. ∗ represents the stably contributed features. Abbreviations: SMN, sensorimotor network; DMN, default mode network; FPN, frontoparietal network; VAN, ventral attention network; DAN, dorsal attention network. For the abbreviations of clinical variables see Table 1 and Table S1.

Reliability and reproducibility results

When using Yeo's 17 network parcellation, the first canonical variates for the sCCA analysis of average controllability and clinical data reached a statistical significance (r_mean = 0.56, P_Bonferroni = 0.02 [Permutation test], median RR score = 0.8, proportion of training/test = 50%) and accounted for 21.3% of the highest covariance. All identified stably contributed features with higher canonical weights in the first sCCA mode of average controllability when using the seven-network atlas reproducibly exhibited higher canonical weights than did the other variables in the sCCA analysis when using 17-network atlas (Figure S5) (Pearson correlation r = 0.70, P < 0.0001 for clinical variables), which further validated the reliability and robustness of our results.

Predictive power of average controllability

We found that the canonical score of the average controllability of all seven networks can predict the clinical score (R² = 0.2, RMSE = 1.7, P = 0.002 [Permutation test]). The predictive power of the average controllability of DAN as the only input feature for clinical score was similar (R² = 0.2, RMSE = 1.7, P = 0.005 [Permutation test]).

Discussion

The current study employed an integrated multivariate model to examine the covariation between functional brain controllability and clinical features in first-episode medication-naïve patients with MDD. Among the clinical variables, depressed mood, suicidal ideation and behaviour, impaired work and activities, and gastrointestinal symptoms were significantly negatively associated with average controllability, and reduced cognitive flexibility was associated with reduced average controllability. We found that the average controllability of the DAN and the visual network were correlated with the integrative assessment of clinical symptoms and cognition, indicating that the disrupted control of the brain reconfiguration of the DAN and the visual network in their state transitions may thus represent a core brain mechanism for the behavioural impairments observed in MDD.

These specific associations have particular clinical interests. First, our findings support the notion that dysphoric symptoms, including depressed mood, suicidal behaviour or ideation, and disrupted work and activities⁵¹ were associated with disruptions in the dynamic organisation and maintenance of brain functional states. Patients with MDD had a more rigid neural temporal dynamic of emotion regulation than healthy controls, presenting with impaired flexibility and ineffective emotion regulation.⁵² Suicidal behaviour⁵³ and ideation⁵⁴ were also correlated with aberrant temporal neural dynamics. A lack of work activities (anhedonia) was associated with a decreased frequency of dynamic FC states in MDD.⁵⁵ Gastrointestinal symptoms in patients with MDD have been related to altered brain FC⁵⁶ and structure.⁵⁷ Our findings thus provide evidence for the association between gastrointestinal symptoms and brain network controllability in MDD, particularly in the control of brain states in specific brain networks. Second, cognitive flexibility impairment has been demonstrated in MDD.^11,58 Our study showed that average controllability was positively correlated with cognitive flexibility, thus confirming the association between the brain and cognitive flexibility. This finding parallels a recent study showing that altered brain state dynamics were related to cognitive flexibility problems in elderly individuals.⁵⁹ Previous studies have shown a negative association between average controllability and synchronizability which is another predictor of brain dynamics that supports the same temporal dynamical pattern.³ Furthermore, there was increased average controllability along with reduced synchoronizability during the process of optimal brain function with age.³ Thus, controllability may represent a promising approach for investigating the pathophysiology of MDD, given that the findings obtained through this method revealed significant covariation with clinical symptoms and cognition.

The DAN is responsible for the top-down control of attention,⁶⁰ and disrupted control of attention shifts may contribute to emotional and cognitive dysfunction.⁶¹ We quantified the controllability of brain networks using a brain control model and demonstrated an association between the average controllability of the DAN and cognition as well as clinical symptoms in MDD. A lower average controllability indicates that functional brain state transitions are less readily driven and are unable to support modulation of emotions and effective cognition which contributes significantly to the symptom severity and cognitive dysfunction of MDD. Associations between brain state transitions and both symptoms and cognition have been reported previously in MDD.^62,63

Brain state transitions are crucial aspects of FC in the brain, and densely connected brain areas generally have higher average controllability.¹ Consistent with our findings, previous connectome-related studies revealed relationships between the DAN and the abovementioned symptoms and cognition aspects in MDD. Altered FC in the DAN was related to several aspects of emotional, visceral, and autonomic dysregulation, such as gastrointestinal symptoms, in patients with MDD.⁶⁴ Additionally, decreased FC within the DAN has been related to anhedonic symptoms such as a lack of interest in work or enjoyable activities in MDD.²⁸ The DAN was shown to have decreased FC with the FPN, thus causing decreased external attention and, in turn, increased negative rumination, which was associated with suicidal ideation in MDD.^65,66 Decreased FC between the cortex in the intraparietal sulcus, a core DAN hub, and the orbitofrontal cortex was related to impaired cognitive control of emotion processing in MDD.^67,68 DAN-related FC has also been used to predict response to psychotherapy⁶⁷ and electroconvulsive therapy in MDD,⁶⁹ further highlighting the critical role of the DAN in MDD. In addition to those static fMRI studies, Simeonova et al.⁷⁰ found altered brain activation in regions within DAN during the Stroop n-back task in MDD. Our findings, together with those previous studies, have demonstrated the important role of the DAN in the neuropathology of MDD.

The visual network is another brain functional network related to clinical symptoms and cognition. Abnormal visual processing has been highlighted in patients with MDD,⁷¹ and dysfunction of the visual network has been correlated with depressive symptoms.⁷² In accordance with our observation of lower average controllability of the visual network, Chen et al. reported decreased dynamism in this brain network in MDD.⁷³ Additionally, altered visual perception has been linked to increased access to mood-congruent information, contributing to the persistence of depressed mood in MDD.⁷⁴ The visual network is regulated by top-down control by the DAN, and patients with MDD have shown disrupted top-down regulation of visual perception.⁷⁵ Several studies have reported abnormalities within and between the DAN and the visual network in MDD.^75,76 Accumulating evidence and our findings have indicated that both the DAN and the visual network are involved in the pathophysiology of MDD and add to existing knowledge by showing the importance of the mechanisms of brain state transition.

Notably, we did not find a correlation between modal controllability and cognition and symptoms, which indicated that the clinically relevant brain changes in depression are rooted more in the driving of easy-to-reach transitions than in high-energy demanding changes to hard-to-achieve transitions. A large number of brain functional states are easily reachable, so as to quickly initialise or be ready to move to other task states when needed and relax back after completing tasks.² This ability can help maintain basic human physics, such as emotion, and facilitate human optimal behaviour and high-order cognition.⁷⁷ The lower flexibility of switching between these easy-to-reach states could thus contribute to the dysregulation of emotion, suboptimal behaviour, and disrupted cognitive and perceptual activities.⁷⁷ This finding was in accordance with our observation that average controllability was related to the canonical score of symptoms and cognitive function in MDD. However, the regions with higher modal controllability in the resting state were the areas involved in cognitive control.¹ Modal controllability in task states had a higher predictive power for cognitive task performance than did average controllability in task states.¹ A previous study showed that modal controllability in patients with schizophrenia was not correlated with clinical symptoms.²⁴ Modal controllability is predominantly responsible for difficult-to-reach states, particularly those activities involved in cognitive control, therefore, that may be exclusively related to cognition but not symptom as prior observations demonstrating that higher cognitive functions are relatively stable regardless of acute illness severity.^78,79

Limitations

Our study has certain limitations. First, the current results are only preliminary insights, given the relatively small sample size, although strict mode selection and reliability analyses were conducted to ensure that the obtained findings were reliable and replicable. Second, as in most studies of brain controllability,^1,2 the control framework was built on a linear model, while neural dynamics are nonlinear. Future studies are needed to determine whether developing nonlinear control models can provide further insights into alterations in brain function related to MDD. Third, the relationships between controllability and brain multimodal features such as anatomy and function,80, 81, 82, 83 and between controllability and other brain dynamic metrics such as those in sliding window analyses,^84,85 need to be explored in future studies, to make controllability more understandable when we implement the theoretical approach to experimental applications in translational and clinical psychiatry.⁸⁶ Fourth, the predictive power of average controllability for clinical scores was not high, and future studies with a larger sample size and deep learning methods are warranted to reproduce and expand upon our findings. Finally, importantly, a longitudinal study is warranted to determine whether the observed pattern of findings persists after treatment or recovery and whether the observed pattern of findings can predict treatment response or clinical course.

Conclusions

In summary, we found that brain average controllability, especially that in the DAN and the visual network was correlated with both clinical symptoms and cognition in first-episode medication-naïve patients with MDD. The results suggest that both the DAN and the visual network are involved in the pathophysiology of MDD and add to existing knowledge by showing the importance of the mechanisms of brain state transition in MDD.

Contributors

Fei Li and Qiyong Gong conceptualised and designed the study. Qian Li, Weihong Kuang, Fei Li and Qiyong Gong have directly accessed and verified the underlying data reported in the manuscript. Qian Li, Youjin Zhao, Yongbo Hu, Yang Liu, Yaxuan Wang, Qian Zhang, Fenghua Long, Yufei Chen, Yitian Wang, Haoran Li, Eline M P Poels, Astrid M Kamperman, John A. Sweeney, Weihong Kuang, Fei Li and Qiyong Gong contributed to data collection, preprocessing and analyses, interpretation and drafted the manuscript. Qian Li, Youjin Zhao, Eline M P Poels, Astrid M Kamperman, John A. Sweeney, Qiyong Gong, and Fei Li critically revised the manuscript. All authors approved the final version of the manuscript.

Data sharing statement

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Declaration of interests

All authors declare no biomedical financial interests or potential conflicts of interest.