Causal connectivity alterations of cortical-subcortical circuit anchored on reduced hemodynamic response brain regions in first-episode drug-naïve major depressive disorder

Some efforts were done to investigate the disruption of brain causal connectivity networks involved in major depressive disorder (MDD) using Granger causality (GC) analysis. However, the homogenous hemodynamic response function (HRF) assumption over the brain may disturb the inference of temporal precedence. Here we applied a blind deconvolution approach to examine the altered HRF shape in first-episode, drug-naïve MDD patients. The regions with abnormal HRF shape in patients were chosen as seeds to detect the GC alterations in MDD. The results demonstrated significantly decreased magnitude of spontaneous hemodynamic response of the orbital frontal cortex (OFC) and the caudate nucleus (CAU) in MDD comparing to healthy controls, suggesting MDD patients likely had alterations in neurovascular coupling and cerebrovascular physiology in these two regions. GC mapping showed increased/decreased GC in OFC-/CAU centered networks in MDD. The outgoing GC values from OFC to anterior cingulate cortex and occipital regions were positively correlated with Hamilton Depression Scale (HAMD) scores, while the incoming GC from insula, middle and superior temporal gyrus to CAU were negatively correlated with HAMD scores of MDD. The abnormalities of directional connections in the cortico-subcortico-cerebellar network may lead to unbalanced integrating the emotional-related information for MDD, and further exacerbating depressive symptoms.

Increasing neuroimaging evidence has emphasized major depressive disorder (MDD) as a network-level neural disorder, associated with the dysregulation of a distributed brain network emcompassing the cortical-subcortical-cerebellar circuit [1][2][3][4][5] . Important progress has been made in understanding the pathogenesis of MDD by investigating the brain's intrinsic functional connectivity networks of resting-state functional magnetic resonance imaging (fMRI) data 3,4,6 . Other than functional connectivity method which measures statistical dependencies of time-series between distinct units, effective connectivity or causal connectivity investigates the influence one neuronal system exerts over another 7 . Granger causality (GC) analysis is one of the powerful and widely applicable techniques to detect the effective connectivity between even remote brain regions 8,9 . Specifically, GC analysis has recently been increasingly employed in the studies of depression to identify the effective connectivity abnormality in MDD 1,2,10,11 . Using GC analysis, increased excitatory effect from hippocampus to anterior

Results
Brain regions with HRF abnormality in MDD. Only the HRF response height showed the significant difference between MDD and HC groups. Figure 1 demonstrates brain regions with significantly different response height of HRF between the groups (p < 0.05, AlphaSim corrected). Compared with HC group, MDD group showed significantly decreased HRF height in the superior frontal gyrus orbital (ORBsup), the middle frontal gyrus orbital (ORBmid) and the inferior frontal gyrus orbital (ORBinf) of the orbital frontal cortex (OFC) extending to the rectus gyrus, and the caudate nucleus (CAU). Significantly higher HRF height in MDD group than HC group was not found in the present study.
Seeds regions of interest (ROI) based on the two sample t-test results of the hemodynamic response were selected to calculate the effective connectivity mapping between the seeds regions and the remaining brain regions. They were centered at the peaks t-values in left superior frontal gyrus orbital (ORBsup), right middle frontal gyrus orbital (ORBmid) and left inferior frontal gyrus orbital (ORBinf), and left CAU. Table 1 illustrated the details of the chosen seeds, including the Montreal Neurological Institute (MNI) coordinates of peak voxels, cluster sizes and statistical t-values.
Partially conditioned Granger causality mapping of left ORBsup seed. Figure 2 demonstrates the brain regions which exhibit significantly different effective connections both to and from left ORBsup in MDD patients comparing to HC group. Partially Conditioned Granger Causality Mapping of right ORBmid seed. The brain regions which exhibit significantly larger effective connections both to and from right ORBmid for MDD are illustrated in Fig. 3. Stronger outgoing information transfer was found from right ORBmid seed to right medial frontal gyrus orbital and left ACC; Stronger incoming information transfer was found from the following regions to the seed: left superior frontal gyrus, left medial frontal gyrus orbital, left inferior frontal gyrus triangular, left Inferior frontal gyrus opercular, left middle temporal gyrus temporal pole and bilateral middle cingulate gyrus (See Table 3 for the details). Specifically, correlation analysis further showed that the GC values from right ORBmid to left ACC were positively correlated with HAMD scores of MDD patients (p < 0.05, Alphasim corrected). Table 4 summarized the results of the voxel-wise correlation analysis, including the MNI coordinates of peak voxels, cluster sizes, mean correlation coefficient r, std r and mean p values.
Partially conditioned Granger causality mapping of left ORBinf seed. Figure 4 demonstrates the brain regions with significantly higher information transfer from left ORBinf in MDD compared with HC. Regions details are showed in Table 5. MDD patients exhibited significantly higher information transfer from the left ORBinf to the distributed regions including right superior frontal gyrus, bilateral precentral gyrus, superior parietal gyrus, fusiform gyrus, middle occipital gyrus and calcarine fissure. The GC values from left ORBinf to left calcarine fissure, left and right fusiform gyrus have positive correlations with HAMD scores of MDD (See Table 4 for details). Significant difference in the incoming effective connectivity was not found.
Partially conditioned Granger causality mapping of left CAU seed. Figure 5 shows the brain regions with significantly lower effective connections to left CAU for patients comparing to HC group. Significant difference in the outgoing effective connection was not detected. Table 6 summarized the results in details. For the case of left CAU, significantly lower causal inflow was reported from bilateral superior temporal gyrus, middle   Table 4 for details).

Discussion
In this work, we used a blind deconvolution approach to reconstruct the brain HRF at voxel-level to detect the altered HRF shape in first-episode, drug-naïve MDD patients, and further explored the deconvolved BOLD level effective connectivity networks centered at regions with HRF response height abnormalities using partially conditioned GC approach 9 . The method allowed us to explore the spontaneous hemodynamic response shape and to detect the neural causality among brain regions using fMRI data more effectively. Comparing to HC, MDD patients had significantly reduced HRF response height in left ORBsup, right ORBmid, left ORBinf and left CAU. HRF has been found to be related to key physiological factors including cerebral blood flow and the cerebral metabolic rate of oxygen 9,19-21 . Especially, this parameter of HRF response height has been found to have striking correlation with cerebral blood flow 21 . Thereby our results suggested that MDD patients likely have decreased brain metabolism and changed cerebrovascular physiology in these regions 20,22 .   ORBsup, ORBmid and ORBinf covered the great majority of OFC. OFC and CAU are two critic nodes in affective network, especially in the frontostriatal network 6,23-26 . The frontostriatal network has been identified as having important roles in the regulation and modulation of emotion, behavior and cognition [23][24][25][26][27] . OFC, which is also referred to as the ventrolateral prefrontal cortex, is an important area in MDD pathology 27 . This frontal region has a role in the "top-down" and volitional regulation of affect, and is associated with extensive cognitive and affective processes such as decision making, sensory and social information integration, mood regulation and emotional recognition [27][28][29][30][31] . Neuroimaging techniques have showed that OFC in patients with MDD have reduced volume 32 , decreased amplitude of low-frequency oscillations 16 , lower biophysical integrity of macromolecular proteins 23 and lower activation in reward processes 33 . As the dorsal part of striatum, CAU is involved in a range of functions including processing of reward and sensory, the regulation and modulation of affect and mood 23,25 . Depressive patients have structural and functional abnormalities in CAU including reduced volumes 34 , decreased regional homogeneity 35 , lower brain metabolism 36 , lower biophysical integrity of macromolecular proteins 23 , aberrant network topological properties 3,6 , and lower activation in reward-related processing 37,38 . Comparing to HC,  significantly lower regional cerebral blood flow were also revealed in the prefrontal cortex and CAU in drug-naïve MDD [39][40][41] . There has been evidence that reduced regional cerebral blood flow reflects brain metabolism and is related to reduced neuronal size or abnormal vascular factors in these regions 42,43 . In the present study, we further demonstrated the decreased height of spontaneous hemodynamic response of OFC and CAU in resting state in MDD. Concerning OFC, the disturbed hemodynamic response may result in the inefficient suppression of maladaptive emotional responses, and may further impair decision making or disrupt the mood regulation and   emotional recognition 27,28,31 . Concerning CAU, the reduced magnitude of hemodynamic response can account for the dysfunction associated with the lack of motivation, ahedonia and altered emotional integration 23,37 . Taken together, the decreased magnitude of spontaneous hemodynamic response of OFC and CAU observed in our patients may underlie the disturbances of the emotion-processing circuit, which would contribute to the persistent negative feelings experienced by MDD 16,28 . With the HRF parameter of response height, our results provided further evidence for a role of OFC and CAU in the pathophysiology of MDD. The seed-based GC networks between these regions and the rest of the brain were further computed for MDD and HC. Our results showed significantly stronger information transfer between the seeds in OFC and the distributed brain regions including the frontal, temporal, parietal, occipital and limbic regions in MDD. In addition, the outgoing GC values from right ORBmid to left ACC, and from left ORBinf to left calcarine and bilateral fusiform gyrus were positively correlated with HAMD scores of MDD. Concerning CAU, significantly lower causal inflow was reported from the temporal gyrus, INS and the cerebellum to left CAU in MDD. The incoming GC values from right INS, middle and superior temporal gyrus to the seed were negatively correlated with HAMD scores in MDD. The results demonstrated the causal connectivity disruptions in the brain network associated with cognitive and affective processes. The contrary alterations in the cortico-subcortical network may lead to unbalanced integrating the cognitive and affective information for MDD, and the depressive severity became heavier as the alterations went bigger.
OFC plays a key part in circuit-based models of several psychiatric and neurological disorders because it is extensively connected with diverse neural areas that underlie its participation in a broad range of psychological functions such as processing emotion, reward learning, and decision making 29,30 . Previous studies have demonstrated that OFC is not homogenous, but can be parcellated into several cytoarchitectural subregions that possess distinct connections with cortical and subcortical structures 29,30,44 . In the present study, three subregions in OFC were chosen by comparing the differences of the response height of HRF in the two groups: left ORBsup, right ORBmid and left ORBinf. The three effective connectivity subnetworks centered in the three seeds were further compared between MDD and HC groups.
Chosen left ORBsup as seed, all aberrant outgoing and incoming information transfer for MDD was reported within the frontal network, especially in the dorsal lateral prefrontal cortex. The dorsal lateral prefrontal cortex with its connections to OFC is regarded as the origination point of the volitional/cognitive regulatory arc 31 . Increased functional connectivity has been found between the dorsal lateral prefrontal cortex and OFC in MDD 28 . Our results further demonstrated aberrantly increased forward and backward effective connections between ORBsup and the dorsal lateral prefrontal cortex in MDD, which might affected the modulations of executive control associated with externally induced emotional states in first-episode, drug-naïve MDD patients.
Chosen right ORBmid as seed, MDD patients exhibited stronger effective connections mainly from left PFC and bilateral middle cingulate gyrus to ORBmid. This may lead to improperly modulate the incoming cognitive information in ORBmid for the patients. Stronger outgoing information transfer from right ORBmid to right medial frontal gyrus orbital and left ACC was found. Specifically, the GC values from right ORBmid to left ACC were positively correlated with HAMD scores of MDD patients. ACC, as one of the primary hubs in the salience network, plays a key role in cognitive-emotional integration and ongoing monitoring of behavior 31 . It has been implicated as a focus of dysfunction in depression 5 . Previous functional connectivity studies has found increased connections between ACC and the prefrontal cortex 5,45 , indicating the dysfunction in the volitional/external regulatory network for MDD 31 . Our results further demonstrated the abnormal directed connections from ORBmid to ACC, and suggested that ORBmid exhibited increased overdrive of top-down emotional processing to ACC with depression severity in MDD patients.
For the case of ORBinf, there was significant stronger causal outflow to the distributed regions in the frontal, parietal and occipital lobe. The GC values from left ORBinf to left calcarine and bilateral fusiform gyrus had positive correlations with HAMD scores of MDD. These visual processing regions were known to be involved in the perception of emotions in visual stimuli and selective attention 5,46 . The disruptions in the resting-state functional  connectivity patterns in the visual cortical area for MDD have been reported before, such as aberrant connectivity of the visual cortical areas with regions associated with visual processing 5,47 , or abnormal topological property of nodal centrality and nodal degree in the occipital region 3,6 . The disruption may contribute to the dysfunction of visual recognition and selective attention in MDD 3,5,35 . Previous GC analysis demonstrated abnormal effective connectivity in fronto-occipital network 48 . Our results further found the abnormally elevated causal influence from ORBinf to the visual sensory areas, and this aberrant influence increased progressively over the illness severity in MDD. This may contribute to the abnormal top-down effects on the visual cortex during the cognitive information processing from higher cognitive regions to visual sensory regions in MDD 5,48,49 .
In sum, our study found that OFC have increased abnormal effective connections with a distributed brain regions in first-episode, drug-naïve MDD patients. Specifically, ORBsup demonstrated the aberrant causal connections with the prefrontal cortex. ORBmid has additional abnormal information transfer mainly with the cingulate gyrus; especially the causal outflow to ACC was positively associated with patients' illness severity. ORBinf showed increased effective connections to the occipital-parietal regions, and the positive correlations were found between the GC values in the visual processing regions and disease severity in MDD. The results demonstrated that different patterns of OFC subnetworks may be associated with specific cognitive and affective processes 29 , and thereby result in differential causal connectivity disruptions in different subregions of OFC in MDD. The further parcellation of OFC in the future OFC-centered network studies will provide better understanding of the psychopathology of MDD.
The effective connectivity mapping reported significantly lower effective connections from the temporal lobe and cerebellum to left CAU for patients comparing to HC group. Incoming GC values from right INS, MTG and STG to the seed were negatively correlated with HAMD scores in MDD patients. The temporal lobe and INS are two critical regions in the affective network 5,50 . In addition to primary auditory function, the temporal lobe also plays important roles in social cognition and emotional processing 16,50 . INS is regarded as an integration center in bottom-up and top-down information processing including autonomic, visceromotor, emotional, and interceptive information 10,51 . The abnormal modulation of the cortical-subcortical loop in the affective network has been observed in MDD 5,16 . Our results further indicated decreased causal connectivity from right superior and middle temporal gyrus and INS to left CAU in first-episode, drug-naïve MDD, suggesting inefficient emotional regulation in the cortical-subcortical loop at the early stages of the illness. The abnormal inflow to CAU may result in improperly integrating the incoming information in CAU, and may further worsen the dysfunctional cognitive-emotional processing for the MDD patients 1 . This ability of information integration was found to been worse with the illness severity, for in our study the incoming GC values from the temporal lobe and INS to the CAU had negative correlation with HAMD scores in MDD.
The cerebellum has anatomical connections with the frontal and limbic regions, suggesting that cerebellum contributes to certain emotion and cognitive processing 52,53 . Abnormalities in the cerebellum in association with depression have been consistently found, indicating the involvement of cerebellar dysfunction in MDD 4,5,54-57 . Altered functional connectivity in resting-state has been observed between cerebellum and the default mode network, the executive control network, the limbic and paralimbic regions 5,55,56 . Specifically, using GC analysis, Guo et al. found the inhibitory effect from the temporal gyrus to the cerebellum in MDD 1 . In the present study, we found the decreased inflow from the cerebellum and the temporal gyrus to CAU, suggesting the dysfunction of emotional modulation in the cortico-subcortico-cerebellar network centered around CAU in MDD. The aberrant cerebellar effective connectivity in MDD need to been explored more deeply in the future study.
Interestingly, chosen seeds located in OFC, significantly stronger information transfer were found in MDD; while for the case of left CAU, significantly less information transfer was detected in patients. This suggested hyper-connections of executive control loops centered around OFC which were associated with externally induced emotional states, and hypo-connections of subcortical networks centered around CAU which were related to motivation modulation and emotional integration in first-episode, drug-naïve MDD patients. We speculated that the contrary alterations in the cortico-subcortical network may lead to unbalanced integrating the emotional-related information for MDD, and further exacerbating depressive symptoms.

Methods
Participants. Twenty-five first-episode, drug-naïve patients were recruited from the Mental Health Center, West China Hospital of Sichuan University. They all met DSM-IV criteria for MDD according to the diagnostic assessment by the Structured Clinical Interview for DSM-IV-Patient Edition (SCID-P) and were with scores of 18 or greater on the 17-item HAMD. Patients comorbided with other Axis I and Axis II psychiatric disorders such as schizophrenia, bipolar affective disorder, personality disorders and substance abuse or dependence were excluded according to the SCID-I and SCID-II assessment. Twenty-four HC with no history of neurological or psychiatric diseases were also recruited. All subjects were right-handed and without severe or acute medical conditions physically based on clinical evaluations and medical records. The present study protocol was approved by the Ethical Committee of the West China Hospital of Sichuan University and was carried out in accordance with the approved guidelines. The informed written consents were obtained from all participants. There were no significant differences of age, gender, and education between MDD group and HC group. The clinical and demographic data of the recruited subjects were shown in Table 7. Data preprocessing. Data preprocessing was performed using the Statistical Parametric Mapping software (SPM8, http://www.fil.ion.ucl.ac.uk/spm). The first five volumes were discarded to ensure steady-state longitudinal magnetization. The remaining 200 resting-state fMRI images were first corrected by the acquisition time delay among different slices, and then realigned to the first volume for head-motion correction. The dataset with translational or rotational parameters exceeding ± 1.5 mm or ± 1.5° would be excluded, remaining the data sets of twenty-two MDD patients and twenty-two HC subjects for further analysis. The images were further spatially normalized into a standard stereotaxic space at 3 × 3 × 3 mm 3 , using the Montreal Neurological Institute template in SPM8. No spatial smoothing was applied in order to avoid artificially introducing local spatial correlation 58 . Since recent studies have showed that functional connectivity analysis is sensitive to gross head motion effects 59 , we further evaluated the framewise displacement (FD) with the suggested threshold of 0.5 to express instantaneous head motion 59 . The largest FD of all subjects was less than 0.2 mm. Two-sample t-test showed there was no significant difference of FD between the two groups (0.057 ± 0.031 for HC and 0.048 ± 0.022 for MDD; p = 0.301).Images were then corrected by linear regression to remove the possible spurious variances including six head motion parameters, the white matter, and the ventricular signals averaged from a white matter mask and a ventricular mask respectively. Since removing the global signal may introduce a shift in the distribution of resting-state functional connectivity data and make biological interpretation ambiguous 60,61 , we did not remove the global signal in the present study. The residuals of these regressions were temporally band-pass filtered (0.01 < f < 0.08 Hz) to reduce low-frequency drifts and physiological high-frequency respiratory and cardiac noise, and linearly detrended for further analysis.
Partially conditioned Granger causality analysis. In the current study, the partially conditioned GC method in deconvolved BOLD level for effective connectivity analysis was performed using an in-house MATLAB toolbox (http://guorongwu.weebly.com/software.html). Firstly, the voxel-specific HRFs for deconvolution were obtained by a blind deconvolution approach in which resting fMRI was considers as 'spontaneous events' 9 . These spontaneous events can be detected by point process analysis (PPA), picking up BOLD fluctuations of relatively large amplitude. The HRF for each voxel was then reconstructed by fitting raw BOLD signal using a canonical HRF (two gamma functions) and two derivatives (multivariate Taylor expansion: temporal derivative and dispersion derivative) 9 . Three parameters of the HRF which estimated the potential changes in neuronal activity were calculated including response height, time to peak and full-width at half-max 9,62 . The group differences of the three parameters at voxel-level were computed respectively by a two-sample t-test in SPM8, to obtain the seeds ROIs for calculating the effective connectivity mapping. Statistical significance was estimated via a Monte Carlo simulation (Alphasim) for multiple comparisons with p < 0.05 (single voxel threshold of p < 0.04 and cluster size > 190). The REST toolbox (www.restfmri.net) was used to determine the cluster size for Alphasim correction significance, and a brain mask for gray matter was used. ROIs were defined as spheres centered at the local maxima of the statistic t map with radius of 6 mm radius.
Then, the causal influence was then calculated between the seeds ROIs and the remaining brain using signals at the neural level by Wiener deconvolution for each subject. The partially conditioned Granger causality approach allows to compute Granger causality conditioned to a limited number of variables in the framework of information theory 14,15 . Briefly, consider n covariance-stationary variables ( ) = , ,  , which maximizes the mutual information , β I X Z { } 14,15 . Finally, the difference of the effective connectivity networks between MDD and HC groups was further analyzed using two sample t-test with age, sex and FD as covariates. The statistically significant threshold was set for multiple comparisons at the cluster level with p < 0.05 using AlphaSim correction (single voxel threshold of p < 0.04 and a minimum cluster size of 190 voxels). To investigate whether GC values were correlated with  symptom severity and course of disease in MDD patients, relationships between GC values in regions showing significant group differences and HAMD scores/illness duration were further detected by voxel-wise correlation analysis (AlphaSim corrected p < 0.05, with single voxel threshold of p < 0.04).