Changes in brain network activity patterns after ketamine treatment in major depression

Fifty-eight depressed individuals (average age = 40.7 years, n = 28 female) participated in this naturalistic clinical trial. The data acquired for this investigation formed part of a larger NIMH-supported Connectomes Related to Human Disease project, where subjects were followed prospectively while receiving different fast-acting antidepressant interventions. The ketamine arm of the study was pre-registered in clinicaltrails.gov (NCT02165449). Depressed participants received four serial intravenous ketamine infusions over the course of 2 weeks. Data, which included MRI and clinical assessments, were collected less than a week before the first infusion (baseline) and 24 hours after the last (4th) ketamine infusion.

To be eligible for participation, individuals had to meet criteria for TRD as defined by a lack of response to at least two antidepressant trials of appropriate dosage and duration, and experiencing continuous depression for at least 6 months. Additional eligibility criteria included meeting DSM-5 criteria for major depression (First et al., 2015), and having moderate-to-severe depressive symptoms (Hamilton Depression Rating Scale (HDRS) 17-item ≥ 17; Hamilton, 1960), no history of psychotic reactions to medications, alcohol or drug abuse in the past, and no other physical or clinical contraindications to ketamine. Exclusion criteria encompassed unstable medical or neurological conditions, current substance abuse or dependence (confirmed by laboratory testing), or substance abuse within the past 3 months, current or history of psychosis, schizophrenia, intellectual disability, or other developmental disorder, a diagnosis of dementia, and any contraindications to MRI scanning (such as metal implants or claustrophobia).

In addition, 56 HCs not experiencing depression as confirmed by clinical interview, but otherwise meeting the same exclusion criteria as the depressed participants were assessed at baseline (average age = 32.6, female = 31). A subset of 18 HCs (average age = 29.5, female = 10) received repeated assessments after approximately 2 weeks to mirror the time frame of ketamine treatment in patients. Recruitment of all subjects was conducted in the Los Angeles area through advertisements, clinician referrals, or through clinicaltrials.gov. All participants provided written informed consent in accordance with procedures approved by the UCLA Institutional Review Board (IRB). A graphical overview of our study design is shown in Figure 1.

The administration of ketamine occurred two to three times per week and involved delivering a sub-anesthetic dose (0.5 mg/kg) of racemic ketamine through an intravenous pump, diluted in 60 cc of normal saline. Participants were allowed to continue taking their approved monoaminergic antidepressant medications if the dosage had not changed in the preceding 6 weeks. Benzodiazepines were discontinued more than 72 hours prior to all study visits including scanning sessions.

At each time point, overall depression severity was assessed using the 17-item Hamilton Depression Rating Scale (HDRS) (Hamilton, 1960). Remission of depression is defined as an HDRS score of ≤7 post-SKI treatment. Rumination was assessed at each time point using the Rumination Response Scale (RRS) (Treynor, 2003), which includes three sub-scales measuring brooding, reflection, and reappraisal.

Imaging sequences were acquired at the UCLA Brain Mapping Center using a Siemens 3T Prisma MRI system with a 32-channel head coil, and were identical to those used by the Human Connectome Project (HCP) Lifespan studies for Aging and Development (Harms et al., 2018). Structural sequences included T1-weighted (voxel size (VS) = 0.8 mm isotropic; repetition time (TR) = 2500 ms; echo time (TE) = 1.81:1.79:7.18 ms; inversion time (TI) = 1000 ms; flip angle (34) = 8.0^o; acquisition time (TA) = 8:22 minutes) and T2-weighted acquisitions (VS=0.8 mm isotropic; TR = 3200 ms; TE = 564 ms; TA = 6:35 minutes), both with real-time motion correction (Tisdall et al., 2016). Resting-state fMRI used a multi-band EPI sequence with opposite phase encoding directions over two runs (VS = 2 mm isotropic; TR = 800 ms; TE = 37 ms, FA = 52^o, MB accl. factor = 8; phase enc. direction = AP(run1)/PA(run2); TA = 13:22 minutes), along with two sets of spin echo images used for distortion correction (Andersson et al., 2003; Smith et al., 2004). Participants viewed a fixation cross throughout the duration of the resting-state fMRI scan. Additionally, motion censoring was performed for 60 seconds prior to the acquisition of resting-state fMRI scans using the Framewise Integrated Real-time MRI Monitoring (FIRMM) software (Dosenbach et al., 2017), in order to determine the level of motion expected for each subject and provide feedback if excessive motion was present.

Imaging data were preprocessed using the HCP minimal processing pipeline (Glasser et al., 2013). Processing of rsfMRI data included ICA-based denoising using FSL's multi-run FIX (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FIX↗), and surface alignment using MSMAll. Further fMRI denoising and processing included global signal regression (GSR), band pass filtering [0.01–0.1 Hz], and smoothing with a 4 mm kernel. GSR, which is shown effective at reducing a variety of fMRI artifacts (Ciric et al., 2018; Power et al., 2017) especially motion correlated artifacts in HCP data (Burgess et al., 2016), was performed to mitigate the effects of the global artifacts in the detection of CAP states. Since there remains a lack of consensus over whether GSR should be used in preprocessing, given that it can remove neurologically relevant signals from the brain (X. Liu et al., 2018), we additionally performed CAP analysis with the same protocol described below without using GSR (Supplementary Fig. S4) to determine its effect on our results.

After denoising, rfMR images were parcellated using 400 cortical regions of interest (ROIs) from the Schaefer 7 network atlas (Schaefer et al., 2018), as well as 54 subcortical ROIs from a functional atlas derived using functional connectivity gradients (Tian et al., 2020). Once parcellated, the ROI time series data were z-scored to determine the relative states of activation and deactivation for all regions across the brain. Motion was assessed by examining average framewise displacement (FD) across subjects, and since all subjects moved less than 0.5 mm FD on average, all subjects were retained for analysis. Additionally, the average FD was assessed within each cluster to ensure CAP states were not clustered on the basis of head motion.

CAP analysis was performed in MATLAB by running k-means clustering of fMRI data time points across all subjects and longitudinal scanning sessions using correlation distance to determine the spatial similarity of fMRI time points (1 – rho(X,Y) for data points X and Y) (J. E. Chen et al., 2015; Huang et al., 2020; X. Liu et al., 2013). A maximum of 1,000 iterations until convergence was used with 100 replicates in order to repeat the clustering algorithm with varying initialized cluster centroids, since k-means is highly affected by the initial conditions. By default, MATLAB chooses the replicate with the most optimal clustering score across replicates to ensure robust clustering results. Since k-means requires the number of clusters to be defined, we followed an approach to find the optimal number of clusters by determining the elbow point across a range of clustering scores, as in prior studies from our group and others (Allen et al., 2014; Goodman et al., 2021; Kupis et al., 2021; Marshall et al., 2020). Here, we ran k-means from k = 2 clusters to k = 50 clusters using data from all subjects and time points, and for each clustering solution, the cluster validity index was calculated, defined as ratio of within cluster to between cluster distances (Allen et al., 2014). Once the cluster validity index was computed for all clustering solutions, an L-curve was fitted to the plot of clustering scores using least squares regression to determine the location of the elbow point (Supplementary Fig. S1). Another common approach to find the optimal number of clusters involves computing the Silhouette Score for each cluster (Supplementary Fig. S2).

Next, CAP metrics were computed within each subject's scanning session, using the optimal number of clusters that were determined to characterize brain dynamics of these states. CAP metrics included (1) the fraction of time spent in a CAP state (FT) (also referred to as the occurrence fraction (C. Liu et al., 2023), dominance (Kaiser et al., 2019), and fractional occupancy (Niu et al., 2025)), defined as the proportion of time points spent in one CAP state relative to all other states, (2) mean dwell time (MDT) or persistence of a CAP, defined as the average amount of TRs one state persists, and (3) the transition probability (TP) from one CAP to another CAP, defined as the number of transitions from states A to B over the number of occurrences of state A, for every pair of states (A, B). TP was modeled as a Markov process, but was calculated after removing repeating CAPs in the time series to control for autocorrelation due to each CAPs persistence (Cornblath et al., 2018; Huang et al., 2020).

Although all data are clustered together to permit meaningful comparisons between CAP metrics across groups and time, we also performed clustering on each individual TRD time points and the HC group to determine the similarity between the clusters formed from each group individually and the clusters formed from all subjects together (Supplementary Fig. S3). While discrepancies may arise between CAP states derived from a sub-group compared with the entire sample, it should be noted that CAP states are cluster centroids defined by the spatial positions of rfMRI time points across a group of subjects. If strong differences in state dynamics (such as the amount of time spent in a CAP state) exist between groups of subjects, then removing the additional groups to identify CAP states from a single group alone would result in a different spatial distribution of rfMRI time points used in clustering, which would likely produce a different set of CAP states. Therefore, it would be hard to disentangle the differences in CAP states between groups and differences in state dynamics within a shared set of CAP states between groups. As such, all primary analyses used the CAP states defined across all groups and focused on examining differences in state dynamics. It should also be noted that most CAP studies focus on examining differences in CAP dynamics rather than the spatial patterns of CAP states, including studies looking across many different clinical groups (Niu et al., 2025).

Furthermore, to determine the effect of cluster number on our results, we performed additional analyses using k_optimal + 1 and k_optimal - 1 (Supplementary Figs. S5 and S6). Lastly, we performed an analysis looking at changes in static-FC following SKI in TRD to compare the changes in dynamic CAP metrics with traditional static-FC (Supplementary Table S1). Static-FC was computed on the parcellated fMRI time series data, by computing the Pearson correlation coefficient in MATLAB between all pairs of ROI time series, producing a symmetric ROI by ROI correlation matrix.

Statistical analyses were performed on the computed CAP metrics and on clinical mood scales. Paired t-tests were used to determine whether significant improvements occurred in symptom scales following SKI. To investigate CAP changes in TRD following SKI, paired t-tests were performed across time for each subject. To determine whether changes in TRD following SKI were significantly associated with improvements in depressive symptoms, Pearson correlations were computed between changes in CAP metrics and percentage change in mood scores ((post-SKI mood score - baseline mood score)/post-SKI mood score) that were significantly improved following ketamine treatment. Statistical analyses of CAP metrics were performed on FT and MDT across all states, and TP was assessed only for states which showed significant effects of FT or MDT, by analyzing transitions from all other states. Multiple comparisons were addressed using Bonferroni correction across all states tested (p < 0.05/6 = p < 0.0083) and all state transitions tested (p < 0.05/10 = p < 0.005) in order to determine the significance level for each set of tests on the CAP data, however, the p-values that are displayed are unadjusted from the tests that are performed. Follow-up analyses were performed on CAP metrics showing significant effects over time in TRD by investigating cross-sectional differences between HC and TRD at baseline, using independent samples t-tests controlling for age and sex, to determine whether changes in TRD moved in the direction of patterns seen in HCs. Secondary analyses focusing on confirming no significant differences existed between HCs and TRD post-SKI, again used independent samples t-tests controlling for age and sex, in order to establish TRD changes as normalizing toward HCs. An additional supplementary analysis was performed using the subsample of longitudinal HCs to test whether longitudinal CAP changes occur in the absence of treatment, such as from arbitrary changes in the scanning environment, in order to ensure the fidelity of longitudinal changes in TRD. Finally, to ensure that results were similar in CAP states derived from TRD patients only, we performed the same set of analyses investigating treatment effects using CAP metrics derived from TRD patients only.

Decreases in the fraction of time spent in a state with activation in VN and DAN were observed in TRD following SKI. The VN includes primary and secondary visual association areas (striate and many extrastriate regions), while the DAN includes the frontal intraparietal sulcus and frontal eye fields as well as the middle temporal region and superior parietal lobule and receives important connections from the VN important for orienting and visuospatial attention (Fox et al., 2006; Kincade et al., 2005; Seitzman et al., 2019; Vossel et al., 2014). Although the visual cortex and visual processing are not classically thought of as contributing to MDD, there is increasing and overlooked evidence that the visual cortex plays a role in depression and antidepressant efficacy (Wu et al., 2023). This evidence suggests that MDD patients show impairments in emotional processing within the primary visual cortex, abnormalities in filtering of irrelevant information, and show more autonomy in the ventral and dorsal streams of visual processing (Wu et al., 2023). Our finding also shows correspondence with prior ketamine studies in depression that have examined brain activation or functional connectivity with standard fMRI methods. For example, observations from our own group using fMRI, perfusion (pseudo-continuous arterial spin labeling, pCASL), or diffusion MRI in an overlapping TRD sample have repeatedly identified changes in visual and visual attention systems as being involved in ketamine response. Specifically, we have previously reported increases in VN and DAN cerebral blood flow (Sahib, Loureiro, Vasavada, Kubicki, Joshi, et al., 2020), decreases in VN and DAN BOLD activity related to a response inhibition task (Sahib, Loureiro, Vasavada, Kubicki, Wade, et al., 2020), changes in VN white matter microstructure (Taraku et al., 2023), and increases in habenular–VN functional connectivity (Taraku et al., 2024) following ketamine treatment. Changes in most of these neurobiological markers related to symptomatic improvements post-ketamine treatment. Several independent ketamine studies have also implicated the involvement of visual networks, mostly including secondary visual association regions. For example, Evans et al.(2018) observed increased resting-state functional connectivity between the DMN and VN, and DAN nodes as well as the insula (a key node of the SN) pre- to post-ketamine treatment in MDD. Rivas-Grajales et al. (2021) similarly found increased habenular–VN resting-state connectivity, which is associated with antidepressant response post-ketamine treatment. Further, Reed et al. (2018) reported changes in brain activation for an emotional valence and congruency task in occipital regions that were also associated with antidepressant response in MDD patients receiving ketamine versus placebo. Gilbert et al. (2021) showed that ketamine treatment in TRD accelerated the transmission of GABA and NMDA in the early regions of the visual cortex, and led to inhibition in the early visual cortex and inferior frontal gyrus.

With respect to dynamic changes in functional circuitry, one cross-sectional study showed more functionally optimized, independently organized, and less regulated dorsal and ventral VNs compared in MDD with controls, which included a higher mean variability in dynamic connectivity in MDD (H. Chen et al., 2019). Further in healthy controls, following ketamine administration, ketamine plasma levels were found to have a significant negative influence on dynamic resting-state functional connectivity across VN connections (Spies et al., 2019), adding further evidence that ketamine affects visual and visual–attention network dynamics. Our results together with prior imaging evidence thus suggests that ketamine's response mechanisms involve a perturbation of connectivity, activation, and dynamics in VN and DAN. These effects are unlikely related to the acute psychomimetic side effects of ketamine, given that these side effects subside 2 hours after administration and post-treatment assessments occurred at least 24 hours after drug administration for the studies cited.

We observed significant increases in the fraction of time spent in the CEN following ketamine treatment. The CEN, composed primarily of the dorsolateral prefrontal cortex (dlPFC) and posterior parietal cortex (PPC), is often considered to be involved in executive function, cognitive tasks, attention, and working memory, and, unlike the DMN, is active during goal-oriented tasks (Mulders et al., 2015). The CEN is also considered to be involved in emotion regulation via strategies such as cognitive reappraisal, which involves top–down regulation of emotion processing regions (Buhle et al., 2014; Morawetz et al., 2017; Rive et al., 2013). The role the CEN plays in cognitive control and emotion regulation has made regions of this network promising targets for antidepressant treatment (Park et al., 2019). Using conventional and dynamic fMRI analysis approaches, the CEN has been widely implicated in depression pathophysiology. For example, prior observations include reports of decreased functional and effective connectivity between the CEN and DMN (Li et al., 2021; Zheng et al., 2015), decreased functional connectivity within regions of the CEN, including the dlPFC and the PPC (Kaiser et al., 2015), and decreased occurrences of a coactive dynamic CEN and DMN state (Dini et al., 2021) in MDD patients compared with controls.

Following ketamine treatment, increased resting-state functional connectivity has been found to occur between the CEN with subcortical limbic structures, including the hippocampus and amygdala, in patients with MDD; early changes in connectivity were shown to relate to clinical improvements (Vasavada et al., 2021). Another study found increased resting-state functional connectivity between the dlPFC (a key node of the CEN) and subgenual anterior cingulate cortex, which showed positive associations with antidepressant response post-ketamine treatment (Gärtner et al., 2019). Increased functional connectivity between frontal nodes of the CEN and the DMN posterior cingulate node was also reported in patients treated with ketamine versus placebo (Evans et al., 2018). However, a study using either a low (0.2 mg/kg) or a standard dose (0.5 mg/kg) of IV ketamine found a reduction in functional connectivity between frontal CEN nodes and the dorsal anterior cingulate at the standard dose, and even more pervasive reductions in connectivity between fronto-parietal CEN nodes at low dose in patients with MDD pre- to post-ketamine treatment where changes were shown to associate with improvements in depressive symptoms and/or suicidality (M.-H. Chen et al., 2019). Taken together, these findings seem to suggest that ketamine serves to decrease connectivity within the CEN, but serves to increase functional connectivity between the CEN and other major resting-state networks including the DMN and limbic regions.

Prior studies investigating the molecular mechanism of ketamine's antidepressant action have observed increases in glutamate cycling and extracellular glutamate in the prefrontal cortex (PFC) (Abdallah et al., 2016). This is associated with NMDA receptor blockade on GABAergic interneurons, leading to a disinhibition of glutamate and subsequent surge in AMPA receptor activity in the PFC, which ultimately facilitates post-synaptic neuroplasticity signaling pathways (Abdallah et al., 2016). In the context of this neurobiological model, it is likely that the increase in PFC glutamate and neuroplasticity results in greater and more frequent activity in the PFC, which facilitates frequent co-activation of networks involving PFC regions, including the CEN. Thus, this signaling cascade from NMDA receptor blockage likely results in more frequent co-activation of networks involving the PFC, thus leading to greater dominance of executive control networks during rest.

Disruptions in CEN circuitry may also contribute to cognitive deficits commonly seen in depression (Disner et al., 2011), such as executive function and working memory impairments (Rose & Ebmeier, 2006; Snyder, 2013). Subanesthetic ketamine treatment in MDD has been shown to enhance aspects of neurocognitive function including in the domains of working memory, processing speed, and episodic memory (Zavaliangos-Petropulu, McClintock, et al., 2023), which suggests that ketamine may affect systems involved in both the depressive and cognitive deficits associated with MDD. These findings together with our observations suggest that ketamine treatment alleviates MDD-related within-network CEN hypoactivity, which may affect a range of deficits associated with depression. Whether ketamine-induced CEN changes lead to downstream effects contributing to improvements in cognitive and emotional control remains a goal for future study.

Our analysis of state transitions revealed significant increases in the transition probability from the salience network state to the CEN state, as well as significant decreases in the transition probability from the salience network state to visual network state. The salience network, consisting of nodes primarily in the anterior insula and dorsal anterior cingulate cortex, is believed to play a role in subjective salience, emotional judgment, and facilitates a causal role in the dynamic switching between the DMN and CEN, thus also playing a bottom-up role in attentional control (Menon & Uddin, 2010). The anterior insula node of the salience network has been described as a gatekeeper of executive control due to its role in orchestrating brain network dynamics (Molnar-Szakacs & Uddin, 2022; Uddin, 2015). It is also believed to play a role in emotional processing and emotional control through extensive subcortical connections (Mulders et al., 2015). The salience network, along with the DMN and CEN, together comprises a set of large-scale networks which encompass the triple network model, which posits that dynamic interactions between these networks underlie a wide range of psychopathologies (Menon, 2011). Prior studies investigating resting-state networks in MDD have found abnormalities between the SN and DMN, including hypoconnectivty between SN and DMN (H. Chen et al., 2019; Kaiser et al., 2015), as well as increased dynamic FC fluctuations between SN and DMN (Kaiser et al., 2016). Since our findings revealed increased occurrences of CEN states and decreased occurrences of VN states, our follow-up analyses of transition probability imply that the SN is driving these changes following ketamine treatment. Partially in line with our findings, another fMRI study of ketamine in MDD showed a normalization of connectivity between the SN with the DMN and CEN, to support the triple network dysfunction model of MDD and its modulation by ketamine. Together these findings might suggest that ketamine treatment targets SN-related circuitry, which in turn modulates broader brain circuitry related to emotional, cognitive, and sensory processing.

When examining the association between changes in brain state dynamics and improvements in clinical symptoms, we found a significant association between decreases in the occurrences of an SN state and improvements in reflective rumination, though did not find significant associations for overall mood scales. Rumination is a prominent symptom of depression, characterized by repetitive and passive self-focused thoughts on one's negative feelings and symptoms of distress (Parola et al., 2017). However, recent theories suggest rumination is a multidimensional construct consisting of two distinct components, which include brooding, considered maladaptive and links to depressive symptoms, and reflection, which is "a purposeful turning inward to engage in cognitive problem-solving to alleviate one's depressive symptoms" (Parola et al., 2017). Although prior research on rumination has focused primarily on brooding in depression, one study found that after performing a factor analysis on RRS scores, a distinction between reflection and brooding could only be found in formerly depressed and never depressed individuals, but did not obtain this distinction in currently depressed individuals. Furthermore, this study identified another factor in depressed individuals which they labeled "intentional rumination" (Whitmer & Gotlib, 2011). Therefore, the precise interpretation of our finding that reflective rumination is reduced following SKI and how this is distinguished from brooding is yet to be determined, since rumination may be more complex in depressed individuals than previously theorized. Further research is needed to establish how ketamine treatment alters the dimensions of rumination in depressed individuals.

The SN is involved in responding to externally salient cues, and acts as a dynamic switch between the DMN involved in negative, self-referential processing, and the CEN, involved in executive function. Rumination has often been associated with dysfunction in the DMN (H.-X. Zhou et al., 2020). Many studies have also linked the SN and intrinsic connectivity between nodes of the SN and DMN with rumination (Burrows et al., 2017; Ordaz et al., 2017), including in MDD specifically (Cooney et al., 2010; Paul et al., 2013; Ray et al., 2005). Furthermore, one particular study examining the topological features of large scale brain networks found that the SN is expanded nearly two-fold in patients with depression (Lynch et al., 2024). At the same time, previous studies have suggested that overactivity of the SN relates to abnormal affective processing (Schimmelpfennig et al., 2023), including negative response bias to affective stimuli in MDD (Hamilton et al., 2013). Since our results show that time spent in the CEN is significantly greater after SKI, and that transition probability from SN to CEN increased, it is possible that less time spent in the SN state may reflect a restoration of the regulation of negative emotional states, as a consequence of more time spent in the CEN. These findings align with prior CAP findings of increased dwell time in a state composed of co-active SN and DMN as well as increased transitions between SN–DMN states to prototypical DMN states, which both associate with greater rumination in depression (Kaiser et al., 2019).

We note that although significant improvements in HDRS occurred following SKI, we did not observe significant correlations between HDRS changes and changes in CAP metrics over time. While correlations between brain changes and changes in clinical measures can provide insight into potential neurobiological mechanisms that induce symptom improvements, it is possible that the measured brain changes here only describe certain brain–symptom relationships, given the distributed pathophysiology of MDD (Buch & Liston, 2020). In this case, while ruminative symptoms in this study as well as several previous studies (Belleau et al., 2022; Kaiser et al., 2019; C. Liu et al., 2023) are correlated with CAP dynamics, it is possible that the neurobiological mechanisms underlying overall depressive symptoms measured by HDRS may have distinct neurobiological drivers independent of CAP dynamics. For example, it may be the case that connectivity in particular functional circuits are correlates of HDRS symptom improvements rather than functional dynamics across the whole brain. It is also possible that more statistical power is needed to identify relationships between CAP dynamics and HDRS changes, which would imply that the link between these symptoms and CAP dynamics is weaker than that of rumination and CAP dynamics.

By performing post hoc cross-sectional analyses focusing on the dynamic states showing significant effects following ketamine treatment, we found significant differences between TRD and HC at baseline in the occurrences of a CEN state, SN state, and the transitions from a SN state to both the CEN and VN state. Comparisons between TRD–post-SKI and HCs suggest that most of the observed changes following ketamine treatment normalize aberrant network dynamics present in depression, with patterns showing more similarity to HCs following treatment. These cross-sectional findings are also supported by prior research which has found differences in CEN (Kaiser et al., 2015; C. Liu et al., 2023), SN (Manoliu et al., 2014), and VN (H. Chen et al., 2019) between depressed individuals and controls.

We acknowledge several limitations of this investigation. We note that this study was not a randomized clinical trial and did not include a placebo control condition. However, this study was designed as a mechanistic study focused on imaging outcomes and delineating relevant neural circuitry, rather than on the efficacy of SKI. We also note that patient participants were allowed to continue concurrent stable antidepressant medication. However, patients served as their own controls in longitudinal analyses, with our study focusing on within-subject changes. The antidepressant effects of ketamine are reported to peak at around 24 hours post-infusion and to be greater after serial therapy (Marcantoni et al., 2020). However, the temporal relationships between brain circuit-level neuroplasticity and therapeutic response may be more complex and the timing of post-treatment assessments may influence results.

We also acknowledge that there are limitations associated with the CAP method used here, which includes the identification of specific clusters used to identify brain states, which can sometimes be arbitrarily defined and differ across studies. However, despite brain states not always being identical across studies, the brain states identified here show correspondence with commonly identified and frequently studied resting-state brain networks (Uddin et al., 2019), overlap with states identified in prior studies (Belleau et al., 2022), and our findings regarding the dynamics of the CEN also show correspondence with prior CAP studies (Kaiser et al., 2019; C. Liu et al., 2023). Furthermore, we show that when using a different number of clusters (k-1 and k+1), we find similar brain states and also show similar significant results with k-1 clusters, suggesting that the dynamic activity patterns we are seeing are not contingent on one configuration of clusters. One limitation with this approach was that we could not replicate any results using k+1 clusters, suggesting that the observed changes in brain dynamics cannot be identified when using a set of seven states. Although CAP analysis requires a discrete number of brain states to analyze brain dynamics, other studies have modeled brain dynamics as a smooth manifold of brain activity patterns (Rué-Queralt et al., 2021). Thus, a discrete number of brain states may be an approximation to the landscape of brain dynamics, and future research should reconcile findings from CAP studies with other methods to estimate brain dynamics.

In addition to these points, we acknowledge that we observed minor differences in the spatial patterns of CAP states generated from HC alone compared with our entire sample. However, we note that our aim was not to identify differences in CAP states across groups but rather to examine differences in CAP dynamics between a common set of states across groups and treatment. Additionally, our analysis of subject-level similarity to cluster centroids in Table 2 suggests that each group of subject and time points does not significantly differ in their spatial patterns to each cluster centroid, implying that these CAP states are well representative of the entire sample. Further research is needed to examine how CAP states differ across diagnostic groups and how this relates to differences in CAP dynamics and static FC. Lastly, in a supplementary analysis, we show that these CAP metrics are relatively stable across controls scanned twice without intervention, suggesting that these observed changes likely represent brain changes induced by ketamine treatment.

Serial ketamine treatment produces changes in resting-state functional dynamics of large-scale brain networks including the CEN, VN/DAN, and SN. Changes in the SN in particular relate to improvements in rumination symptoms. The observed changes induced by ketamine treatment also normalize aberrant patterns of brain network dynamics seen in depressed patients when compared with healthy controls. These findings provide further insight into the functional brain systems-level mechanisms accounting for the therapeutic effects of ketamine in MDD.