Resting state functional connectivity patterns as biomarkers of treatment response to escitalopram in patients with major depressive disorder

Identification of biomarkers allowing prediction of antidepressant response to individual drugs and allocation of the right treatment to a given patient is a key priority in psychiatric research. Currently, finding such treatment follows a ‘trial and error’ process, based on guidelines derived from the knowledge of populational efficacy of drugs; this process however does not necessarily work for a particular individual (Cipriani et al. 2018; Shinohara et al. 2019). In fact, only one-third of depressed patients respond to their first treatment, and finding an effective one often takes months or longer (Warden et al. 2007; Rush et al. 2009). Despite significant research efforts, clinically applicable biomarkers have not yet been identified.

Such research has been largely facilitated by the development of non-invasive in vivo imaging techniques, which put forward some potential biomarkers, such as increased activity of the pgACC as a general predictor of good treatment response (Fu et al. 2013; Godlewska et al. 2018). However, after years of research, it has become increasingly unlikely that an aberrant function of one particular structure in the brain could reliably predict treatment response. This is in line with depression being conceptualized as a network disorder, with meaningful aberrations related to the networks of brain structures rather than individual regions (Li et al. 2018).

Neuroimaging has been the key in the understanding of network activity. A commonly used method is an assessment of temporal correlation of the activity of brain regions—so-called functional connectivity (FC)—with functional magnetic resonance (fMRI) technique, either during task performance or when an individual is awake yet not engaged in any external tasks (so-called resting state). Assessment of resting-state functional connectivity (rsFC) has some important benefits: it is task independent; hence, results are not dependent on the selection of an appropriate task, data collection is relatively easy as it does not depend on the ability to perform a task, which may be impaired in depression, and it allows an assessment of networks predominantly activated during rest. Resting state analysis can be performed with different methods, all having one assumption in common: the definition of functional connectivity is based on a statistical dependency between time series, expressing co-activation of brain structures. Conventional methods of analyzing resting-state FC (rsFC) can roughly be divided into region-of-interest (ROI)–based and data-driven methods (Yang et al., 2020). While both approaches are beneficial, ROI-based methods require a pre-existing hypothesis as to the role of a region, and a choice of a ‘seed’ can be quite subjective with slightly different seeds giving different results. A seed-based analysis is furthermore restricted by having to predefine the region a priori which does not allow for explorative analysis. An alternative is defining nodes based on the data in a data-driven parcellation by, for example, independent component analysis (ICA) (Yang et al., 2020). A benefit of this method is that new regions can be discovered from which new hypotheses can be formulated.

Both types of analysis provided support for models of depression focused on dysfunctional communication between brain regions and helped to explore mechanisms of antidepressant action at the network level. Two prominent models focused on the role of inadequate cognitive control by the central executive network (CEN) over the affective and default mode networks (AN and DMN, respectively), resulting in inadequate top-down regulation of attention and emotion regulation in depressed patients regardless of their treatment status (Drevets 2001, Mayberg 2003, Wang et al. 2012, Kaiser et al., 2015, Zhong et al. 2016).

In line with these models, antidepressants, including a number of serotonin reuptake inhibitor (SSRIs), serotonin-norepinephrine reuptake inhibitor (SNRIs) and an norepinephrine reuptake inhibitor (NRI) reboxetine were shown, through both data- and hypothesis-driven analyses, to modulate disrupted network connectivity in the fronto-limbic networks and between the frontal and DMN regions (McCabe et al. 2011; Dichter et al. 2015; Vai et al. 2016; Brakowski et al. 2017), with potentially key role of the posterior DMN (Li et al. 2013; Posner et al. 2013). Some findings are promising as potential biomarkers, such as, for example, low connectivity between the lateral PFC and sgACC being a predictor of good treatment response, which was shown for both typical antidepressant drugs, such as an SSRI sertraline (Chin Fatt et al. 2020) and a new glutamatergic treatment ketamine (Gartner et al. 2019). SgACC connectivity may also play a role in non-pharmacological methods, such as TMS (Philip et al. 2018). At this point, however, there are no clinically useful biomarkers of response, and although the work on the proposed markers needs to continue, it is similarly important to explore brain connectivity in a data-driven manner, so that potentially relevant connections are not missed because of restriction related with an a priori selection of brain regions included in the analysis.

Therefore, this study employed a whole-brain data-driven independent component analysis, in order to explore differences in patterns of baseline (pre-treatment) resting-state functional connectivity (rsFC) in patients with major depressive disorder (MDD) who responded to 6 weeks of treatment with an SSRI escitalopram, patients who did not respond to such treatment and healthy volunteers.

The main aim of the study was to explore baseline (i.e. pre-treatment) patterns of resting-state functional connectivity (rsFC) associated with response to SSRI treatment. To achieve this, we compared pre-treatment rsFC in depressed individuals who responded to 6 weeks of treatment with escitalopram, depressed individuals who did not respond to such treatment and rsFC in untreated healthy individuals without depression.

Using an unbiased ICA approach, we identified a number of patterns. Treatment responsivity was associated with enhanced rsFC of the right FPN—with the posterior DMN (bilateral precuneus), and with the SMN and somatosensory association cortex. At the same time, treatment non-responders appeared to be characterized by reduced rsFC: of the bilateral FPN with the contralateral SMN, of the right FPN with the posterior DMN (bilateral precuneus), and of the wide sensorimotor-auditory area (including the SMN, auditory cortex, posterior insula, central and parietal operculum, MCC, and SMA) with ANG/SMG and the posterior DMN (precuneous and PCC). We also observed reduced rsFC of the posterior DMN with ANG/SMG in treatment responders, although only when compared with HC.

This brief summary of results points at interactions between three networks—the DMN, the FPN, and primary sensory/motor regions—as being key to response to escitalopram in our study.

Our results suggest that the DMN, and in particular its posterior part, may be significant for treatment response, both when its connectivity as a network, and connectivity of its elements with other networks, is considered. Firstly, treatment R, as compared to NR, showed greater rsFC between the right FPN and the posterior DMN hub, precuneus, while there was no difference between R and HC (Fig. 2). Secondly, significantly reduced rsFC of the posterior DMN with the right ANG/SMG was observed in R compared to HC (Fig. 4), and was generally reduced in patients compared to HC (results not presented). Thirdly, significantly reduced rsFC of the extended somatomotor-auditory area including—mostly primary—sensory and motor regions, with the posterior DMN hub in the precuneus/PCC, and with the ANG/SMG, was seen in NR compared to HC (Fig. 3c), and generally in all patients versus HC (results not presented).

The DMN, in which abnormalities have been consistently observed in depression, is a network involved in a wide array of inward-directed mental activities, including self-referential processing, imagery and memory (Buckner et al. 2008). The involvement of the DMN is one of the most consistent findings regarding the physiopathology of depression (Dutta et al. 2014) and normalization of increased rsFC between key regions of the DMN may underlie antidepressant treatment response (McCabe et al. 2011). The DMN is organized into subsystems, which converge on the mPFC (anterior DMN) and precuneus/PCC (posterior DMN) (Buckner et al. 2008). While both parts of the DMN are involved in self-related processing, the anterior DMN shows greater bias towards self-other relations, and the posterior DMN—towards self-centred processing; it also has a role in involuntary (bottom-up) attentional processing, in particular related to visual and sensorimotor stimuli. The posterior DMN might have a particular role in treatment response, as suggested by our study and research by other groups, which observed normalization of the posterior DMN connectivity after successful antidepressants treatment (Li et al., 2013; Posner et al., 2013; Wang et al. 2014; Shen et al. 2015), while that of the anterior DMN remained unaffected (Li et al. 2013).

The FPN, composed of lateral prefrontal and posterior parietal cortices, is highly integrated with multiple brain networks, which allows a flexible and coordinated higher order modulation of cognitive and emotional processes (Marek and Dosenbach 2018, 2019). Our results show that greater temporal coherence between the FPN and the precuneus, similar to healthy individuals, was associated with favourable treatment response, which suggests that sustained FC between these structures may be essential for antidepressant response. This may be linked to control that the FPN exerts over the DMN (Cavanna and Trimble 2006; Chen et al. 2013; Marek and Dosenbach, 2019). FPN-DMN interactions are facilitated by rich cortico-cortical connections between the frontal and parietal lobes and, specifically, precuneus (Cavanna and Trimble 2006). Our findings are in line with already published research (e.g. Posner et al. 2013).

Good response to treatment was also associated with reduced connectivity of the DMN with the right ANG/SMG. The ANG/SMG are parts of the inferior parietal lobule (IPL), a complex region involved in complex functions across cognitive and emotional domains and supporting inter-network interactions thanks to hosting the nodes of major brain networks (Seghier 2013, Ingelström and Graziano, 2017). The area identified in our study corresponds to the location of the DMN node. The ANG and DMN activities are often found to be correlated, and our findings suggest that a decrease in this correlation may be meaningful for treatment response. The IPL shows strong lateralization of function, with the right IPL, identified in our study, particularly involved in attentional processing (Ingelström and Graziano, 2017). One hypothesis explaining better treatment response is that this reduced connectivity might facilitate directing attention away from the DMN and inward-directed processing, and towards other functions supported by the IPL. The role for the right ANG in treatment response has been reported in previous studies (Guo et al. 2013; Korgaonkar et al. 2014), with one showing its best performance in classifying the early improvement status among other identified, DMN and frontal, regions.

The third finding involving DMN in our study was that treatment NR were characterized by significantly decreased FC of the sensorimotor-auditory component extending across the SMN, primary auditory cortex (Heschl’s gyrus and planum temporale), posterior insula, central and parietal operculum, MCC and SMA. Reduced temporal correlation was observed with the posterior DMN (the precuneus and PCC), and the right ANG/SMG in the cluster partly overlapping with the one described in the preceding paragraph. The component extends across canonical resting-state networks and it might be argued that it would further subdivide if dimensionality was increased. However, the common functional denominator of the areas involved is their role in sensory and motor processing, which supports treating it as a functional entity. The areas constituting this component include, mostly primary, somatosensory and somatomotor areas, which, through perception and integration of the internal and external sensory data (related to touch, temperature, pain, proprioception, interoception, visceral sensations, as well as auditory, olfactory and gustatory stimuli) facilitate planning of complex and coordinated motor movements, their production, execution and control (Northoff 2016). These areas are also strongly involved in production of emotional states (Kropf et al. 2019). It has been proposed that sensory information from both internal and external environments is one of the key elements in emotion formation and that internal body sensations influence the subjective experience of emotion (Levenson 1999; Wiens 2005; Pace-Schott et al. 2019). Our results suggest that a certain level of sustained connectivity of these sensory and motor regions, with the posterior DMN and the ANG/SMG cluster in the IPL, is important for treatment response. This may be related to the role of posterior DMN and the right IPL in involuntary attentional processing, crucial for appropriate processing of the sensorimotor stimuli (Seghier 2013). Additionally, this component was also identified in another recent study using a data driven-approach (Martens et al. 2021). Additionally, this component was also identified in another recent study using a data driven-approach (Martens et al. 2021).

Importance of the sensory and motor processing, as well as of the FPN, for treatment response was again supported by an association of the failure to respond to escitalopram with reduced rsFC of the bilateral FPN with contralateral precentral and postcentral gyri, as compared to both R and HC, and with somatosensory association cortex, as compared to R. The FPN provides flexible modulation of the activity of other brain regions and its inadequate connectivity with primary and associative somatosensory and motor regions may affect how well this control will be exerted (Marek and Dosenbach, 2018, 2019). In depression, reduction in the SMN resting-state neuronal variability was linked to decreased modulation of spontaneous activity in the SMN by external stimuli (Martino et al., 2016; Conio et al. 2019). Apart from the importance of somatosensory processing for emotion formation (Kropf et al. 2019), at the behavioural level, this might translate into reduced responsiveness and psychomotor activation, common symptoms of depression (Martino et al. 2016). Aberrant FC of the CEN (corresponding to the FPN) with the pre- and postcentral gyri in depression was observed using the ICA approach (Dichter et al. 2015), and the role of FPN-SMN connectivity for treatment response was recently shown in a large multicentre study (EMBARC) (Chin-Fatt et al. 2020). Our study adds to the growing body of results suggesting a role for the somatosensory and motor processing in antidepressant response.

Previous research on the role of resting-state connectivity for treatment response in depression yielded a range of results, and although they are largely heterogeneous, a few general themes were suggested (for detailed review of findings see Dichter et al. 2015; Brakowski et al. 2017). The most consistent pattern was an association of treatment response with an increase in connectivity between frontal and limbic areas, in line with the importance of cognitive control over emotional circuits. We did not observe this in our study. Other themes included the role for visual recognition circuits and subgenual cingulate cortex connectivity. The review also suggested that treatment-sensitive patients may have lower connectivity within the DMN (Guo et al., 2012; Ma et al., 2012). More recent studies strongly provided further support for the role of the DMN and FPN in treatment response; interestingly, an involvement of the SMN and sensory regions has been a more common finding, reported, for example, by the large multicentre studies iSPOT (Korgaonkar et al. 2019) and EMBARC (Chin-Fatt et al. 2020). ISPOT indeed identified functional connectivity between the DMN, the FPN and the SMN as particularly important for treatment responsiveness, suggesting that elevated connectivity was generally associated with better treatment response, while hypoconnectivity at baseline distinguished NR from HC. This was in line with our findings, as failure to respond was generally associated with reduced connectivity as compared to HC. Despite emerging themes, findings vary largely between the studies. This heterogeneity may be caused by group differences regarding, for example, symptomatic pictures, chronicity, treatment status (e.g. drug-naive, drug-free, under treatment), or the length and type of treatment. Not less important are technical aspects, such as differences in data collection methods (e.g. the length of resting-state data acquisition, eyes-open vs. eyes-closed paradigms), and differences in data analysis approaches (e.g. ICA or seed-based, ReHo, graph theory). The picture becomes even more complex when not only network involvement is considered, but also details of the affected connectivity, e.g. related to the structures, with which network connectivity are altered. To address these questions, a larger number of studies, providing all their findings, followed by meta-analyses, are needed.

In this study, we performed post hoc correlational analysis to explore potential relationships between the extracted functional connectivity measures and subjective measures of mood and anxiety. The primary aim of both BDI and STAI is to measure the global construct of, respectively, depression and anxiety. Previous research however showed that BDI contains items that can be grouped as separate factors, which not only may be more sensitive to change (Beck et al., 1996) but, in line with dimensional approach, may also have different neural correlates, which may have bearing on treatment response (Nicholson and Sommer, 2018). Given that HAMD score change was used to define response, and FC measures assessed already survived multiple comparisons at the whole-brain level, significant correlations were expected for the change in HAMD score—and strongly correlated with it total BDI score. At the same time, correlations between change in scores in BDI subscales (affective, cognitive and somatic), and STAI-Trait total score, have the potential to provide new information about the relationship between distinctive groups of symptoms, their improvement and particular connections. Our data points at the particular role of both the right and left FPN connectivity—likely reflecting better cognitive control—for response across BDI dimensions and STAI. Interestingly, we were unable to show any relationship between the sensorimotor-auditory network and anxiety or somatic dimension of BDI. The results suggest that considering separate factors of multidimensional scales may have value for a better understanding of the role of particular connections in the context of specific symptomatic dimensions. It needs to be noted, however, that this is a post hoc exploratory analysis on a small group, and the subject would need to be explored in studies specifically designed to test it, using scales validate for testing some groups of symptoms, such as State-Trait Inventory for Cognitive and Somatic Anxiety (STICSA) (Grös et al. 2007).

Our study has both strengths and limitations. The main limitation is the small size of the group, which reduces the power and increases the likelihood of type 2 error. Also, prediction analysis, e.g. cross-validation, was not performed due to the limited sample size, which restricts the interpretation of the data. In the future, a replication of findings in a larger independent sample will be needed. An important strength was that the group consisted of carefully selected drug-free individuals with depression, who underwent an assessment by the same psychiatrist to increase homogeneity. Treatment was standardized, with the same medication and dose for all participants. This is a strength of the study, although it also means that conclusions can be drawn for escitalopram only. Another strength was related to analytical methods chosen. Our study is one of the few using ICA rather than seed-based approach. This allowed for an unbiased exploration of functional networks, not restricted by existing hypotheses. A limitation of such an approach is however the lack of exploration of rsFC of the structures, which role has been shown by other studies, yet which may not pass the statistical threshold, which is more stringent in whole-brain analysis than in a seed-based approach. The stringent statistical approach using nonparametric permutation testing and TFCE (Threshold-Free Cluster Enhancement) allows further confident interpretation of the results. However, it needs to be noted that due to the exploratory nature of this study, correction for multiple comparisons related to inclusion of multiple RSNs was not performed. One general limitation, shared by all functional connectivity studies, is related to the very nature of this approach, where the estimation of the functional connectivity is based on the temporal correlation of activities in various brain regions and does not provide information about anatomical connections nor directionality of the effect or causality. Hypotheses regarding the meaning of findings may however be proposed based on known anatomy and function, as well as interactions of structures involved.

In summary, patterns of resting-state functional connectivity related to escitalopram treatment response focused primarily on three networks, the DMN, the FPN and the SMN. This adds to the body of results supporting importance of network connectivity for treatment response. Future studies are needed that will meta-analyze data from different studies, accounting for potential sources of heterogeneity of the results.