Almost two decades ago, it was proposed that the pathophysiology of psychosis is more closely related to abnormal functional integration between specific brain regions rather than to changes in local brain activity (Friston and Frith, 1995; Friston, 1998). Numerous experimental investigations using functional magnetic resonance imaging (fMRI) and diffusion tensor imaging (DTI) have supported the disconnection hypothesis by showing abnormal structural and functional connectivity between a wide range of brain regions and across different stages of the psychosis spectrum (Kyriakopoulos and Frangou, 2009; Peters et al., 2010; Pettersson-Yeo et al., 2011; Yao et al., 2013; Samartzis et al., 2014). These findings provide evidence that abnormal brain connectivity develops even before the onset of psychosis and evolves along the psychosis continuum, which suggests that the assessment of the brain connectivity pattern may permit the detection of the early phases of the illness (Schmidt and Borgwardt, 2013). One of the most consistent findings is a reduction in FP connectivity (whether structural or functional) across all stages in the psychosis spectrum (Pettersson-Yeo et al., 2011), and this suggests that abnormalities in this pathway reflect a vulnerability marker for emerging psychosis (Schmidt et al., 2013, 2014a). Functional and structural interactions between FP brain regions are crucial for successful working memory (WM) processing (Gazzaley et al., 2004; Cole and Schneider, 2007; Gazzaley and Nobre, 2012) and it is therefore conceivable that abnormal FP connectivity underlies WM impairment in psychosis. These deficits in WM processing are of particular interest, as they precede the onset of psychosis by many years and in the absence of any psychotic symptom. They may therefore offer valuable predictions about the longitudinal course of the disease (Fusar-Poli et al., 2012a).

In the present article, we selectively review neuroimaging findings of FP connectivity across different stages of the psychosis continuum, including chronic psychosis, early psychosis, clinical high-risk patients, genetic high-risk patients, as well as the general population with psychotic symptoms (see Table 1 for the different phenotypes across the psychosis continuum). We used the term “psychosis” throughout the paper to refer to patients with schizophrenia spectrum psychoses, i.e., schizophrenia, schizophreniform disorder, schizoaffective, as well as other psychoses including unipolar or bipolar depression with psychotic symptoms. We first report structural FP connectivity findings (by means of fractional anisotropy values derived from DTI), as this may help the reader to better understand the findings on abnormal functional connectivity during the resting state and in WM-induced paradigms, which are presented subsequently. In the next brief section, we then provide evidence that abnormal FP connectivity may result from glutamatergic disturbances, in particular from changes in N-methyl-D-aspartate receptor (NMDAR)–dependent synaptic plasticity.

There are two main concepts as to how brain connectivity can be measured and modeled: one is based on graph theory describing the brain's network topology in terms of undirected connections (correlations) derived from structural or functional MRI. Graph theory analysis has emerged as a very helpful approach to infer complex network properties of the brain from a more global perspective. In this framework, brain networks can be understood as graphs that are composed of nodes representing neural elements, such as neurons or brain regions, that are linked by edges denoting structural (anatomical links) or functional connections. Their topological properties or patterns of connectivity can be expressed by a variety of parameters, such as node degree, clustering coefficient, path length, connection density, or modularity (Bullmore and Sporns, 2009; Rubinov and Sporns, 2010). For a comprehensive review of the contribution of modern network theory to the understanding of neurological diseases see Stam (2014). One of the major contributions of graph theory to our understanding of neurological and neuropsychiatric diseases has been in highlighting the important role of hubs, which are nodes of the network with an unusually high degree (number of connections). For details, see Rubinov and Bullmore (2013b), Van Den Heuvel and Sporns (2013), Crossley et al. (2014). These special nodes are often highly connected to each other, thus giving rise to the so-called rich-club organization. The second approach constitutes model-based formulations of context-specific effective (directed) connectivity that are based on the biophysics of neuronal interactions, such as dynamic causal modeling (DCM) (Park and Friston, 2013). In particular, DCM is a Bayesian identification scheme based on a model of neuronal interactions and an observation model (for fMRI, the haemodynamic model is used) (Friston et al., 2003). This is particularly useful in testing neurobiologically informed a priori hypotheses (different DCMs), which can be compared with model comparison procedures. Both of these approaches,—model-based investigations of effective connectivity and more global data-driven correlation approaches across the whole brain with graph theory—have already demonstrated network connectivity abnormalities in psychosis (Fornito et al., 2012; Dauvermann et al., 2014; Van Den Heuvel and Fornito, 2014) and are thus promising approaches to detect robust network connectivity endophenotypes for different stages of the psychosis continuum, although a number of challenges remain. In order to address these current limitations and challenges, in the last part of this paper we suggest some potential approaches in developing a network connectivity-driven classification of the psychosis continuum on the basis of non-invasive neuroimaging data.

Numerous studies provide evidence for reduced functional and structural FP connectivity across the psychosis spectrum, suggesting that this pathway is one of the most vulnerable for the emergence of psychosis. It is crucial that functional and structural FP connectivity measures in these regions are often progressively reduced from healthy controls via high risk subjects to patients with first episode psychosis and then with established psychosis. Such a symptom-based dissection of the psychosis continuum has recently been supported by more sophisticated whole-brain network analyses. However, several investigations also reported increased FP connectivity, mainly in genetic high risk subjects and non-help-seeking individuals with psychosis-like experiences. Further studies must investigate whether this increase reflects a compensatory mechanism. Table 2 gives a selective overview of functional and structural FP connectivity findings in each domain in the psychosis continuum.

In this brief chapter, we will discuss a potential pharmacological mechanism underlying aberrant FP connectivity in psychosis. The reductions in WM-induced modulation of FP connectivity have often been interpreted as being a result of the abnormal regulation of N-methyl-D-aspartate receptor (NMDAR)–dependent synaptic plasticity (Deserno et al., 2012; Schmidt et al., 2013). This interpretation is based on theoretical reflections proposing that disordered brain connectivity in psychosis results from abnormal regulation of NMDAR–dependent synaptic plasticity by modulatory transmitters like dopamine, acetylcholine, or serotonin (Stephan et al., 2009a). This regulatory effect of modulatory transmitters on NMDAR-dependent synaptic plasticity (for specific mechanisms underlying this modulation see Stephan et al., 2009a) distinguishes the dysconnection hypothesis from other pathophysiological theories of schizophrenia that postulate impaired NMDAR function alone (Javitt and Zukin, 1991; Olney and Farber, 1995). Numerous genes implicated in schizophrenia also converge on glutamatergic pathways (Allen et al., 2008; Walsh et al., 2008). For example, the glutamate-related genes NRG1 and DTNBP1 are involved in both building long-range connections during development and in regulating synaptic plasticity (Harrison and Weinberger, 2005). This is important, because any impairment in synaptic plasticity would affect the way long range connections are established in the developing brain, given that the strength of functional coupling between two neurons determines whether their connection survives developmental pruning (Hua and Smith, 2004).

With regard to the FP pathway, a recent study suggests that psychotic patients have glutamate-related dysregulation of the brain network supporting cognitive control (Falkenberg et al., 2014). The NMDAR antagonist ketamine induces psychotic-like experiences that are highly reminiscent of those observed in schizophrenia (Krystal et al., 1994, 2005; Schmidt et al., 2012), so it is intriguing that this also reduces diffusivity in the superior longitudinal fasciculus (Edward Roberts et al., 2014). It has further been demonstrated that ketamine reduced the negative correlation between the FP system and the default network in healthy volunteers, while the degree of this disruption predicted task performance (Anticevic et al., 2012). Furthermore, ketamine reduced accuracy on a spatial WM task and connectivity in a prefrontal network bilaterally, and region-specific reductions in connectivity were related to WM performance in healthy subjects (Driesen et al., 2013). These findings underpin the importance of the NMDAR for WM processing (Karlsgodt et al., 2011), in keeping with animal models reporting aberrant WM function after the inhibition of glutamatergic receptors (Timofeeva and Levin, 2011; Arnsten et al., 2012). However, recent data suggest that acute ketamine-induced alterations in brain network connectivity do not parallel those seen in chronic psychosis (Dawson et al., 2014). More evidence is needed to draw valid conclusions about glutamatergic synaptic physiology across different psychosis stages, in particular changes in NMDAR functioning.

Despite the considerable effort made in neuroimaging studies to detect abnormal connectivity patterns across different psychosis stages, there is no reliable neurobiological marker of any clinical utility in the psychosis spectrum to date. This can probably be explained on the one hand by significant differences in the analytical approaches across studies and by problems in the interpretation of the results on the other. We briefly present some of these limitations, although it is beyond the scope of this paper to comprehensively address all of them. In graph analysis, the selection of the included brain regions or nodes —the parcellation of the brain —is one of the most critical steps and hampers comparisons among studies. The nodes should represent distinct, functionally homogeneous neural elements or brain regions. However, in the absence of any gold standard for large scale parcellation of the brain, the nodes are typically arbitrarily defined by a variety of methods (Fornito et al., 2012). The nature of links (i.e., connections), whether binary or weighted, also influence connectivity findings (Rubinov and Sporns, 2010). Another general issue is that graph theory provides a bunch of connectivity indexes that can be used in brain networks; it will be important to understand the relations between them and how they are changed in psychosis (Bullmore and Vértes, 2013). Furthermore, structural findings (either analyzed with conventional or graph theory analysis) in patients are difficult to interpret due to the non-specific nature of FA values, i.e., it is unclear whether they reflect neuroinflammation or axonal degeneration (for a comprehensive review on pitfalls in terms of methods and interpretations of DTI findings see O'donnell and Pasternak, 2014). Free-water imaging on DTI data might provide a possible approach to elucidate the exact underlying microstructural basis of white matter changes in psychosis (Pasternak et al., 2012). Many investigations of functional connectomics in psychosis have been performed in the resting state (Alexander-Bloch et al., 2013; Argyelan et al., 2014) and only a few during tasks (Lord et al., 2011). It thus important to disentangle which (dys)connectivity only occurs in a specific psychological context (i.e., task-specific) and which are context-independent and reflect more general impairments (Fornito et al., 2012). In contrast to whole-brain graph analysis, the limitations of DCM analyses lie in the simplified neuronal network model, which often includes only a relatively small number of brain regions. The regions are included on the basis of task-induced activity and effective connectivity findings from DCM analysis and can thus only be compared across studies if exactly the same paradigms have been conducted. In other words, model-based assays can never fully capture a comprehensive neural network that fully considers a specific psychological process. The numbers of included regions are often limited, as model fitting is computationally demanding (Penny et al., 2004). Nevertheless, the restricted model space in DCM analysis can also be an advantage if the included regions comprise highly sensitive clinical data, which are not confounded by unimportant and noisy data from other regions (which is a potential danger of whole-brain graph analysis).

Furthermore, inconclusive findings from patient studies may often reflect the clinical heterogeneity of patient samples, the fact that the patients enrolled were treated with antipsychotic drugs (for different periods) (Fusar-Poli et al., 2013c) and other confounding factors such as comorbidities (Modinos et al., 2014). This is particularly relevant for ultra-high risk samples, as these exhibit high clinical heterogeneity (Fusar-Poli and Van Os, 2013; Simon et al., 2013, 2014; Fusar-Poli et al., 2013b). Furthermore, patient and control samples are often not well matched for intellectual functions and educational level, which might influence imaging results. Cross-sectional designs further hampered inferences about how brain connectivity abnormalities develop over the course of the illness. There are additional problems relating to publication and reporting biases of neuroimaging studies (Fusar-Poli et al., 2014a).

Further analytical steps are also needed to tackle these limitations and to improve the biological- and model-based diagnosis of psychiatric spectrum diseases (Cuthbert and Insel, 2010; Stephan and Mathys, 2014), with the incorporation of the findings on FP connectivity (cf. Table 2). Figure 1 gives a speculative graphical overview of these critical key steps to achieving a network connectivity-driven classification of the psychosis continuum.

In the first and very important step, we have to elucidate whether functional dysconnectivity in psychosis has a structural basis and how such a putative structure-function relation characterizes each domain of the psychosis continuum, i.e., chronic psychosis, early psychosis, clinical high risk, genetic high risk and general population with psychotic experiences. For this structural-functional approach, multiple imaging modalities are required to explore not only connectivity properties within a specific network such as the FP system, but also to assess comprehensive connectivity maps across the whole brain. For model-based investigations of effective connectivity, i.e., DCM analysis, the definition of the model space should not only be motivated by task-induced brain activity and between-group differences in these activities, but also by evidence for abnormalities in structural connectivity in specific regions. This is important, as it is likely that functional connectivity over long time periods of 5 to 10 min reflects underlying structural connectivity (Park and Friston, 2013). Graph theoretical network analysis also offers promising perspectives for this purpose (Bullmore and Sporns, 2009; Filippi et al., 2013). For example, a recent network study found increased coupling between structural whole-brain connectivity and resting state functional connectivity in patients with psychosis. The increased correlation may suggest that the illness leads to functional interactions that are more directly related to the underlying anatomical connectivity of the brain and may be indicative of more stringent and less dynamic brain function in patients (Van Den Heuvel et al., 2013). Another study found that psychotic patients showed decreased functional connectivity and impaired white matter integrity in a distributed network encompassing frontal, temporal, thalamic, and striatal regions (Cocchi et al., 2014). Along these lines, recent work summarizing results from structural correlation studies, diffusion-imaging tractography studies, and functional correlation studies found strong evidence for network abnormalities of prefrontal hubs, and moderate evidence for network abnormalities of limbic, temporal, and parietal hubs (Rubinov and Bullmore, 2013b).

In a second step after the detection of connectivity pathways that are vulnerable with respect to structure and function, specific imaging methods can be applied to elucidate glutamatergic transmission in these pathways. For instance, positron emission tomography challenges might be used to explore specific receptor densities at putative vulnerability connections (e.g., the NMDAR profile at frontal hubs). Within this perspective, it has been shown that functional connectivity from the prefrontal cortex to thalamus and nucleus accumbens correlated positively with prefrontal glutamate concentrations. The correlations involving prefrontal glutamate and prefrontal-related functional connectivity were mirrored by correlations with structural connectivity (Duncan et al., 2013). Linking such experimental investigations with computational models may help to bring together evidence from the receptor to the system levels in psychosis (Adams et al., 2013; Anticevic et al., 2013a).

In a third step, it should be studied how these comprehensive network connectivity phenotypes are associated with behavioral, cognitive and genetic phenotypes across each domain of the psychosis continuum. This may allow the detection of potential candidate network circuits with particularly high clinical sensitivity. Such studies have already been performed, for example on the relation between abnormal brain signals, cognition and chemistry in early stages of psychosis (Schmidt and Borgwardt, 2014). Having in mind that the psychosis continuum describes a trajectory with increasingly severe symptoms, including cognitive impairments (Yung et al., 2012; Fusar-Poli et al., 2013a), a robust relation between aberrant network organization and cognitive functioning might facilitate the dissection of different illness stages. For example, altered network organization among frontal and parietal brain areas in schizophrenia patients has been related to WM performance (Bassett et al., 2009), whereas global connectivity in the lateral prefrontal cortex, involving connections both within and outside the FP network, predicts cognitive control and intelligence (Cole et al., 2012). Variation at one RGS4 single nucleotide polymorphism that has been previously associated with psychosis (rs951436) impacts fronto-parietal and fronto-temporal network coupling during WM and results in regionally specific reductions in the structural volume of gray and white matter in individuals carrying the A allele (Buckholtz et al., 2007). Another psychosis-related gene, the ZNF804A rs1344706, is also associated with prefrontal brain connectivity (Mothersill et al., 2012).

In a final determining step, we propose exploration of whether the pattern of network connectivity differs between domains in the psychosis continuum. One possible approach for disease classification and prediction would be to use supervised learning algorithms such as support vector machine (SVM), which have already been applied in early psychosis detection (Koutsouleris et al., 2012, 2014; Borgwardt et al., 2013). However, applying multivariate methods with whole-brain connectomics data requires some way of reducing dimensionality, as the connectomes are massively large. This problem has been addressed in a recent study, that employs a multivariate approach based on the SVM and regularization methods that exploit the 6-D spatial structure of the functional connectome (Watanabe et al., 2014). In this regard, a very recent study demonstrated that structural connectomes from DTI data in combination with functional connectomes from resting state fMRI data provided classification accuracy of 100% between healthy controls and chronic psychotic patients (Zhu et al., 2014). Another conceivable procedure, which actually only requires an additional step before running classification algorithms, is to use all these neurophysiological features (structure, function and chemistry) and formulate biophysiologically informed generative models (e.g., DCMs) (Stephan and Mathys, 2014), which, with formal selection procedures such as Bayesian model selection (Stephan et al., 2009b), might lead to dissection of the psychiatric spectrum diseases. For example, one could build different FP network models in which cognitive load modulates connectivity strength between the regions in the model at different sites. These models could then be fed into a multivariate classification algorithm. At best, the unsupervised classifier may find some connectivity clusters, which separate subjects into clinically meaningful subgroups. A very recent study provides a concrete example for such an approach by demonstrating how psychotic patients can be dissected into three subgroups defined in terms of neurophysiological mechanisms specified by a generative model of network dynamics (DCMs). It is impressive that the three neurophysiologically defined subgroups mapped onto three clinically distinct subgroups, distinguished by significant differences in negative symptom severity (Brodersen et al., 2013). In other words, instead of simply entering all possible neurophysiological features into the classification algorithm, all this information should be included in a generative model, such as DCM. Given that functional connectivity is highly constrained by structural connectivity (Park and Friston, 2013) and that structural connectivity can be influenced via NMDAR-dependent synaptic plasticity (Stephan et al., 2006), one should include only brain regions in the generative model from which such alterations have been reported in psychosis patients. In such an approach, all neurophysiological parameters are embedded in one model. By applying multivariate classifiers to measures of gray matter volumes, more than 80% accuracy has been achieved in distinguishing clinical high-risk subjects with later transition to psychosis from those without transition (Koutsouleris et al., 2012). Given that the accuracy of distinction between healthy controls and psychotic patients increases when using DCM-based effective connectivity estimates (78%), rather than functional connectivity (62%) or local activity estimates within these regions (55%) (Brodersen et al., 2013), feeding formally described mechanisms of network connectivity phenotypes into classification approaches should be more accurate than using simple isolated features. Replication and large-scale longitudinal studies are then required to validate diagnostic reliability and treatment responses.

Brain imaging may provide a powerful tool to improve the specificity and validity of an early diagnosis and to sustain preventive intervention prior to the onset of illness (Fusar-Poli et al., 2012b). Although no reliable neuroimaging marker of any clinical utility in the psychosis spectrum has yet been established, the assessment of network connectivity information from fMRI and DTI data might bring us one step closer to this goal. The present paper tried to demonstrate this potential using the example of FP network connectivity. One promising approach for the detection of network connectivity endophenotypes is the computation of whole-brain connectivity patterns using graph theory. This technology allows the study of network properties from a global perspective that may help to detect potential network endophenotypes for psychosis spectrum disorders (data-driven approach) (Van Den Heuvel and Fornito, 2014). The organization of network hubs seems to be especially important, given that recent studies have suggested that abnormal connectivity of brain hubs may be a core aspect of the disorder (Rubinov and Bullmore, 2013a,b; Van Den Heuvel and Sporns, 2013). An alternative strategy is to search endophenotypes in terms of biophysically constrained models of effective connectivity (e.g., DCM). This procedure is particularly useful in testing synaptic mechanisms that might underlie aberrant patterns of correlations (hypothesis-driven approach). This is important, as pharmacological treatments exert their effect at the synaptic level. It is important that these two concepts should be applied in a complementary fashion (Park and Friston, 2013). For example, graph theoretical analysis of DTI data can be used to detect affected brain circuitries in terms of abnormal structural connectivity properties, which at best also show a relation to clinical endophenotypes. These clinically relevant findings can then be used to formulate the model architecture of generative network models, i.e., the model of effective connectivity is constrained by structural connectivity evidence from graph theory. This improves the biological realism of the model and facilitates a mechanistic understanding of the signals that may increase the accuracy of machine learning engines, as recently suggested (Varoquaux and Thirion, 2014). It is quite clear that extensive work is needed to test the predictive and diagnostic validity of these network techniques, including longitudinal study design and large and well-matched study samples.