Peripheral Inflammatory Subgroup Differences in Anterior Default Mode Network and Multiplex Functional Network Topology are Associated with Cognition in Psychosis

Abstract

INTRODUCTION:

High-inflammation subgroups of patients with psychosis demonstrate cognitive deficits and neuroanatomical alterations. Systemic inflammation assessed using IL-6 and C-reactive protein may alter functional connectivity within and between resting-state networks, but the cognitive and clinical implications of these alterations remain unknown. We aim to determine the relationships of elevated peripheral inflammation subgroups with resting-state functional networks and cognition in psychosis spectrum disorders.

METHODS:

Serum and resting-state fMRI were collected from psychosis probands (schizophrenia, schizoaffective, psychotic bipolar disorder) and healthy controls (HC) from the B-SNIP1 (Chicago site) study who were stratified into inflammatory subgroups based on factor and cluster analyses of 13 cytokines (HC Low n=32, Proband Low n=65, Proband High n=29). Nine resting-state networks derived from independent component analysis were used to assess functional and multilayer connectivity. Inter-network connectivity was measured using Fisher z-transformation of correlation coefficients. Network organization was assessed by investigating networks of positive and negative connections separately, as well as investigating multilayer networks using both positive and negative connections. Cognition was assessed using the Brief Assessment of Cognition in Schizophrenia. Linear regressions, Spearman correlations, permutations tests and multiple comparison corrections were used for analyses in R.

RESULTS:

Anterior default mode network (DMNa) connectivity was significantly reduced in the Proband High compared to Proband Low (Cohen’s d=−0.74, p=0.002) and HC Low (d=−0.85, p=0.0008) groups. Inter-network connectivity between the DMNa and the right-frontoparietal networks was lower in Proband High compared to Proband Low (d=−0.66, p=0.004) group. Compared to Proband Low, the Proband High group had lower negative (d=0.54, p=0.021) and positive network (d=0.49, p=0.042) clustering coefficient, and lower multiplex network participation coefficient (d=−0.57, p=0.014). Network findings in high inflammation subgroups correlate with worse verbal fluency, verbal memory, symbol coding, and overall cognition.

CONCLUSION:

These results expand on our understanding of the potential effects of peripheral inflammatory signatures and/or subgroups on network dysfunction in psychosis and how they relate to worse cognitive performance. Additionally, the novel multiplex approach taken in this study demonstrated how inflammation may disrupt the brain’s ability to maintain healthy co-activation patterns between the resting-state networks while inhibiting certain connections between them.

INTRODUCTION

Peripheral inflammation has been linked to clinical symptoms, cognition, and neuroanatomical changes in psychotic disorders (Kose et al., 2021; Morrens et al., 2022). The underlying pathophysiology of peripheral inflammation in psychotic disorders is theorized to be initiated through either microglial induction of pro-inflammatory responses in the brain (Bishop et al., 2022, p. 2022; Marques et al., 2019; Monji et al., 2009), breakdown of the blood-brain barrier (BBB) (Futtrup et al., 2020; Pong et al., 2020) or the blood-cerebral spinal fluid barrier (Bitanihirwe et al., 2022; Lizano et al., 2019) through inflammatory effects on brain microvascular endothelial cells or choroid plexus epithelial cells, respectively.

Peripheral markers of inflammation have been previously associated with cognition in psychosis, but the underlying etiology and mechanism of this effect is unclear. A recent meta-analysis of psychotic and mood disorders (n=75 studies), including 42 studies on schizophrenia, found that higher levels of C-reactive protein (CRP) correlated with worse global cognitive performance, visual memory, verbal memory, working memory, reasoning and language, while higher levels of interleukin-6 (IL6) were associated with worse global cognitive performance and processing speed (Morrens et al., 2022). Our group found similar results for CRP and IL6 using the composite scores in the Brief Assessment of Cognition in Schizophrenia (BACS) (Lizano et al., 2021). The association between inflammation and psychosis symptom severity has been mixed, but a recent systemic review found that higher levels of CRP and IL6 were related to worse clinical outcomes and greater symptomatic deterioration (Kose et al., 2021), while our group did not identify a significant correlation between inflammation and symptom severity (Lizano et al., 2021). Investigations of the neuroanatomical consequences of peripheral inflammation in psychosis have identified heterogeneity in structural alterations depending on the specific markers and brain regions examined. Kose et al conducted a recent meta-analysis concluding that elevated levels of specific inflammatory markers were associated with smaller hippocampal and amygdala volumes and reduced cortical thickness in frontal, temporal, and occipital regions (Kose et al., 2021). Our research group identified associations between elevated levels of IL6 and larger choroid plexus volume (Lizano et al., 2019), as well as a correlation between higher levels of CRP and lower V2 (extrastriate visual cortex) thickness (Türközer et al., 2022). These findings support hypotheses that dysregulated inflammation may be linked to cognitive and neuroanatomical changes in psychosis.

In recent years, there has been accumulating evidence supporting the existence of a subgroup of individuals with psychosis with exacerbated pro-inflammation status. This has become a consistent finding, particularly when using multiple inflammatory markers to assess a broader landscape of inflammation pathways as compared to analyses focusing on single cytokines (Bishop et al., 2022; Fillman et al., 2013, 2014, 2016; Tamminga et al., 2021). Several cytokine approaches have been used to define inflammatory subgroups using serum, plasma, or white blood cells from patients, with additional studies using postmortem tissue, which have collectively been essential in parsing the heterogeneity of psychosis. Subgrouping approaches to date have involved the use of composite scores (Noto et al., 2019), recursive two-step clustering (Fillman et al., 2013), principal component analysis (Elkjaer Greenwood Ormerod et al., 2022), canonical correlation analysis (Sæther et al., 2022), machine learning (Enrico et al., 2023), or a combination of factor analysis with hierarchical clustering (Hoang et al., 2022; Lizano et al., 2021) on a combination of cytokine markers to create subgroups with a focus on immune, inflammatory, infection, BBB, and/or microglial function. Using these approaches, it has been consistently estimated that approximately 30–50% of people with psychosis can be classified into an elevated inflammatory subgroup (Boerrigter et al., 2017; Fillman et al., 2013, 2014, 2016; Hoang et al., 2022; Lizano et al., 2021). Additionally, subgroups with higher inflammation have been demonstrated to have worse cognitive performance compared to their low inflammation counterparts in chronic psychosis (Fillman et al., 2016; Lizano et al., 2021; Sæther et al., 2022), but not first episode psychosis (FEP) (Enrico et al., 2023; Hoang et al., 2022), which may have been due to the small sample size.

A potential underlying mechanistic explanation for these cognitive deficits may be related to neuroanatomical alterations and related functional consequences. We previously demonstrated that a psychosis subgroup with a high pro-inflammatory signature performed significantly worse on visuo-spatial working memory, response inhibition, and overall cognition. These findings overlapped with neuroanatomical changes associated with these cognitive domains, including elevated hippocampal, amygdala, putamen and thalamus volumes, and gray matter thickening in fronto-temporal-parietal regions compared to the low inflammatory psychosis group (Lizano et al., 2021). These structural changes were replicated in a FEP study, where the high inflammatory FEP group showed thickening of the parahippocampal, caudal anterior cingulate, and banks of the superior temporal sulcus, but the cognitive alterations between subgroups found in our chronic psychosis population were not replicated in the FEP study sample (Hoang et al., 2022; Lizano et al., 2016, 2017, 2021). We further expanded the neuroanatomical findings in FEP inflammatory subgroups by performing graph theoretical analyses and found that the high inflammation group had greater centrality (important brain hubs) and a mix of high and low segregation (ability for specialization) and integration (ability to rapidly synthesize information) compared to the low inflammation group (Hoang et al., 2022). While these findings have been critical in our understanding of the potential effects of peripheral inflammation on brain structure, the impact that inflammatory signatures have on resting-state and functional network dysfunction in psychosis has not yet been investigated. This represents an opportunity to clarify the biological mechanisms underlying behavioral consequences of altered inflammatory processes in psychotic disorders.

The premise for the potential importance of functional network dysfunction relationships with inflammation alterations have been established via immune challenge studies in healthy individuals which found relationships with a variety of behavioral (Eisenberger et al., 2009; Inagaki et al., 2012), cognitive (Eisenberger et al., 2010; Kullmann et al., 2013, 2014; Moieni et al., 2019; Muscatell et al., 2016) and resting-state network changes (Labrenz et al., 2016, 2019; Lekander et al., 2016). Studies of transient BBB disruption induced with focused ultrasound also identified network changes (Meng et al., 2019; Todd et al., 2018). Changes within the default mode network (DMN) have been associated with greater IL6 or CRP (Dev et al., 2017; Marsland et al., 2017). Three depression studies have determined that higher levels of CRP are associated with lower DMN connectivity (Aruldass et al., 2021; Felger et al., 2016; Kitzbichler et al., 2021), while only one study used a multi-cytokine approach in patients with post-traumatic stress disorder (PTSD) (Kim et al., 2020). While these studies are promising they have yet to be examined in patients with psychotic spectrum disorders. We examined the dimension of psychosis since our group has previously demonstrated the validity and importance of transdiagnostic approach that transcend traditional diagnostic boundaries of psychotic disorders (Reininghaus et al., 2019).

In this present study, psychosis spectrum disorder patients from the Bipolar-Schizophrenia Network of Intermediate Phenotypes 1 (B-SNIP1) Chicago site were examined to determine the association between inflammation and functional connectivity alterations (Lizano et al., 2021; Tamminga et al., 2013). As noted above, our group has previously identified a high inflammatory psychosis subgroup associated with lower cognitive performance and thickening of cortical and subcortical structures as compared to the low inflammatory psychosis subgroup (Lizano et al., 2021). Separately, our group has demonstrated that nine resting-state networks were affected in psychosis spectrum disorder patients compared to healthy controls (Meda et al., 2016). This study is the first to investigate relationships between inflammatory subgroups in psychosis and brain functional network connectivity based on our prior work (Meda et al., 2016) along with their cognitive consequences in psychosis. Furthermore, we incorporate a novel Multilayer (network) analysis approach to investigate the relationship between the positive and negative functional connections, which are traditionally examined separately (Mijalkov et al., 2022). This approach allows for a greater understanding of the network topology associated with inflammation in psychosis spectrum disorders. We hypothesized that the high inflammation psychosis subgroup will be associated with lower DMNa activation, DMNa functional connectivity, and negative network topology, which will correspond with worse verbal fluency, verbal memory, working memory, and overall cognition compared to the low inflammatory psychosis subgroup.

MATERIALS AND METHODS

Study Participants

Participants were recruited from the Chicago site of the B-SNIP study, where they provided blood samples and underwent clinical, cognitive, and neuroimaging assessments (Tamminga et al., 2013). All participants gave written informed consent before entering the study. The study was approved by the institutional review board and was conducted in accordance with the declaration of Helsinki. For inclusion into the study, participants met the following criteria: 1) age between 18 and 65 years old, 2) no history of substance dependence in the prior 6 months, 3) negative urine toxicology screen and 4) no history of major neurological disorders. Consensus diagnoses for probands, either schizophrenia spectrum illness (SZ) or bipolar 1 disorder with psychosis (BPP), were obtained after reviewing the Structured Clinical Interview for DSM-IV-TR (SCID-IV), clinical interviews and psychiatric histories. Psychosis probands were clinically stable without major medication changes for at least 4 weeks. Healthy controls had no personal history of psychotic disorders and no current depression or history of recurrent depression. Healthy controls were excluded if they had any known family history of psychotic disorders.

Individual status on cardiometabolic disease (CMD) status was collected through patient report and/or medical record information. CMD status was determined based on the presence of cardiovascular or metabolic disorders including, coronary heart disease, hypertension, diabetes, hyperlipidemia, or other related disorders. Information about cardiometabolic disease status was collected because of its reported correlations with peripheral inflammation (Lopez-Candales et al., 2017). While Body Mass Index was also previously correlated with cardiometabolic disease and peripheral inflammation, that information was not available in this study (Ebron et al., 2015; Lyall et al., 2017).

Genetic ancestry was determined as an objective measure in addition to collecting self-reported race, which may have variability in accuracy. Participants’ ancestry was determined through high density genotype data available on the participants collected and analyzed in other B-SNIP studies (Alliey-Rodriguez et al., 2019; Lencer et al., 2017). The first two genetic ancestry principle components (predominantly African and European ancestry) were used as covariates in the peripheral inflammatory marker analyses (Lizano et al., 2021).

All participants underwent a medication history interview for both prescribed and unprescribed medications. To standardize antipsychotic dosage for comparison, average daily doses were calculated into chlorpromazine equivalents (Andreasen et al., 2010). Medication status (yes/no) but no dosage information was collected for anti-depressants, lithium, and non-steroidal anti-inflammatory drugs (NSAIDs).

Blood samples were collected and processed from 200 participants at the B-SNIP1 Chicago site, which included participants with schizophrenia spectrum illness (n=79; SZ n=50, schizoaffective disorder (SAD) n=29), bipolar 1 disorder with psychosis (n=61), and healthy controls (HC, n=60). Healthy controls were similar to the psychosis group in terms of sex and race.

Clinical Scales

Symptom ratings for psychosis probands were collected through the Positive and Negative Syndrome Scale (PANSS), the Young Mania Rating Scale and the Montgomery-Asberg Depression Rating Scale (Kay et al., 1987; Montgomery & Asberg, 1979; Young et al., 1978). Symptom ratings for all participants, including both healthy controls and probands, included the Global Assessment of Functioning (GAF) and the Birchwood Social Functioning Scale (SFS) (Birchwood et al., 1990; Jones et al., 1995).

Cognitive Scales

In addition to clinical assessments and symptom ratings, participants underwent a brief cognitive assessment. The Wide Range Achievement Test-IV (WRAT-IV): Reading subtest was used to estimate premorbid abilities (Gladsjo et al., 1999). The BACS was used as the primary cognition outcome for the analyses and was age and sex adjusted. The BACS composite score is based on measures of verbal memory and fluency, motor speed, attention, working memory, and executive functions, all of which are considered impaired in patients with schizophrenia. Several studies have previously indicated that probands perform worse than healthy controls on the BACS (Hill et al., 2013, 2015; Keefe et al., 2004; Reilly et al., 2017).

Peripheral inflammatory markers

Each participant gave a 10ml blood sample in Becton Dickinson vacutainer red top tubes used for serum isolation. A standard protocol for serum isolation was followed, involving 1) 30-minute incubation at room temperature, 2) centrifugation at 3000 rpm for 15 minutes at 4°C, 3) pipetting serum into 300 L aliquots, and 4) storing samples at −80°C until they were assayed. Information was collected on hemolysis score and number of storage days to ensure validity of samples. The peripheral inflammatory markers assayed and analyzed for this study were selected based on previous studies identifying dysregulation related to disorders of psychosis (Buckley, 2019; Buckley et al., 2007; Fernandes et al., 2016; Goldsmith et al., 2016; Lizano et al., 2016, 2021; Miller et al., 2014; Misiak et al., 2018; Sekar et al., 2016).

All assays were performed blind to diagnostic status, run at recommended dilutions and performed in duplicate. Patient samples were randomized with constraints to 1 of 5 multiplex plates and were balanced for sex, ancestry and diagnosis. V-Plex sandwich immunoassays from Sector 6000 Microplate enzyme-linked immunosorbent assay (ELISA) system and Meso Scale Discovery were used to determine serum concentrations of both microvascular and inflammatory markers: IL1α, IL1β, IL2, IL4, IL6, IL8, IL10, IL12/IL23p40, IL12p70, interferon gamma (IFNγ), Tumor Necrosis Factor alpha (TNFα), TNFβ, CRP, Fms Related Receptor Tyrosine Kinase 1 (Flt-1), VEGF, VEGFC, VEGFD, and transforming growth factor beta 1 (TGFβ1) (Lizano et al., 2021). These assays were sourced from Meso Scale Diagnostics in Rockville, Maryland. In addition to these immunoassays, a solid-phase sandwich ELISA was used to determine Complement 4 (C4a) concentrations. The solid-phase sandwich ELISA was sourced from Beckton, Dickinson and Company BD Biosciences in San Jose, California.

We used the same inflammatory marker pre- and post-processing steps described in our prior study (Lizano et al., 2021). In brief, 19 markers were examined, but IL1α, IL2, IL4, IL12p70, VEGFC, and TGFβ1 were excluded due to quality control metrics. The remaining 13 markers were adjusted using the intercept method for sample set, storage days, and hemolysis scores. These markers underwent winsorization to 3 standard deviations to normalize any outliers. Some markers were associated with age, sex, or ancestry and these variables were used as covariates. Medications and CMD status were not significantly associated with the peripheral inflammatory markers and thus were not used as covariates for subsequent analyses. Inflammatory subgroups were identified using a combination of principal component analysis (PCA) and subsequent hierarchical clustering (Hoang et al., 2022; Lizano et al., 2021) on the 13 markers (IL1β, IL6, IL8, IL10, IL12/IL23p40, IFNγ, TNFα, TNFβ, CRP, Flt-1, VEGF, VEGFD and C4a). A five principal component solution was identified and Factor 1 (loaded highly by CRP, IFNγ, IL1β, IL8, IL10, TNFα, and VEGF) offered the best clustering solution using the Wards method (agglomerative coefficient of 0.999) and confirmation with the Silhouette method (score = 0.59).

Functional neuroimaging acquisition and processing

A subset of the sample (32 HC Low, 65 Proband Low, and 29 Proband High) had complete resting-state fMRI data after removing 7 HC High individuals due to the small sample size in that group. A single 5-minute run of resting-state fMRI was performed using a 3T MRI scanner (GE Signa HDX) while participants focused their eyes on a crosshair on a monitor and remained still. Head motion inside the MRI scanner was restricted by using a head coil cushion. Image processing techniques were previously described (Allen et al., 2011; Meda et al., 2014, 2016).

FMRI data preprocessing was performed using a combination of toolboxes (AFNI (http://afni.nimh.nih.gov/), Statistical Parametric Mapping program (SPM8; http://www.fil.ion.ucl.ac.uk.spm), and custom code written in Matlab. We performed rigid body motion correction using the INRIAlign toolbox in SPM to correct for subject head motion followed by slice-timing correction to account for timing differences in slice acquisition. Then the fMRI data were despiked using AFNI3s 3dDespike algorithm to mitigate the impact of outliers. The fMRI data were subsequently warped to a Montreal Neurological Institute (MNI) template and resampled to 3 × 3 × 3 mm isotropic voxels. Instead of Gaussian smoothing, we smoothed the data to 6 mm full width at half maximum (FWHM) using AFNI’s BlurToFWHM algorithm which performs smoothing by a conservative finite difference approximation to the diffusion equation. This approach has been shown to reduce scanner specific variability in smoothness providing “smoothness equivalence” to data across sites (Calhoun et al., 2004; Friedman et al., 2008). The subject specific TCs obtained from group Independent Component Analysis (ICA) were detrended, orthogonalized with respect to estimated subject motion parameters, and then despiked. The despiking procedure involved detecting spikes as determined by AFNI’s 3dDespike algorithm and replacing spikes by values obtained from third order spline fit to neighboring clean portions of the data. The despiking process reduces the impact/bias of outliers on subsequent FNC measures (Allen et al., 2011).

Then an ICA was applied to the preprocessed fMRI data using the GIFT toolbox version 1.3f. This approach was used since it offers a more flexible, comprehensive, and unbiased approach for identifying brain networks compared to choosing a priori templates. Nine functional connectivity networks, including the cuneus-occipital (Visual Network), right frontoparietal (Attentional Working Memory Network, rAWN), left frontoparietal (lAWN), cerebellar-occipital (Visuo-Motor Integration Network, VMIN), anterior DMN (DMNa), inferior posterior DMN (DMNip), superior posterior DMN (DMNsp), fronto-temporal motor (Somatosensory Network, SSN), and salience (SAN) networks were used in this study based on previously published results from our group and others (Kim et al., 2020; Meda et al., 2016). We used Kim et. al’s inter-network technique (2020), where the loading coefficients for the networks were Fisher z-transformed to develop a 9 × 9 correlation matrix for each subject. The values within this matrix correspond to the correlation coefficients between the networks, where higher values relative to the overall sample represent greater inter-network connectivity and vice versa. Positive and negative values represent positive or negative coupling of networks. Moderator analyses identified age, framewise displacement (FD power), and signal to noise ratio (SNR) as significantly (p<0.05) different between groups, and for this reason they were used as covariates in our models (Supplementary Table 1). CMD status, age of disease onset, and medications did not correlate with resting-state networks and were not used as covariates in subsequent analyses.

Multiplex Analysis

The field of network neuroscience is increasingly utilizing complex network measures to determine brain functional connectivity (van den Heuvel & Hulshoff Pol, 2010). Network topology is typically examined by considering only the positive correlations and excluding the negative correlations from the analyses (Fornito et al., 2013), but the BRAPH software (Mijalkov et al., 2017) allows examining positive and negative connections separately as well as together, by including them as separate layers in a multilayer network. Our group recently used this technique to examine sex-related differences in aging in healthy individuals (Mijalkov et al., 2022).

Whole-brain functional connectivity patterns were modeled for each participant as a nine-node network, with nodes representing the different resting-state functional networks and the connection between the nodes being determined by positive and negative correlations between their activation signals. First, whole-brain networks were divided into two separate networks consisting of only positive connections (positive correlations or co-activations) or only negative connections (anti-correlations). The topological organization of these networks was then analyzed independently using the measures of global efficiency and clustering coefficient, which provide information about the degree to which the network promotes functional integration and segregation respectively (Mijalkov et al., 2017). Functional integration describes the process by which the different brain regions communicate with each other and coordinate their activity to perform cognitive tasks, including perception, attention, memory, as well as decision making. On the other hand, functional segregation refers to the specialized functions of the distinct brain regions in order to process specific types of information and perform specific functions (Bullmore & Sporns, 2009; van den Heuvel & Hulshoff Pol, 2010). To further understand relationships between the networks of positive and negative connections, and whether they can offer additional insights into the brain organization of different groups, we integrated them as separate layers into a two-layer multiplex network (Mijalkov et al., 2022). The organization of this multiplex network was examined using the multiplex participation coefficient, which provides an estimate of the similarity of the connectivity patterns in both layers (Battiston et al., 2014). Since the values of the topological measures in weighted networks depend on the number and strength of connections within the network (Fornito et al., 2013), all topological measures were normalized (divided by the number/strength of the connections) by both quantities before performing the group comparisons and correlation analyses. Furthermore, as network measures are not defined for negative weights (Fornito et al., 2013), the weights in the network of negative connections were substituted by their absolute values, thus network measures were equivalent for both positive and negative connections networks. Self-connections were excluded from the analysis by setting them to zero. Both single-layer and multiplex measures were calculated with the BRAPH 2.0 software (Mijalkov et al., 2017)(http://braph.org/).

Network organization of single-layer networks

Clustering coefficient and global efficiency were used because they reflect two important network organization properties: integration and segregation. While there are many more graph measures to assess, these two properties were selected because they are the most commonly used and would give an overall picture of network organization.

Clustering coefficient

The clustering coefficient of a given node reflects the proportion of that node’s neighbors that are also connected to each other (Mijalkov et al., 2017; Onnela et al., 2005). It can be calculated as the ratio of the number of closed triangles (connections between three nodes) that are present around the node, to the total number of possible triangles that could be theoretically formed around the node. As such, the nodal clustering coefficient provides an estimate of the density of local connections (for example, it increases with the increasing number of local connections) and how these connections are organized around that node. The network (or global) clustering coefficient is the average of the clustering coefficients for all the nodes in the network. It can be interpreted as a measure of network segregation that reflects the network’s ability to perform specialized tasks due to the presence of densely connected local clusters.

Global efficiency

The distance between any two nodes in a network is the shortest path that needs to be taken to travel between the two nodes. In a weighted network, the shortest path is defined as the path with the smallest sum of edge weights, which may not necessarily be the path with the fewest number of edges. Then, the global efficiency of a given node is calculated as the average of the inverse distances from that node to all other nodes in the network (Latora & Marchiori, 2001; Mijalkov et al., 2017). Therefore, a node with a higher global efficiency is considered to be well-integrated within the network and can participate in a more efficient transfer of information by having shorter paths to other network nodes. The global efficiency of a network is calculated as the average of the global efficiencies of the individual nodes. It is a measure of network integration and is an estimate of the ability of the network to facilitate efficient communication and information processing between any pair of network nodes.

Network organization of multiplex networks

For each participant, a two-layer multiplex network was created with one layer representing the network of positive connections and the other layer representing the network of negative connections. The two layers were connected by inter-layer connections that were allowed only between the same nodes in both layers (Mijalkov et al., 2022). Multiplex participation coefficient was used to evaluate the balance between the positive and negative connections in the multiplex network (Battiston et al., 2014). The nodal participation coefficient measures the heterogeneity of the connectivity patterns of a node across layers, determining whether the node’s connections are evenly distributed across the layers or concentrated in one layer. Specifically, the multiplex participation coefficient can take on values from 0 to 1, with a value of 1 indicating that the node has evenly distributed connections in both layers and a value of 0 indicating that the node has no overlapping connections in both layers. The multiplex participation coefficient for a network is defined as the average of the nodal participation coefficients for all the nodes in the network.

Statistical analysis

All statistical analyses were performed in R version 4.1.2. Demographic variables were compared across groups using one-way analysis of variance (ANOVA) or chi-squared tests. The demographic variables included ancestry, age, sex, CMD status and medication use of NSAIDs, selective serotonin uptake inhibitors (SSRI), lithium, antipsychotic medication generation (first or second) and corresponding chlorpromazine equivalents. Additionally, ANOVA was used to test for possible moderating effects of important clinical and technical variables with significant (p<0.05) measures being used as a covariate in later statistical tests (e.g., age, sex, FD Power, and SNR). General linear models were used to compare differences between subgroups on 9 resting state networks, 8 functional connectivity networks involving DMNa, and 5 multiplex measures while controlling for confounding variables. Permutation testing, with 10,000 permutations, was used to compare the multiplex network measures between subgroups.

A false discovery rate correction (herein q values) was applied to relevant post-hoc analyses to adjust for multiple comparisons (Benjamini & Hochberg, 1995). Resting-state networks that survived multiple comparison correction from the resting state network group comparisons were used for subsequent functional connectivity analyses. Partial Spearman correlations were used to quantify associations between cognition (covaried for age, sex, and ancestry) and network measures (covaried for age, sex, FD Power, and SNR) for the resting state network and functional connectivity networks that survived multiple comparison correction above within each subgroup and orthogonal differences between correlations were measured using Fisher’s r-to-z transformation. Additionally, exploratory analyses were performed examining correlations between the 13 cytokines plus 5 factor loadings with the 9 resting state networks or 8 functional connectivity measures.

RESULTS

Demographic and clinical characteristics

Figure 1 shows the study design. After matching inflammatory subgroups with resting-state fMRI data there were 32 HC Low, 65 Proband Low and 29 Proband High subjects. In this study sample, 31% of probands fell into the Proband High group, which is consistent with other studies including studies from our group (Boerrigter et al., 2017; Fillman et al., 2013, 2014, 2016; Hoang et al., 2022; Lizano et al., 2021). Similar to our prior study, the Proband High group had higher pro-inflammatory marker levels compared to the Proband Low or HC Low groups and there were no biomarker differences between the Proband Low and HC Low groups (Supplementary Figure 1). The HC Low mirrored the Proband High and Low groups for age, sex, ancestry, and use of NSAIDs, but not for CMD, GAF and SFS scores (Table 1). The Proband High and Low groups did not differ on any of the clinical measures. We previously demonstrated that there were no significant relationships identified between inflammatory markers and symptom severity (Lizano et al., 2021).

Figure 1: Study Design.

B-SNIP1 study participants, including psychosis spectrum disorder probands and healthy controls (HC) underwent cytokine level measurements at baseline for 19 markers. Inflammatory subgroups were determined by unsupervised exploratory factor analysis followed by hierarchical clustering using 13 markers that passed quality control metrics. Resting state functional MRI networks in the inflammatory groups underwent multiple analyses to determine resting state network activity, functional connectivity between networks, and multiplex approaches using positive and negative network connections.

Table 1:

Sociodemographic and Clinical Measures Across Study Subgroups

	HC-Low (n=32)	Proband-Low (n=65)	Proband-High (n=29)	P-value (all groups)	P-value (Probands only)
Age (mean, SD)	34.5 (13.9)	34.1 (13.2)	35.1 (14.1)	0.947	0.742
Sex (Male/Female)	20/12	38/27	15/14	0.691	0.702
Race (AA, CA, OT)	8/17/7	18/39/8	11/17/1
Ancestry (AF/EU)	9/23	19/46	10/19	0.841	0.789
Diagnosis Group (SZP/BPP)	-	34/31	16/13		0.973
Antipsychotic (Yes/No)	-	4/61	4/25		0.409
1st Generation Antipsychotic	-	47/18	20/9		0.933
2nd Generation Antipsychotic
Daily CPZ equivalent (mean, SD)	-	328.1 (268.3)	442.3 (372.2)		0.179
Lithium (Yes/No)	-	12/53	5/24		1
SSRI (Yes/No)	-	21/44	6/23		0.366
Age of onset (mean, SD)	-	18.8 (8.0)	18.5 (6.6)		0.831
Illness Duration (mean, SD)		15.3 (12.1)	16.6 (12.6)		0.622
NSAID (Yes/No)	9/23	10/55	6/23	0.332	0.738
CMD (Yes/No)	0/29	19/40	8/21	0.003	0.845
GAF (mean, SD)	84.3 (5.5)	52.4 (12.7)	50.3 (12.7)	<0.001	0.453
SFS total (mean, SD)	160.0 (16.1)	135.2 (19.4)	134.2 (18.7)	<0.001	0.840

Note: SD = standard deviation; AA = African American; CA = Caucasion; OT = Other Races; EU = European; AF = African; HC = healthy control; SZP = schizophrenia and schizoaffective; BPP = bipolar disorder with psychosis; CPZ = chlorpromazine; SSRI = Selective Serotonin Reuptake Inhibitor; NSAID = non-steroidal anti-inflammatory agent; CMD = cardiometabolic disorder; GAF = global assessment of function; SFS = Birchwood social functioning scale. N=7 HC-High were excluded from analyses due to the small sample size of the group.

Network activity among inflammatory subgroups and their relation to cognition

The Proband High group demonstrated lower resting-state network connectivity in the DMNa as compared to the Proband Low (d=−0.74, p=0.002, q=0.024) and HC Low groups (d=−0.85, p=0.0008, q=0.007) (Figure 2A–B, Supplementary Table 2). There were no other significant resting-state network differences between subgroups (Figure 2A). The functional network connections between the DMNa and other resting-state networks were examined between subgroups. The connection between the rAWN and the DMNa had lower functional network connectivity in the Proband High group compared to the Proband Low group (d=−0.66, p=0.004, q=0.035) (Figure 2C–D, Supplementary Table 3). The Proband High group also showed greater VMIN to DMNa connectivity (d=0.50, p=0.02, q=0.07), but this effect did not survive multiple comparison corrections (Supplementary Table 3). Post-hoc analyses testing the effect of chlorpromazine equivalents on resting state and functional connectivity between the Proband High and Proband Low group revealed that the DMNa findings remained significant (d=−0.67, p=0.024), but the functional connectivity measure between rAWN to the DMNa (d=−0.49, p=0.09) and VMIN to the DMNa (d=0.55, p=0.055) did not. Correlations between BACS scores and DMNa or DMNa-rAWN were examined, and lower DMNa activation was related to worse BACS verbal fluency performance (r=0.558, p=0.002, q=0.027, Figure 2E, Supplementary Table 4A) in the Proband High subgroup that was orthogonal to the Proband Low (z=2.67, p=0.007) and HC Low subgroups (z=2.50, p=0.012, Supplementary Table 4B, C).

Figure 2: Differences in Resting State and Functional Network Connectivity Amang Inflammatory Subgroups.

A-B) Forrest plot and box plot of pairwise contracts of 9ring state catwork cavity between Prohard High, Proband Low, and Healthy Control (HC) Lo groups. C-D) Forrest plot and box plot of pairwise contrasts of 8 fark connectivity between Proband High, Proband Low, and HC Low props. E) Group-based correlation between DMN activity and verbal fancy performance on the brief at of coration (BACS) tank Fisher's 22.67, p=0.007 comparing Proband High to Proband Low. Fisher’s Superior Posterior Default Made Network; SSN-Frognal Motor Network or Sctory Network; SAN-Salice Network; VN-C Occipital or Vinal Network, AWN- Right Frontopistal network or Right Attentional Working Merry Network; LAWN- Left Frotoparietal network or Left Attentional Working Marry Network VAIN-Cerebellar Occipital network or Visto-motor Integration Network.

Exploratory analyses investigating the relationships between inflammatory markers or factor loadings and resting-state (Supplementary Table 5A) or functional network connectivity (Supplementary Table 5B) were performed to examine these measures in a continuous dimension (vs subgroups). In this section, we focus on the individual markers and factor 1, but the remaining correlations are found in supplementary table 5. In psychosis probands lower DMNa activation correlated with higher IL1β (r=−0.231, p=0.025), IL6 (r=−0.212, p=0.040), and Factor 1 loading (r=−0.273, p=0.008), but not with other individual inflammatory markers. Whereas, lower rAWN to DMNa network connectivity correlated with higher IFNγ (r=−0.221, p=0.033), TNFβ (r=−0.234, p=0.024), IL12/IL23p40 (r=−0.306, p=0.003), and factor 1 (r=−0.210, p=0.043).

Differences in single-layer and multiplex topology among inflammatory subgroups and their relationship to cognition

To explore how functional network topology differs across inflammatory subgroups, global clustering coefficient and global efficiency were calculated for the network of negative and positive connections independently. Networks in the Proband High subgroup had lower clustering coefficients in both networks (positive network: d=0.49, p=0.042; negative network: d=0.54, p=0.021, Figure 3A) compared to the Proband Low group. There were no global efficiency differences between subgroups (Figure 3B).

Figure 3: Between-group differences in single-layer and multiplex measures.

A) Illustration of clustered connectivity around a given node (denoted in green). The red connections constitute clusters around the node, while the black connection does not contribute to the clustering coefficient as it is not part of a closed triangle. Between-group differences in the clustering coefficient are shown for the network of B) negative (Proband Low to Proband High: d=0.54, p=0.021) and C) positive connections (Proband Low to Proband High: d=0.49, p=0.042). D) Illustration of distance between two nodes (denoted by connections in red). Global efficiency is calculated as the average of the inverse distances between a node and all other nodes in the network. Differences in global efficiency between all groups in E) negative and F) positive networks. G) Illustration of multiplex participation. Interlayer connections are shown in broken lines (---). The blue node has high participation coefficient as most of its connections are the same in both layers (shown in red). H) Between group differences in multiplex participation coefficients (Proband low to HC Low: d=0.63, p=0.004,; Proband High to Proband Low d=−0.57, p=0.014;. In all plots, black dots denote individual values, whereas the red points indicate the group average values. The boxplots extend between 25^th and 75^th percentile of the group data and the asterisks denote significant between-group differences (p<0.05).

To assess the relationship between the positive and negative networks, the networks were combined into a single multiplex network and the multiplex participation coefficient was calculated to determine, on average, how evenly the nodes are connected in both layers. These analyses revealed that the multiplex participation was higher in the Proband Low group compared to the HC Low group (d=0.63, p=0.004, Figure 3C), and lower in the Proband High (d=−0.57, p=0.014, Figure 3C) compared to Proband Low group.

In the Proband High group, a lower negative clustering coefficient was associated with worse verbal memory scores (r=0.422, p=0.025, Figure 4A). In the Proband High group, higher global efficiency of the negative network was correlated with worse scores on verbal fluency (r=−0.420, p=0.026, Figure 4B) and symbol-coding tests (r=−0.439, p=0.022, Figure 4C), while lower positive global efficiency was associated with worse BACS composite scores (r=0.410, p=0.037, Figure 4D). The correlation between symbol-coding test scores and the negative network in the Proband High group was significantly different from the correlation in the Proband Low group (z=2.02, p=0.043, Figure 4C), while the correlation between BACS composite scores and the positive network in the Proband High group was significantly different from the correlation in the HC Low group (z=−2.3, p=0.021, Figure 4D). Overall, high global efficiency in the positive network is associated with better performance, while high global efficiency in the negative network is associated with worse performance. Furthermore, high cognitive performance is supported by highly segregated systems within the negative network.

Figure 4: Correlation between BACS test scores and multiplex network measures.

Plots showing the best line fit between test scores and global measures for: A) clustering coefficient and BACS verbal memory scores, as well as global efficiency and B) BACS verbal fluency, C) BACS symbol-coding and D) BACS composite scores. The different lines represent the best line fits for the different groups separately, while the dots indicate individual values.

DISCUSSION

To the best of our knowledge, this is the first study to report on the relationships between data-driven inflammatory subgroups or individual peripheral markers of inflammation with resting-state, functional connectivity, and multiplex network measures in psychosis spectrum disorders with the goal of determining the role of peripheral inflammation on functional network dysfunction and related cognitive dysfunction. We found that a subgroup of individuals with psychosis and high inflammation have lower DMNa activation, DMNa to rAWN functional connectivity, as well as lower negative network and positive network clustering and multiplex participation compared to the low inflammation psychosis group. The associations between inflammation and network dysfunction were best explained by the use of multivariate markers as compared to individual markers. Additionally, we expanded on these findings to show that lower network measures correlated with worse verbal fluency, verbal memory, digit sequencing and overall cognitive performance. These findings suggest that low-grade inflammation may increase the risk of psychosis through disruption in the functional coupling of higher-order cognitive network systems such as the DMNa and rAWN.

We have previously demonstrated that using a combination of markers may be more clinically and pathophysiologically informative than single measures of inflammation or BBB disruption (Bishop et al., 2022; Lizano et al., 2021; Zhang et al., 2022), and this idea was further validated in the current study by demonstrating the presence of network related differences. Moreover, inflammation-related network dysfunction has been previously implicated in studies with healthy adults, depression, and PTSD. For example, perturbation experiments used to induce an immune reaction, increase inflammation, or disrupt the BBB have demonstrated resting-state or functional connectivity changes. Endotoxemia in healthy humans, which increases systemic inflammation, has been associated with changes in social pain (Eisenberger et al., 2009) and threat (Inagaki et al., 2012), reward (Eisenberger et al., 2010; Moieni et al., 2019), emotion processing (Kullmann et al., 2013), and social cognitive related neural responses (Kullmann et al., 2014; Muscatell et al., 2016), as well as resting-state functional MRI changes (Labrenz et al., 2016, 2019; Lekander et al., 2016). Whereas, focused ultrasound, which can induce a transient disruption of the BBB, has been associated with reduced functional connectivity in rats (Todd et al., 2018) and patients with Alzheimer’s disease (Meng et al., 2019) and it remained unchanged for 3 months in the latter study. In healthy adults not receiving an immune challenge, lower DMN connectivity was associated with higher levels of IL6 (Marsland et al., 2017) and/or CRP (Dev et al., 2017; Marsland et al., 2017). Few studies have examined the relationship between individual inflammatory markers and functional network alterations in psychiatric disorders. Of the few existing studies, three focused on patients with depression and found that higher levels of CRP were associated with lower DMN connectivity (Kitzbichler et al., 2021), ventral attention-DMN (Aruldass et al., 2021), and ventral striatum-to-ventromedial prefrontal cortex connectivity (Felger et al., 2016). One study of patients with schizophrenia demonstrated that higher levels of IL6 correlated with lower connectivity between the parietal cortex and precuneus (King et al., 2021), which we were not able replicate in our resting state network or functional connectivity analyses (Supplementary Table 5A,B). Only one study to date has incorporated a multi-cytokine approach by using the median split of a composite score of IL6, TNFα, and IL1β in patients with PTSD (Kim et al., 2020). Findings indicated that low-grade inflammation was associated with reduced functional connectivity in a triple network consisting of the DMNa, central executive network (CEN), and SAN. This was consistent with the present finding indicating that functional network dysfunction was associated with low-grade inflammation.

The DMN was coined by Raichle et al in 2001 and refers to the brain’s “resting-state” when it is alert and awake, but not actively engaged in a task (Raichle et al., 2001). The DMN consists of the posterior cingulate cortex/precuneus, medial prefrontal cortex, mesial and inferior temporal lobe, and inferior parietal lobe, whereas the DMNa consists primarily of the medial prefrontal cortex (Meda et al., 2016). In a recent meta-analysis, authors demonstrated that individuals with SZ have large-scale hypoconnectivity of the DMN, which is associated with deficits in information processing (Li et al., 2019). Additionally, they identified reduced connectivity for the triple network model in psychosis which includes the DMN, CEN, and SAN (Li et al., 2019). Similarly, Kim et al 2020 identified a relationship between low-grade inflammation and disruption in this triple network in PTSD (Kim et al., 2020). In the present study, we found reduced network connectivity between the DMNa and the rAWN, which is part of the frontoparietal network and CEN, but not between the DMNa and SAN. Furthermore, lower DMNa connectivity in SZ has been correlated with poorer attention, concentration, and memory (this composite measure included symbol coding) (Camchong et al., 2011), which was reflected in the association we found between DMNa and verbal fluency in the high inflammation group. Interestingly, a Finnish population study of healthy midlife adults showed that elevated baseline levels of IL6 and TNFα, but not CRP, were associated with poorer verbal fluency and verbal memory performance at 10 years follow up (Kipinoinen et al., 2022). We previously demonstrated that higher IL6 and VEGF correlated with worse verbal fluency and symbol coding, respectively, in a psychosis spectrum population (Lizano et al., 2021). Using multiple peripheral inflammatory markers, researchers found a relationship between elevated inflammatory signature and worse verbal fluency scores (Fillman et al., 2016). Further supporting this relationship are studies demonstrating that immunomodulatory drugs can result in mild improvements in working memory (Zhang et al., 2019), while a biologic agent targeting IL6 demonstrated improvements in symbol coding (Miller et al., 2016), however, these findings remain mixed in the literature (Girgis et al., 2018; Jeppesen et al., 2020). Thus, our current findings suggest that elevated inflammation may disrupt the functional coupling of higher-order cognitive network systems such as the DMNa and CEN (Siman-Tov et al., 2016; van den Heuvel & Hulshoff Pol, 2010) to disrupt cognitive processes commonly observed in psychosis or with later cognitive decline.

The observed lower clustering in the Proband High compared to the Proband Low suggests a reduced density of local connectivity between the network nodes (Bullmore & Sporns, 2009). This indicates that when engaged in specialized cognitive tasks the networks in the Proband High group exhibit greater independence and decreased synchronization (Anderson & Cohen, 2013; Chan et al., 2014; Dong et al., 2018; Karbasforoushan & Woodward, 2012; Liu et al., 2008; Woodward et al., 2011). Additionally, the absence of significant differences in global efficiency between the Proband High group and the other groups indicates that the Proband High group preserves its global network organization, thus facilitating efficient interregional communication (Fornito et al., 2013). These findings suggest a compensatory reorganization of the functional connectivity within the Proband High group, which compensates for the sparse local connections between resting-state networks by preserving the overall global efficiency of the network.

We used the multiplex coefficient to quantify the relationship between positive and negative correlations, which can be understood as the brain’s ability to maintain healthy co-activation among resting-state networks while inhibiting specific connections between them (Battiston et al., 2014; Mijalkov et al., 2022). As they originate from different neurovascular mechanisms (Goelman et al., 2014) the presence of both positive and negative connections is crucial and contributes to balanced communication between resting-state networks ultimately enhancing cognitive performance (Mijalkov et al., 2022; Saberi et al., 2021). In psychosis, studies have shown a suppression of the between-network negative correlations in the DMNa as well as its connectivity with other resting-state networks (Ramkiran et al., 2019; Shao et al., 2018; Uddin et al., 2009; Williamson, 2007). Consequently, the hyper-balance observed in the Proband Low group could reflect the reorganization of the functional connectivity patterns in response to this diminished negative connectivity. However, the Proband High group displayed a disruption in this balance, consistent with previous studies indicating that neuropsychiatric disorders are associated with disruptions in the equilibrium between task-related activation and suppression of functional connectivity (Anticevic et al., 2012; Whitfield-Gabrieli et al., 2009).

Cognitive performance is known to be affected by the topological organization of the connections between different brain regions, which has been previously assessed using patterns of co-activation or positive connections between them (Bressler & Menon, 2010; Fornito et al., 2013; Park & Friston, 2013; van den Heuvel & Hulshoff Pol, 2010). Previous research has found that increased cognitive performance is associated with increased local functional segregation, which supports more specialized processes through short connections, and global functional integration, which enables higher cognitive processes through long-range connections (Bullmore & Sporns, 2009; Cohen & D’Esposito, 2016; Stam & Reijneveld, 2007; van den Heuvel et al., 2009; van den Heuvel & Hulshoff Pol, 2010). Moreover, studies have also indicated that more segregated networks are associated with better cognitive performance (Chan et al., 2014; Cohen & D’Esposito, 2016), suggesting that optimal network organization is the one that can simultaneously promote high segregation and high integration (Liao et al., 2017). Our findings are in line with these results, as they demonstrate that reduced global efficiency in the positive network is associated with worse performance on the BACS composite scores. It is important to note that the organization of negative correlation networks have not been fully explored and there is a lack of understanding of how high integration and high segregation apply to cognitive functioning. Nevertheless, our negative clustering coefficient finding suggest that high segregation is also needed for negative networks to achieve optimal network performance, and therefore, better cognitive function. Conversely, we showed that an excessive integration of such negative networks is associated with poorer performance. This hyper-integration may lead to excessive attenuation of the connectivity patterns between resting-state networks, ultimately impacting their coactivation and resulting in the observed impaired cognition. Additionally, higher cognitive performance in the Proband High group was associated with a network of anti-correlated patterns that have lower global efficiency, but a high density of local connections within the network. These findings suggest that cognitive processes can be maintained by opposite patterns of connectivity between the networks of positive and negative connections, which indicates that an optimal balance between the two networks is necessary for efficient information processing, in line with our earlier results on multiplex networks.

LIMITATIONS

The present study had several limitations to consider when interpreting the results. Confounding variables, such as body mass index, lifetime antipsychotic exposure, smoking, trauma and sleep-related issues could impact the findings in this paper (de Jager et al., 2009; Leng et al., 2008). Previous exposure to infectious agents or autoimmune antibodies was not measured and could potentially contribute to inflammatory results (Upthegrove & Barnes, 2014). Additionally, blood was not processed for other basic laboratory testing such as lipid profiles, thyroid levels, glucose, or insulin levels. In future studies, we suggest incorporating these potential confounding factors along with examining laboratory-based factors and autoimmune and infectious exposure status. Dosage information of lithium, anti-depressants and NSAIDs were not collected in this study which may potentially play a role in cytokine levels and/or network connectivity. However, based on sensitivity analyses from our prior work (Lizano et al., 2021) and in the current study, we did not find that medications, including NSAIDs, had any effect on cytokine levels or network measures. Also, the healthy control high inflammation group was too small (n=7) to be used for analyses. This group could elucidate acute inflammatory results in a healthy population and should be studied further.

CONCLUSION

The results described herein expand on our understanding of the potential effects of peripheral inflammatory signatures and/or subgroups on network dysfunction in psychosis and how they might relate to worse cognitive performance. Additionally, the novel multiplex approach taken in this study demonstrated how inflammation may disrupt the brain’s ability to maintain healthy co-activation patterns between the resting-state networks while inhibiting certain connections between them.

HIGHLIGHTS.

• Data driven inflammatory subgroups show connectivity differences in psychosis

• High inflammation subgroup has lower DMNa activity vs the low inflammation group

• High inflammation subgroup has lower DMNa functional connectivity vs the low group

• High inflammation subgroup has lower multiplex participation vs the low group

• Network findings in high inflammation subgroups correlate with worse cognition