Glutamate and Dysconnection in the Salience Network: Neurochemical, Effective-connectivity, and Computational Evidence in Schizophrenia

Journal Pre-proof Glutamate and Dysconnection in the Salience Network: Neurochemical, Effectiveconnectivity, and Computational Evidence in Schizophrenia Roberto Limongi, Peter Jeon, Michael Mackinley, Tushar Das, Kara Dempster, Jean Théberge, Robert Bartha, Dickson Wong, Lena Palaniyappan PII:

S0006-3223(20)30056-1

DOI:

https://doi.org/10.1016/j.biopsych.2020.01.021

Reference:

BPS 14114

To appear in:

Biological Psychiatry

Received Date: 16 September 2019 Revised Date:

6 January 2020

Accepted Date: 27 January 2020

Please cite this article as: Limongi R., Jeon P., Mackinley M., Das T., Dempster K., Théberge J., Bartha R., Wong D. & Palaniyappan L., Glutamate and Dysconnection in the Salience Network: Neurochemical, Effective-connectivity, and Computational Evidence in Schizophrenia, Biological Psychiatry (2020), doi: https://doi.org/10.1016/j.biopsych.2020.01.021. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2020 Published by Elsevier Inc on behalf of Society of Biological Psychiatry.

Glutamate and Dysconnection in the Salience Network: Neurochemical, Effective-connectivity, and Computational Evidence in Schizophrenia Running title: Three-level Hypothesis of Dysconnection

Roberto Limongi1, Peter Jeon1, Michael Mackinley1, Tushar Das1, Kara Dempster1, Jean Théberge1,2,3, Robert Bartha1, Dickson Wong1, Lena Palaniyappan1

1

University of Western Ontario, London, Ontario St. Joseph`s Health Care London, London, Ontario 3 Lawson Health Research Institute, London, Ontario 2

Correspondence concerning this article should be addressed to Roberto Limongi, Robarts Research Institute, 1151 Richmond St. N, UWO, London, Ontario, Canada, N6A 5B7. Phone: +1 226 977 6527. E-mail: [email protected]

ABSTRACT BACKGROUND Functional dysconnection in schizophrenia is underwritten by a pathophysiology of the glutamate neurotransmission that affects the excitation-inhibition balance in key nodes of the salience network. Physiologically, this is manifest as aberrant effective connectivity in intrinsic connections involving inhibitory interneurons. In computational terms, this produces a pathology of evidence accumulation and ensuing inference in the brain. Finally, the pathophysiology and aberrant inference would partially account for the psychopathology of schizophrenia as measured in terms of symptoms and signs. We refer to this formulation as the three-level hypothesis. METHODS We tested the hypothesis in core nodes of the salience network (the dorsal anterior cingulate cortex, dACC, and the anterior insula, AI) of 20 first-episode psychosis (FEP) and 20 healthy control (HC) subjects. We established three-way correlations between the MRS measures of glutamate, effective connectivity of resting state fMRI, and correlations between measures of this connectivity and estimates of precision (inherent in evidence accumulation in the Stroop task) and psychopathology. RESULTS Glutamate concentration in the dACC was associated with higher and lower inhibitory connectivity in the dACC and in AI respectively. Crucially, glutamate concentration correlated negatively with the inhibitory influence on the excitatory neuronal population in the dACC of FEP subjects. Furthermore, aberrant computational parameters of the Stroop task performance

were associated with aberrant inhibitory connections. Finally, the strength of connections from the dACC to the AI correlated negatively with severity of social withdrawal. CONCLUSIONS These findings support a link between glutamate-mediated cortical disinhibition, effectiveconnectivity deficits, and computational performance in psychosis. Key words: dysconnection hypothesis; glutamate hypothesis; effective connectivity; dynamic causal models; predictive coding; schizophrenia.

The glutamate hypothesis (1-4) has been a central focus of interest in the study of the psychopathology of schizophrenia for more than 20 years. It states that dysfunction of glutamatergic neurotransmission is associated with signature symptoms of schizophrenia. This dysfunction is likely caused by NMDA-receptor hypofunction (i.e., glutamate hypofunction) in inhibitory GABAergic interneurons which would lead to an increase in the synaptic gain of excitatory neurons (the disinhibition hypothesis, 5, 6). Abnormal increase of synaptic gain is also conceptualized as a disruption of the excitation-inhibition balance which, as described below, rests at the core of the dysconnection hypothesis (7) within the theoretical framework of “the Bayesian brain”1 (8). The dysconnection hypothesis suggests that the psychopathology of schizophrenia should be studied at three levels of analysis: neurochemical, effective-connectivity (network-connectivity), and computational levels. At the neurochemical level, NMDA hypofunction would increase the synaptic gain in deep and superficial pyramidal cells, reflecting a decrease in the strength of intrinsic inhibitory connections (7, 9). At the effective-connectivity level, a predictive coding algorithm (10), namely, hierarchical message passing between lower and higher cortical levels, would be altered in terms of aberrant backward and forward interregional connectivity strength (11). At the computational level, a suboptimal Bayesian brain (12) would overly afford confidence or precision (i.e., inverse variance) to its predictions about the external stimuli and

1 This Bayesian account tends to suggest the existence of an inner controller (e.g., a central executive) which inevitable would call upon the homunculus fallacy. However, the dysconnection hypothesis is nested within the active inference theory under the free energy formulation. From this perspective, the homunculus fallacy dissolves in terms of autopoietic self-organization.

1

would overestimate the reliability of the prediction errors (PE), leading to false inferences and failures in cognitive control (13-15). To test this hypothesis, in this work we study the relationship between cognitive-control dysfunction in schizophrenia, the effective connectivity between core nodes of the salience network (the right dorsal anterior cingulate cortex, dACC, and the right anterior insula, AI), and the concentration of 1H-MRS glutamate in the dorsal anterior cingulate cortex ([Glu]dACC). At the computational level, we compare the performance of first-episode psychosis (FEP) and healthy control (HC) subjects in the Stroop task (16), which reliably engages the dACC–AI network (17, 18). Suboptimal Stroop computations in FEP are reflected in long reaction time (RT) and low response accuracy (19, 20). We show that a drift-diffusion model, as a specific case of a Bayesian decision making (21), explains these suboptimal computations. At the effectiveconnectivity level, we show that the computational parameter associated with aberrant predictions in FEP maps onto extrinsic connections. Ultimately, we demonstrate at the neurochemical level that [Glu]dACC drives the dysconnection within the dACC–AI network which, as we detail below, is an appropriate anatomical and functional motif (c.f., 22) to evaluate our three-level hypothesis. The dACC functionally specializes in conflict monitoring during the Stroop task (23) and is anatomically connected to the AI (24-26). Crucially, psychosis is associated with consistent structural deficits (27) as well as resting-state functional dysconnectivity in this network (28, 29). Given that the AI is particularly sensitive to descending afferents from the dACC (30), we expected variations of excitation-inhibition balance resulting from the glutamatergic abnormalities within the dACC to induce dysconnectivity in the network, affecting both intrinsic inhibitory connections of the dACC and the extrinsic connections with the AI. 2

Negative symptoms are regarded as the core drivers of functional deficits in schizophrenia. Nevertheless, the evidence relating glutamate to negative symptoms has been conflicting to date (31). From previous experimental work linking reduced precision of priors (relative to precision of PE) and negative symptoms (32-34), we expected negative association between effective connectivity within the two-node network and higher negative symptoms. In summary, we used a multilateral approach to test a specific dysconnection hypothesis that has three related aspects: first, functional dysconnection in schizophrenia is underwritten by a pathophysiology of the glutamate neurotransmission that affects the excitation-inhibition balance in key nodes of the salience network. Physiologically, this is manifest as aberrant effective connectivity in intrinsic connections involving inhibitory interneurons. In computational terms, this produces a pathology of evidence accumulation and ensuing inference in the brain —as manifest in terms of psychophysics. Finally, the pathophysiology and aberrant inference partially account for the psychopathology of schizophrenia as measured in terms of symptoms and signs. We will refer to this formulation as the three-level hypothesis and establish its validity by looking for three-way correlations between the MRS measures of glutamate, dynamic causal modelling (DCM) measures of directed connectivity, and correlations between measures of this connectivity and estimates of precision (inherent in evidence accumulation) on the one hand — and psychopathology on the other. METHODS AND MATERIALS

Subjects Twenty FEP and 20 HC subjects participated in the study (Supplemental Methods Table S1 and S2). Glutamate MRS assessment could not be performed in one FEP subject. Data from this subject were not analyzed. Subjects were recruited from the “Prevention and Early Intervention 3

Program for Psychosis” in London, Ontario. Criteria for inclusion in the FEP group included (i) first clinical presentation with psychotic symptoms and (ii) DSM-5 (35) criteria A for schizophrenia satisfied (ii) less than 2 weeks of lifetime antipsychotic exposure. All patients with FEP received a consensus diagnosis from 3 psychiatrists (LP/KD and the primary treatment provider) after approximately 6 months on the basis of the best estimate procedure, as described in (36), and the Structured Clinical Interview for DSM-5. All patients satisfied criteria for Schizophrenia-Spectrum Disorders, with 15 patients satisfying DSM-5 criteria for Schizophrenia and 3 for Schizoaffective Disorder. 1 subject lacked follow-up clinical data at 6 months, with the available baseline data suggesting a diagnosis of Schizophreniform Disorder. We use the term FEP to describe the patient group to capture all the schizophrenia-spectrum disorders as above. HC participants did not report a personal history of mental illness or family history of psychotic disorders. Informed consent from participants was obtained according to the approval by Western University’s Human Ethics Committee. Symptoms assessment was performed using Positive and Negative Syndrome Scale-8 items version (37, Supplemental Methods Table S3). As detailed below, participants underwent resting-state 1H-MRS, a color version of the Stroop task, and resting-state fMRI (in this order) inside a Siemens MAGNETOM 7-Tesla MRI scanner using an 8-channel transmit/ 32-channel receive, head-only, radiofrequency coil. 1

H-MRS

We measured a four-minute block of resting-state 1H-MRS in A 2.0 x 2.0 x 2.0 cm (8cm3) 1HMRS voxel was placed on the dACC (Fig. 1 and Supplemental Methods Figure S1) where activation was expected based on our previous work using the Stroop task (20). To locate the voxel, we used a two-dimensional anatomical imaging sequence in the sagittal direction (37 slices, TR = 8000 ms, TE = 70 ms, flip-angle (α) = 120°, thickness = 3.5 mm, field of view = 4

240×191 mm, voxel control is described in the Supplemental Methods). A 32 channel-combined water-suppressed spectral average was acquired using the semi-LASER 1H-MRS pulse sequence (repetition time = 7500 ms) with a long echo time (100 ms) — this long echo time improves glutamate measurement in the human brain (38). Water suppression was achieved using the VAPOR preparation sequence and a water-unsuppressed spectrum was also acquired for spectral post-processing (Supplemental Methods).

Resting-state fMRI We acquired 360 resting-state whole-brain functional images. A gradient echo-planar-imaging sequence was used with phase-encoding direction = A > > P, repetition time = 1000 ms, echo time = 20 ms, flip angle = 30 deg, field of view = 208 mm, field of view phase = 100 %, voxel dimension = 2 mm isotropic, slice thickness = 2 mm, multi-band acceleration factor = 3, acquisition time = 6 min. 26 s, and number of slices = 63 (interleaved slice order).

Computational Model of the Stroop Performance Based on our previous work (20), we were interested in the incongruent condition of the Stroop task (Supplemental Methods) in which we expected lower accuracy and longer RT in FEP than in HC subjects. We fit a hierarchical drift-diffusion model to the RT and accuracy data (Supplemental Methods). In this model, participants accumulate information and trigger a response after reaching an accumulation threshold. The model comprises four basic parameters representing the accumulation threshold, the starting-point of the accumulation process, the accumulation (or drift) rate, and the non-decision processes. Predictive coding maps elegantly onto the drift diffusion model (13, 21, 39). Evidence accumulation corresponds to the accumulation of presynaptic afferent activity from neuronal 5

populations encoding PE (i.e., superficial pyramidal cells), the drift rate represents precision of ascending PE, and the starting point represents prior beliefs. Formally, corrected prior beliefs (i.e., a change in the starting point parameter) correlates with a larger drift rate —indicating more precise PE. We expected larger drift rate (i.e., aberrant precision of PE) and larger starting point (aberrant prior beliefs) in the FEP group than in the HC group. Although, we did not have apriori expectations about differences in the decision threshold and in non-decision processes, we report the differences in these parameters for completeness and post-hoc interpretation. We report between-groups difference in the posterior probability (PP) of parameter estimates.

Effective Connectivity DCM. We estimated the resting-state effective connectivity within the dACC–AI network by fitting a two-state spectral DCM (40) to the fMRI data (41) (Supplemental Methods). We identified regions with blood oxygen level fluctuations within frequencies ranging from 0.0078 to 0.1 Hz (42) via an F-contrast. We extracted the timeseries that summarized the activity within spheres (8-mm radius) in the dACC and in the AI (Fig. 1 and Supplemental Methods Figure S1). MNI coordinates of these spheres were defined based on our previous works (20, 43). Two-state DCM assumes excitatory and inhibitory populations of neurons within a region. Each population comprises self-inhibition connections (which are fixed parameters). Crucially, two free parameters are fit to the fMRI data: interregional excitatory-to-excitatory connections and within-region inhibitory-to-excitatory connections (Fig. 1). Since we aimed at demonstrating the relationship between the effect of [Glu]dACC on inhibitory-to-excitatory connections and the ensuing consequences in the entire network, this two-state model was sufficient to test our hypothesis despite not capturing all the circuitry constraints of cortical columns.

6

For each participant, a fully connected DCM, with no exogenous inputs (c.f., task fMRI) was specified and inverted using spectral DCM. At a group level, we estimated parameters of six parametric empirical Bayes models (44, 45) aiming to test our three-level hypothesis against five alternative hypotheses. The three-level model comprised (i) group, (ii) [Glu]dACC, (iii) precision of PE, (iv) prior beliefs, (v) [Glu]dACC × Group, (vi) [Glu]dACC × Prior Beliefs, and (vii) [Glu]dACC × Precision of PE as covariates. All but the “behavioral model” (see below) were reduced versions of the three-level model. A “no-glutamate” model represented the hypothesis that [Glu]dACC did not affect the network’s effective connectivity. neither the main effect of [Glu]dACC nor the [Glu]dACC × Group interaction was included in the model. A “no-computational” model did not include the effects of prior beliefs, precision of PE, Prior Beliefs × Group, and Precision of PE × Group. It represented the hypothesis that prior beliefs and precision of PE do not affect connectivity. A “behavioral model” substituted (in the three-level model) the behavioral responses and their interactions with group for the hierarchical drift-diffusion model parameters. This model represented the hypothesis that the observed responses better account for network connectivity than the computational parameters. For completeness, we fit a “group model” (with only the main effect of group representing the hypothesis that only differences between groups would cause changes in the network connectivity) and a “null model” (representing the hypothesis of no effect of covariates on the network). To adjudicate between models, we performed a Bayesian model comparison (44). To evaluate our ‘three-level’ hypothesis, we report the effect sizes; i.e., parameters of the between-subject parametric empirical Bayesian model associated with each covariate along with the posterior probabilities. A data analysis pipeline is detailed in the Supplemental Methods Figure S2. 7

RESULTS Participants performed the task as instructed. As expected, in the incongruent condition, FEP subjects were less accurate and took longer than HC subjects (Table 1 and Supplemental Results Table S4). The hierarchical drift-diffusion model showed that the drift-rate parameter in the HC group (M = 2.83, SD = 0.18) was larger than in the FEP group (M = 1.83, SD = 0.2). The difference in PP was = 0.99. The starting-point in the FEP group was larger (i.e., closer to the decision boundary, M = 0.44, SD = 0.03) than in the HC group (M = 0.32, SD = 0.03), difference in PP = 0.99. The decision threshold was lower in the FEP group (M = 2.06, SD = 0.32) than in the HC group (M = 2.52, SD = 0.09), difference in PP = 0.96. Finally, the non-decision processes took longer in the FEP group (M = 0.64, SD = 0.11) than in the HC group (M = 0.46, SD = 0.09), difference in PP = 0.99 (Supplemental Results Figure S3 and Tables S5-S6).

[Glu]dACC Affects Intrinsic Inhibitory Connections in Core Nodes of the Salience Network The three-level model outperformed all the alternative models (Fig. 2 and Supplemental Results Table S7). The mean estimates across groups revealed inhibition in both regions (dACC, M = 2.35, PP =1.0; AI, M = 2.51, PP = 1.0) and an attenuation in the strength of excitatory connections (dACC→AI, M = -0.51, PP = 0.99; AI→dACC, M = -0.45, PP = 0.99). Fig. 3 shows no main effect of group on parameters. However, main effect of [Glu]dACC on self-connections in both regions was detected. Intrinsic inhibition increased in the dACC (β = 0.11, PP = 0.99) and decreased in the AI (β = -0.07, PP = 0.96). Prior beliefs did not affect extrinsic connections. However, precision of PE strengthened inhibitory connections (dACC, β = 0.92, PP = 1; AI, β = 1.1, PP = 1).

8

[Glu]dACC Correlates Negatively with Intrinsic Inhibitory Influence in the dACC of FEP Subjects, Accounting for Aberrant Prior Beliefs and Aberrant Precision of PE In the dACC, [Glu]dACC was associated with weaker inhibitory connections in FEP than in HC (β = -0.1, PP = 0.98), and the effect of prior beliefs on inhibitory connections was stronger in FEP than in HC (dACC, β = 40.01, PP =1; AI, β = 42.48, PP = 1, Fig. 4). Surprisingly, the effect of prior beliefs on excitatory connections was weaker in FEP than in HC (dACC→AI, β = -8.72, PP = 0.99; AI→dACC, β = -7.76, PP = 0.99). Finally, we found that the effect of precision on inhibitory connections was weaker in FEP than in HC participants (dACC, β = -1.28, PP = 1; AI, β = -1.48, PP = 1). In summary, [Glu]dACC differentially affected the network’s effective connectivity which was differentially associated with the parameter estimates of the hierarchical drift–diffusion model in the presence of FEP.

Connectivity Strength Correlates Negatively with Severity of Social Withdrawal We searched for negative association between the severity of blunted affect, social withdrawal, and lack of spontaneity and the strength of connections in the FEP group. We made three comparisons per connection (Bonferroni correction, P = 0.017). Only the effect of backward connections on social withdrawal survived correction. Severity of social withdrawal increased as the connectivity strength of dACC→AI connections decreased (Fig. 5). For completeness, we also explored the association between hallucinations and delusions and connectivity strength (two comparisons per connection with Bonferroni correction, P = 0.025). None of the comparisons reached statistical significance at this level of correction. DISCUSSION

9

We have demonstrated that low computational performance during cognitive-conflict resolution in FEP is explained by aberrant resting-state effective connectivity (dysconnection) in core nodes of the salience network (dACC and AI). This dysconnection was associated with glutamate hypofunction in the dACC, as proven by the fact that a model without the effect of [Glu]dACC had low probability of explaining the fMRI data. [Glu]dACC was associated with opposite intrinsic connectivity strength in both the dACC and the AI, indicating that [Glu]dACC affected the entire network and confirming the sensitivity of the AI to changes in the dACC excitation-inhibition balance. Crucially, in the FEP group the inhibitory influence on the excitatory population of the dACC decreased as a function of [Glu]dACC. This finding is the first imaging evidence directly linking the glutamate hypofunction to the cortical disinhibition hypothesis in schizophrenia. Decreased intrinsic inhibition and decreased extrinsic connectivity have been previously reported in the default (46) and dorsal attention networks (47) of schizophrenia subjects. The fact that three independent groups have shown the same relationship between decreased intrinsic and extrinsic connections in the default, attention, and salience networks provide strong support to the dysconnection hypothesis of schizophrenia. Crucially, our results point to glutamate hypofunction and aberrant computations of sensory and prior precision as critical causes of dysconnection and are in line with recently reported data showing that the effect of glutamate hypofunction should be observed at the network level (48). These results also demonstrate that establishing the relationship between glutamate hypofunction, the disinhibition hypothesis, and the pathophysiology of schizophrenia requires the integration of the neurochemical, effective-connectivity, and computational levels of analysis — since Bayesian model comparison afforded low probability to any model not comprising all three levels. Specifically, parameter estimates of resting-state aberrant connectivity (driven by 10

resting-state [Glu]dACC) in the dACC–AI network account for computational parameters of cognitive dysfunction in FEP. Furthermore, an increase in the effect of the inhibitory population on the excitatory population in both regions explains the precision subjects afforded to ascending information, in line with the known role of GABA interneurons in pyramidal circuits (49). The three-level hypothesis also explains the computational compensatory effect of aberrant prior beliefs to aberrant precision of ascending information in FEP subjects. They needed to accumulate less information than HC (i.e., lower decision threshold), relied more on their prior beliefs (i.e., at the beginning of each trial they were closer to the decision boundary) than on sensory information, and were less cautious when resolving cognitive conflicts ―they tended to jump to conclusions (50). At a group level, we did not find evidence for a larger precision of PE in the FEP group than in the HC group. On the contrary, the HC group showed a steeper drift rate. However, within the FEP group aberrant precision PE was associated with a decrease in the strength of inhibitory connections —in both regions (Fig. 4-B). Crucially, an increase in the inhibitory activity was associated with an increase in prior beliefs (Fig. 4-C), suggesting (at the effective-connectivity level) that more precise prior beliefs compensate for aberrant precision of PE —as predicted by the dysconnection hypothesis (7). The current results are in concert with a previous study that used a four-neuronal-population DCM to model GABA connections to superficial pyramidal cells (9). In the referred work, DCM parameters were associated with behavioral performance, showing that establishing a relationship between synaptic dysfunction and behavior requires modeling effective connectivity. However, our three-level model outperformed a model including behavioral observations, suggesting that symptoms of schizophrenia emerge from the interaction between neurochemical, network-connectivity, and computational variables. 11

Another previous work has reported no relationship between [Glu]dACC and deficits in Stroop performance in FEP (20). However, unlike our study, it included neither the network nor the computational level of analysis. These two levels allowed us to observe the indirect effect of [Glu]dACC on Stroop computations through the aberrant effective connectivity within the dACC– AI network. Similarly, a recent Bayesian study has suggested that prior beliefs might not influence behavioral performance in the absence of sensory stimulation (51). Our FEP subjects did not expect sensory stimulation. However, the weaker the connectivity strength in descending connections (representing prior beliefs) were in resting state the more biased were their responses towards the decision threshold in the Stroop task. Therefore, it is possible that the effect that prior beliefs might exert on behavioral performance depends not only on the sensory stimulation but also on the resting-state connectivity strength of descending connections. As per the dysconnection hypothesis, the negative correlation between the dACC→AI connectivity strength and severity of social withdrawal appears counterintuitive since, on the one hand, overly afforded confidence in prior beliefs would be expected to correlate positively with increased strength in descending connections. On the other hand, increased connectivity strength should correlate positively with negative symptom severity. It is important to note that we observed a relationship between strong prior beliefs and reduced dACC→AI connectivity that relates to social withdrawal. In fact, stronger priors relate to both reduced dACC→AI connectivity as well as increased intrinsic inhibitory tone of dACC and AI. This relationship is consistent with an effective connectivity study that relates intrinsic inhibitory connections within primary visual cortex and negative symptoms. Moreover, recent models of effortful control (52) suggest that strong priors along with top-down dysconnectivity may relate to a reduced effort signal from the dACC to the AI, and thus present as motivational deficits indexed by social

12

withdrawal. While our data do not support the predictions of a direct relationship between strong prior beliefs and negative symptoms severity, it is worth noting that our resting-state experimental set up did not allow for the suboptimal Bayesian agent (the patient) to withdraw from a social environment that is rich in sensory stimulation. From the perspective of active inference under the free-energy formulation, there is no social withdrawal in a dark room (53). Within the context of these previous works, a negative correlation between connectivity strength and negative symptoms appears to be explained not only by the general active inference theory under the free-energy formulation but also in terms of utility models (as special cases of active inference, 54). However, because the number of works showing an association between high prior beliefs, weak connections, and negative symptoms is scarce and equivocal —e.g., a recent functional connectivity work found no association between connections in the salience network and negative symptoms (55)— this association should be taken as a working hypothesis.

Limitations and Future Directions Our three-level hypothesis was tested via hierarchical Bayesian inference which considers the data hierarchy and uses informative prior distributions and shrinkage. Therefore, within the assumptions of Bayesian statistics the posterior probabilities of our estimates showed large and reliable effect sizes. Nevertheless, the current sample size could be regarded as small for classical inferences such as those pertaining to the between-group differential effect of [Glu]dACC and the association between symptoms and connectivity strength. With resting-state fMRI, we controlled for task as a confounding factor of the effect of [Glu]dACC on inhibitory connections. Since the current results shows this effect, future studies could address the Task × Glutamate interaction. By combining task fMRI with four neuronal population DCM,

13

it would be possible to set the effect of [Glu]dACC on inhibitory connections as priors while modeling the effect of glutamate in other regions. One region that could be added to the current network is the ventrolateral prefrontal cortex (vLPFC) which, along with the AI and the dACC, is engaged by interference and cognitive control processes during the Stroop task. In the context of hierarchical message passing, four neuronal population DCM would allow to investigate the hierarchical relationship between the dACC and vLPFC2, by testing different hierarchical connections between superficial and deep pyramidal cells across regions. Conclusion This work provides evidence to the hypothesis that the glutamate hypofunction relies on the disinhibition hypothesis and manifests itself in the effective connectivity of key nodes of the salience network. The three-level hypothesis provides compelling explanation to deficits in cognitive control and negative symptoms in schizophrenia. The findings support a link between glutamate-mediated cortical disinhibition, deficits in effective connectivity, and computational performance in FEP. ACKNOWLEDGMENTS This study was funded by CIHR Foundation Grant (375104/2017) to LP; Schulich School of Medicine Clinical Investigator Fellowship to KD; AMOSO Opportunities fund to LP; Bucke Family Fund to LP; BrainSCAN to RL; Grad student salary support of PJ by NSERC Discovery Grant (No. RGPIN2016-05055) to JT; Canada Graduate Scholarship to KD. Data acquisition was supported by the Canada First Excellence Research Fund to BrainSCAN, Western University

2

We thank an anonymous reviewer for suggesting this future study.

14

(Imaging Core); Innovation fund for Academic Medical Organization of Southwest Ontario; Bucke Family Fund, The Chrysalis Foundation and The Arcangelo Rea Family Foundation (London, Ontario). We thank Dr. Joe Gati, Mr. Trevor Szekeres and Dr. Ali Khan for their assistance in data acquisition and archiving. We thank Dr. Maria Densmore for producing supplementary SPM outputs. We thank Dr. William Pavlovsky for consultations on clinical radiological queries. We thank Drs. Raj Harricharan, Julie Richard, Priya Subramanian and Hooman Ganjavi and all staff members of the PEPP London team for their assistance in patient recruitment and supporting clinical care. We gratefully acknowledge the participants and their family members for their contributions. These data have been published as preprint in https://www.biorxiv.org/content/ 10.1101/828558v1. DISCLOSURES LP reports personal fees from Otsuka Canada, SPMM Course Limited, UK, Canadian Psychiatric Association; book royalties from Oxford University Press; investigator-initiated educational grants from Janssen Canada, Sunovion and Otsuka Canada outside the submitted work. All other authors report no biomedical financial interests or potential conflicts of interest. REFERENCES 1.

Coyle JT (1996): The glutamatergic dysfunction hypothesis for schizophrenia. Harv Rev

Psychiatr. 3:241-253. 2.

Olney JW, Farber NB (1995): Glutamate receptor dysfunction and schizophrenia.

Archives of general psychiatry. 52:998-1007.

15

3.

Snyder MA, Gao WJ (2013): NMDA hypofunction as a convergence point for

progression and symptoms of schizophrenia. Frontiers in cellular neuroscience. 7:31. 4.

Snyder MA, Gao WJ (2019): NMDA receptor hypofunction for schizophrenia revisited:

Perspectives from epigenetic mechanisms. Schizophr Res. 5.

Homayoun H, Moghaddam B (2007): NMDA Receptor Hypofunction Produces Opposite

Effects on Prefrontal Cortex Interneurons and Pyramidal Neurons. The Journal of Neuroscience. 27:11496. 6.

Gonzalez-Burgos G, Lewis DA (2008): GABA Neurons and the Mechanisms of Network

Oscillations:

Implications for Understanding

Cortical

Dysfunction

in

Schizophrenia.

Schizophrenia Bulletin. 34:944-961. 7.

Friston KJ, Brown HR, Siemerkus J, Stephan KE (2016): The dysconnection hypothesis

(2016). Schizophrenia Research. 176:83-94. 8.

Friston KJ (2012): The history of the future of the Bayesian brain. Neuroimage. 62:1230-

1233. 9.

Shaw AD, Knight L, Freeman TCA, Williams GM, Moran RJ, Friston KJ, et al. (2019):

Oscillatory, Computational, and Behavioral Evidence for Impaired GABAergic Inhibition in Schizophrenia. Schizophrenia Bulletin. 10.

Spratling MW (2017): A review of predictive coding algorithms. Brain and Cognition.

112:92-97.

16

11.

Bastos AM, Usrey WM, Adams RA, Mangun GR, Fries P, Friston KJ (2012): Canonical

microcircuits for predictive coding. Neuron. 76:695-711. 12.

Aitchison L, Lengyel M (2017): With or without you: predictive coding and Bayesian

inference in the brain. Curr Opin Neurobiol. 46:219-227. 13.

Limongi R, Bohaterewicz B, Nowicka M, Plewka A, Friston KJ (2018): Knowing when

to stop: Aberrant precision and evidence accumulation in schizophrenia.

Schizophrenia

Research, pp 386-391. 14.

Powers AR, 3rd, Gancsos MG, Finn ES, Morgan PT, Corlett PR (2015): Ketamine-

Induced Hallucinations. Psychopathology. 48:376-385. 15.

Powers AR, Mathys C, Corlett PR (2017): Pavlovian conditioning–induced

hallucinations result from overweighting of perceptual priors. Science. 357:596-600. 16.

Stroop JR (1935): Studies of interference in serial verbal reactions. Journal of

Experimental Psychology. 18:643-662. 17.

Wang X, Wang T, Chen Z, Hitchman G, Liu Y, Chen A (2015): Functional connectivity

patterns reflect individual differences in conflict adaptation. Neuropsychologia. 70:177-184. 18.

Shapira-Lichter I, Strauss I, Oren N, Gazit T, Sammartino F, Giacobbe P, et al. (2018):

Conflict monitoring mechanism at the single-neuron level in the human ventral anterior cingulate cortex. Neuroimage. 175:45-55.

17

19.

Taylor R, Théberge J, Williamson PC, Densmore M, Neufeld RWJ (2016): ACC Neuro-

over-Connectivity Is Associated with Mathematically Modeled Additional Encoding Operations of Schizophrenia Stroop-Task Performance. Frontiers in psychology. 7:1295-1295. 20.

Taylor R, Neufeld RWJ, Schaefer B, Densmore M, Rajakumar N, Osuch EA, et al.

(2015): Functional magnetic resonance spectroscopy of glutamate in schizophrenia and major depressive disorder: anterior cingulate activity during a color-word Stroop task. NPJ schizophrenia. 1:15028-15028. 21.

Bitzer S, Park H, Blankenburg F, Kiebel S (2014): Perceptual decision making: drift-

diffusion model is equivalent to a Bayesian model. Front Hum Neurosci. 22.

Sporns O, Kötter R (2004): Motifs in Brain Networks. PLOS Biology. 2:e369.

23.

Pardo JV, Pardo PJ, Janer KW, Raichle ME (1990): The anterior cingulate cortex

mediates processing selection in the Stroop attentional conflict paradigm. Proceedings of the National Academy of Sciences of the United States of America. 87:256-259. 24.

Evrard HC, Forro T, Logothetis NK (2012): Von economo neurons in the anterior insula

of the macaque monkey. Neuron. 74:482-489. 25.

Mesulam MM, Mufson EJ (1982): Insula of the old world monkey. III: Efferent cortical

output and comments on function. J Comp Neurol. 212:38-52. 26.

Mufson EJ, Mesulam MM (1982): Insula of the old world monkey. II: Afferent cortical

input and comments on the claustrum. J Comp Neurol. 212:23-37.

18

27.

Palaniyappan L, Liddle PF (2012): Does the salience network play a cardinal role in

psychosis? An emerging hypothesis of insular dysfunction. J Psychiatry Neurosci, 2011/06/23 ed, pp 17-27. 28.

Palaniyappan L, Simmonite M, White TP, Liddle EB, Liddle PF (2013): Neural primacy

of the salience processing system in schizophrenia. Neuron. 79:814-828. 29.

Moran LV, Tagamets MA, Sampath H, O'Donnell A, Stein EA, Kochunov P, et al.

(2013): Disruption of anterior insula modulation of large-scale brain networks in schizophrenia. Biol Psychiatry. 74:467-474. 30.

Limongi R, Sutherland SC, Zhu J, Young ME, Habib R (2013): Temporal prediction

errors modulate cingulate-insular coupling. Neuroimage, pp 147-157. 31.

Merritt K, McGuire P, Egerton A (2013): Relationship between Glutamate Dysfunction

and Symptoms and Cognitive Function in Psychosis. Frontiers in psychiatry. 4:151-151. 32.

Adams R, Stephan K, Brown H, Frith C, Friston K (2013): The Computational Anatomy

of Psychosis. Frontiers in Psychiatry. 4. 33.

Valton V, Romaniuk L, Douglas Steele J, Lawrie S, Seriès P (2017): Comprehensive

review: Computational modelling of schizophrenia. Neuroscience and biobehavioral reviews. 83:631-646. 34.

Jardri R, Denève S (2013): Circular inferences in schizophrenia. Brain : a journal of

neurology. 136:3227-3241.

19

35.

Association AP (2013): Diagnostic and Statistical Manual of Mental Disorders. Arlington

991: American Psychiatric Publishing, Inc. 36.

Leckman JF, Sholomskas D, Thompson WD, Belanger A, Weissman MM (1982): Best

estimate of lifetime psychiatric diagnosis: a methodological study. Archives of general psychiatry. 39:879-883. 37.

Andreasen NC, Carpenter WTJ, Kane JM, Lasser RA, Marder SR, Weinberger DR

(2005): Remission in schizophrenia: proposed criteria and rationale for consensus. American Journal of Psychiatry. 162:441 - 449. 38.

Wong D, Schranz AL, Bartha R (2018): Optimized in vivo brain glutamate measurement

using long-echo-time semi-LASER at 7 T. NMR in Biomedicine. 31:e4002. 39.

Fard PR, Park H, Warkentin A, Kiebel SJ, Bitzer S (2017): A Bayesian Reformulation of

the Extended Drift-Diffusion Model in Perceptual Decision Making. Frontiers in Computational Neuroscience. 11. 40.

Marreiros AC, Kiebel SJ, Friston KJ (2008): Dynamic causal modelling for fMRI: A

two-state model. Neuroimage. 39:269-278. 41.

Razi A, Kahan J, Rees G, Friston KJ (2015): Construct validation of a DCM for resting

state fMRI. Neuroimage. 106:1-14. 42.

Kahan J, Urner M, Moran R, Flandin G, Marreiros A, Mancini L, et al. (2014): Resting

state functional MRI in Parkinson's disease: the impact of deep brain stimulation on 'effective' connectivity. Brain : a journal of neurology. 137:1130-1144.

20

43.

Limongi R, Perez FJ, Modrono C, Gonzalez-Mora JL (2016): Temporal Uncertainty and

Temporal Estimation Errors Affect Insular Activity and the Frontostriatal Indirect Pathway during Action Update: A Predictive Coding Study. Frontiers in Human Neuroscience. 10:14. 44.

Friston KJ, Litvak V, Oswal A, Razi A, Stephan KE, van Wijk BCM, et al. (2016):

Bayesian model reduction and empirical Bayes for group (DCM) studies. NeuroImage. 128:413431. 45.

Zeidman P, Jafarian A, Seghier ML, Litvak V, Cagnan H, Price CJ, et al. (2019): A guide

to group effective connectivity analysis, part 2: Second level analysis with PEB. NeuroImage. 200:12-25. 46.

Bastos-Leite AJ, Ridgway GR, Silveira C, Norton A, Reis S, Friston KJ (2015):

Dysconnectivity within the default mode in first-episode schizophrenia: a stochastic dynamic causal modeling study with functional magnetic resonance imaging. Schizophr Bull. 41:144-153. 47.

Zhou Y, Zeidman P, Wu S, Razi A, Chen C, Yang L, et al. (2018): Altered intrinsic and

extrinsic connectivity in schizophrenia. NeuroImage Clinical. 17:704-716. 48.

Bryant JE, Frölich M, Tran S, Reid MA, Lahti AC, Kraguljac NV (2019): Ketamine

induced changes in regional cerebral blood flow, interregional connectivity patterns, and glutamate metabolism. Journal of Psychiatric Research. 117:108-115. 49.

Tremblay R, Lee S, Rudy B (2016): GABAergic Interneurons in the Neocortex: From

Cellular Properties to Circuits. Neuron. 91:260-292.

21

50.

Evans SL, Averbeck BB, Furl N (2015): Jumping to conclusions in schizophrenia.

Neuropsychiatr Dis Treat. 11:1615-1624. 51.

Valton V, Karvelis P, Richards KL, Seitz AR, Lawrie SM, Seriès P (2019): Acquisition

of visual priors and induced hallucinations in chronic schizophrenia. Brain. 142:2523-2537. 52.

André N, Audiffren M, Baumeister RF (2019): An Integrative Model of Effortful

Control. Frontiers in Systems Neuroscience. 13. 53.

Friston K, Thornton C, Clark A (2012): Free-energy minimization and the dark-room

problem. Frontiers in psychology. 3:130-130. 54.

Friston KJ, Daunizeau J, Kiebel SJ (2009): Reinforcement Learning or Active Inference?

PLOS ONE. 4:e6421. 55.

Supekar K, Cai W, Krishnadas R, Palaniyappan L, Menon V (2019): Dysregulated Brain

Dynamics in a Triple-Network Saliency Model of Schizophrenia and Its Relation to Psychosis. Biol Psychiatry. 85:60-69.

22

TABLES Table 1. Summary statistics of accuracy and RT in the Stroop task Accuracy (proportion) Group

RT (s)

Condition M

SD

M

SD

Color-only

0.99

0.07

1.06

0.27

Congruent

0.98

0.14

0.98

0.27

Incongruent

0.92

0.27

1.19

0.31

Word-only

0.97

0.17

0.95

0.26

Color-only

0.99

0.05

0.89

0.20

Congruent

0.99

0.05

0.82

0.21

Incongruent

0.98

0.14

1.04

0.25

Word-only

1.00

0.00

0.79

0.18

FEP

HC

23

FIGURES LEGENDS Fig. 1. Voxel positioning for 1H-MRS measurements, sphere positioning for fMRI timeseries extraction, and two-neuronal-population DCM. In the dACC, the fMRI sphere falls within the MRS voxel (black box). The grey box shows the two-neuronal-population DCM of the dACC–AI network. Each region comprises one population of excitatory neurons (E) and one population of inhibitory neurons (I). Parameters of effective connectivity represent the influence of inhibitory to excitatory connections (IE, assumed to be GABAergic neurons), the influence of excitatory to inhibitory connections (EI), the influence of self-inhibitory connections within each population (SE, SI), and the influence of excitatory population of one region on the excitatory population of the other region (EE, assumed to be glutamatergic connections). Whereas EI, SE, and SI parameters are fixed in the model, IE and EE are free parameters. The small white box shows a sample spectrum obtained using 1H-MRS semi-LASER (TE = 100 ms). A frequentist two-sample t-test did not reveal statistically significant difference between the mean [Glu]dACC in the FEP group (M = 6.90 mM, SD = 1.35) relative to the HC group (M = 6.38 mM, SD = 1.3); t(36.72) = 1.21, p = 0.88. Fig. 2. Bayesian model comparison results and mean parameter estimates in the dACC–AI network across groups. PP (posterior probability), * PP ≈ 0. Fig. 3. Main effects of the winning three-level parametric empirical Bayes model. In A, bars represent differences between groups as defined by the effect coding (FEP = 1, HC = -1), [Glu]dACC, prior beliefs, and precision of PE were mean centered, PP (posterior probability) Fig. 4. Interactions between group and covariates in the winning three-level parametric empirical Bayes model. In A, B, and C, positive values represent stronger effect in the FEP group, and vice versa, PP (posterior probability).

24

Fig. 5. Association between Extrinsic Connectivity Strength and Severity of Social Withdrawal. β = -15.59, SE = 5.61, t(16) = 2.78, p = 0.013, 95% CI [27.43, 3.75].

25

(X = 1, Y = 16, Z = 38

Fixed Parameters

Forward E E

Estimated Parameters

I Backward

I

dACC dACC (fMRI)

AI

(X = 1, Y = 16, Z = 38)

9.00

Data Fit [Glu]dACC mM

Residuals

7.50 6.00 4.50 3.00

HC

(X = 0.8, Y = 10.3, Z = 42.1)

FEP

dACC (MRS)

1.50 0.00

Other metabolites

AI (fMRI) (X = 38, Y = 20, Z = -4)

Glutamate

Group

4.00 3.00 2.00

-1.00 -2.00

90 % Highest Density Interval

0.00

PP > 0.95

Posterior Estimate

1.00

Connection dACC

dACC→AI

AI→dACC

AI

1.00 0.80

PP

0.60 0.40 0.20 *

*

Null Model

No-glutamate

*

*

*

Group

Behavioral

0.00 Three-level PEB Model

No-computational

A

C

Group

0.20

1.00

0.00

0.50

-0.20

0.00

-0.40

-0.50

-0.60 dACC

B

dACC→AI

AI→dACC

AI

[Glu]dACC

0.30

Connection

90 % Highest Density Interval

1.50

PP < 0.95

0.40

PP > 0.95

2.00

Posterior Estimate

0.60

-1.00 dACC

D

dACC→AI

AI→dACC

AI

Prior Beliefs

16.00

0.20

12.00

0.10

8.00

0.00

4.00

-0.10

0.00

-0.20

-4.00

-0.30

Precision of PE

-8.00 dACC

dACC→AI

AI→dACC

AI

dACC

dACC→AI

AI→dACC

AI

A

C

Group × [Glu]dACC

Group × Prior Beliefs

0.30

100.00

0.20

80.00

0.10

60.00

0.00

40.00

-0.10

20.00

-0.20

0.00

-0.30

-20.00 dACC

B

dACC→AI

AI→dACC

dACC

AI

dACC→AI

Group × Precision of PE

-0.50 -1.00

Connection

-1.50 -2.00 dACC

dACC→AI

AI→dACC

AI

90 % Highest Density Interval

0.00

PP < 0.95

Posterior Estimate

0.50

PP > 0.95

1.00

AI→dACC

AI

6

Severity of Social Withdrawal

5

4

3

2

1

0 0.65

0.7

0.75 0.8 dACC→AI Connectivity Strength

0.85

0.9

KEY RESOURCES TABLE Resource Type

Specific Reagent or Resource

Add additional rows as needed for each Include species and sex when applicable. resource type

Source or Reference Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.

Software; Algorithm

SPM12

Wellcome Centre for Human Neuroimaging https://www.fil.ion.ucl.ac.uk/spm/software/spm12/

Software; Algorithm

JMP 14

JMP Statistical Discovery from SAS https://www.jmp.com/en_us/software/newrelease/preview-jmp14.html

Software; Algorithm

HDDM 7.0.1

http://ski.clps.brown.edu/hddm_docs/

Key Resource Table

Limongi et al.

Glutamate and Dysconnection in the Salience Network: Neurochemical, Effective Connectivity, and Computational Evidence in Schizophrenia Supplemental Information

1

Supplemental Methods Table S1. Demographic information of participants. (C, Caucasian), (A / ME, Asian or Middle Eastern), (NSSEC, national statistics socioeconomic status), (SES, socioeconomic status), (SOFAS, social and occupational functioning assessment scale), (N/A non reliable information). Age FEP male M = 21.75, SD = 3.59; age FEP female M = 21.71, SD = 3.15; age HC male M = 21.09, SD = 4.23; age HC female M = 21.44, SD = 3.17. No group difference in age was detected across groups (p > |t| = 0.43). Subject

Group

Ethnicity

SOFAS

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC HC FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP

C A / ME C A / ME C A / ME C C A / ME C C C A / ME C A / ME C A / ME C C C C C C C C C A / ME C C A / ME C C A / ME C C C C C C

73 85 79 85 87 81 83 86 80 79 79 83 80 80 85 80 85 85 80 85 40 37 40 60 30 51 34 50 25 40 33 65 25 44 20 55 50 40 45

Cannabis Use (selfendorsed) 0 0 0 0 0 0 0 0 0 1 0 0 0 1 1 0 1 0 0 0 1 1 1 1 1 0 0 1 1 1 1 0 1 1 0 1 1 0 1

Parental SES (NSSEC)

Age at Study Date

Gender

3 4 2 5 1 5 2 2 3 2 2 1 4 2 3 2 5 5 5 2 4 5 2 2 4 4 5 2 2 2 2 3 2 3 2 4 1 5 4

26 23 17 23 16 25 16 16 29 22 23 20 20 20 20 20 27 22 18 22 19 20 19 17 18 17 24 21 25 28 20 23 23 24 23 20 27 26 19

Male Female Male Male Female Male Male Male Male Female Female Male Male Male Male Female Female Female Female Female Male Male Male Male Male Female Male Male Male Male Female Female Male Female Female Male Male Female Female

Age Psychosis Onset (years) — — — — — — — — — — — — — — — — — — — — 17 20 19 17 18 16 23 21 20 28 18 N/A 22 24 N/A 19 21 26 19

Duration of Untreated Psychosis (months) — — — — — — — — — — — — — — — — — — — — 24 4 1 2 2 9 12 1 59 3 14 N/A 6 1 N/A 9 72 1 1

2

Medication FEP patients with less than 2 weeks of lifetime antipsychotic exposure were selected for our study, and roughly 40% of patients were not exposed to any antipsychotic at the time of assessment (Table S2). Defined Daily Dose exposure was calculated for these patients and the mean antipsychotic Defined Daily Dose in our sample was 1.05 (suggesting the average patient had had 1-day worth of the minimum maintenance dose at the time of assessment). HC subjects did not report any antipsychotic medication. Table S2. Participants medication at the day of assessment. (N/A non reliable information) Subject 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Group

Days of Antipsychotic

Antipsychotic

Dose (mg)

Duration (days)

Defined Daily Dose (DDD)

FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP

0 0 7 0 7 0 0 0 7 0 0 5 0 7 0 0 0 7 10

0 0 Paliperidone 0 Aripiprazole 0 0 0 Paliperidone Olanzapine 0 Risperidone 0 Olanzapine 0 0 0 Risperidone Aripiprazole

0 0 6 0 5 0 0 0 3, 6 10 0 1.5 0 7.5 0 0 0 1 5

0 0 7 0 7 0 0 0 3, 4 1 0 5 0 7 0 0 0 7 10

0 0 7 0 2.6 0 0 0 5.5 1 0 1.5 0 5.25 0 0 0 1.4 3.333

3

Table S3. Symptoms scores assessed via the 8-item positive and negative symptoms scale (PANSS-8). HC subjects scored 1 in all metrics (P1, delusions; P2, conceptual disorganization; P3, hallucinations; N1, blunted affect; N4, social withdrawal; N6, lack of spontaneity; G5, mannerisms; G9, unusual thoughts). Subject

Group

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP FEP

PANSS-8 P1 4 5 4 5 5 5 6 5 7 6 6 4 5 5 7 5 7 5 5

P3 4 5 4 5 4 5 6 4 5 2 2 5 5 2 6 5 2 5 1

G9 4 4 3 4 3 2 5 4 6 3 4 3 5 5 6 3 3 3 4

P2 4 4 5 1 3 1 5 4 3 4 3 2 6 4 4 1 3 1 3

G5 2 3 1 1 1 3 4 1 2 1 1 1 3 1 1 1 1 1 1

N1 4 2 5 3 2 3 1 2 4 1 1 3 1 1 1 4 5 1 3

N4 5 3 4 4 3 5 1 4 3 1 1 1 3 1 1 3 4 5 3

N6 3 1 3 3 3 3 1 2 1 1 1 2 4 1 1 1 3 1 3

PANSS Negative 12 6 12 10 8 11 3 8 8 3 3 6 8 3 3 8 12 7 9

PANSS Positive 12 14 13 11 12 11 17 13 15 12 11 11 16 11 17 11 12 11 9

PANSS-8 Total 30 27 29 26 24 27 29 26 31 19 19 21 32 20 27 23 28 22 23

Anatomical voxel location There was a threefold control for voxel location: First, we placed the voxel based on anatomical landmarks. The center was placed on the junction of the right cingulate sulcus with the paracentral sulcus (angled at the antero-posterior line, tangential to the corpus callosum), with the base of the voxel resting on the grey matter fold immediately above the corpus callosum in the sagittal sections. In transverse sections, the voxel was centered on the interhemispheric fissure. Second, we confirmed our MRS voxel location by creating a probabilistic map in the MNI space. This was done by co-registering our MRS file to MP2RAGE image using Gannet 3.0 (1) and then normalizing the MP2RAGE image to an MNI template, with applying deformation matrix to MRS voxel binary map, using SPM12. After repeating this for all participants, an average voxel binary map was calculated for all HC and all FEP participants and the average MRS voxel binary map was overlay onto a normalized image from a single-subject (Figure S1). Third, we corrected for frequency and phase drifts during post-processing, as described in (2). 4

Figure S1. MRS voxel and fMRI sphere overlapping and previous task activation map in the dACC. Overlaps (top) are shown separately for FEP and HC groups. The task activation map (bottom) corresponds to a previously reported work (3). In that work, three groups (schizophrenia, HC, and major depressive disorder) performed the four conditions of the Stroop task (the same four conditions our participants performed in the current work). The design matrix shows a task (collapsed across the four conditions) vs. baseline contrast only for the schizophrenia group. The task activation map shows a dACC cluster which reliably comprises our dACC ROI. In the current work, this regional engagement was confirmed by means of Bayesian model comparison. Spectral post-processing The spectral acquisition was analyzed to measure glutamate concentration using methods previously developed in our laboratory (4-7). Specifically, the 32 spectral acquisition was corrected for frequency and phase drifts as described in Near, Edden (2). The corrected spectra were then averaged into one spectrum to represent the entire four-minute resting-state block. The averaged spectrum underwent post-processing using combined QUALITY Eddy Current Correction (8), using 400 QUALITY (quantification improvement by converting line shapes to the Lorentzian type) points, to reduce linewidth distortions as well as Hankel Singular Value Decomposition (HSVD) water removal to remove any residual water signal between 4.2 ppm and 5.7 ppm before being fit with fitMAN (4, 5), a time-domain fitting algorithm that uses a non-linear, iterative Levenberg-Marquardt minimization algorithm to echo time-specific prior knowledge templates. The metabolite fitting template included eighteen brain metabolites: alanine, aspartate, choline, creatine, 𝛾𝛾-aminobutyric acid (GABA), glucose, glutamate, glutamine, glutathione, glycine, lactate, myo5

inositol, N-acetyl aspartate, N-acetyl aspartyl glutamate, phosphorylethanolamine, scyllo-inositol, and taurine. As well, a single peak was used to fit the water unsuppressed spectrum. Using Barstool (4), tissue-specific (gray matter, white matter, and CSF) T1 and T2 relaxations were corrected through partial volume segmentation calculations of voxels mapped onto T1-weighted images acquired using a 0.75 mm isotropic MP2RAGE sequence (repetition time = 6000 ms, TI1= 800 ms, TI2 = 2700 ms, flip-angle 1 (α1) = 4°, flip-angle 2 (α2) = 5°, field of view = 350 mm × 263 mm × 350 mm, acquisition time = 9 min. 38 s, iPATPE = 3 and 6/8 k-space). Finally, glutamate concentration quantification was calculated using the postprocessed water suppressed and unsuppressed spectra for each participant along with voxel-appropriate, tissue-specific relaxation time adjustments (9). To assess the quality of our acquired 1H-MRS data, signal-to-noise ratio (SNRNAA) was calculated by dividing the amplitude of the NAA CH3 peak in the frequency domain by the standard deviation of the noise in the last 32 most upfield points of the spectrum. The linewidth was also calculated using waterunsuppressed spectra and measuring the FWHM of the water peak in Hertz (Hz). To assess the reliability of the negative finding, we now provide an estimate of the largest real difference that would have not been detected given the observed amount of measurement variability. This estimate is the smallest detectable difference (SDD) between the two groups. Based on Théberge, Williamson (10), we calculated the SDD and computed the percent size of the smallest detectable between-groups difference (18.77%) using zα(0.05, one-tailed) = 1.645 and zβ(0.8, two-tailed) = 1.28. Preprocessing and general linear model Functional images were realigned, normalized to the Montreal Neurological Institute (MNI) space, and spatially smoothed using a 4 mm (full width at half maximum) Gaussian kernel. We fit a general linear model to the images and included six head movement parameters and time series corresponding to the white matter and cerebrospinal fluid as regressors. We included a cosine basis set with frequencies ranging from 0.0078 to 0.1 Hz (11). Images were high-pass filtered to remove slow frequency drifts (< 0.0078 Hz). 6

Stroop task We implemented the same Stroop task we used in (3). By pressing a key in a four-key response pad, participants identified as quickly and as accurate as possible the color of the ink of a string of letters (on a gray background) representing either the name of a color (yellow, green, blue, or red) or a series of ‘Xs’ (‘XXXX’) projected on screen visible via a mirror located in the scanner’s coil. The task comprised four conditions: word-only (a word in white ink saying a color name, e.g., the word ‘yellow’ written in white ink), color-only (an ‘XXXX’ string with the color ink), congruent (the ink color matching the meaning of the word, e.g., the word ‘red’ in red ink), and incongruent (the ink color differing from the meaning of the word, e.g., the word ‘red’ written in blue ink). Outside the scanner, participants rehearsed the task until reaching 80 % accuracy —collapsed across conditions. Inside the scanner, participants performed 20 trials per condition. They were allotted 2 s to respond with an interstimulus trial interval of 1 s. Hierarchical drift diffusion model (HDDM) Using Bayesian and Markov chain Monte Carlo methods (18), we estimated the parameters of the driftdiffusion model from participants of each group separately. We assumed that subjects of a particular group were alike. Therefore, individuals’ parameters were constrained by groups’ parameters —which allowed us to estimate between-groups difference without over or underestimating the true group-level variance (19). Prior distributions were informative as per the default option used in the HDDM package (18). The chain length was 200,000. The number of burn-in iterations was 2000, and the chains were generated with thinning = 20. Data analysis pipeline Analysis was performed at three levels: neurochemical (MRS), effective connectivity (fMRI), and computational (Stroop). All levels informed the “three-level” model tested via parametric empirical Bayes and Bayesian model comparison (Figure S2).

7

Figure S2. Data analysis pipeline relevant to the three level of analyses: computational, networkconnectivity, and neurochemical.

Supplemental Results Regional engagement in task We confirmed the engagement of the dACC and AI regions in the task via model selection. A model including the task-elicited computations had the highest probability of explaining the fMRI data (please see variance accounted for by the model in the Table S2). Crucially, if the task had not engaged these regions, either the null or the no-computational model would have performed better than the three-level model. Within the context of previous reports, the region of interest (ROI) corresponding to the dACC falls within a larger cluster reported by our group (3) in which we observed significant activation in schizophrenia subjects performing the same Stroop paradigm as in the current study (see Figure S1). Similarly, our insular ROI falls within a cluster that has been shown active during the Stroop task in schizophrenia, as reported by a metanalysis (12). Interestingly, our effective connectivity evidence of regional engagement in task expands previous works linking (resting state) functional connectivity and task-evoked measurements (1317). Note, however, that from the perspective of DCM task-evoked activation differs from task-related regional engagement. Certainly, although our DCM proves regional engagement in task it does not comprise task-context as driving inputs to regions.

8

Stroop task Accuracy and RT (separately) were regressed against condition, group, and the Condition × Group interaction via a mixed-effects linear model (52). Participants were included as random effects. As expected, FEP participants performed worse than HC participants (Table S4). Planned F contrasts confirmed that they were less accurate (F(1, 3026) = 39.19, p < 0.001) and took longer (F(1, 3023) = 72.77, p < 0.001) in the incongruent condition. This analysis was performed in JMP 14 (Key Resources Table). Furthermore, the recovered parameter estimates of the DDM (Table S5 and Figure S3) accurately reproduce the observed response accuracy and reaction times (Table S6). DDM analysis was performed in HDDM 7.0.1 (Key Resources Table). Table S4. Parameter estimates of the mixed effects model fit to the response accuracy and RT of the Stroop task. Estimate

SE

DF

t

P

Intercept

0.980

0.006

19.209

162.19

<.0001

95 % CI Lower Upper 0.967 0.992

Group [FEP]

-0.014

0.002

3036.937

-5.81

<.0001

-0.019

-0.009

Condition [Color-only]

0.016

0.004

3018.308

3.89

0.0001

0.008

0.024

Condition [Congruent]

0.009

0.004

3018.325

2.23

0.0257

0.001

0.017

Condition [Incongruent]

-0.030

0.004

3018.493

-7.24

<.0001

-0.039

-0.022

Group [FEP] × Condition [Color-only]

0.012

0.004

3018.354

2.97

0.0030

0.004

0.020

Group [FEP]× Condition [Congruent]

0.005

0.004

3018.330

1.32

0.1884

-0.003

0.014

Group [FEP] × Condition [Incongruent]

-0.016

0.004

3018.624

-3.92

<.0001

-0.025

-0.008

Intercept

0.968

0.015

19.088

62.85

<.0001

0.936

1.000

Group [FEP]

0.081

0.004

3032.809

18.52

<.0001

0.072

0.089

Condition [Color-only]

0.012

0.007

3018.130

1.60

0.1100

-0.003

0.027

Condition [Congruent]

-0.068

0.007

3018.138

-9.08

<.0001

-0.082

-0.053

Condition [Incongruent]

0.152

0.008

3018.222

20.13

<.0001

0.137

0.167

Group [FEP] × Condition [Color-only]

0.004

0.007

3018.153

0.54

0.5871

-0.011

0.019

Group [FEP]× Condition [Congruent]

0.001

0.007

3018.140

0.19

0.8476

-0.013

0.016

Group [FEP] × Condition [Incongruent]

-0.006

0.008

3018.287

-0.80

0.4252

-0.021

0.009

Parameter

Accuracy

RT

9

Table S5. Parameter estimates of the HDDM (subject level). Parameters: a (decision threshold), z (starting point or prior beliefs), v (drift rate or precision of prediction errors), t (nondecision processes). Parameter Group

HC

FEP

a Subject 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

M 2.535 2.558 2.650 2.538 2.535 2.678 2.373 2.592 2.541 2.439 2.475 2.462 2.431 2.344 2.510 2.656 2.509 2.507 2.505 2.474 2.365 2.619 1.371 1.916 1.797 1.943 1.854 2.173 1.707 2.006 2.580 2.439 1.950 2.460 2.411 1.974 1.666 1.905 2.147

v SD 0.259 0.275 0.299 0.236 0.241 0.295 0.248 0.266 0.258 0.241 0.235 0.237 0.263 0.282 0.236 0.300 0.245 0.272 0.244 0.245 0.486 0.548 0.173 0.368 0.243 0.263 0.275 0.390 0.268 0.221 0.457 0.393 0.257 0.452 0.350 0.285 0.214 0.274 0.317

M 2.872 2.884 2.735 2.845 2.881 2.561 2.841 2.863 2.871 2.795 2.554 2.902 3.051 2.926 2.912 2.701 2.767 3.011 2.891 2.804 2.305 1.954 1.322 2.524 1.633 1.784 1.202 2.422 2.259 0.746 1.797 1.877 1.654 2.248 1.621 2.052 1.290 2.021 1.839

t SD 0.263 0.272 0.253 0.260 0.267 0.285 0.263 0.270 0.265 0.254 0.291 0.268 0.331 0.288 0.275 0.256 0.262 0.319 0.266 0.262 0.403 0.377 0.366 0.434 0.335 0.321 0.350 0.421 0.404 0.280 0.338 0.322 0.323 0.404 0.305 0.343 0.302 0.355 0.332

M 0.455 0.540 0.524 0.300 0.363 0.563 0.377 0.447 0.454 0.455 0.548 0.360 0.458 0.503 0.324 0.486 0.590 0.582 0.389 0.528 0.821 0.739 0.533 0.743 0.642 0.543 0.800 0.652 0.640 0.681 0.573 0.463 0.707 0.567 0.546 0.395 0.739 0.599 0.733

z SD 0.057 0.064 0.065 0.038 0.044 0.058 0.050 0.052 0.055 0.054 0.064 0.048 0.059 0.067 0.041 0.062 0.064 0.068 0.050 0.059 0.082 0.090 0.029 0.063 0.056 0.050 0.095 0.056 0.046 0.065 0.067 0.051 0.057 0.054 0.053 0.042 0.063 0.048 0.065

M 0.326 0.326 0.325 0.326 0.326 0.324 0.327 0.326 0.326 0.326 0.325 0.327 0.327 0.327 0.326 0.325 0.325 0.326 0.326 0.325 0.443 0.442 0.441 0.444 0.443 0.444 0.443 0.443 0.443 0.442 0.442 0.443 0.442 0.442 0.441 0.445 0.445 0.443 0.442

SD 0.034 0.034 0.034 0.034 0.034 0.034 0.034 0.034 0.034 0.034 0.034 0.035 0.035 0.035 0.034 0.034 0.034 0.035 0.034 0.034 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035 0.035

10

Figure S3. Posterior probability (PP) and visual representation of the drift diffusion model. Parameters: a (decision threshold), z (starting point or prior beliefs), t (non-decision processes), v (drift rate).

11

Goodness of fit and convergence assessment of the drift diffusion model Model convergence was assessed by computing the R-hat statistic (20) on five chains of the model, 0.99 ≤ R-hat ≤ 1. Goodness of fit was assessed via posterior predictive checks (21) (Table S6). Table S6. Goodness of fit of the drift diffusion model. Accuracy, mean proportion of response accuracy; Mean = mean RT to reach the relevant boundary; ub, upper boundary, lb, lower boundary Group

FEP

HC

Statistic

Observed

Estimated

SD

SEM

MSE

Credible

Quantile

Mahalanobis

Accuracy

0.919

0.932

0.098

0.000

0.010

TRUE

30.663

0.124

Mean_ub

1.184

1.228

0.211

0.002

0.047

TRUE

45.305

0.208

SD_ub

0.300

0.339

0.132

0.001

0.019

TRUE

44.895

0.292

10q_ub

0.817

0.895

0.148

0.006

0.028

TRUE

29.842

0.533

30q_ub

0.987

1.015

0.171

0.001

0.030

TRUE

44.979

0.166

50q_ub

1.153

1.146

0.204

0.000

0.042

TRUE

55.011

0.033

70q_ub

1.352

1.322

0.256

0.001

0.067

TRUE

59.547

0.119

90q_ub

1.607

1.650

0.363

0.002

0.134

TRUE

50.421

0.120

Mean_lb

-1.263

-1.132

0.349

0.017

0.139

TRUE

28.215

0.377

SD_lb

0.359

0.139

0.198

0.048

0.088

TRUE

87.181

1.108

10q_lb

0.767

1.011

0.322

0.060

0.163

TRUE

19.290

0.758

30q_lb

1.028

1.057

0.328

0.001

0.108

TRUE

55.071

0.089

50q_lb

1.287

1.108

0.348

0.032

0.153

TRUE

76.957

0.513

70q_lb

1.487

1.178

0.387

0.095

0.245

TRUE

82.231

0.797

90q_lb

1.747

1.270

0.467

0.227

0.445

TRUE

85.639

1.019

Accuracy Mean_ub SD_ub 10q_ub 30q_ub 50q_ub 70q_ub 90q_ub Mean_lb SD_lb 10q_lb 30q_lb 50q_lb 70q_lb 90q_lb

0.980 1.045 0.256 0.770 0.887 1.003 1.133 1.408 -1.016 0.103 0.915 0.920 0.995 1.100 1.135

0.987 1.057 0.254 0.794 0.902 1.008 1.140 1.372 -0.751 0.019 0.736 0.743 0.751 0.759 0.767

0.028 0.131 0.071 0.113 0.120 0.131 0.149 0.202 0.222 0.058 0.221 0.220 0.222 0.225 0.232

0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.001 0.070 0.007 0.032 0.031 0.060 0.117 0.136

0.001 0.017 0.005 0.013 0.015 0.017 0.022 0.042 0.119 0.011 0.081 0.080 0.109 0.167 0.190

TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE TRUE

22.240 48.530 56.930 42.750 46.470 49.870 50.240 60.810 9.847 93.121 84.577 84.487 89.164 93.031 93.345

0.242 0.091 0.033 0.214 0.131 0.032 0.051 0.180 1.193 1.475 0.807 0.801 1.103 1.516 1.587

12

Variance accounted for by the PEB model The mean proportion of variance accounted for by the model was 0.17 (SD = 9. 73, 95 % CI [12.67, 21.79]) in the HC group and 0.23 (SD = 12.93, 95% CI = [16.49, 28.94]) in the FEP group. —a proportion of a least 0.10 of variance accounted for by the model (averaged across subjects) is considered acceptable (22). Subject-wise values are shown in Table S6. Table S7. Percentage of variance accounted for by the three-level model Subject

HC

FEP

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 10 20

5.1 6.21 18.23 32.19 32.55 11.97 31.64 15.28 19.77 12.18 5.76 21.81 24.01 6.48 30.61 27.6 11.82 7.61 15.3 8.51

21.94 9.48 37.41 20.83 20.13 18.03 3.6 8.37 5.39 31.89 38.74 46 27.92 31.96 36.86 31.09 9.82 6.09 26.32 -

Data Files Two folders (“group” and “subjects”) and an excel file are available at doi:10.17632/pn94pydhv7.1. The “group” folder comprises all competing PEBs and a demo script. After saving this folder in the SPM directory in MATLAB, the interested reader can reproduce the Bayesian model comparison procedure by running the relevant script. Furthermore, glutamate data and parameters of sensory PE and prior beliefs can be extracted from any PEB model comprising the relevant regressor. The “subjects” folder comprises all subjects’ dynamic causal models (including the fMRI timeseries for the relevant regions of interest). The excel file comprises the four conditions of the Stroop task performed by the subjects. 13

Supplementary References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. 22.

Edden RAE, Puts NAJ, Harris AD, Barker PB, Evans CJ (2014): Gannet: A batch-processing tool for the quantitative analysis of gamma-aminobutyric acid–edited MR spectroscopy spectra. Journal of Magnetic Resonance Imaging. 40:1445-1452. Near J, Edden R, Evans CJ, Paquin R, Harris A, Jezzard P (2015): Frequency and phase drift correction of magnetic resonance spectroscopy data by spectral registration in the time domain. Magnetic Resonance in Medicine. 73:44-50. Taylor R, Neufeld RWJ, Schaefer B, Densmore M, Rajakumar N, Osuch EA, et al. (2015): Functional magnetic resonance spectroscopy of glutamate in schizophrenia and major depressive disorder: anterior cingulate activity during a color-word Stroop task. NPJ schizophrenia. 1:15028-15028. Wong D, Schranz AL, Bartha R (2018): Optimized in vivo brain glutamate measurement using long-echo-time semi-LASER at 7 T. NMR in Biomedicine. 31:e4002. Bartha R, Drost DJ, Williamson PC (1999): Factors affecting the quantification of short echo in-vivo 1H MR spectra: prior knowledge, peak elimination, and filtering. NMR Biomed. 12:205-216. Ryan K, Wawrzyn K, Gati J, Chronik BA, Wong D, Duggal N, et al. (2018): ¹H MR Spectroscopy of the Motor Cortex Immediately following Transcranial Direct Current Stimulation at 7 Tesla. bioRxiv.322941. Penner J, Bartha R (2015): Semi-LASER 1H MR spectroscopy at 7 Tesla in human brain: Metabolite quantification incorporating subject-specific macromolecule removal. Magnetic Resonance in Medicine. 74:412. Bartha R, Drost DJ, Menon RS, Williamson PC (2000): Spectroscopic lineshape correction by QUECC: Combined QUALITY deconvolution and eddy current correction. Magnetic Resonance in Medicine. 44:641645. Gasparovic C, Song T, Devier D, Bockholt HJ, Caprihan A, Mullins PG, et al. (2006): Use of tissue water as a concentration reference for proton spectroscopic imaging. Magnetic Resonance in Medicine. 55:1219-1226. Théberge J, Williamson KE, Aoyama N, Drost DJ, Manchanda R, Malla AK, et al. (2007): Longitudinal greymatter and glutamatergic losses in first-episode schizophrenia. British Journal of Psychiatry. 191:325-334. Kahan J, Urner M, Moran R, Flandin G, Marreiros A, Mancini L, et al. (2014): Resting state functional MRI in Parkinson's disease: the impact of deep brain stimulation on 'effective' connectivity. Brain : a journal of neurology. 137:1130-1144. Minzenberg MJ, Laird AR, Thelen S, Carter CS, Glahn DC (2009): Meta-analysis of 41 functional neuroimaging studies of executive function in schizophrenia. Archives of general psychiatry. 66:811-822. Cole MW, Ito T, Bassett DS, Schultz DH (2016): Activity flow over resting-state networks shapes cognitive task activations. Nature Neuroscience. 19:1718-1726. Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, et al. (2009): Correspondence of the brain's functional architecture during activation and rest. Proceedings of the National Academy of Sciences. 106:1304013045. Tavor I, Jones OP, Mars RB, Smith SM, Behrens TE, Jbabdi S (2016): Task-free MRI predicts individual differences in brain activity during task performance. Science. 352:216-220. Cole Michael W, Bassett Danielle S, Power Jonathan D, Braver Todd S, Petersen Steven E (2014): Intrinsic and Task-Evoked Network Architectures of the Human Brain. Neuron. 83:238-251. Ito T, Kulkarni KR, Schultz DH, Mill RD, Chen RH, Solomyak LI, et al. (2017): Cognitive task information is transferred between brain regions via resting-state network topology. Nature communications. 8:1027-1027. Wiecki TV, Sofer I, Frank MJ (2013): HDDM: Hierarchical Bayesian estimation of the Drift-Diffusion Model in Python. Front Neuroinform. 7:14. Boehm U, Marsman M, Matzke D, Wagenmakers E-J (2018): On the importance of avoiding shortcuts in applying cognitive models to hierarchical data. Behavior Research Methods. 50:1614-1631. Gelman A, Rubin DB (1992): Inference from Iterative Simulation Using Multiple Sequences. Statistical Science. 7:457-472. Kruschke JK (2013): Posterior predictive checks can and should be Bayesian: Comment on Gelman and Shalizi, ‘Philosophy and the practice of Bayesian statistics’. British Journal of Mathematical and Statistical Psychology. 66:45-56. Zeidman P, Jafarian A, Corbin N, Seghier ML, Razi A, Price CJ, et al. (2019): A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI. NeuroImage. 200:174-190. 14