Glutamatergic modulation of auditory cortex connectivity with attentional brain networks in unpredictable perceptual environment

In a stable environment the brain can minimize processing required for sensory input by forming a predictive model of the surrounding world and suppressing neural response to predicted stimuli. Unpredicted stimuli lead to a prediction error signal propagation through the perceptual network, and resulting adjustment to the predictive model. The inter-regional plasticity which enables the model-building and model-adjustment is hypothesized to be mediated via glutamatergic receptors. While pharmacological challenge studies with glutamate receptor ligands have demonstrated impact on prediction-error indices, it is not clear how inter-individual differences in the glutamate system affect the prediction-error processing in non-medicated state. In the present study we examined 20 healthy young subjects with resting-state proton MRS spectroscopy to characterize glutamate + glutamine (rs-Glx) levels in their Heschl’s gyrus (HG), and related this to HG functional connectivity during a roving auditory oddball protocol. No rs-Glx effects were found within the frontotemporal prediction-error network. Larger rs-Glx signal was related to stronger connectivity between HG and bilateral inferior parietal lobule during unpredictable auditory stimulation. We also found effects of rs-Glx on the coherence of default mode network and frontoparietal network during unpredictable auditory stimulation. Our results demonstrate the importance of Glx in modulating long-range connections and wider networks in the brain during perceptual inference.

The neurophysiological underpinnings of the auditory change detection have been related to glutamate, the major excitatory neurotransmitter. Glutamatergic neurotransmission has been suggested to implement the top-down predictions and carry information about the bottom-up prediction errors 3 . Consequently, individuals with high glutamate levels could have more efficient prediction error signaling, as suggested by a combined MRS-EEG study 11 .
How glutamate modulates the signaling between the prediction error generators distributed within the network consisting of superior temporal plane and inferior frontal regions, as well as with the wide-spread brain networks supporting attention is an important question. Blocking glutamatergic NMDA-receptors with ketamine has been shown to reduce long-distance inter-regional functional connectivity within the auditory network 3 . Further, in the context of predictive modeling, ketamine has been suggested to reduce specifically the higherorder prediction errors in the later post-stimulus time window 12 , which has been associated with fronto-parietal activation. Additional line of evidence comes from imaging unperturbed glutamate using magnetic resonance spectroscopy (MRS). The MRS-measured resting-state level of Glx (composite of glutamate and glutamine) in auditory cortex has been shown to be related to faster mismatch response for certain types of deviant events (duration-and gap-deviants), which have been hypothesized to require more complex frontal network 11,13 .
Taken together, glutamatergic signaling is clearly important for the information exchange between the auditory areas and higher-level, frontoparietal networks in order to represent the confidence in the environment in which the standards and deviants are embedded.
Another important question is the behavior of widespread brain networks during unpredictable auditory stimulation, and how this is modulated by glutamatergic neurotransmission. Ketamine has been shown to reduce deactivation in the task-negative default mode network and activation in the task-positive frontoparietal regions during a working memory task 14 . In agreement with this it has been shown that rs-Glu and GABA measured in posterior cingulate cortex predict task-induced deactivations in the entire (task-negative) default mode network during a working memory task 15 , while rs-Glx in dorsal ACC predicted task-induced BOLD response in inferior parietal (task-positive) areas during high cognitive control demands 16 . The findings generally emphasize that the regional glutamate concentration effects are shown exclusively in inter-regional connections (outside the measured area), possibly reflecting the long-range projections from glutamatergic excitatory neurons 17 .
In the present study we examined how rs-Glx affects signaling within the frontotemporal auditory prediction error network, and relationship of the auditory cortex with wider networks (frontoparietal attention network and default mode network). We used a paradigm previously used to examine brain responses to auditory unpredictability in fMRI 8 . Subjects were exposed to blocks of tones in predictable order (ABAB) and unpredictable order (alternating A-and B-tone sequences of varying length, so-called 'roving standard' protocol). Two different deviating-feature types were used in separate conditions (frequency and duration). Glutamate was quantified from a resting-state measurement of Glx with 1 H-MRS in Heschl's gyrus (HG) before fMRI was used to measure the BOLD response to the auditory stimulation.
We hypothesized that: (1) during the unpredictable blocks individuals with higher rs-Glx show increased signaling in the frontotemporal auditory hierarchical prediction network consisting of bilateral HG, planum temporale (PT), posterior STG (pSTG) and inferior frontal gyrus (IFG) (Fig. 1, right); (2) during the unpredictable blocks there should be increased connectivity between auditory cortex and areas involved in the frontoparietal attention network in individuals with higher rs-Glx;  www.nature.com/scientificreports/ (3) rs-Glx levels relate to strength of the anticorrelation between task-negative (DMN) and task-positive (frontoparietal) networks during the unpredictable blocks, and increased within-network connectivity in the frontoparietal network.

Methods
Subjects. All methods were carried out in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. The experimental protocol was approved by the regional ethics committee of Western Norway (REK-Vest). The subjects were 20 healthy young right-handed adults (9 male, 11 female, age 20-38 years, mean 25.5, SD 4.8), psychiatrically and neurologically healthy by self-report, recruited from community by convenience sampling. The subjects received oral and written information about the study before participation and signed an informed consent form. The subjects were asked to avoid consuming nicotine and caffeine within 5 h before the experiment. The subjects' hearing acuity was measured in both ears for frequencies 250, 500, 1,000, 2000 and 3,000 Hz; exclusion criterion for hearing loss was acuity level larger than 25 dB SPL for all tested frequencies in either ear. The first principal component of the hearing acuity was used as a covariate for all analyses.
Auditory paradigm. The stimulation paradigm is presented in Fig. 1  MRS data analysis. The MRS data were analyzed using LCModel (version 6.3-1L) (Provencher, 1993) with a standard basis set (Ala, Asp, Cr, PCr, GABA, Glc, Gln, Glu, GPC, PCh, GSH, Ins, Lac, NAA, NAAG, Scyllo, Tau), scaling the metabolite estimates to the internal water reference. For water-scaling, voxel-specific estimates calculated based on the tissue composition of the MRS voxel. Briefly, the estimates were corrected to account for the differences in water concentration and water relaxation times and partial volume effects depending on the tissue class inside the voxel (WM, GM, CSF). The fractional proportion of the GM, WM and CSF within the spectroscopy voxels was estimated using SPM12 routines to segment the anatomical image into grey matter, white matter and CSF images. The proportion of GM, WM and CSF was measured within the MRS voxel mask. The correction was implemented for each metabolite (term metab) according to the equation where frac indicates fraction of the tissue class in voxel, conc water,TC is the concentration of water in the tissue class (conc GM/WM/CSF = 4,330/36,080/54,840 mM respectively, Ernst 1993). The parameters for the metabolites of interest were: T1/T2glu: 1,270/180 18,19 , T1/T2cre: 1,460/152 18 . Glx was expressed as creatine ratio (Glx/ Cre) for consistency with previous reports on the relationship between Glx and neurocognitive variables 17 . To verify that significant effects are not due to creatine, we examined the significant connections in a regression analysis where we entered the water-scaled concentration of Glx and Cre separately, full details are reported in Supplementary Materials. fMRi data preprocessing. The preprocessing was conducted with SPM12 functions incorporated in CONN toolbox processing pipeline. The functional images were realigned, normalized to MNI space, and smoothed with a Gaussian kernel with 6 mm full width at half maximum. FSL Harvard-Oxford atlas was used to define the following ROIs: bilateral Heschl's gyrus (L_HG and R_HG), bilateral pSTG (L_pSTG and R_pSTG), bilateral planum temporale (L_PT, R_PT) and bilateral inferior frontal gyrus, pars opercularis (L_IFG, R_IFG). The ROI data was extracted from non-smoothed volumes as the mean timeseries of the voxels within the ROI mask. Further preprocessing for the connectivity analysis consisted of removing the subject-specific movement parameters estimated during realignment step, the white matter and CSF signal, and the main effects of the task blocks (modeled as boxcar functions for the duration of the stimulation block, convolved with the hemodynamic response function), finally the data were filtered with a bandpass filter from 0.008 to 0.09 Hz. www.nature.com/scientificreports/ Data extraction and statistical analysis. A post hoc power analysis of the achieved statistical power using G*Power software (3.1.9.7) indicates that our sample size (n = 20) enables moderate power for mediumsized effects in a 2-way repeated measures design (power = 0.6), and adequate power for large-sized effects (power = 0.9). Within each subject, the connectivity was calculated as bivariate correlation between the timeseries of all ROIs. For whole-brain seed-to-voxel analysis, the connectivity between the timeseries L_HG and R_HG and the rest of the voxels in the entire brain within each subject was calculated.
For the second-level analyses, we examined the change in the connectivity in both the seed-to-voxel and ROI-to-ROI analyses between the predictable and unpredictable blocks (unpred > pred contrast), as well as modulation of that change by rs-Glx. For this, the rs-Glx measures in the left and right hemisphere were entered into the second-level model as covariates.
The ROI-to-ROI analysis examined the effect of left and right rs-Glx on any of the connections between all the ROIs in the frontotemporal network during unpred > pred contrast (FDR-corrected at p < 0.05 for all connections in the analysis). Additionally we performed a planned test (one-sided uncorrected p < 0.05) for increased connectivity in the connections between L_HG and L_PT as well as L_HG and L_STG, due to earlier studies showing specific modulation of connection between HG and posterior STG by ketamine 3 .
To explore connectivity changes due to rs-Glx outside of the frontotemporal network we performed a seedto-voxel analysis, where we examined the effect of ipsilateral rs-Glx on the connectivity between each of the two HG ROIs and the rest of the brain during unpred > pred contrast (height threshold uncorrected p < 0.005 and cluster extent threshold FWE-corrected p < 0.01).
Finally, we tested whether the DMN and FPN change their properties depending on the predictability of the blocks, and whether this is modified by rs-Glx levels.  45). We tested the within-network connectivity and between-network connectivity, defined as mean correlation across all nodes within each network, and mean correlation between all the nodes of two networks, respectively. The within-and between-network connectivity was calculated using the withinbetweenROI tool in the CONN toolbox. The statistical model, consisting of within-subject factors predictability and condition with between-subject covariates left rs-Glx and right rs-Glx as above was estimated using SPSS. For these three analyses, a Bonferroni correction was applied for the significance level of 0.05, giving the corrected alpha level of 0.017.

ROI-to-ROI analysis: the effect of rs-Glx on the frontotemporal network. First, we examined
the effect of predictability on the frontotemporal network (main effect of unpred > pred). The contrast showed negative connectivity between the temporal ROIs (Fig. 2), While both during unpredictable and predictable blocks there was positive connectivity within the auditory network consisting of the temporal ROIs, this was lower in the unpredictable blocks (see Supplementary Fig. 2, left). Additionally, there was increased connectivity between L_IFG and L_HG (Fig. 2); this was due to more positive-going connectivity in unpredictable blocks ( Supplementary Fig. 2, right). There was no significant condition*predictability interaction (all p > 0.95 FDR, all t < 2.2), indicating that the observed connectivity changes due to unpredictability were common for Duration and Frequency conditions. Next, we examined the effect of left and right rs-Glx on the frontotemporal network during unpred > pred. For the connectivity between all ROIs we did not find any effect of either left or right rs-Glx on the main effect of predictability (all p > 0.5 FDR, all abs(t) < 2.1), or the condition*predictability interaction (all p > 0.6 FDR, all abs(t) < 2.6).
The planned comparison concentrating on the connections between the L_HG and L_pSTG, L_PT did not show any rs-Glx effects on the main effect of predictability (all p > 0.9 FDR, all abs(t) < 0.95), or the condition*predictability interaction (all p > 0.7, all abs(t) < 0.49).
Seed-to-voxel analysis: the effect of rs-Glx on whole-brain connectivity. For the whole-brain connectivity analysis of the unpred > pred effect with the seed in the L_HG we found an effect of left rs-Glx in bilateral inferior parietal lobule (MNI x, y, z = 50, -48, 32 and -40, -58, 42), see Fig. 3. To verify that the effect was not due to creatine-scaling, we performed a hierarchical regression analysis with the water-scaled values of rs-Glx and rs-Cre, including the hearing acuity in the model. This control analysis (see Supplementary Materials) showed that water-scaled Glx was significantly related to the unpred > pred effect while controlling for the other variables (standardized beta = 0.83, t = 3.54, p = 0.003).
Using the SPM Anatomy toolbox 20 to localize the significant clusters showed that in the right hemisphere, the activation was in areas PGa (29.6% of the anatomical area covered), PGp (14.4% area covered), intraparietal sulcus area hIP1 (18% area covered), and PFm (6% area covered). In the left hemisphere the activation was in areas PGa (32.4% area covered), PGp (7% area covered), PFm (8.7% area covered), and intraparietal sulcus area hIP1 (10.8% area covered). We did not find any effects for the analysis in the right hemisphere (seed R_HG).
The effect of rs-Glx on DMN and FPN within-network connectivity. We hypothesized that higher rs-Glx level should lead to larger unpred-pred difference in the connectivity within the examined networks (FPN and DMN), reflecting the brain's increased ability to react to unpredictable stimulation by recruiting the FPN while disengaging the DMN.  19). There was a trend-level, small effect of left rs-Glx on the interaction between feature type and deviation (p = 0.019, F(1,16) = 6.75, η 2 = 0.05). To explore this effect, we calculated the unpred > pred difference for both duration and frequency condition, and tested for the relationship between the unpred > pred difference and left rs-Glx. The post-hoc regression analysis, including hearing accuracy and right rs-Glx as covariates, showed that the effect of left rs-Glx on unpred > pred difference was different in the two conditions (Fig. 4), being a weaker negative relationship in duration (standardized beta = − 0.34, t = − 1.47) and a stronger positive relationship in frequency (standardized beta = 0.41, t = 1.81).
In FPN, in all four conditions there was significant positive connectivity within the network (duration pred, duration unpred, frequency pred, frequency unpred: Cohen's d 3.55, 2.42, 2.84, 2.54). There was a small effect which did not pass the corrected significance level of right rs-Glx on the interaction between feature type and Top left: Unpred > pred main effect was significant in several connections (p < .05, two-sided, FDR-corrected across all connections in the analysis) in the frontotemporal network; we observed reduced connectivity during unpred blocks in the temporal areas, and increased connectivity during unpred blocks between left HG and left IFG. Top right: The full connectivity profile, including significant and non-significant results for the unpred > pred contrast, shows that overall the frontal-to-temporal connections were positive (increased connectivity during unpred blocks), while temporal-to-temporal connections were negative (reduced connectivity during unpred blocks). Bottom: The right and left HG connectivity profile was overall similar, with emphasis on the reduced connectivity during unpred blocks among the temporal areas. Left HG specifically showed increased connectivity with ipsilateral IFG during the unpred blocks.

The effect of rs-Glx on DMN and FPN between-network connectivity.
We examined betweennetwork connectivity, to test whether there was increased anticorrelation between DMN and FPN with increased rs-Glx. In all of four conditions (duration pred, duration unpred, frequency pred, frequency unpred), there was a small but significant positive connectivity between the networks (Cohen's d: 0.81, 0.71, 0.84, 0.61). We did not find a significant unpred > pred contrast (F(1,16)  The only significant effect we did observe was a small main effect of condition (F(1,16) = 5.2, p = 0.037, η 2 = 0.04), due to higher positive connectivity between the networks during duration than frequency condition.

Discussion
Glutamate has been related to coordinating the work of large brain networks during stimulus processing and cognitively demanding tasks. Glutamatergic receptors have been related to mediating the plasticity seen in the learning during oddball tasks: forming the model of the expected stimulation so that the deviance from the model could be detected. In this study we examined the effect of glutamate level in the primary auditory cortex on the brain activity during unpredictable auditory stimulation. We presented unpredictable blocks where the standard was repeatedly altered, which according to the predictive coding theories should lead to prediction error signaling. We observed that during the unpredictable blocks there was reduced connectivity within the www.nature.com/scientificreports/ auditory network and increased connectivity with prefrontal cortex, which is consistent with increased input from hierarchically higher levels during prediction error signaling. The rs-Glx level in the left auditory cortex was correlated to increased connectivity between the left auditory cortex and bilateral inferior parietal lobule. Further, we saw that in agreement with our hypotheses the rs-Glx level affected the coherence of both DMN and FPN during unpredictable stimulation: higher rs-Glx was associated with more coherence in FPN, whereas lower rs-Glx was associated with more coherence in DMN. However, this effect was specific to duration-deviant blocks, and depended on the hemisphere where rs-Glx was measured. Contrary to our hypotheses we did not find an effect of rs-Glx on connectivity within the hypothesized frontotemporal network based on source modeling of electrophysiological studies. We note that while this study had sufficient statistical power to detect large effects, we do not have sufficient power to discuss negative findings; it is possible that there exists a small to moderate effect. The areas in the inferior parietal lobe which showed connectivity with the left HG depending on the glutamate level were the areas in the caudal portion of the inferior parietal lobule (areas PGa and PGp). The rostral and caudal portion of IPL each exhibit a distinct cytoarchitectonic profile 21 . Also, these areas show functional segregation, with the rostral part (regions termed PF) related to motor and sensory functions, while the caudal part (PG) shows activation during complex cognitive abilities involved in language as well as attention 22 . The anatomical connectivity in the caudal areas is selectively oriented towards lateral occipital and temporal areas, following the inferior longitudinal fascicle; these regions are also connected to inferior frontal cortex, and show similarity in the distribution of glutamatergic and other receptors 22 . In particular, the caudal portion shows strong and preferential anatomical connectivity with areas in the temporal lobe, including the primary auditory cortex and planum temporale 23 . A significant left-right asymmetry in these connections has been found, with preferential connectivity between left primary auditory cortex and caudal IPL.
The involvement of IPL has been observed in several paradigms related to the current experiment. The IPL regions have been shown to be related to MMN for both duration-and frequency-deviance 8 . In the study by Molholm and colleagues, hemispheric differences were noted: the frequency-deviant blocks predominantly activated www.nature.com/scientificreports/ the right hemisphere, whereas the duration-related activation was bilateral or left-dominant. In our study, we did not find significant interaction effects suggesting the predominance of either hemisphere, however we did observe stronger effect in the right hemisphere, underlining its importance for both frequency-and duration-deviance. The important causal role for the parietal lobe in particular for duration-MMN has also been demonstrated by a TMS study showing reduced duration-MMN amplitude following a parietal TMS stimulation 24 . While there have been repeated attempts to relate the ventral attention network to oddball processing via the activations seen in the right inferior parietal region, this association deserves some careful anatomical consideration. The inferior portion of the region PGa (PGi), proposed based on the parcellation from the HCP project 25,26 , is suggested to be a part of the ventral attention network 27 . This region, however, appears to be more ventral than the areas found here; thus there is no clear evidence from this study of increased signaling between primary auditory cortex and ventral attention network during auditory deviance. Taking a more general view relative to the ventral attention network role of the IPL, Ptak has suggested that the role of the right IPL is related to shifting and maintaining of attention, interacting with a priority map which is held in the nearby inferior parietal sulcus 10 .
Beyond attention, the parietal areas found here have been considered significant in representing predictionrelated activity. During temporally unpredictable situations, right parietal activations have been observed, independent of WM load 28 , and posterior parietal cortex has been associated with generating temporal expectancies 29 . In a recent meta-analysis, bilateral inferior parietal lobule (focus in IPS) was found to be sensitive to the magnitude of surprise in a reinforcement learning situation 30 . There appear to be fine-grained distinctions in the inferior parietal region, with the IPS being more sensitive to trial-specific surprise, whereas PGp is more activated for between-trial updating 31 . Intraparietal sulcus has been associated with relative uncertainty, which is maximal after a change in action-outcome mapping 32 . Also, uncertainty reduction via belief updating has been found to activate IPS together with superior frontolateral regions 33 . The present results suggest that the glutamatergic neurotransmission in sensory cortex plays an important role in the representation of expectancy, by influencing the connectivity between the expectancy-related brain areas and sensory brain regions.
Regarding the findings on the FPN and DMN, we predicted that across both deviance types there would be increased anticorrelation between the networks during unpredictable blocks, which should be stronger in participants with higher rs-Glx levels. We expected this anticorrelation to result from stronger connectivity within the FPN and weaker connectivity within DMN with higher rs-Glx levels. This hypothesis did not find support in the data. As above, we are aware that the sample size limited our statistical power to detect small-to moderate-sized effects. We did however find small, trend-level effects suggesting that during duration-deviant blocks there may be effects of rs-Glx in the predicted direction on the within-network connectivity in both DMN as well as FPN. Here, the exploratory analysis suggested that direction of the effects was as predicted: higher rs-Glx was related to higher FPN and lower DMN within-network connectivity during duration-deviant blocks, suggesting that individuals with higher rs-Glx mobilize their executive network during demanding perceptual environment, and correspondingly deactivate their DMN. However, the effects were dependent on the hemisphere where the rs-Glx was measured. In FPN, higher level of rs-Glx in the right hemisphere predicted higher within-network connectivity during the duration-deviant compared to duration-standard blocks. In DMN, higher levels of rs-Glx in the left hemisphere predicted lower within-network connectivity. Frontoparietal brain areas have been found to be engaged in estimating the environmental certainty, with increased environmental uncertainty leading to increased frontoparietal activation. Regularities, by contrast, have been shown to reduce metabolic demands in various areas. The dorsal frontoparietal network shows decreased activation for regular stimulation, where the prediction error is lower 34 . Also, predictability has been associated with greater DMN activity 35 . The importance of the right-sided rs-Glx may be related to different roles of the hemispheres for processing deviance in the stimulus features used here. Generally, representing certainty has been associated more with right than left frontoparietal network regions 36 . However, the frontoparietal areas in the right hemisphere in particular have been suggested to be important for duration discrimination 37 ; also other studies have noted the particular relationship between parietal areas in the right hemisphere and duration-discrimination [36][37][38][39][40] , also in brain-damaged patients 41,42 .
Taken together, our results show the importance of rs-Glx on the brain activation during unpredictable auditory stimulation. The primary auditory cortex connectivity with caudal portion of inferior parietal lobule (regions PGa and PGp and IPS) was higher in subjects with higher rs-Glx levels. This was observed independently of the deviating feature (duration or frequency). Rs-Glx did not have effect on the local connectivity within the temporal lobes. This agrees with previous findings of MRS-measured Glx: the effects are visible on long-range communication, not on local activation levels 17 .
We suggest that the finding of increased communication between left HG and bilateral IPL reflects the increased calculations relevant to updating the representations of the auditory stimulation dependent on the surprise. Individuals with higher rs-Glx level in auditory cortex are more sensitive to the mismatch between predicted and actual stimulation as the increased glutamatergic neurotransmission allows more efficient signaling of prediction errors. This leads to more surprise-related activity in the hierarchically higher regions in the inferior parietal lobule.
We also demonstrate that the rs-Glx level has different effects on duration-deviance and frequency-deviance. Only the duration-deviant blocks were related to the network-wide activation as predicted, with higher rs-Glx related to increased connectivity within the FPN and decreased connectivity in DMN. The duration-representation has been shown to be particularly sensitive to conditions affecting the glutamatergic system, including high-risk to psychosis 43 , and more sensitive to wider-spread calculations in the brain, including input from frontal sources 13 . The present results underline the difference in learning to predict different auditory features.
While this work is done in healthy adults with no pharmacological challenge, it has relevance for patients with schizophrenia. The ketamine model of psychosis, where the blockade of NMDA receptors leads to psychosis symptoms, has highlighted the importance of glutamatergic system in psychosis. There is a potential Scientific RepoRtS | (2020) 10:15059 | https://doi.org/10.1038/s41598-020-72044-0 www.nature.com/scientificreports/ contradiction between results from MRS studies with healthy adults and clinical patients, where higher rs-Glx levels are associated with positive symptoms (see e.g. 44 ), which in turn are related to reduced ability to process deviating or unpredictable information 45 . However, the data on schizophrenia patients suggests an imbalance in regional metabolite concentration 46,47 , with dynamic changes over the course of the illness 48 , emphasizing that the dysfunction in the glutamate system may manifest as a dysregulation due to impaired NMDA receptors, leading to various effects throughout the brain, rather than reduction of glutamatergic signaling. Further studies are required to establish the relationship between in vivo measurements of glutamate and stimulus processing in schizophrenia; improved measurement techniques such as rapid event-related MRS combined with fMRI and EEG may help to provide the important data needed for fuller understanding of the glutamatergic system in psychiatric disorders.

Data availability
The data is available from the corresponding author upon request.