Predictive acoustical processing in human cortical layers Authors: Lonike K. Faes, Isma Zulfiqar, Luca Vizioli, Zidan Yu, Yuan-hao Wu, Jiyun N. Shin, Martijn A. Cloos, Ryszard Auksztulewicz, Lucia Melloni, Kamil Uludag, Essa Yacoub, Federico De Martino Published: 2026 ABSTRACT In our dynamic environments, predictive processing is vital for auditory perception and its associated behaviors. Predictive coding formalizes inferential processes by implementing them as information exchange across cortical layers and areas. With laminar-specific blood oxygenation level-dependent imaging we measured responses to a cascading oddball paradigm, to ground predictive auditory processes on the mesoscopic human cortical architecture. Using a modeling approach, we derive putative changes in neural dynamics while accounting for draining effects. We show that the violations of predictions are potentially hierarchically organized and associated with responses in superficial layers of the planum polare and middle layers of the lateral temporal cortex. Moreover, we relate the updating of the brain’s internal model to changes in deep layers. Our results support the role of temporal cortical architecture in the implementation of predictive coding and highlight the ability of laminar fMRI to investigate mesoscopic processes in a large extent of temporal areas. Introduction The ability to form predictions about incoming sensory stimuli is crucial for everyday behavior. Sounds that reach our ears are often noisy and incomplete and by anticipating what we will hear next, the brain can efficiently filter and prioritize behaviorally relevant sounds, facilitating for example auditory stream segregation and understanding speech in noise1–3. Aberrant predictive processing is even hypothesized to underlie phantom perception such as tinnitus4 or hallucinations5–7. A widely referenced theory of perceptual processing is predictive coding (PC8,9), suggesting that this process of inference occurs through the hierarchical exchange of information across brain areas, attributing specialized roles to feedforward and feedback streams, by capitalizing on the known anatomical separation across cortical layers10, but see11. In PC, descending feedback is hypothesized to convey predictions of lower-level representations that are compared to sensory input at lower stages of the hierarchy. Conversely, instead of forwarding the sensory input, the discrepancy between the prediction and this sensory input (i.e., the prediction error) is fed forward to higher-order brain areas12–14. The relevance of superficial layers to the processing of contextual (top-down) information is also underscored by their entrainment to acoustic patterns that can be separated solely based on their statistical regularity15. Evidence in favor of distinct computational roles of cortical layers for PC mainly comes from invasive animal studies16–21. Nevertheless, it remains an open question whether this version of PC is translatable to humans22. Invasive electrophysiological measures in patients offer a potential means of bridging the gap between animal and human literature. Although direct recordings using laminar electrodes in humans are feasible, to date, these electrodes still offer recordings from a limited number of areas across the cortex and have not yet been used to test hypotheses related to PC23–25. Ultra-high field functional magnetic resonance imaging (UHF fMRI) offers an alternative to non-invasively investigate the laminar information processing in the human brain. The use of UHF fMRI has yielded insights into the specific contributions of cortical layers in predictive processing. A wealth of studies has centered on the visual26–29 and the somatosensory cortex30 and have broadly confirmed the hypotheses regarding the mesoscopic implementation of PC. PC has been proposed as a general framework for understanding how perceptual processing occurs in the brain. As such, it is commonly assumed that mechanistic implementations generalize across senses; yet, its implementation in the human auditory cortex remains untested. Moreover, an important assumption of PC is the existence of a hierarchical organization (for a review see ref. 13). This has already been explored using invasive electrophysiology in both rodents18,31 and non-human primates16,21,32–34, and humans35–37. However, due to its invasiveness, these studies are limited to patient populations. Therefore, the laminar profile of hierarchical information exchange remains largely unexplored in humans (but see ref. 27). In this work, we present the first human laminar fMRI results investigating how the violation of expectation modulates responses in multiple regions of the auditory cortex probed by a cascading oddball paradigm. To this end, we collected high spatial resolution (0.8 mm isotropic) fMRI data using gradient-echo blood oxygenation level-dependent (GE-BOLD). The data was analyzed using a previously proposed laminar BOLD model38, to disentangle underlying neuronal activity from (unrelated) vascular changes within the dynamic causal modeling (DCM) framework39. We show distinct roles of cortical laminae in processing unpredictable and mispredicted tone sequences compared to predictable sequences. More specifically, for prediction violations, we find error-related signals in superficial layers in a lower-level auditory region, whereas we find middle layer modulation in a higher-order temporal area. Furthermore, a widespread deep layer effect is present across the temporal cortex, possibly indicating global updating of the internal model of the brain both in mispredicted and unpredictable tone sequences. Results Task and fMRI Ten healthy volunteers passively listened to sequences of four tones. Their local and global predictability was manipulated while the BOLD response was measured across multiple hierarchical auditory areas at submillimeter resolution (0.8 mm isotropic) with gradient-echo echo-planar-imaging (GE-EPI). The first three ‘contextual’ frequency tones (493.9, 659.3, and 987.8 Hz) were common to all sequences, and could either be ascending (3 sequences) or descending (3 sequences). We expected that the three successive ascending or descending tones would create a strong expectation, while minimizing local adaptation driven by e.g., the repetition of the same tone within each trial as in classical oddball paradigms13. The fourth tone in the sequences could either fulfill (predictable condition) or violate (mispredicted condition) these expectations (see Fig. 1A). Finally, in a third condition (unpredictable), the order of the first three tones was rearranged (see Fig. 1A) to reduce the contextual expectation of the last tone, thereby reducing local predictability of the fourth tone and designed to result in a reduced magnitude of local error signals.Fig. 1Timing and conditions of the experimental paradigm.a Three types of conditions were presented. The first three tones either followed a pattern (ascending/descending) that elicited a strong local prediction (predictable and mispredicted) or the order of the sequence was scrambled, decreasing the predictability of the last tone (unpredictable). The fourth tone could either fulfill the prediction (predictable) or violate it (mispredicted). Ascending or descending sequences could end in a low- or high-frequency tone. Across the experiment, predictable conditions were presented more often than the mispredicted and unpredictable conditions (60-20-20) creating a global expectation that the predictable condition was more likely. The figure was adapted from ref. 80. b Each session consisted of 6 to 8 runs. In each run, 36 trials were presented in a slow event-related design with an inter-trial interval of 5 to 7 TRs. The order of conditions was randomized. Timing and conditions of the experimental paradigm. a Three types of conditions were presented. The first three tones either followed a pattern (ascending/descending) that elicited a strong local prediction (predictable and mispredicted) or the order of the sequence was scrambled, decreasing the predictability of the last tone (unpredictable). The fourth tone could either fulfill the prediction (predictable) or violate it (mispredicted). Ascending or descending sequences could end in a low- or high-frequency tone. Across the experiment, predictable conditions were presented more often than the mispredicted and unpredictable conditions (60-20-20) creating a global expectation that the predictable condition was more likely. The figure was adapted from ref. 80. b Each session consisted of 6 to 8 runs. In each run, 36 trials were presented in a slow event-related design with an inter-trial interval of 5 to 7 TRs. The order of conditions was randomized. Moreover, local predictability (within a sequence) was manipulated alongside global predictability (within the experiment). To induce the latter, the predictable condition was presented more frequently (60%, 20%, and 20% respectively for the three condition types, see Fig. 1A). As a result, we expected that mispredicted and unpredictable responses would reflect putative global error signals due to the rarity of these conditions. While we cannot separately analyze local and global errors in our experimental paradigm, as local violations were always accompanied by global violations, we did expect local errors to be reduced for unpredictable conditions due to the reduced contextual expectation elicited by the first three tones. Each session consisted of 6 to 8 runs that lasted about 6.5 min each. Within each run, the conditions were randomized and presented using a slow event-related design with an inter-trial interval ranging between 8–11.2 s (5–7 TRs - Fig. 1B). Submillimeter cortical auditory responses Our functional data slab encompassed the bilateral auditory cortex (Fig. 2A). The anatomical data (0.75 mm isotropic) were first upsampled (0.4 mm isotropic), automatically segmented into white matter (WM), gray matter (GM) and cerebral spinal fluid (CSF) and finally manually corrected to accurately delineate the gray matter ribbon in the bilateral temporal cortices (Fig. 2B). Using cortical depth estimation, a mid-gray matter surface was created from this high-resolution segmentation, on which we defined five regions of interest (ROIs) based on macro-anatomical landmarks (ref. 40 – Fig. 2C). Per ROI, the cortical depth of every voxel, together with the local curvature information led to the assignment of each voxel to different cortical depths using the equivolume approach (Fig. 2D). First, we localized the voxels responding to our tone sequences (all conditions versus baseline, t > 2.0, p < 0.05 uncorrected; Fig. 2E). Subsequently, the analysis proceeded with generating (in each individual subject data and ROI layer) event-related averages (ERA) per condition. These ERAs are obtained by averaging all single trials per condition, time-locked to the start of the stimulus presentation. For illustration purposes, Fig. 2F shows the ERAs of one of the ROIs in a single volunteer, averaged across all layers. See Fig. S1 for activation maps and the parcellation of ROIs of all volunteers, and Fig. S2 for an illustration of the draining effect in individual subjects.Fig. 2Preprocessing before modeling.a Coverage of the functional slab overlaid on the high-resolution anatomy. b Manually edited segmentation of white matter, gray matter and cerebral spinal fluid. c Example of the regions of interest drawn on a slightly inflated mid gray matter surface. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. d The gray matter curvature was divided into equivolume layers. Illustration of the depth estimates of HG in one volunteer. e We then selected voxels that responded to auditory stimulation at p < 0.05 uncorrected. f Per ROI, we average the response across trials of each condition. Here, we average all trials of the conditions ending in a low frequency (PredL, MispredL and UnpredL). Preprocessing before modeling. a Coverage of the functional slab overlaid on the high-resolution anatomy. b Manually edited segmentation of white matter, gray matter and cerebral spinal fluid. c Example of the regions of interest drawn on a slightly inflated mid gray matter surface. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. d The gray matter curvature was divided into equivolume layers. Illustration of the depth estimates of HG in one volunteer. e We then selected voxels that responded to auditory stimulation at p < 0.05 uncorrected. f Per ROI, we average the response across trials of each condition. Here, we average all trials of the conditions ending in a low frequency (PredL, MispredL and UnpredL). GE-BOLD acquisitions are commonly used to survey the cortex at laminar resolution due to their high functional contrast-to-noise ratio (fCNR) and large coverage41,42. This large coverage is especially beneficial for investigating the auditory cortex, since the temporal lobe can be imaged bilaterally43. Yet, a downside of GE-BOLD is its higher sensitivity to (unwanted) macrovasculature effects (e.g., draining veins44), that shows as an increase of percent signal change of BOLD towards the surface (see Fig. 3).Fig. 3Vascular draining towards the surface.For each of the three conditions separately, we illustrate the group average profiles of the percent signal change of the BOLD response before modeling (n = 10). In these laminar plots, an increase from deep towards superficial layers is visible. Note that the x-axis (illustrating percent signal changes) are different across ROI and conditions. Dots in circles indicate the median value and bars illustrate the interquartile range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). For individual plots that illustrate the vascular draining, see Supplementary Fig. 2. Source data are provided in the Source Data file. Vascular draining towards the surface. For each of the three conditions separately, we illustrate the group average profiles of the percent signal change of the BOLD response before modeling (n = 10). In these laminar plots, an increase from deep towards superficial layers is visible. Note that the x-axis (illustrating percent signal changes) are different across ROI and conditions. Dots in circles indicate the median value and bars illustrate the interquartile range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). For individual plots that illustrate the vascular draining, see Supplementary Fig. 2. Source data are provided in the Source Data file. To mitigate this effect, we have applied a validated45 generative laminar BOLD model38, incorporated within the DCM framework. By considering neuronal and vascular effects within a generative model, this approach aims at estimating the difference between conditions at the neural level, having considered the effect of draining veins (see e.g39., and Fig. 4A). Using this approach, we compared the event-related response of mispredicted and unpredictable conditions to the response of predictable ones (in every participant and ROI, see Fig. 4B). DCMs general approach involves the implementation of two distinct manipulations. First, the input driving the neuronal response (common across conditions) is defined, and then contextual manipulations (i.e., experimental conditions) or task demands are defined46. Here, the driving input (i.e., a sequence of four tones) was defined for all conditions, while the experimental manipulation (change in condition) was described by additional modulatory input (see Fig. 4C). In the mispredicted condition, the modulatory input occurred at the time point at which the prediction was violated (i.e., when the fourth tone was presented). In the unpredictable condition, the modulatory input was modeled at the third and fourth tones to capture the modulatory effect both at the moment at which the context was disrupted (i.e., the first two tones follow a different pattern than the last two tones) and at the end of the sequence to capture the effects of contextual disruption on the processing of the fourth tone. In the DCM framework, driving input results in the estimation of an average value across conditions of the free model parameters. Modulatory inputs, on the other hand, allow the estimation of changes in (selected) model parameters to explain the difference between conditions (or task demands). Here, the target of the modulatory input was the self-excitation connection of the excitatory neuronal population(s). Thus, the self-excitation connection of the excitatory neuronal population(s) was estimated once together with all other free parameters to explain average effects across conditions, while a change in the self-excitation connection of the excitatory neuronal population(s) was estimated to explain the difference between conditions. We limited modulatory input to affect only the self-excitation of excitatory neuronal population(s) as this has been previously shown to modulate changes in the overall amplitude of the BOLD response (Fig. 4D, and see39). Moreover, in electrocorticography (ECoG) data, the self-excitatory parameter has been linked to gain modulation in auditory areas when content predictability was manipulated47. Note that due to the relatively short inter-stimulus interval, we only modeled the effect of the modulation of self-excitation and refrained from modeling the BOLD undershoot as we did not capture a sufficient number of time points to model it. While DCM is typically used to perform effective connectivity analyses between regions, here we used the Bayesian framework of DCM (free energy minimization48) to estimate underlying putative laminar neuronal dynamics for each region independently. In addition, the neuronal model we used does not include any intralaminar connections. Within the DCM framework, we compared eight models, each hypothesizing distinct neuronal layer(s) targeted by the modulatory input (Fig. 4E) in each region independently. Apart from the winning model (see Supplementary Fig. 3B), this approach allowed us to estimate (in each participant and ROI) the strength of the modulatory input across the three neuronal depths induced by mispredicted and unpredictable conditions (see Supplementary Figs. 4 and 5 for data without modeling and individual data points). This modulatory input is akin to the deveined differential response between conditions (i.e., the difference between conditions after having considered the vascular draining).Fig. 4Schematic of the modeling approach.a Schematic of the biophysical model (adapted with permission from39). The number of neuronal depths was fixed to 3. The neuronal and hemodynamic models are linked through neurovascular coupling (NVC). The data was modeled at several vascular depths. b Example event-related averages (ERAs) of both data and feedforward (generated) estimation of the laminar BOLD response for three conditions visualized at 3 depths. The model fit illustrates one iteration of the expectation maximization process for one of the tested models. c Schematic illustration of the hypothesis-driven timing of the model inputs used to estimate BOLD responses in panel b. The driving input is the same in all three conditions as this indicates the presentation of an auditory stimulus irrespective of its predictive content. The modulatory input is varied across conditions, occurring at the specific timing of the violation of the prediction (in the mispredicted condition) and at the moment at which the context was disrupted (third tone) and at the end of the sequence to capture the effects of contextual disruption on the processing of the fourth tone in the unpredictable condition. The modulatory input is defined in the mispredicted and unpredictable conditions to estimate the change in model parameters that explains the difference between these conditions and the predictable conditions (acting as baseline). d More detailed schematic of the neuronal model to illustrate the spatial target(s) of the driving and modulatory input. Each layer contains both an excitatory and inhibitory neuronal population. Driving input arrives at all neuronal depths in the excitatory populations. The modulatory input targets the self-excitatory connection of the excitatory population. We test different hypotheses about which layer is targeted by this modulatory input. e A total of 8 models were generated, illustrating our hypotheses about which layer or combination of layers was targeted by the modulation. Driving input (not illustrated here) arrived at all excitatory populations. The red arrows represent the location of modulatory input at the time illustrated by the red arrows in Fig. 4C. Model 8 signifies our null model with no modulatory input targeting any cortical depth. Schematic of the modeling approach. a Schematic of the biophysical model (adapted with permission from39). The number of neuronal depths was fixed to 3. The neuronal and hemodynamic models are linked through neurovascular coupling (NVC). The data was modeled at several vascular depths. b Example event-related averages (ERAs) of both data and feedforward (generated) estimation of the laminar BOLD response for three conditions visualized at 3 depths. The model fit illustrates one iteration of the expectation maximization process for one of the tested models. c Schematic illustration of the hypothesis-driven timing of the model inputs used to estimate BOLD responses in panel b. The driving input is the same in all three conditions as this indicates the presentation of an auditory stimulus irrespective of its predictive content. The modulatory input is varied across conditions, occurring at the specific timing of the violation of the prediction (in the mispredicted condition) and at the moment at which the context was disrupted (third tone) and at the end of the sequence to capture the effects of contextual disruption on the processing of the fourth tone in the unpredictable condition. The modulatory input is defined in the mispredicted and unpredictable conditions to estimate the change in model parameters that explains the difference between these conditions and the predictable conditions (acting as baseline). d More detailed schematic of the neuronal model to illustrate the spatial target(s) of the driving and modulatory input. Each layer contains both an excitatory and inhibitory neuronal population. Driving input arrives at all neuronal depths in the excitatory populations. The modulatory input targets the self-excitatory connection of the excitatory population. We test different hypotheses about which layer is targeted by this modulatory input. e A total of 8 models were generated, illustrating our hypotheses about which layer or combination of layers was targeted by the modulation. Driving input (not illustrated here) arrived at all excitatory populations. The red arrows represent the location of modulatory input at the time illustrated by the red arrows in Fig. 4C. Model 8 signifies our null model with no modulatory input targeting any cortical depth. Error signaling and model updating for mispredicted stimuli We first investigated how mispredicted stimuli are represented in the mesoscopic cortical architecture compared to predictable stimuli. We hypothesized this comparison to highlight the strongest violation effects as well as the global update of the internal model (in response to the representation of a rare event). To this end, we averaged across sequence-type (i.e., ascending or descending sequences) and hemispheres. Figure 5 illustrates (across all subjects and ROIs) the layer-dependent modulation in self-recurrent excitation that best explains the difference between mispredicted and predictable conditions (see Supplementary Fig. 6 for the same plot with individual subject data points). Statistical significance was assessed using a non-parametric t-test (i.e., the null distribution was estimated using permutations obtained by sign flipping the modulation values49) and all p-values were corrected for multiple comparisons using false discovery rate (FDR; across regions and layers) using the Benjamini-Hochberg procedure. The differences were considered significant at q < 0.05. For the results both prior to modeling and after modeling in the same plot, see Supplementary Fig. 7. Note that, by presenting the difference between conditions on the data before modeling (pre-deveining) and the estimated self-recurrent modulation (difference between conditions post-deveining), this Supplementary Fig. allows the evaluation of the effect that modeling has on the data.Fig. 5Neuronal modulation of mispredicted compared to predictable stimuli.For each ROI and cortical layer, the modulation values of the σ parameter (arb. units) that best explain the difference between mispredicted and predictable sequences is plotted across participants (n = 10). Dots in circles indicate the median value and bars illustrate the interquartiles range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). All ROIs exhibit significant modulation in the deep layers in response to mispredicted stimuli; while a superficial layer effect was present in planum polare and a middle layer effect was visible in posterior superior temporal gyrus. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. * indicates q < 0.05. Source data are provided in the Source data file. Neuronal modulation of mispredicted compared to predictable stimuli. For each ROI and cortical layer, the modulation values of the σ parameter (arb. units) that best explain the difference between mispredicted and predictable sequences is plotted across participants (n = 10). Dots in circles indicate the median value and bars illustrate the interquartiles range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). All ROIs exhibit significant modulation in the deep layers in response to mispredicted stimuli; while a superficial layer effect was present in planum polare and a middle layer effect was visible in posterior superior temporal gyrus. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. * indicates q < 0.05. Source data are provided in the Source data file. Our results demonstrate a distinct role of cortical depths in processing the violation induced by mispredicted sequences. In particular, we observed a significant effect in deep layers in all five ROIs (HG: t(9) = 4.88, q = 0.008, BF10 = 45.93; PP: t(9) = 6.53, q = 0.005, BF10 = 267.65; PT: t(9) = 4.37, q = 0.005, BF10 = 25.07; aSTG: t(9) = 6.90, BF10 = 384.73, q = 0.005; pSTG: t(9) = 5.88, q = 0.005, BF10 = 138.19). Interestingly, we observed an effect in the superficial layers in planum polare (t(9) = 3.75, q = 0.005, BF10 = 11.77) and an effect in middle layers of the posterior superior temporal gyrus (t(9) = 4.69, q = 0.005, BF10 = 36.73). See Table S1 for an overview of the t-values, the uncorrected p-values, q-values (after FDR-correction) and Bayes Factors of all ROIs and layers, including the ones that did not reach significance. Note that, the superficial layer effect in pSTG, albeit not significant after FDR-correction (q > 0.05), reaches a Bayes Factor of 4.14, making it worth mentioning50. Processing of unpredictable sequences We then assessed the modulation induced by unpredictable sequences (compared to predictable ones). We hypothesized that the unpredictable sequences would have reduced expectations for the last tone because of the scrambled order of the contextual ones and thus reduced (local) errors, while preserving the need for a global update of the internal model because of the rarity of the events. Using a similar procedure as for the analysis of modulation induced by mispredicted stimuli (averaging across hemisphere and sequence-type, significance testing using permutations and FDR-correction for multiple comparisons), we observed significant effects in deep layers in HG (t(9) = 3.41, q = 0.049, BF10 = 7.57), PP (t(9) = 3.57, q = 0.049, BF10 = 9.30) and aSTG (t(9) = 4.70, q = 0.049, BF10 = 37.05). In the middle and superficial layers, we found no evidence for a condition-specific modulation (Fig. 6). See Table S2 for an overview of the t-values, uncorrected p-values, q-values (after FDR-correction) and Bayes Factors for all ROIs and layers, including the ones that did not reach significance. Note that, the deep layer effect in PT, albeit not significant after FDR-correction, reached a Bayes Factor of 5.80, making it worth mentioning according to50. The results of both the data before and after modeling can be found in Supplementary Fig. 8. Note that, by presenting the difference between conditions on the data before modeling (pre-deveining) and the estimated self-recurrent modulation (difference between conditions post-deveining), this Supplementary Fig. allows the evaluation of the effect that modeling has on the data.Fig. 6Neuronal modulation of unpredictable compared to predictable stimuli.A condition-specific modulation of the σ parameter (arb. units) is present in deep layers in Heschl’s gyrus, planum polare and anterior superior temporal gyrus. This figure illustrates the modulation values of sigma that best explain the difference between unpredictable and predictable sequences across participants (n = 10). Dots in circles indicate the median value and bars illustrate the interquartiles range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). There is no evidence of modulation in the middle and superficial layers. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. * indicates q < 0.05. Source data are provided in the Source Data file. Neuronal modulation of unpredictable compared to predictable stimuli. A condition-specific modulation of the σ parameter (arb. units) is present in deep layers in Heschl’s gyrus, planum polare and anterior superior temporal gyrus. This figure illustrates the modulation values of sigma that best explain the difference between unpredictable and predictable sequences across participants (n = 10). Dots in circles indicate the median value and bars illustrate the interquartiles range (IQR) across participants, with whiskers extending to one time the IQR beyond the limits of the IQR (roughly representing 83% of the data). There is no evidence of modulation in the middle and superficial layers. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. * indicates q < 0.05. Source data are provided in the Source Data file. Discussion To map the mesoscopic circuitry relevant for the processing of the violation of predictions in the auditory cortex, we measured cortical depth-dependent responses to stimuli that adhered to, reduced or violated expectations. Departing from previous research, to maximize neuronal specificity, we employed a modeling approach to account for venous drainage in the most commonly used laminar fMRI acquisition (i.e., GE-EPI). The use of GE-EPI allowed us to explore the hierarchical organization of layer-specific signals across a large extent of the temporal lobe. Our findings revealed modulation in superficial layers in planum polare and in the middle layers of the posterior STG in response to the violation of the internal model. Moreover, we found consistent modulation in deep layers in all our ROIs when contrasting predictable to mispredicted stimuli. In contrast, the unpredictable stimuli modulated the response in deep layers of HG, PP and aSTG, but did not elicit a notable modulation in the superficial or middle layers. In the sections below, we offer an interpretation of our results that considers the implementation of standard PC across the mesoscopic architecture, following the hypothesized roles of distinct layers in the processing of feedforward prediction errors and feedback predictions. We demonstrated that mispredicted sounds elicit a significant modulation in the superficial layers of PP. The involvement of superficial layers in the processing of errors has been hypothesized by the proposed microcircuit for predictive coding12. Evidence in favor of this hypothesis so far has been gathered by invasive research in animals, that has identified superficial layer activity associated with prediction errors in the visual cortex of mice17–20. In the auditory modality, invasive recordings in non-human primates have also identified supragranular layers of the primary auditory cortex to be responsive to the violation of predictions induced by the repetition of the same tone and disentangled from stimulus-specific adaptation21, and to be contextually modulated by top-down influences for the detection of auditory patterns15. In humans, the relevance of the mesoscopic architecture to the processing of errors is so far supported by a study conducted in the visual domain demonstrating that deviant stimuli are only decodable in superficial layers of V129. Here, we contribute to the existing body of research by extending its scope to the mesoscale of the human auditory cortex. It offers novel results showing the involvement of superficial cortical layers in processing prediction errors. In PP, this superficial layer effect we observe could be due to the outputted error when feedback targets only deep layers (which through interlaminar connections arrives in superficial layers for error computation) or it could be due to the combination of modulatory feedback targeting superficial layers and the resulting error that is outputted from superficial layers. In the absence of simultaneous electrophysiological recordings, these two hypotheses cannot be disentangled. In our data, errors are signaled by superficial layers of PP, indicating that the more complex acoustic rule (compared to repeating the same tone) of our cascade stimuli (minimizing stimulus-specific adaptation within a trial) may move the relevant signaling to a higher hierarchical level compared to previous research in non-human primates21. The involvement of non-core areas for the detection of local violations has also been shown in human invasive recordings that investigated prediction violations using speech sounds37. Interestingly, the larger coverage afforded by UHF laminar fMRI also highlighted a modulation in the middle layers of pSTG in response to prediction violations. Within the PC framework, middle cortical layers at higher hierarchical levels receive errors fed forward by preceding processing stages, suggesting the routing of information from PP to pSTG in response to the violation of predictions. The hierarchical nature of predictions and prediction errors has been previously reported in rodents31, non-human primates21,32,33 and humans35,37, while the relevance of superficial and deep layers for errors and predictions (and their updates) respectively has been evidenced in non-human primates16. To our knowledge, a modulation in response to prediction violations has not been previously shown in auditory areas. So far, one previous high-resolution fMRI study did show a middle layer effect in the somatosensory cortex that was driven by predictable sensory stimuli when only sensory input was isolated by contrasting conditions. However, this data was collected using vascular space occupancy (VASO), an acquisition method that is more sensitive to changes in the microvasculature (see e.g30.). In addition to the middle layer effect, in superficial layers of pSTG, albeit not significant after FDR-correction, we found a substantial effect (according to Bayes Factors50) of the modulation of the self-recurrent excitation in response to mispredicted sequences. This result may indicate that pSTG is an intermediate area in the processing violations for the type of stimuli we presented and that errors may ultimately be resolved in higher hierarchical regions. Considering the alignment of the architectonic model developed in non-human primates to the human temporal cortex (see e.g51.,) we can interpret the hierarchical organization of cortical error processing for our type of stimuli. PP is considered to be the human homologue of the belt middle medial (MM) and rostro medial (RM) areas52–54, while the pSTG is thought to be the homologue to the caudal parabelt (CP). In non-human primates, these areas have been shown to be connected, suggesting a similar connectivity in humans55. It is thus conceivable that error signals could travel from superficial layers of PP to middle layers of pSTG (see Figs. 7A, B). Formal evidence for this hypothesis can only be obtained by considering the influence that functional connectivity has on laminar cortical responses.Fig. 7Hypothesized information exchange across the auditory cortical hierarchy.a Schematic overview of hypothesized feedforward and feedback streams across different cortical areas. Feedforward streams are thought to convey a prediction error, whereas the feedback information represents a global model update signal. b Schematic overview of the hypothesized laminar information exchange. Feedforward prediction error is conveyed from superficial layers in planum polare to the middle layer of posterior superior temporal gyrus. The global model update signals are conveyed to deep layers of all areas. Gray arrows indicated the other anatomical connections described in previous research55,86. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. Hypothesized information exchange across the auditory cortical hierarchy. a Schematic overview of hypothesized feedforward and feedback streams across different cortical areas. Feedforward streams are thought to convey a prediction error, whereas the feedback information represents a global model update signal. b Schematic overview of the hypothesized laminar information exchange. Feedforward prediction error is conveyed from superficial layers in planum polare to the middle layer of posterior superior temporal gyrus. The global model update signals are conveyed to deep layers of all areas. Gray arrows indicated the other anatomical connections described in previous research55,86. HG Heschl’s gyrus, PP planum polare, PT planum temporale, aSTG anterior superior temporal gyrus, pSTG posterior superior temporal gyrus. We also found a robust effect in deep layers in response to mispredicted compared to predictable stimuli, which cannot be ascribed to local prediction differences, as the local context is the same across the two conditions. This deep layer neuronal modulation to mispredicted stimuli was not reported in a previous laminar fMRI study that contrasted expected and unexpected stimuli in the visual domain29. Yet, we suggest that this result could indicate similar mechanisms as shown by electrophysiological studies in animals32,33. Using a local-global paradigm, these studies have shown a feedforward sweep of two components of increased high gamma-band activity, representing local and global errors, followed by a late decrease in the alpha/beta power band. This decrease in beta power was suggested to be indicative of a widespread global model update32,33, allowing feedback information entering a cortical area to update their respective internal models. Considering the relationship between high and low frequency oscillations with errors and predictions in the canonical microcircuit12, and the negative correlation of low frequency oscillations to the BOLD response56,57, the increased responses to mispredicted sounds in deep layers could be related to the modulation of beta power caused by feedback to the cortical area. Thus, the increased modulation to mispredicted stimuli in deep layers, suggests that deep layers may play a critical role in representing an updated internal model (see Figs. 7A, 7B), which at the highest level may be stored in frontal cortical areas58,59. These oscillatory dynamics are often related to spiking of neuronal populations12,16, whereas the BOLD response has been shown to be more correlated to synaptic input60. However, BOLD responses are indicative of the metabolic demands of neurons that increase in both cases. Therefore, in our hypothesized mechanism we give suggestions about areas that receive and send information based on standard PC (Fig. 7B). The analysis of the responses elicited by unpredictable stimuli provided further evidence to support our interpretations. We hypothesized that the presentation of unpredictable tone sequences would result in a reduced magnitude of errors due to the rearrangement of the three contextual tones. In addition, as both the mispredicted and unpredictable sequences were presented proportionally less often than the predictable sequences, we hypothesized them to both result in the updating of the internal model. In line with these hypotheses, we found no evidence for superficial and middle layer effects, yet a similar deep layer effect was present in three areas of the temporal cortex. In deep layers of PT, albeit not significant after FDR-correction, we found a substantial effect (according to Bayes Factors50) of the modulation of the self-recurrent excitation in response to unpredictable sequences. This result would support our hypothesized widespread model update. Currently, we have interpreted our results based on the distinct laminar asymmetry of predictions and prediction errors, as hypothesized by standard PC12,13. However, differences in the experimental paradigm employed by us and previous studies using invasive electrophysiology32,33,35–37, impede the one-to-one comparison of results. In our cascading oddball paradigm, used to mitigate stimulus-specific adaptation within each trial, local and global violations always occurred simultaneously, hindering the distinction between local and global errors. To better understand the contribution of different cortical layers (and areas) to local and global effects, future studies may employ tailored experimental designs (e.g., the local-global paradigm61). In addition, while we focused on the inclusion of temporal areas, a logical extension of the coverage would be to include frontal regions, such as the inferior frontal gyrus that was not systematically included in our slab coverage (it was only fully covered in 2 out of 10 participants bilaterally). Collecting data with high-resolution fMRI (including frontal areas), human laminar electrophysiology or ECoG with the same cascading oddball paradigm, or local-global paradigms (e.g32,33,37,61.), would allow testing our hypothesized contribution of local and global effects to the superficial error signals we identify, as well as verifying that deep layer modulations are related to a global model update by combining temporal and spatial information. In this study, we have opted for the use of a relatively simple, passive auditory paradigm. From the experimental point of view, future research could build on our results in several ways. First, while (short) events, separated in time, like the stimuli we have used here, are necessary for the modeling approach we have used, the understanding of predictive processing could benefit from the use of dynamic soundscapes with changes in entropy62,63. Future work may extend the modeling approach to these more dynamic stimulation paradigms. Second, while the pre-attentive nature of (local) deviance detection justifies the use of simple passive exposure, precision weighting exerted by cognitive processes such as attention is hypothesized to affect predictive processing (see13 for a review and references therein). Future studies could employ active paradigms in combination with laminar fMRI and modeling of the ensuing BOLD response to better understand how precision weighting affects the mesoscopic cortical processing that we have started to uncover here. To study predictive processing in the temporal cortex, we used GE-BOLD, which enables efficient, sensitive imaging of the bilateral auditory cortex43. However, GE-BOLD is influenced by macrovascular contributions, causing spatial displacement of neuronal activity38,64. Other acquisition and analysis approaches have been suggested to reduce draining vein contributions in GE-BOLD65. For example, alternative acquisition techniques, such as spin-echo EPI66, 3D gradient-echo and spin-echo (3D-GRASE67,68), and VASO69,70, reduce macrovascular effects but suffer from limited sensitivity and coverage for auditory cortex imaging43. On the contrary, previous studies that did collect GE-BOLD data, used simple subtraction methods to mitigate macrovascular effects71 or decoding strategies27,29. This assumption, however, has been shown to be inaccurate in certain applications72. We addressed this limitation using biophysical modeling to account for the ‘draining vein’ effect38. Apart from its initial validations39,45, this model has not previously been used in neuroscientific studies. Our current modeling approach reduces the influence of macrovascular draining veins in laminar fMRI data. As both middle and superficial layers are affected by vascular draining39,64, the modeling approach effectively reduced non-neuronal contributions at these depths, while leaving the deeper layers largely unaffected (see Supplementary Figs. S7 and S9 for the results obtained without the use of the model in the same plot as the modeling results). That is, when condition differences in deep layers are present before modeling, they are typically preserved after our modeling approach (see Supplementary Fig. S10 for simulations of null effects in deep layers). In contrast, our analysis shows a large effect of the modeling approach in middle and superficial layers, in line with the hypothesized effect of draining. These results indicate that, in our data, there is not sufficient evidence to rule the BOLD effects in middle and superficial depths in some of the ROIs as coming from true neuronal effects that are originating in those layers. The persistence of a superficial layer effect in PP and a middle layer effect in pSTG can, on the other hand, be interpreted as stronger evidence towards a possible neuronal effect in these layers, rather than a vascular draining artifact. In the absence of additional measures (e.g., cerebral blood volume) it is not possible for us to speculate on whether middle and superficial signals are overcorrected in this particular dataset. However, in previous publications of this model, the behavior of the model has been characterized extensively39, for example, cerebral blood volume38 and arterial spin labeling45 measurements have been used and confirmed the accuracy of the draining vein bias correction. While this first implementation offers a foundational framework, several expansions can be proposed for future studies. The modeling approach we have used, implemented within the DCM framework, requires a stable selection of activated voxels across analyses. This precluded reselecting voxels on independent splits of the data. In absence of an independent localizer, we therefore used an activation-based selection strategy. As our subsequent analysis is based on event-related averages rather than GLM regressors, the potential bias induced by this approach is minimized73. In future work, it would be relevant to evaluate the validity of Bayesian averaging when different selected voxels are selected across independent splits of the data. In addition, the neuronal model that is currently implemented encompasses three neuronal depths, each with an excitatory and an inhibitory neuronal population without any interlaminar connections. However, more complex models such as the canonical microcircuit12 integrate intra- and inter-areal connectivity as well as more neuronal layers. Furthermore, while we have currently interpreted our results in the standard hierarchical predictive coding framework, assigning error computation to superficial layers, and predictions and model updating to deep layers, alternative frameworks have been proposed such as dendritic error computation74,75 and the presence of error units in deep layers76. For example, an alternative cortical model for PC has recently been put forward that may allow a different reading of our results. In this model and in line with previous electrophysiological research19, prediction errors occur in both L2/3 and L5. According to this hypothesis, our results may indicate error signals in both the superficial and deep layers76. As stated above, the design we used in this study is not suited to disentangle local errors from global model updating, and future studies may investigate the role of deep layers in updating representations versus computing error using, for example, a local-global paradigm61. In addition, L5 pyramidal neurons with long apical dendrites have been shown to be able to integrate feedforward and feedback information streams arriving at different cortical layers75,77, a computation that may be of particular relevance for PC. Our results in superficial layers of PP thus could also reflect the integration of forward and backward information in the superficial compartment of these neurons (for the computation of errors). The deep layer modulations we observe could then be explained by either considering the role of feedback connections targeting deep layers, or may be explained by activity backpropagated from the superficial to the deep compartment. To test these alternative hypotheses at the neuronal level, future studies could compare the simple model we used here to alternative, more biologically grounded neuronal models such as the canonical microcircuit12, a model of dendritic error computation74 or others76,78. Moreover, as the DCM framework also allows the inference of effective connectivity, another extension of our modeling approach could entail the inclusion of inter-regional connections to test our hypotheses about the suggested routing of information (see Fig. 7B). However, it is imperative to acknowledge that enhancing model complexity comes with its own caveats and does not necessarily result in a better understanding of a system. For instance, the presence of interdependencies among parameters or an excessive number of parameters that can adequately explain a multitude of responses can compromise interpretability and obscure the underlying mechanics79. In addition, fMRI data has to be acquired with large coverage at laminar spatial resolution to include all brain areas needed to model inter- and intra-areal effective connectivity. In conclusion, the investigation of the violation of expectations across a large extent of the temporal cortex, with cortical depth-dependent fMRI, revealed error signals in superficial cortical layers, and the relevance of deep layers for broadcasting model updates. The large spatial coverage, in particular, allowed us to ground these results within the known hierarchy of temporal cortical areas - with superficial errors in belt regions feeding into middle layers in the parabelt. Our study provides potential valuable insights into the relevance of mesoscopic deviance processing in a small sample. Despite our focus on low-level acoustic processing, for which we do not expect to find large intersubject variability, this limitation of sample size should be taken into account when interpreting the results. In future research, it is important that the mechanism of laminar auditory predictive processing should be extended to larger populations or confirmed by other studies investigating similar mechanisms using different auditory paradigms. Notwithstanding, these findings advance our understanding of how contextual sensory information and expectations contribute to the processing of sounds and how this processing may be embedded within the human cortical (laminar) architecture. Methods The data presented in this study were previously used for the evaluation of a denoising technique and published by Faes and colleagues80. The raw data is available on 10.18112/openneuro.ds004928.v1.0.0. Participants Ten participants took part in this fMRI study (5 females, mean age = 37.6 years, SD = 14.2 years) and signed informed consent before partaking in this study. None of the participants had a history of neurological or hearing disorders. Data were collected at two different locations: eight participants were scanned at the Center for Magnetic Resonance Research in Minneapolis (CMRR) and two participants (S04 and S05) were scanned at New York University (using identical imaging protocol with the exception of a slight difference in TR = 1650). Each local Institutional Review Board approved the experiment. Experimental design and stimuli Participants completed 6 to 8 functional runs in the scanner while they passively listened to tone sequences. These sequences were designed to manipulate predictability based on contextual information. Each sequence consisted of four sounds (see Fig. 1) in which 100 ms tones were intersected with a 400 ms gap resulting in a total stimulus length of 1.6 s. The initial three contextual frequencies (493.9, 659.3, and 987.8 Hz) were consistently present across all conditions, though their presentation order varied depending on the specific condition (ascending, descending or scrambled). In the predictable conditions, the fourth tone adhered to the prediction generated by the three preceding sounds. On the contrary, in the ‘mispredicted’ conditions, we violated the local context by presenting a “deviant” frequency. The unpredictable conditions were designed to generate a weaker local prediction by scrambling the order of the first three frequencies (e.g., the first two tones are ascending and the last two tones were descending). With reduced predictions, we also expected the local error signal to be reduced. It’s important to note that the fourth frequency was always either 329.6 Hz or 1318.5 Hz in all conditions. The auditory stimuli were presented concomitantly with the scanner noise (i.e., no silent gap for sound presentation was used). The sequences were presented in a slow event-related design with an average inter-trial interval of 6 TRs (ranging between 5 and 7 TRs). Besides local predictability of the fourth tone, we also manipulated the global predictability by making the predictable sequences more probable (presented about 60% of the time) than the mispredicted and unpredictable sequences (each being presented about 20% of the time). Sounds were presented in the MRI scanner using MRI-compatible S14 model earbuds of the Sensimetrics Corporation (www.sens.com). During each run a fixation cross was presented in the middle of the screen. MRI acquisition At both sites, data were collected with a 7 T Siemens Magnetom System with a single-channel transmit and 32-channel receive NOVA head coil (Siemens Medical Systems, Erlangen). Whole-brain anatomical T1-weighted images were collected using a Magnetization Prepared 2 Rapid Acquisition Gradient Echo (MP2RAGE) sequence81 at a resolution of 0.75 mm isotropic (192 slices, TR = 4300 ms, TE = 2.27 ms). The functional data were acquired with 2D gradient-echo (GE) simultaneous multi-slice (SMS)/multiband (MB) echo planar imaging (EPI)42,82 (0.8 mm isotropic, 42 slices, TR = 1600 ms, TE = 26.4 ms, MB factor 2, iPAT factor 3, 6/8 Partial Fourier, bandwidth 1190 Hz, field of view: 170 × 170 mm, matrix size: 212 × 212, phase encoding = anterior to posterior). At NYU, all acquisition parameters were the same as at CMRR with the exception of a TR of 1650 ms. In each session (except for S01), we also obtained 5 functional volumes using the opposite phase encoding (posterior to anterior). Data Analysis Processing anatomical data Both the anatomical and functional data were analyzed using BrainVoyager (BV - v21.4 unless otherwise specified, Brain Innovation, Maastricht, The Netherlands). The anatomical data were upsampled to 0.4 mm isotropic, corrected for inhomogeneities, and rotated in such a way that the anterior commissure (AC) and posterior commissure (PC) were on the same plane (ACPC space). A high-resolution segmentation of the white matter (WM) and gray matter (GM) was created using a deep neural network implemented in BV (v22.0). The initial segmentation of the temporal lobes was manually corrected in ITK snap (83 v3.8). Mid-GM surface meshes were created in BV using this corrected segmentation. Pre-processing functional data Preprocessing of the functional data consisted of slice scan time correction (sinc interpolation) and 3D motion correction (rigid body with 6 parameters) using intrasession alignment to the run closest in time to the collection of opposite phase encoding (run 1 in most participants, run 4 in two participants). Moreover, we applied temporal filtering to remove low frequencies (high-pass filtering with 7 cycles per run) and high frequencies (temporal Gaussian smoothing with a full-width half-maximum kernel of 2 data points). In two participants, we slightly diverged from the general processing pipeline (for 2 or 3 runs) because after motion correction they displayed geometric distortions (across runs). In these participants (and for these runs), we corrected for these distortions using nonlinear alignment using Advanced Normalization Tools (ANTs), prior to correcting for the distortions induced by phase encoding. In all (but one) participants, anatomical distortions introduced by phase encoding were corrected using reversed phase polarity acquisitions and Topup (FSL, v6.0.4). Finally, the functional data were aligned to the upsampled anatomical data using boundary-based registration. Regions of interest The anatomical data were processed by defining five regions of interest (ROIs) on each individual’s cortical mesh surface of each hemisphere40. The temporal lobe was subdivided in the Heschl’s gyrus (HG), Planum Polare (PP), Planum Temporale (PT), anterior superior temporal gyrus (aSTG) and posterior superior temporal gyrus (pSTG). These ROIs were projected back in volume space extending 3 mm inwards and outwards from the mid-GM surface. To maintain an accurate delineation of the GM ribbon, the ROIs were once more intersected with the high-resolution segmentation to identify the GM portion of each ROI. These ROIs were used as the starting point for the modeling of the hemodynamic responses (see below). Modeling and statistical analysis The CSF-gray matter and gray matter-white matter boundaries as identified in volume space by the BrainVoyager segmentation were additionally inspected with custom MATLAB (The MATHWORKS Inc., Natick, MA, USA) code and ITK-snap83. An equivolume approach was used to estimate the normalized voxel-wise depth estimate in the gray matter. The depth map was then downsampled to match the voxel size of the functional data (0.8 mm isotropic). We analyzed the functional data by fitting a general linear model (GLM) in BV with the 6 conditions as predictors. For each ROI, the analysis was confined to the voxels with the strongest response to our auditory stimuli (F-statistic; p < 0.05 uncorrected) at the individual subject level. Note that any potential bias introduced by this voxel selection73 is minimized in our case because we do not select on the basis of a differential effect and our subsequent analyses use event-related averages rather than the GLM beta estimates. Supplementary Fig. 11 shows that, in pure noise simulations, voxel selection followed by beta-based analysis introduces a large bias (Supplementary Fig. 11A), while ERA-based analysis (Supplementary Fig. 11B) reduces this bias greatly. The procedure we followed for modeling the responses of each ROI, layer and participant followed a methodology previously introduced39,45. In particular, we restricted the further modeling analysis to ROIs that contained 300 or more active voxels to ensure a sufficient number of voxels (in each depth) for model fitting. This restriction resulted in the exclusion of 4 ROIs (PP LH in one participant and aSTG LH in 3 participants). For all subjects, event-related averages were computed per voxel by averaging all trials per condition. The event-related averages comprised 7 time points and were baseline corrected with respect to the onset of the auditory stimulation. We identified outlier voxels by comparing each voxel’s percent signal (in each ROI and for each condition) to a distribution obtained considering all participants’ voxels within all ROIs and across all trials. Outliers were identified as voxels whose percent signal change of a trial in a voxel was below the 0.01st or above the 99.99th percentile of the overall distribution. With this step, we excluded voxels with an extremely high or low percent signal change in certain trials, often located at the edge of the slab (influenced by motion). Importantly, the outlier voxels were then excluded for all conditions. After the selection of the voxels, we identified the contribution of each voxel to several depths (7, 9, 10 and 11 bins) in each ROI using normalized equally-spaced bell-shaped weighted probability maps (see Fig. 6B in39). Using this method, voxels are not assigned to only one depth, but voxels are weighted more towards a certain depth depending on the distance from equivolume sampled layers. For every cortical depth, the hemodynamic response (e.g., the event-related average) of two conditions was concatenated in order to use the generative model of laminar BOLD. This was done for the predictable condition followed by either the mispredicted or the unpredictable conditions, separately for the ascending and descending sequences. To ensure the model fit would come back to baseline, we separated the two hemodynamic responses by 20 time points of white noise, randomly sampled (between −0.1 and 0.1) from a normal distribution. That is, we added an artificial ‘rest’ period to avoid overlap of the responses to predictable and mispredicted/unpredictable conditions. These added data points did not affect the modeling results (see model inversion details below). To estimate laminar neuronal activity underlying the measured BOLD responses, we used the generative model of the laminar BOLD signal38. This model estimates BOLD responses using a combination of venous microvasculature and ascending veins in the DCM framework. We fixed the number of neuronal depths to 3. These were then linked independently to four larger numbers of vascular depths (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K\in$$\end{document}K∈ [7, 9, 10, 11]) using neurovascular coupling. The number of vascular depths was chosen based on previous simulations38 recommending more than six depths for the BOLD signal even if the neuronal depths are fixed to three. The model was applied per ROI and the different regions were not connected. We generated a model space comprising eight models for each of the vascular depths. These eight models were based on the hypotheses regarding external modulatory input(s) targeting specific neuronal depths. This included the null hypothesis that responses in both conditions were the same (i.e., no influence of modulatory input was captured by the model), to modulatory input targeting only one, any two, or all three neuronal depths (superficial, middle, and deep - see Fig. 4D for an illustration of the 8 models). At every layer, the model of Havlicek and Uludag38 offers two mechanisms to modulate the response at the neuronal level, changes in the self-excitation connection (σ) or changes in the inhibitory-excitatory connection (μ). The modulation of μ has been shown to capture the changes in the post-stimulus BOLD undershoot84. As our experimental design did not consider long inter-trial intervals, it didn’t capture the post-stimulus undershoot. Thus, in our analyses, we limited the modulatory input to target only the self-excitation connection. The parameter σ controls the self-excitation of excitatory neurons. A higher σ corresponds to stronger (and more sustained) neuronal activity and therefore changes the overall amplitude and shape of the generated BOLD response. In the modeling approach, σ is estimated for both conditions that are compared, indicative of neuronal responses to auditory stimulation. However, we look at the modulation of the σ parameter, which changes the amplitude and shape of the generated BOLD response by scaling excitatory-inhibitory and inhibitory-excitatory connections (see84) specifically in response to our experimental manipulation. The modulation is a free parameter and can be either positive or negative. Positive modulation values indicate increased neuronal (i.e., BOLD) responses, while negative modulation values indicate decreased neuronal activation (i.e., BOLD). Note that this implementation of the neuronal model has no interlaminar connections and thus the modulatory inputs were modeled independently between depths. Within this framework, the predictable condition was taken as the baseline, while the modulatory input was used to estimate the change in model parameters that would best explain the difference between the predictable condition and the second condition (mispredicted or unpredictable, modeled separately) received both driving and modulatory input. The driving input followed the structure of the input sound stimulus (4 tones, 1600 ms duration, 8 − 11.2 s inter-trial interval). The modulatory input was hypothesized to overlap (onset and duration) with driving input at the break of prediction i.e., at the fourth tone (at 1500 ms after stimulus onset) of the quartet for the mispredicted condition. For the unpredictable condition, this was modeled both at the moment at which the context was disrupted at the third tone (1000 ms after stimulus onset) and fourth tone (1500 s in ms stimulus onset) to capture the effect of contextual disruption on the processing of the fourth tone. Note that we only compare the predictable condition to the mispredicted and unpredictable conditions independently. Comparing the mispredicted and unpredictable conditions would necessitate distinguishing BOLD effects elicited by stimuli that are separated by 400 milliseconds, which is challenging with a TR of 1.6 s and it would also result in a much larger space. The neuronal activity at each depth was transformed into changes in cerebral blood flow (CBF), and subsequent cerebral blood volume (CBV). For each value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document}K, the simulated BOLD responses were convolved with a unique spatial blurring kernel (Gaussian point spread function, standard deviation, and scaling constant adjusted for the number of vascular depths, the curvature of the ROI, and the voxel size). We optimized neuronal and vascular parameters per vascular depth (see Table 1 for estimated parameters and prior distributions). The fixed parameters are reported in Table 2.Table 1Model parameters and priors (only the ones estimated during model inversion)DescriptionPrior meanPrior varianceNeuronal Model C (-)Strength of driving input to an excitatory population[1, 1, 1][1, 1, 1] σ (Hz)Excitatory self-connection3e−2 B (Hz)Modulatory effects on σ0e0.5Hemodynamic Model SAV (-)Slope of CBV0 increase towards surface (AV)0e1 αAV (-)Exponent relating CBF to CBV during steady-state (AV)0.35e−5Table 2Fixed model parameters (see ref. 39 for the remaining parameters)DescriptionValueNeuronal Model μ (Hz)Inhibitory Excitatory connection−0.8 λ (Hz)Inhibitory gain factor1.8Hemodynamic Model Τ (Tesla)Field strength7 TR (s)Repetition time1.6 TE (s)Echo time0.02 V0 (mL)Total baseline blood volume (MV + AV)3 ωMV (-)Fraction of total CBV0 represented (MV)0.5 nMV (-)Ratio between CBF and CMRO2 (MV)3 Table 1Model parameters and priors (only the ones estimated during model inversion)DescriptionPrior meanPrior varianceNeuronal Model C (-)Strength of driving input to an excitatory population[1, 1, 1][1, 1, 1] σ (Hz)Excitatory self-connection3e−2 B (Hz)Modulatory effects on σ0e0.5Hemodynamic Model SAV (-)Slope of CBV0 increase towards surface (AV)0e1 αAV (-)Exponent relating CBF to CBV during steady-state (AV)0.35e−5 Model parameters and priors (only the ones estimated during model inversion) Table 2Fixed model parameters (see ref. 39 for the remaining parameters)DescriptionValueNeuronal Model μ (Hz)Inhibitory Excitatory connection−0.8 λ (Hz)Inhibitory gain factor1.8Hemodynamic Model Τ (Tesla)Field strength7 TR (s)Repetition time1.6 TE (s)Echo time0.02 V0 (mL)Total baseline blood volume (MV + AV)3 ωMV (-)Fraction of total CBV0 represented (MV)0.5 nMV (-)Ratio between CBF and CMRO2 (MV)3 Fixed model parameters (see ref. 39 for the remaining parameters) The model inversion was performed using a slightly modified version of the standard Variational Bayes approach implemented in spm12 (“spm_nlsi_GN.m”). This modification included considering only the data time points and not the white noise inserted between the data from the two conditions while performing expectation-maximization. The estimated parameters were pooled across the different values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K$$\end{document}K (7, 9, 10 and 11) by performing Bayesian parameter averaging (“spm_bpa.m”) for each of the eight models. After Bayesian averaging, the eight models were compared using the Free Energy metric48. The winning model was used to obtain the estimate of modulatory input (σ) across the three neuronal depths for each subject and ROI - representing the modulation induced by the second condition (mispredicted or unpredictable sequence) with respect to the first (predictable sequence). To assess the significance of modulatory input, for each ROI and layer, at the group level (across subjects), we performed two-sided t-tests (against zero) and used a non-parametric permutation approach for significance testing49. Specifically, we generated all 210 possible sign permutations of the observed data, recalculated the t-statistics for each permutation, and determined p-values based on how often the observed t-value differed from this null distribution (note that this can result in identical p-values). Finally, we applied False Discovery Rate (FDR) correction using the Benjamini-Hochberg procedure across all ROIs and layers to account for multiple comparisons. This uses a stepwise procedure that can result in identical q-values. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data