The mismatch negativity (MMN) is an ERP component seen in response to unexpected “novel” stimuli, such as in an auditory oddball task. The MMN is of wide interest and application, but the neural responses that generate it are poorly understood. This is in part due to differences in design and focus between animal and human oddball paradigms. For example, one of the main explanatory models, the “predictive error hypothesis”, posits differences in timing and selectivity between signals carried in auditory and prefrontal cortex (PFC). However, these predictions have not been fully tested because (1) noninvasive techniques used in humans lack the combined spatial and temporal precision necessary for these comparisons and (2) single-neuron studies in animal models, which combine necessary spatial and temporal precision, have not focused on higher order contributions to novelty signals. In addition, accounts of the MMN traditionally do not address contributions from subcortical areas known to be involved in novelty detection, such as the amygdala. To better constrain hypotheses and to address methodological gaps between human and animal studies, we recorded single neuron activity from the auditory cortex, dorsolateral PFC, and basolateral amygdala of two macaque monkeys during an auditory oddball paradigm modeled after that used in humans. Consistent with predictions of the predictive error hypothesis, novelty signals in PFC were generally later than in auditory cortex and were abstracted from stimulus-specific effects seen in auditory cortex. However, we found signals in amygdala that were comparable in magnitude and timing to those in PFC, and both prefrontal and amygdala signals were generally much weaker than those in auditory cortex. These observations place useful quantitative constraints on putative generators of the auditory oddball-based MMN and additionally indicate that there are subcortical areas, such as the amygdala, that may be involved in novelty detection in an auditory oddball paradigm.
An organism's ability to extract patterns from the world and to quickly detect deviation from these predictions is key to survival (Friston, 2009). The auditory oddball task is one of the dominant paradigms used to examine mechanisms of novelty detection (Näätänen et al., 2012; Nelken, 2014). In this task, an auditory stimulus is presented repeatedly as a “standard” and is infrequently interleaved with a different “deviant” or “oddball” auditory stimulus. The response to the deviant is greater than the response to the standard or to the response of the deviant stimulus when played as a standard. This early difference in response is referred to as the “mismatch negativity” (MMN) in scalp-based ERPs. Though the neural basis of this potential is poorly understood, the MMN is of wide use and interest as an index of early sensory processing and novelty sensitivity and is of clinical interest as a potential index of symptoms of psychiatric disorders.
To examine the basis of the auditory oddball mediated MMN, human studies using noninvasive measures (fMRI, EEG, MEG) have suggested that it has both temporal and frontal cortical generators. According to the predictive error hypothesis of the MMN, sensory and frontal generators have distinct functions (Garrido, Kilner, Stephan, & Friston, 2009; Näätänen, Teder, Alho, & Lavikainen, 1992). While the standard is being repeated, a stimulus-specific memory is built up in sensory (auditory) cortex, and this memory drives an expectation. On the other hand, frontal cortex represents whether there is a difference between the expected and actual stimulus (Garrido et al., 2009; Näätänen et al., 1992; Giard, Perrin, Pernier, & Bouchet, 1990). This hypothesis predicts that frontal novelty signals should arise later than those in auditory cortex (AC) and additionally that the signal in prefrontal cortex (PFC) be abstracted from any stimulus-selective activity seen in AC. It has been difficult to test this hypothesis in humans using noninvasive techniques because they do not combine the precise temporal and spatial resolution needed to compare time courses with certainty (Tse & Penney, 2008; Deouell, 2007; Rinne, Alho, Ilmoniemi, Virtanen, & Näätänen, 2000). In addition, to compare stimulus specificity, one must be able to compare the novelty response between stimulus types across areas, which is not commonly done. Human electrocorticography studies have also been used to examine the correlates of oddball and novelty detection and have confirmed frontal and sensory components. However, heterogeneity of clinical electrode placement within and between patients makes it difficult to systematically compare magnitude and time course of early novelty signals between areas (Durschmid, Edwards, et al., 2016; Edwards, Soltani, Deouell, Berger, & Knight, 2005; Rosburg et al., 2005). In contrast, animal models of auditory oddball paradigms do have the necessary spatial and temporal specificity to compare signals across areas. However, these studies have primarily been concerned with mechanisms of oddball in early auditory processing (cochlear nucleus through AC) and not activity from other areas (Parras et al., 2017; Yarden & Nelken, 2017; Fishman, 2014; Ayala, Perez-Gonzalez, Duque, Nelken, & Malmierca, 2012; Ulanovsky, Las, & Nelken, 2003; Javitt, Steinschneider, Schroeder, Vaughan, & Arezzo, 1994). In addition, the majority of animal studies choose neuron-specific “standard” and “deviant” stimuli based on the receptive fields of the neuron being recorded at the time. This tuned design has been undeniably powerful for other questions (e.g., mechanisms of adaptation and effects of pharmacological manipulations at a single-neuron level) but may not be ideal for connecting to the larger body of human literature where, by definition, identical standard and deviant stimuli are used for all areas. This distinction appears especially important when questions pertain to the comparison of time course and magnitude of novelty signal between areas.
Largely separate from accounts involving the cortical substrates of the MMN in oddball, a body of work has established correlates of novelty and salience detection in subcortical nuclei such as the amygdala (AMY) and nucleus accumbens (NAc; Bradley et al., 2015; Balderston, Schultz, & Helmstetter, 2013; Zaehle et al., 2013; Blackford, Buckholtz, Avery, & Zald, 2010). For example, a human intracranial study suggested robust auditory oddball-based MMN in NAc (Durschmid, Zaehle, et al., 2016). AMY activation in auditory oddball has been reported in at least one fMRI study (Czisch et al., 2009), but an intracranial study suggested that auditory oddball novelty signal may only arise in the AMY during active detection (Kropotov et al., 2000). There has thus been renewed interest in nuclei that can affect, either directly or indirectly, the MMN, with the understanding that source localization techniques used in human noninvasive studies can miss subcortical generators. However, it is unclear whether these areas' signals are as fast or as robust as those of the hypothesized cortical generators. The human intracranial studies described above are suggestive but cannot systematically compare the signals in AMY to that of other areas due to the necessary heterogeneity of clinical placement. Studies in animal models would be able to address magnitude and timing differences between these and more established generators, but extant studies have largely focused on examining oddball responses in early auditory areas. The goal of the present experiment, therefore, was to systematically compare the single neuron correlates of an early auditory oddball signal in AC, dorsolateral PFC, and basolateral AMY. This approach, drawing on a substantial number of neurons (∼600 per area afforded by a multichannel approach), allows us to address predictions of the predictive error account in AC and PFC and evaluate whether the account should be extended to areas such as the AMY. This comparison, with the advantage that it is modeled after human studies, provides useful quantitative constraints for future accounts of the brain bases of auditory oddball-mediated MMN.
Subjects, Surgical Procedures, and Neurophysiological Data Acquisition
Experiments were carried out on two adult male rhesus macaques (Macaca mulatta). All experimental procedures were approved by the National Institute of Mental Health Animal Care and Use Committee and followed the Guide for the Care and Use of Laboratory Animals. Before data acquisition, monkeys were implanted with titanium head posts for head restraint. In a separate procedure, monkeys were fitted with custom acrylic chambers oriented to allow vertical grid access to the left dorsolateral PFC (dorsal bank of the principal sulcus extending ventral, >1 mm away from arcuate sulcus, roughly 46/8Ad), the lateral portion of the AMY (entire dorsoventral extent, primarily basolateral AMY), and AC (primarily A1 but including small portions of lateral belt areas). Recording areas were verified though a T1 scan of grid coverage with respect to underlying anatomical landmarks (Figure 1B–C), combined with maps of frequency reversals and response latencies of single neurons to determine A1 location and extent (Camalier, D'Angelo, Sterbing-D'Angelo, de la Mothe, & Hackett, 2012). Recordings were made using either 16- or 24-channel laminar “V-trodes” (Plexon, Inc., Dallas, TX; 200–300 μm contact spacing, respectively), which allowed identification of white matter tracts, further allowing identification of electrode location with respect to sulci and gyri. Electrodes were advanced to their target location (NAN Microdrives, Nazareth, Israel) and allowed to settle for at least 1 hr before recording.
Multichannel spike and local field potential recordings were acquired with a 64-channel data acquisition system. Spike signals were amplified, filtered (0.3–8 kHz), and digitized at ∼24.4 kHz. Spikes were initially sorted online on all channels using real-time window discrimination. Digitized waveforms (snippets) and timestamps of stimulus events were saved for final sorting (Plexon offline sorter V3.3.5). The units were also graded according to isolation quality (single or multiple neurons). Single and multiunits were analyzed separately, but the patterns of results were similar and so were combined. The acquisition software interfaced directly with the stimulus delivery system, and both systems were controlled by custom software (OpenWorkbench and OpenDeveloper, controlling a RZ2, RX8, Tucker Davis Technologies System 3, Alachua, FL).
Stimuli, Experimental Design, and Statistical Analysis
We utilized a “flip-flop” auditory oddball design used in intracranial studies (e.g., Ulanovsky et al., 2003; Näätänen, Gaillard, & Mantysalo, 1978). This paradigm controls for differences in activity driven by the stimulus selectivity of neurons unrelated to stimulus type (type: standard vs. deviant; Figure 1A). Because our questions were primarily concerned with comparisons of neural responses in AC and higher order areas, the stimuli used were constant across all areas and neurons and chosen to drive responses in higher order areas such as the AMY and PFC and thus were more complex than pure tones. The stimuli, 300-msec 1- and 8-kHz square waves, are perceptually distinct, spectrally nonoverlapping wideband stimuli likely to activate large parts of the AC. The spectrum of a square wave contains the odd harmonics of the fundamental frequency, such that the 1-kHz square wave contains 1 kHz and harmonics of 3, 5, 7, 9, up to ∼33 kHz, and the 8-kHz square wave contains 8 kHz and a functional harmonic of 24 kHz, as macaques can hear up to about 35 kHz (Hauser, Burton, Mercer, & Ramachandran, 2018; Recanzone, Guard, & Phan, 2000; Jackson, Heffner, & Heffner, 1999). These stimuli have minimal spectral overlap, which is thought to be important for evoking MMN correlates (see Khouri & Nelken, 2015). The sounds were 300 msec in duration (with 5 msec cos2 onset/offset ramps to minimize spectral splatter), and they were presented from a speaker 10 cm from the contralateral (right) ear calibrated to 60 dB SPL.
The “flip-flop” design has two blocks, one in which the 1-kHz pulse is the standard and the 8-kHz is the deviant, and the other in which standard and deviant identities are reversed. The flip-flop design critically allows the ability to dissociate any stimulus specificity of the response from actual novelty; the measure of most interest is the difference between the response to a stimulus when it is presented as a standard versus to the identical stimulus when it is a presented as a deviant (Dev–Std). If this comparison is positive, then the response to a stimulus presented as a deviant is greater than when it is presented as a standard. Each block in the flip-flop design contained 270 standards and 30 deviants (10% deviant probability) presented in randomized order with an interonset interval between 760 and 860 msec. Ten additional standards were added to the beginning of each block to ensure a stable standard trace (Figure 1A). The deviant stimulus identity of the first block (1 or 8 kHz) was switched across days to eliminate any order effects. During the oddball task, head-restrained monkeys sat quietly in a primate chair, watching a soundless video (common in human oddball studies) in a double-walled acoustically isolated sound booth (Industrial Acoustics Company). Oddball data were from 21 sessions in Monkey 1 and 55 sessions in Monkey 2, and each monkey was presented with a single session a day.
All data analysis was performed with custom scripts in MATLAB. Code and data are available upon reasonable request to the corresponding author. Data from the first 10 standards was discarded, and then spikes evoked by each stimulus (1 vs. 8 kHz) and type (standard vs. deviant) combination were calculated over time. For each stimulus, 270 presentations of it as standard and 30 as a deviant were included in analyses (600 presentations in total). For all analyses, mean evoked spike rates were converted to z scores using the baseline mean and standard deviation (−150 to 0 msec prestimulus).
To establish the population of auditory-responsive cells (here defined as responding to any of the stimuli in any context), we performed a 100-msec sliding window ANOVA on the evoked activity of each neuron (time bin advanced from 0 to 250 msec in 20-msec increments), including the factors of Stimulus frequency (1 vs. 8 kHz) and Type (deviant vs. standard) and their interaction. If any factor was significant in any window (p < .01, FDR-corrected for multiple comparisons), the neuron was considered task responsive. For the population-based analysis of oddball selectivity by area and stimulus (Figures 3–5), we used an analysis of a mixed within-between effects ANOVA (3 × 2 × 2; Area × Stimulus × Type). In this case, neuron identity was specified as a “random” effect in the model and was nested under area, where area is the region recorded from. To ensure these effects were not driven by results from an individual monkey, we reran the tests with monkey identity as a factor in the model, and effects did not change (see Results section). To establish oddball selectivity at the level of individual neurons, if a neuron exhibited a significant response to type, and response to deviant > standard, or a significant interaction and one of the stimuli had a deviant response > standard response (p < .01, Bonferroni-corrected as appropriate), then it was considered to exhibit deviance selectivity analogous to the MMN. To be conservative, we refer to this firing rate-based deviance selectivity as simply “oddball selectivity.” It has the same pattern as the MMN measured by continuous methods such as EEG but is not derived from a continuous measure. Note that this measure is similar to “stimulus-specific adaptation” used in other reports, but as the basis of this is probably not strictly adaptation, especially in higher order areas examined here, we prefer to use descriptive nomenclature to avoid confusion.
For the individual neuron-based analysis of time course of deviance detection (Figure 7D), we measured the first time bin that exhibited a selective response to stimulus type for each responsive neuron in each area (Type × Stimulus, 30-msec windows sliding at 10 msec (p < .05, Bonferroni-corrected). For the population-based analysis of the time course of novelty detection (Figure 7A–C), we used a sliding ANOVA (mixed within-between, Stimulus × Type × Neuron Identity) on response in 30-msec bins with 10-msec intervals and looked for periods with a significant effect of Type (p < .01, Bonferroni-corrected). To examine whether adaptation exists and is different between areas (Figure 8), we calculated the evoked response as a function of time (either presentation number since the beginning of each block or presentation number following deviant) for each responsive neuron. To quantify this time course, we fit an exponential decay function (R = a × exp(b × t), where t = presentation number; after Antunes, Nelken, Covey, & Malmierca, 2010; Ulanovsky, Las, Farkas, & Nelken, 2004) to each neuron's time course for each stimulus and then determined whether the population of decay constants (b) were less than 0 via a one-tailed t test. If adaptation was indicated, ANOVAs on the population of b values was done to determine whether adaptation was greater for some stimuli. For the analysis of within-block adaptation, this was done for the first 27 presentations. This covers the majority of the adaptation and includes the same number of stimuli for standards as deviants. For the analyses relative to the last deviant, adaptation was calculated for the first six presentations of standards following a deviant. Consistent with a population-based approach done by other studies (see above), these results were confirmed by a bootstrap test on 1000 fits done to averages of 70 neurons (chosen randomly from population), and the results were identical.
We recorded neurons from primary AC (n = 690), dorsolateral PFC (n = 598), and basolateral AMY (n = 627) while monkeys were presented with a flip-flop auditory oddball paradigm (Figure 1A). In all three areas, we found neurons that responded more strongly to the stimulus when presented as an oddball (responses from exemplar individual neurons in Figure 2). Note that generally both baseline firing rates and responses were weaker in AMY and PFC, so subsequent analyses focus on the normalized differences in firing rates expressed as a z score. To compare novelty responsiveness and stimulus selectivity across areas, we first identified the fraction of neurons that were generally auditory responsive under these oddball task conditions. In AC, 462 neurons were responsive to the task stimuli in any context (67%). Smaller proportions were responsive in PFC (105, 18%) and AMY (105, 17%). This analysis was done using a fixed window (0–250 msec) chosen to correspond to the period of MMN generation in the primate (Gil-da-Costa, Stoner, Fung, & Albright, 2013). To ensure that this was not due to the choice of analysis window, we examined larger or smaller sliding windows (50, 200 msec), and the pattern of results did not change, though the fixed window yielded more conservative estimates of responsiveness.
The MMN measured at the scalp surface presumably represents activity of a neural population. Although there is no direct mapping of neural activity to ERP components, a straightforward approach is to examine the magnitude of population activity of task-responsive neurons, whether that population activity carries a novelty signal, and whether this novelty signal magnitude is different across areas or different with stimulus. For visualization purposes and for a more direct comparison with human studies (which generally do not use a flip-flop design to account for differences in neurons' stimulus responsiveness), we first collapse across all stimulus types to show time course of and magnitude of average oddball response across areas (all deviants–all standards: Figure 3). Note that the overall deviance response was approximately twice as strong in AC than in PFC or AMY (significant effect of Area, F(2, 676) = 15.4; p < .001, between Area post hoc tests AC vs. PFC: F(1, 573) = 14.8, p =.001; AC vs. AMY: F(1, 573) = 16.8, p < .001; PFC vs. AMY area: ns, p = .56).
To better understand what is driving this pattern, we first examined the magnitude of evoked responses and saw that responses were approximately twice as strong in AC than in PFC and AMY (significant effect of Area, F(2, 676) = 7.2, p < .001, between Area post hoc tests AC vs. PFC: F(1, 572) = 7.4, p = .006; AC vs. AMY: F(1, 676) = 7.0, p = .009; PFC vs. AMY area: ns, p = .69). When evoked responses were separated by stimulus and type (Figure 4), note that AC responded more strongly to 1 kHz than to 8 kHz. We next examined the magnitude of novelty signal by stimulus for each area, which, as above, can be measured as Dev–Std. We saw that the novelty signal was dependent on the stimulus in AC—it was stronger for the 1 kHz than for the 8 kHz stimulus (Figure 5). However, this stimulus selectivity was not present in PFC and AMY—novelty signals were the same magnitude irrespective of driving stimulus (overall effect of type, F(1, 676) = 58.2, p < .001; AC stimulus, F(1, 468) = 66.1, p < .001; type, F(1, 468) = 190.8, p < .001; and Stimulus × Type, F(1, 468) = 51.3, p < .001; PFC type, F(1, 104) = 41.0, p < .001; AMY type, F(1, 104) = 25.5, p < .001; all other factors in PFC and AMY, p > .1). These results persist when monkey was included as a factor, and there were no significant interactions with monkey factor (e.g., Stimulus × Type × Monkey; stimulus: overall effect of Type, F(1, 2027) = 11.3, p < .001; AC stimulus, F(1, 1401) = 34.5, p < .001; type, F(1, 1401) = 8.1, p = .004; and Stimulus × Type, F(1, 1401) = 8.8, p = .003; PFC type, F(1, 309) = 9.13, p = .002; AMY type, F(1, 309) = 12.6, p < .001; all other factors in PFC and AMY and critical interactions with monkey, p > .05). Thus, at a population level, the average novelty signal in AC was robust and stimulus dependent, and in AMY and PFC, it was smaller (though significant) and not stimulus dependent.
For a more complete picture of what is driving this population-level signal, we next asked whether the novelty signal is driven by a few or many neurons in each area (e.g., a weak signal in PFC could be a few neurons carrying a signal comparable in magnitude to that of AC or many neurons carrying a weak signal). In AC, a substantial number of neurons show a novelty signal for at least one of the stimuli (Figure 6; 166/462 = 35.9%). However, few single neurons in PFC and AMY showed a similar profile (PFC: 9/105 = 8.6%, AMY: 9/105 = 8.6%), and the magnitude of the signal is again smaller relative to that in AC. Thus, at the level of single neurons, few neurons are carrying a strong signal in the AMY and PFC—it is a weak signal carried by many that leads to the overall novelty population response. Note that, again, these results were not dependent on window choice—they did not change substantially with smaller (0–100 msec) or larger (0–400 msec) fixed or sliding windows.
For a given neuron, there is no predictive relationship between stimulus selectivity and deviance selectivity in these areas. We first examined whether the presence of deviance selectivity (ANOVA significant for Stimulus type; Std vs. Dev) is a function of stimulus selectivity (significant for Stimulus frequency; 1 vs. 8 kHz). In AC, stimulus frequency affects magnitude of deviance response, but does not strongly affect whether deviance is signaled; of the 166 units that show a deviance selectivity, only 21 (12%) show it for only one of the stimuli, split fairly evenly between the two stimuli (13 units for 1 kHz only, 8 units for 8 kHz only). In PFC and AMY, all neurons that show deviance selectivity show it for both stimuli (1 and 8 kHz). Second, we examined the reverse—whether the presence of stimulus selectivity affected deviance selectivity. Again, a neuron exhibiting stimulus selectivity does not correspond to a neuron exhibiting deviance selectivity—in AC, 248 neurons were stimulus selective of which only about half (87, or 54%) were also deviance selective. Additionally, 58 neurons were deviance selective but not stimulus selective, and 21 neurons were deviance selective for only one stimulus (described above). In PFC and AMY, the number of individually responsive neurons is quite low, and in PFC, eight neurons were stimulus selective of which only two of these (25%) were also deviance selective, seven additional neurons were deviance selective but not stimulus selective, and no neurons had deviance selectivity to only one stimulus. In AMY, it was very similar—eight neurons were stimulus selective, of which only one of these (12%) was also deviance selective, eight additional neurons were deviance selective but not stimulus selective, and again no neurons had deviance selectivity to only one stimulus. Thus, although stimulus selectivity affects the magnitude of the single unit response in AC, it is not predictive of whether a neuron exhibits deviance selectivity in AC, AMY, or PFC.
We next wanted to examine a prediction of the predictive error hypothesis of the MMN, whether the novelty signal emerges first from sensory cortex. To do this, we compare the time course of the novelty signal between AC, PFC, and AMY, again using two approaches, one based on population level activity and the other based on single neurons, to more fully understand what is driving the population activity. As an estimate of how novelty signal evolves across the population in each area, we calculated the population average peristimulation time histograms (PSTHs) for each Stimulus × Type combination over all the responsive neurons and examined when response to deviants differed from response to standards at a population level (Figure 7A–C). In AC, the earliest difference was detected in the bin spanning 10–40 msec. In AMY, it was later, 30–60 msec. In PFC, no 30-msec bin reached significance. The lack of an onset time in this analysis in PFC was probably due to the diffuse nature of the deviance signal in PFC and was not due to the size of the analysis bin. We repeated the analysis with shorter and longer bins, and PFC did not reach significance in any bin from 10 to 150 msec.
Although this population approach is informative about when the signal generally is arising, it may obscure heterogeneous but important dynamics (e.g., a smaller population of prefrontal neurons respond before AC but the majority are later). To examine whether there are any fast dynamics that are obscured by averages, we analyzed latencies on a single-neuron basis. Here, we calculated the latency at which the deviance signal (e.g., selectivity to type) arose in each area on an individual neuron basis, using 30-msec bins to increase temporal resolution (Figure 7D, as cumulative distribution functions). The 30-msec bins increases the overall number of neurons that carry a novelty signal in each area (relative to the fixed bins in Figure 6), but importantly, as stated above, the overall trends between areas are the same. According to this single-neuron analysis, sensitivity to type was still earlier in AC than AMY and PFC, and AMY and PFC did not differ (AC = 56.6 msec, n = 332; PFC = 128.9 msec, n = 36; AMY = 137.2 msec, n = 46; ANOVA significant for Area, F(2, 411) = 57.4, p < .001; post hoc t tests AC < PFC, t(366) = −7.6, p < .001; AC < AMY, t(376) = −9.1, p < .001; PFC = AMY, t(80) = 0.45, p = .64). Earliest novelty sensitivity latencies in AC were still well earlier than those in PFC and AMY. Taken together, these analyses suggest that a deviance signal occurs first in AC and later emerges in PFC and AMY and that the population signal is a veridical reflection of dynamics occurring at the single neuron level.
As a final way to examine differences in how the novelty signal evolves across these areas, we compared the dynamics of response adaptation between areas. Neural responses may adapt at the beginning of the block and also after a deviant is presented, and the existence and time course of adaptation may change with level of processing and brain region. For example, responses that are strongly sensory driven may show strong adaptation effects, but one would not expect adaptation in higher order responses that presumably reflect more generalized novelty processing. To address this, we analyzed adaption of the responses to standards and deviants in two ways. First, we examined standard and deviant adaptation within blocks (Figure 8A). Consistent within-block adaptation occurred in AC for three of the stimuli (mean and t test on fit beta values: 1 kHz standard mean = −0.036, t(461) = −10.9, p < .001; 8 kHz standard mean = −0.016, t(461) = −4.7, p < .001; 1 kHz deviant mean = −0.03, t(461) = −4.2, p < .016; 8 kHz deviant mean = −0.006; t test, ns, p = .056). In AC, adaptation was faster for the standards and faster for the 1-kHz stimulus (2 × 2 Stimulus × Type within-neuron ANOVA stimulus, F(1, 900.2) = 16.4, p < .001; type, F(1, 472.5) = 17.2, p < .001; Stimulus × Type, ns, p > .1). Adaptation was not observed in PFC, and it was only observed in AMY for the 8-kHz deviant (8 kHz deviant mean = −0.02, t(104) = −3.6, p < .001; all other, p > .1). Note that this within-block adaptation did not contribute to the novelty signal described above, as the first deviant was not presented until at least 10 standards had been presented (dotted line), after the within-block adaptation had occurred. A second type of adaptation is the effect that the deviant has on subsequent standards (Figure 8B). We found that postdeviant adaptation occurred in AC for the 1-kHz stimulus (mean = −0.02, t(461) = −3.3, p < .001), but not for other stimuli or areas (all other t tests, p > .1). Note that the lower responses in AMY and PFC, as well as AC (8 kHz), may make adaptation difficult to detect. Generally, however, it appears that stimulus-specific characteristics of adaptation seen in AC were not present in PFC or AMY, consistent with the more generalized novelty signal these higher order areas appear to carry.
The overall goal of this experiment was to leverage the spatial and temporal specificity of a primate model to address outstanding questions about novelty processing. Specifically, we compared the magnitude, timing, and stimulus specificity of auditory oddball-based novelty signals in AC, dorsolateral PFC, and basolateral AMY. These data were recorded under identical conditions from a substantial number of neurons in the alert macaque, so even subtle effects could be captured and compared. This approach allows us to test predictions made by the predictive error/sensory memory model of the MMN, which posits that a frontal generator exists that is sensitive to a change in the expected sensory stream, and that this signal should be both abstracted from and later than the one in AC (Garrido et al., 2009; Näätänen et al., 1992; Giard et al., 1990). We find that, compared with the novelty signal seen in AC, the signal seen in dorsolateral PFC and basolateral AMY is smaller in magnitude, longer in latency, and not stimulus dependent. This is generally consistent with predictions of the predictive error hypothesis in that novelty signals in PFC should be generally later than in AC, as well as abstracted from stimulus-specific effects seen in AC. However, the fact that signals in AMY were comparable in magnitude and timing to those in PFC, and both prefrontal and AMY signals were generally much weaker than those in AC, is not wholly consistent with the account.
The approach we present here allows more precise cross-area comparisons than previous noninvasive or electrocorticography approaches. We see that novelty response in PFC and AMY has properties that are strikingly similar to each other, yet do not appear to be simply inherited from the signal seen in AC. First, the stimulus specificity seen in AC was no longer significant in PFC or AMY. Second, the within-block adaptation and postdeviance adaptation seen in AC were also no longer observed. This suggests that PFC and AMY have, to some degree, an abstracted deviance signal distinct from AC. These areas signal “difference,” but they do not reflect low-level properties of the driving stimuli (e.g., spectral overlap). The timing of the difference signal is also consistent with that interpretation. The prefrontal difference signal lags that of AC by an average of ∼100 msec, and the AMY difference signal lags AC by between 20 and 100 msec. Note that a single neuron deviance signal, which emerges in AC starting in a time window of 10–40 msec is quite consistent with estimates that primate MMN (and human) emerges around 50 msec (Gil-da-Costa et al., 2013) as well as early response latencies generally in AC (Camalier et al., 2012). These observations place useful quantitative constraints on putative generators of the auditory oddball-based MMN and indicate that subcortical areas, such as the AMY, may need to be included in future explanatory accounts of auditory oddball.
Previously, there have been only indirect measures suggesting that frontal MMN generators are later than sensory generators (Tse & Penney, 2008; Deouell, 2007; Rinne et al., 2000). Human intracranial studies of auditory oddball paradigms have supported the existence of separate frontal and temporal components (Durschmid, Edwards, et al., 2016; El Karoui et al., 2015; Edwards et al., 2005; Rosburg et al., 2005; Liasis, Towell, Alho, & Boyd, 2001; Kropotov et al., 2000). However, because of the heterogeneity of the clinical placement of monitoring electrodes within and between patients, it has been difficult to compare the magnitude and timing of the deviance signal between these areas. In addition, our study extends our understanding of putative substrates of the oddball-based novelty processing to the AMY. This builds on a hypothesized role of the AMY in salience detection (Blackford et al., 2010). Here, we show that AMY correlates of novelty are comparable in magnitude, timing, and response characteristics to those seen in PFC. Previous accounts may have been biased toward cortical generators because noninvasive scalp-based measures are less sensitive to deep generators. This new finding may require an update of the accounts of the generators underlying the MMN.
Experiments in both rodents (Parras et al., 2017; Yarden & Nelken, 2017; Taaseh, Yaron, & Nelken, 2011; Ulanovsky et al., 2003) and primates (Fishman, 2014; Fishman & Steinschneider, 2012; Javitt et al., 1994) have examined single-neuron correlates of the MMN in the early auditory hierarchy (cochlear nucleus through AC). The novelty signal in AC that we observed in this study had a magnitude, time course, and stimulus-dependent characteristics consistent with what has been seen in population-based analyses of single-neuron studies in AC (Parras et al., 2017; Nieto-Diego & Malmierca, 2016; Farley, Quirk, Doherty, & Christian, 2010; Ulanovsky et al., 2003, 2004). There are two methodological differences worth noting between our study and these studies, driven by the aim of our study, which was to compare the characteristics and timing of two putative higher order generators of the deviance signal to the deviance signal seen in AC. First, the stimuli used in this study were wideband and kept constant for all of the data collected, instead of the previous studies' method of using pure tones whose identity is chosen to carefully flank the individual tuning of the neurons studied (and “population” responses collapsed across stimuli later). Thus, we would expect the population responses in our study to be more comparable to human studies in which all the data are collected using the same stimuli for all of the areas. Second, the majority of the single-neuron studies were performed under anesthesia (but see Parras et al., 2017; Farley et al., 2010). Although deviance detection, especially the MMN, can be elicited under anesthesia, the anesthetic state may affect higher order areas and minimize top–down effects seen from, for example, PFC, which does not exhibit robust responses under deep anesthesia. Such top–down effects have also been seen on the MMN during some kinds of attention manipulations, though the mechanisms of this effect are still hotly debated (Rinne, Särkkä, Degerman, Schröger, & Alho, 2006).
Utilizing the spatial and temporal specificity of a macaque model, this study overcomes limitations in our understanding from earlier findings derived from noninvasive methods and human intraoperative recordings, though there are multiple aspects of these findings that merit further investigation. The MMN-indexed deviance signal is commonly thought to contain an adaptation component (suppression of the expected) and/or a surprise component (enhancement of the unexpected; see Fishman, 2014). Note that detection of deviance is of substantial evolutionary importance, and therefore, it is likely supported by a broad network of areas whose magnitude, network recruitment, and mechanism (e.g., adaptation vs. surprise) may depend on task type, attentional state, behavioral requirements, and even anesthetic depth (Nourski et al., 2018; MacLean, Blundon, & Ward, 2015; Warbrick, Reske, & Shah, 2013). Even in the “simple” passive auditory oddball paradigm used here and in other studies, the relative contributions of each of these mechanisms (adaptation vs. surprise) to the recorded signal has been a matter of intense debate, especially in the early auditory hierarchy (Parras et al., 2017; Fishman, 2014).
Although the “error prediction” model posits that the frontal activation is an important generator of the MMN, the “adaptation only” hypothesis (e.g., Fishman, 2014) would posit that it is epiphenomenal. Our study finds that the signals in AMY and PFC have characteristics described by the error prediction hypothesis, but that the strength in both areas is weaker than that of AC. Therefore, an important future direction will be to establish causal links between PFC and AMY and the deviance signal seen in AC and at the scalp to be able to disambiguate between these accounts, especially using a combination of single unit and local field potential data compared with the extracranial MMN, which has been demonstrated (Gil-da-Costa et al., 2013). An earlier study reports that patients with prefrontal lesions exhibit a smaller scalp-measured MMN (Alain, Woods, & Knight, 1998; Alho, Woods, Algazi, Knight, & Näätänen, 1994). Though this would seem to imply a causal relationship of the PFC to the scalp-recorded MMN, it is difficult to compare the magnitude of ERP components when there is substantial difference in damage to the underlying cortex. Because of its shared primate homology and established MMN (Gil-da-Costa et al., 2013), the macaque model is an ideal model to test these questions using reversible inactivation. In addition, other areas that are hypothesized to play a role should be explored, such as NAc and ACC. Toward that end, these data provide a significant advance in our conceptual framework of how deviance is processed outside the early subcortical to cortical auditory hierarchy and open new directions of investigation in one of the dominant deviance detection paradigms, the auditory oddball paradigm.
The authors would like to thank Dr. Israel Nelken and Dr. Brian Scott for valuable feedback, Dr. Richard Saunders for surgical assistance, Anna Leigh Brown and Jess Jacobs for assistance in data collection, and the National Institutes of Health Section on Instrumentation for assistance in custom manufacture of recording chambers and grid. Research was supported by National Institute of Mental Health DIRP ZIA MH002928-01 to B. A. and ZIA MH001101-25 to M. M.
Reprint requests should be sent to Corrie R. Camalier, Department of Biomedical Engineering, Duke University, 1427 CIEMAS, Box 90281, 101 Science Drive, Durham, NC 27708, or via e-mail: firstname.lastname@example.org.