Neural circuits underlying auditory contrast gain control and their perceptual implications — TEI-tagged view

LohsepersName, BajopersName, KingpersName, WillmorepersName (2020date). “Neural circuits underlying auditory contrast gain control and their perceptual implications.” Nature Communicationstitle·j.

Object of study — whether auditory contrast gain controlterm·var is a cortical computation, as the literature had it, or is already present and self-sufficient in the midbrainterm and thalamusterm

Research question — whether the contrast gain controlterm·var seen in subcorticalterm neurons is inherited from the cortexterm through descending projections, or computed independently there

Contribution — contrast gain controlterm·var is equally strong in midbrainterm, thalamusterm and cortexterm, survives cortical silencingterm·mth intact, slows at each ascending level, and is enough to account for humanterm perceptualterm adaptationterm·dom

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Neural circuits underlying auditory contrast gain controlterm·var and their perceptualterm implications Michael LohsepersName Michael LohsepersName 1 Department of Physiology, Anatomy, and Genetics, University of Oxford, Oxford, OX1 3PT UK 1, ✉ , Victoria M BajopersName Victoria M BajopersName 1 Department of Physiology, Anatomy, and Genetics, University of Oxford, Oxford, OX1 3PT UK 1 , Andrew J KingpersName Andrew J KingpersName 1 Department of Physiology, Anatomy, and Genetics, University of Oxford, Oxford, OX1 3PT UK 1, ✉, # , Ben D B WillmorepersName Ben D B WillmorepersName 1 Department of Physiology, Anatomy, and Genetics, University of Oxford, Oxford, OX1 3PT UK 1, # 1 Department of Physiology, Anatomy, and Genetics, University of Oxford, Oxford, OX1 3PT UK ✉ Corresponding author.

# Contributed equally.

Received 2019 Jul 9; Accepted 2019 Dec 19; Collection date 2020.

© The Author(s) 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptationterm·dom, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ .

PMCID: PMC6965083 PMID: 31949136 Abstract Neural adaptationterm·dom enables sensory information to be represented optimally in the brain despite large fluctuations over time in the statisticsterm of the environment. Auditory contrast gain controlterm·var represents an important example, which is thought to arise primarily from cortical processing. Here we show that neurons in the auditory thalamusterm and midbrainterm of mice show robust contrast gain controlterm·var, and that this is implemented independently of cortical activity. Although neurons at each level exhibit contrast gain controlterm·var to similar degrees, adaptationterm·dom time constantsterm become longer at later stages of the processing hierarchyterm, resulting in progressively more stable representations. We also show that auditory discriminationterm·mth thresholds in humanterm listeners compensate for changes in contrastterm, and that the strength of this perceptualterm adaptationterm·dom can be predicted from physiological measurements. Contrastterm adaptationterm·dom is therefore a robust property of both the subcorticalterm and cortical auditory system and accounts for the short-term adaptability of perceptualterm judgments.

Subject terms: Cortexterm, Midbrainterm, Thalamusterm, Neural encoding Auditory contrast gain controlterm·var helps us perceive sounds as constant despite changes in the environment or background noise. Here, the authors show that neurons in the auditory thalamusterm and midbrainterm of mice display independent contrast gain controlterm·var, not just the cortexterm as previously thought.

Introduction Adaptationterm·dom to stimulus statisticsterm is a fundamental principle of sensory processing 1 – 3 , which enables the brain to represent sensory information in ways that are computationally efficient 3 , 4 and robust to noise 5 , 6 . Certain forms of adaptationterm·dom to stimulus statisticsterm have been well studied and are known to be present at early sensory processing levels. In the visual system, for example, retinal responses adapt to mean light intensity 7 , while in the auditory system, adaptationterm·dom to mean sound levelterm has been demonstrated at the level of the auditory nerve 8 . Nevertheless, it remains poorly understood how adaptationterm·dom to higher stimulus statisticsterm changes as a result of hierarchical processing within the sensory systems or how this links to perception.

In both the visual and auditory systems, stimulus contrastterm·var—the variability of light or sound levelterm—is a stimulus statisticterm that results in neuronal adaptationterm·dom 9 – 11 . Contrastterm adaptationterm·dom may affect multiple neuronal response properties, but is accomplished principally by adjustments in response gain that compensate for the distribution of stimulus levels in a given sensory environment. This specific form of contrastterm adaptationterm·dom is known as contrast gain controlterm·var (or contrastterm normalizationterm·dom). Visual contrast gain controlterm·var is implemented at several stages of the visual system 10 – 17 , and is partially guided by corticofugalterm projections from primary visual cortexterm 18 . The perceptualterm consequences of visual contrastterm adaptationterm·dom are controversial 19 , although one report suggests that this enhances the ability of observers to detect subsequent contrastterm changes 20 . In the auditory system, however, the relative contributions of subcorticalterm and cortical structures and their role in contrast gain controlterm·var have not yet been fully elucidated, and it is not known how contrast gain controlterm·var affects perception.

Contrast gain controlterm·var is a prominent feature of neuronal responses in the auditory cortexterm of mice 21 and ferretsterm 9 , but in ferretsterm it is less robust in the midbrainterm 6 . Although this implies a primary role for auditory cortexterm in contrast gain controlterm·var, other studies have shown that the responses of subcorticalterm neurons are influenced by sensory context 22 – 28 , as well as motor and cognitive demands 29 – 32 . This raises the possibility that subcorticalterm circuits may also contribute to adaptationterm·dom to stimulus contrastterm·var. Furthermore, descending influences from the cortexterm need to be considered: manipulation of auditory corticofugalterm projections can alter the excitabilityterm and tuning properties of neurons in both the thalamusterm 33 – 35 and midbrainterm 34 – 37 , but their involvement in adaptationterm·dom to stimulus statisticsterm remains largely unexplored 27 , 38 .

In this study, we demonstrate the effects of contrastterm adaptationterm·dom on humanterm perception, by showing that acuity in a level discriminationterm·mth task is rapidly adjusted to partially compensate for changes in sound contrastterm. We also show physiologically that auditory contrast gain controlterm·var is present to comparable degrees in the lemniscalterm auditory midbrainterm, thalamusterm, and primary auditory cortexterm of mice, with progressive increases in temporal stability at each ascending processing level. Surprisingly, cortical silencingterm·mth has no effect on subcorticalterm contrast gain controlterm·var, despite significant effects on neuronal excitabilityterm, suggesting that the midbrainterm and thalamusterm implement adaptationterm·dom independently of cortexterm. Finally, we show that the strength of perceptualterm contrastterm adaptationterm·dom in humansterm can be predicted from the physiological contrastterm adaptationterm·dom observed in mouseterm auditory neurons.

Results Humanterm sound levelterm discriminationterm·mth is modulated by contrastterm To examine the perceptualterm consequences of changing the contrastterm of auditory stimuli, we measured the ability of humanterm participants to discriminate the levels of two broadband noise stimuli presented in different contrastterm environments. The stimuli were 100 ms snippets of noise, separated by 250 ms, and flanked by dynamic random chords (DRCsterm) with either 10 or 30 dB contrastterm (Fig. 1a ). We found that level discrimination performanceterm·var improved when the contrastterm of the flanking DRCsterm was low (Fig. 1b ), and that this effect was not the result of small contrast-dependent differences in overall sound levelterm that are inherent to the DRCterm stimuli (Supplementary Fig. 1 ; see Methods). All participants showed this increase in sensitivityterm ( t (7) = 5.2, p = 0.003, n = 8), as measured by the just noticeable difference (JNDterm, the dB difference between the 25% and 75% points on a fitted psychometric curveterm·stat; Fig. 1c ). The JNDterm increased by a mean of 38.8% between low- and high-contrast conditions (a threefold change in stimulus contrastterm·var), corresponding to 28.8% compensationterm for contrastterm change.

Fig. 1. Sensitivityterm to sound levelterm differences in humanterm listeners improves with decreasing auditory contrastterm.

a Spectrogram illustrating two-alternative forced-choice sound levelterm discriminationterm·mth task in different contrastterm environments (dynamic random chords) for humanterm listeners. Participants were instructed to judge whether the target sound (100 ms broadband noise) was “quieter” or “louder” than the reference sound (also 100 ms broadband noise). b Examples of psychometric functions from one participant for sound levelterm discriminationterm·mth in low- (10 dB, blue) and high- (30 dB, red) contrastterm conditions. c Changes in just noticeable difference (JNDterm, difference in dB between 25% and 75% points on psychometric curveterm·stat) across participants. Source data for c are provided as a Source Data file.

Contrast gain controlterm·var in midbrainterm, thalamusterm, and cortexterm In order to understand the role of different sensory processing levels in auditory contrastterm adaptationterm·dom, we recorded extracellular activity from neurons in the lemniscalterm parts of the auditory midbrainterm (central nucleus of the inferior colliculus, CNICterm), thalamusterm (ventral division of the medial geniculate body, MGBvterm), and primary auditory cortexterm (A1term) of anesthetized mice while playing complex spectro-temporal stimuli (DRCsterm, see Methods) with either high- (40 dB) or low- (20 dB) contrastterm (Figs. 2a–c and 3a, b ). We fitted separate spectro-temporal receptive fields (STRFsterm·mth) to the responses of each neuron to high- and low-contrast stimuli and measured various STRFterm·mth properties in both conditions. We found that there were significant systematic changes in STRFterm·mth tuning across the population (Supplementary Fig. 2 ). However, these changes were small enough that the predictionterm quality of a single combined STRFterm·mth was either statistically indistinguishable from (CNICterm, MGBvterm) or better than (A1term) the predictionterm quality of within-condition STRFsterm·mth (Supplementary Fig. 2 ). We therefore fitted a single STRFterm·mth to all the data from each neuron (Figs. 2d–e and 3c ) for subsequent analyses. We then fitted an output nonlinearityterm·mth for each contrastterm condition (Figs. 2f and 3d ). Contrastterm adaptationterm·dom in auditory neurons was assessed by comparing the output nonlinearities in high- and low-contrast conditions (see Methods).

Fig. 2. Stimulus paradigm for electrophysiological experiments and schematic of linear–nonlinear contrast-dependent modelterm·mth of auditory neurons.

a Spectrograms of snippets (1,000 ms duration) of DRCsterm with high (red) or low (blue) contrastterm. b Cross-section through an example frequency channel (top) and time point (bottom) of DRCsterm. Colored bars indicate the sound levelterm range for high (red) and low (blue) contrastterm. c Example peri-stimulus time histograms (PSTHs) during DRCterm stimulation with high- (top) and low- (bottom) contrastterm DRCsterm. d Spectro-temporal receptive field (STRFterm·mth) describing the best-fit linear relationship between stimulus structure and the response of an example neuron (red: positive, white: 0, blue: negative). e Example of 1 s of predicted neuronal response to DRCsterm with high (red) and low contrastterm (blue), based on the linear STRFterm·mth modelterm·mth. f Sigmoidal contrast-dependent output nonlinearities for an example unitterm, modeling the relationship between the actual responses of the unitterm under high- (red) and low- (blue) contrastterm conditions and the predicted responses of the STRFterm·mth linear modelterm·mth.

Fig. 3. Contrastterm adaptationterm·dom in the lemniscalterm auditory pathway.

a Schematic illustrating recordings in A1term and MGBvterm (under anesthesia) and in the CNICterm (in both anesthetized and awaketerm mice). b Confocal image showing DiI-coated electrode tracks in the MGBvterm (Dor, dorsal, Lat, lateral; scale bar, 300 µm). c Example STRFsterm·mth from unitsterm recorded in each brain region (red: positive, white: 0, blue: negative). d Contrast-dependent output nonlinearities for these same four unitsterm. e Magnitude of contrast gain controlterm·var in the auditory pathway, measured as % compensationterm where 100% would indicate a halving of the gain when the contrastterm is doubled. f Contrast-dependent changes in the baselineterm activity (y-offsetterm of the output nonlinearityterm·mth) in the auditory pathway. Colored error bars in e , f , 95% bootstrapped non-parametric confidence intervals (individual data points (excluding outliers) are displayed next to the corresponding error bars). Source data for e , f are provided as a Source Data file.

As predicted from previous studies 9 , 21 , we found that neurons in A1term exhibited strong contrast gain controlterm·var—i.e., the slope of the output nonlinearityterm·mth was adjusted following a change in contrastterm—and that this gain controlterm largely compensated for the difference in stimulus contrastterm·var (Figs. 2f and 3d–e ). In auditory cortexterm, the median degree of compensationterm was 70.2% ( p = 9.6 × 10 −14 , n = 106 unitsterm, ten mice, Wilcoxon signed-rank testterm·stat). Surprisingly, we also found strong compensatory contrast gain controlterm·var in MGBvterm (median = 55%, p = 3.6 × 10 −16 , n = 136 unitsterm, eight mice) and CNICterm (median = 70.8%, p = 1.7 × 10 −64 , n = 499 unitsterm, 13 mice; Fig. 3d–e ). A Kruskal–Wallis testterm·stat between contrast gain controlterm·var in CNICterm, MGBvterm and A1term revealed no significant differences ( p = 0.31). These results show that neurons in CNICterm, MGBvterm and A1term substantially compensate for changes in stimulus contrastterm·var by adjusting the gain of their input–output relationships. These findings did not depend on the assumption that a single STRFterm·mth could appropriately describe each neuron (Supplementary Fig. 3 ). The strength of CGC was not correlated with changes in STRFterm·mth parameters between contrastterm conditions, or the magnitude of those changes (Supplementary Fig. 3 ). These findings were also robust to the specific inclusion criteriaterm·mth used in this study (Supplementary Fig. 4 ).

Rabinowitz et al. 9 found no difference in contrast gain controlterm·var in cortical neurons between awaketerm and anesthetized ferretsterm. We extended this observation by examining whether anesthesia affected contrast gain controlterm·var in the CNICterm. We repeated our recordings in the CNICterm of awaketerm, passively listening, head-fixed mice. We found that contrast gain controlterm·var was robustly present in the CNICterm of awaketerm mice (median = 63.9% compensationterm, p = 1.2 × 10 −50 , n = 380, six mice, Wilcoxon signed-rank testterm·stat), and indistinguishable in magnitude from that exhibited by CNICterm unitsterm under anesthesia ( p = 0.1, Wilcoxon-rank-sum test; Fig. 3d, e ). A control experiment confirmed that these effects could not be attributed to small changes in overall sound levelterm between high- and low-contrast stimuli (Supplementary Fig. 5 ).

We also determined whether the baselineterm firing rateterm during DRCterm stimulation—i.e., the y-offsetterm of the output nonlinearityterm·mth—was altered by contrastterm (Fig. 3f and Supplementary Fig. 4 ). We found that baselineterm firing ratesterm in CNICterm were unaffected by contrastterm in both anesthetized ( p = 0.46, Wilcoxon signed-rank testterm·stat) and awaketerm mice ( p = 0.74). However, significant decreases in baselineterm firing ratesterm were measured in both MGBvterm (−18.5% median change, p = 9.4 × 10 −7 , Wilcoxon signed-rank testterm·stat) and A1term (−8.8% median change, p = 3.1 × 10 −9 ) during high-contrast stimulation, potentially providing an additional mechanism to make overall firing ratesterm invariant to contrastterm at these higher levels of the auditory pathway (Fig. 3f ).

Cortexterm modulates subcorticalterm excitabilityterm and reliabilityterm Although the auditory cortexterm has been found to heavily influence the subcorticalterm processing of simple tones 33 , 35 , little is known about its contribution to the representation of complex sounds in the thalamusterm or midbrainterm. In order to understand the role of descending corticofugalterm projections in the implementation of contrast gain controlterm·var, we first examined the effect of cortical inactivation on the activity of subcorticalterm neurons during continuous DRCterm stimulation.

Transiently silencingterm·mth auditory cortexterm by optogeneticterm·mth activation of inhibitory neurons (Supplementary Fig. 6 ) reduced the mean firing rateterm of MGBvterm unitsterm ( n MGBvterm = 102, five mice) during both high-contrast (−23.6% median change, p = 4.2 × 10 −18 , Wilcoxon signed-rank testterm·stat) and low-contrast (−31.3% median change, p = 3.4 × 10 −18 ) stimulation, as well as the standard deviation of the firing rateterm across time (high contrastterm: –15.8% median change, p = 7.8 × 10 −17 ; low contrastterm: −23.1% median change, p = 2.1 × 10 −17 ) (Fig. 4a, b ). Similar but weaker effects of cortical silencingterm·mth were found in the CNICterm of awaketerm mice (Supplementary Fig. 7 ).

Fig. 4. Silencingterm·mth auditory cortexterm decreases excitabilityterm in MGBvterm, increases reliabilityterm and linearity, but leaves STRFterm·mth parameters unaffected.

a Change in mean firing rateterm in MGBvterm during low-contrast (top, blue) and high-contrast (bottom, red) DRCterm stimulation following optogeneticterm·mth cortical silencingterm·mth. b Change in standard deviation (SD) of firing rateterm in MGBvterm during low-contrast (top) and high-contrast (bottom) DRCterm stimulation following optogeneticterm·mth cortical silencingterm·mth. Light shaded areas in a and b indicate unitsterm that were not significantly modulated by cortical silencingterm·mth, while dark areas represent unitsterm that were affected by cortical silencingterm·mth ( p < 0.05, t -test). c Example STRFterm·mth (example is marked with black circle in panels d – i ) of an MGBvterm unitterm with auditory cortical activity intact (top), or auditory cortexterm optogenetically silenced (bottom) (red: positive, white: 0, blue: negative). d Comparison of the best frequency (BF), i.e., the largest value of the spectral kernel of the STRFterm·mth, of MGBvterm unitsterm between recordings made with auditory cortical activity intact (ACx On) or optogenetically silenced (ACx Off). e Frequency bandwidth (fBW), i.e., the full-width half-maximum (in octaves) around the BF, of MGBvterm unitsterm with and without cortical silencingterm·mth. f Value of the maximum weight of the STRFterm·mth of MGBvterm unitsterm with and without cortical silencingterm·mth. g Temporal bandwidthterm (tBW), i.e., the full-width half-maximum (in ms) around the largest value of the temporal kernel of the STRFterm·mth, of MGBvterm unitsterm with and without cortical silencingterm·mth. h The ratio between noise and signal powerterm (NP/SP) in the MGBvterm with and without cortical silencingterm·mth. i Linear modelterm·mth predictionterm performance within contrastterm (cross-validated correlation between predicted and actual responses) in the MGBvterm with and without cortical silencingterm·mth. Color of points in d – i denotes the predictionterm strength (correlation coefficientterm·stat) of the modelterm·mth on a cross-validated dataset. Black dots are unitsterm excluded from analysis, according to exclusion criteria described in the Methods. Source data are provided as a Source Data file.

Given these strong effects on MGBvterm activity, and to a lesser degree on CNICterm activity, we examined whether corticofugalterm input influenced the structure of the STRFsterm·mth in these subcorticalterm regions (Fig. 4c ). We measured the effects of cortical silencingterm·mth on the best frequency (BF), spectral bandwidth, temporal bandwidthterm, and on the value of the largest weight in the kernel. We found that silencingterm·mth auditory cortical activity had no effect on the shape of the STRFsterm·mth of either MGBvterm unitsterm (Fig. 4d–g ) or CNICterm unitsterm (Supplementary Fig. 7 , Supplementary Fig. 8 ) ( p > 0.05, Wilcoxon signed-rank teststerm·stat).

We measured the reliabilityterm of neuronal responses by taking the ratio of noise power of the responses to the signal powerterm, NP/SP 39 (see Methods). Surprisingly, the reliabilityterm of responses to DRCterm stimuli was increased (i.e., lower NP/SP) in both MGBvterm (−23.8% median change, p = 1.0 × 10 −6 Wilcoxon signed-rank testterm·stat) and CNICterm of awaketerm mice (–11.4% median change, p = 6.0 × 10 −6 ) when cortexterm was silenced (Fig. 4h and Supplementary Fig. 7 ). We also found that after silencingterm·mth auditory cortexterm, neurons were better described by a linear modelterm·mth in the MGBvterm (14.9% median change, p = 8.0 × 10 −6 ; Fig. 4i ) and in the CNICterm of awaketerm mice (4.0% median change, p = 8.0 × 10 −6 ; Supplementary Fig. 7 ).

These results demonstrate that despite providing a strong excitatory input to MGBvterm, and to a lesser extent the CNICterm, the auditory cortexterm does not contribute to the receptive field structure of their neurons, but instead influences the reliabilityterm and linearity of thalamic responses to complex sounds.

Subcorticalterm contrast gain controlterm·var is independent of cortexterm Given the effects of cortical silencingterm·mth on subcorticalterm responses, it is possible that contrast gain controlterm·var in MGBvterm and CNICterm neurons might reflect a context-dependent influence of the extensive corticofugalterm pathways to each of these subcorticalterm structures 40 . Alternatively, subcorticalterm contrastterm adaptationterm·dom could be the result of independent computations in the CNICterm and/or MGBvterm. We addressed this directly by optogenetic silencing of auditory cortexterm·var while recording from the CNICterm and MGBvterm and presenting DRCsterm with either high (40 dB) or low (20 dB) contrastterm (Fig. 5 ). We fitted separate output nonlinearities to each condition (four conditions) from a linear spectro-temporal predictionterm across all conditions (cortexterm silenced or intact, with high- or low-contrast stimuli) (Fig. 5a, b ).

Fig. 5. Contrastterm adaptationterm·dom in the CNICterm and MGBvterm is unaffected by silencingterm·mth of auditory cortexterm.

a Examples of spectro-temporal receptive fields of unitsterm recorded in MGBvterm and CNICterm of anesthetized mice and in CNICterm of awaketerm mice (red: positive, white: 0, blue: negative). b The output nonlinearities of the same unitsterm during high- and low-contrast stimulation, with or without silencingterm·mth of cortexterm. c Summary of effects of cortical silencingterm·mth on contrast gain controlterm·var in unitsterm recorded in MGBvterm and CNICterm of anesthetized mice and CNICterm of awaketerm mice; this was quantified as the % gain change with cortexterm silenced minus the % gain change with cortexterm intact. d Summary of effects of cortical silencingterm·mth on contrast-dependent y-offsetterm adaptationterm·dom in the MGBvterm; this was quantified as % adaptationterm·dom with cortexterm silenced minus % adaptationterm·dom with cortexterm intact. No contrast-dependent y-offsetterm changes were observed in the CNICterm, so the effects of cortical silencingterm·mth are not shown. c , d Colored error bars, 95% bootstrapped non-parametric confidence intervals around the medians (individual data points (excluding outliers) are displayed next to the corresponding error bars). Source data for c , d are provided as a Source Data file.

We found that subcorticalterm contrast gain controlterm·var in anesthetized mice was not affected by transient optogeneticterm·mth cortical silencingterm·mth. This was the case for unitsterm in both MGBvterm ( p = 0.1, n = 99, five mice, Wilcoxon signed-rank testterm·stat) and CNICterm ( p = 0.5, n = 169, five mice) (Fig. 5b, c ). To control for anesthetic stateterm·var, we carried out optogeneticterm·mth cortical silencingterm·mth in awaketerm head-fixed mice while recording from CNICterm. Again, we found no effect on contrast gain controlterm·var in the CNICterm ( p CNICterm_awaketerm = 0.3, n CNICterm_awaketerm = 129, three mice) (Fig. 5b, c ).

We also examined whether auditory cortexterm contributes to the effects of contrastterm on the y-offsetterm in the MGBvterm. Cortical silencingterm·mth had a marginal, non-significant effect on this value in MGBvterm unitsterm ( p MGBvterm = 0.054, n MGBvterm = 99, five mice), suggesting that the contrast-dependent change in y-offsetterm adaptationterm·dom may not depend on cortical activity (Fig. 5d ). These results therefore suggest that auditory cortexterm does not provide the basis for the auditory contrastterm adaptationterm·dom (gain controlterm and y-offsetterm adaptationterm·dom) exhibited by subcorticalterm neurons.

Contrastterm adaptationterm·dom slows along the auditory pathway To assess the dynamics of contrast gain controlterm·var at different levels of the auditory pathway, we collected an additional dataset with recordings (under anesthesia) from CNICterm ( n = 155 unitsterm, four mice), MGBvterm ( n = 56 unitsterm, four mice) and A1term ( n = 73 unitsterm, four mice). We presented DRCsterm whose contrastterm switched between high (40 dB) and low (20 dB) values every 2 s. We modeled responses (Fig. 6a ) to this switching DRCterm using an expanded contrast-dependent LN (Linear–Nonlinear) modelterm·mth, where the parameters of the output nonlinearityterm·mth were allowed to decay exponentially between high- and low-contrast states with a time constantterm τ .

Fig. 6. Increasing time constantsterm of contrastterm adaptationterm·dom along the ascending auditory pathway.

a Mean PSTHs from example unitsterm recorded in A1term, MGBvterm, and CNICterm after switching from low to high contrastterm (red) or high to low contrastterm (blue). The adaptationterm·dom time constantterm for these unitsterm is given above each plot. b Contrast adaptation time constantsterm·var ( τ ) for all unitsterm recorded using continuously switching contraststerm in A1term, MGBvterm, and CNICterm. Black bars indicate a subset of these unitsterm whose modelterm·mth predictionterm performance ( cc predterm·def ) was improved by including an adaptationterm·dom time constantterm in the contrast-dependent LN modelterm·mth. c Temporal bandwidthterm of these unitsterm. Source data for b , c are provided as a Source Data file.

In the CNICterm, time constantsterm were very fast (median τ CNICterm = 28 ms), indicating that substantial adaptationterm·dom occurred during the first chord (duration 25 ms) after each spectro-temporal contrastterm transition (Fig. 6b ). For many CNICterm unitsterm, the inclusion of an adaptationterm·dom time constantterm did not improve predictionsterm over the standard contrast-dependent LN modelterm·mth. This further suggests that adaptationterm·dom was rapid compared to the chord duration. Adaptationterm·dom time increased with each ascending sensory processing step (median τ MGBvterm = 79 ms; median τ A1term = 175 ms) (Kruskal–Wallis testterm·stat, p < 0.001), with post-hoc comparisons (Dunn–Sidak corrected)term·stat demonstrating significantly longer median adaptationterm·dom times from CNICterm to MGBvterm ( p < 0.05) and from MGBvterm to A1term ( p < 0.05) (Fig. 6b ).

In accordance with this increase in adaptationterm·dom time from the midbrainterm to the cortexterm, the inclusion of an adaptationterm·dom time constantterm in the contrast-dependent LN modelterm·mth also became increasingly important. While including adaptationterm·dom time as a parameter in the contrast-dependent LN modelterm·mth improved the predictionterm of neural activity in 14.9% of CNICterm unitsterm, this increased to 25.0% in MGBvterm, and to more than half the unitsterm recorded in A1term (54.8%; black bars in Fig. 6b ). A subset of unitsterm was estimated to have the maximum time constantterm allowed by the modelterm·mth (700 ms, because longer time constantsterm could not be reliably estimated using stimuli whose contrastterm switched every 2 s). This is likely to be a ceilingterm effect, and suggests that a subset of unitsterm have time constantsterm that may be longer than this. Unitsterm estimated to have these long time constantsterm were most frequently found in A1term.

The progressive increase in time constantsterm might result from differences in the temporal resolution of spectro-temporal representations at different processing levels. Indeed, the temporal bandwidthterm (estimated as the full-width half-maximum of the temporal kernel in a separableterm·mth STRFterm·mth) differed between unitsterm recorded at each level (Kruskal–Wallis testterm·stat, p = 1.1 × 10 −12 ; Fig. 6c ). Post-hoc comparisons revealed significantly (Dunn–Sidak corrected) shorter temporal bandwidthsterm in CNICterm relative to both A1term ( p < 0.05) and MGBvterm ( p < 0.05). Unitsterm in MGBvterm had intermediate values between CNICterm and A1term, but these were not significantly different from A1term ( p > 0.05). However, within each auditory structure, we did not find a correlation between temporal bandwidthsterm and contrast adaptation time constantsterm·var (Spearman correlationterm·stat, p > 0.10). Thus, although both parameters increase in value along the auditory pathway, temporal bandwidthterm does not in itself account for the increase in contrastterm adaptationterm·dom time.

Neuronal contrastterm adaptationterm·dom accounts for humanterm performance Having demonstrated that contrastterm adaptationterm·dom can be observed both behaviorally in humansterm and physiologically in mice, we explored the link between the two. To do this, we developed a modelterm·mth that simulated perceptualterm judgments in the sound levelterm discriminationterm·mth task (Fig. 1 ). This incorporated simulated neural responses, where each simulated neuron was based on the contrast-dependent LN modelterm·mth of a real neuron in CNICterm, MGBvterm, or A1term (Fig. 7a and Supplementary Fig. 9 ; see Methods).

Fig. 7. The strength of perceptualterm contrastterm adaptationterm·dom can be predicted from contrastterm adaptationterm·dom in auditory neurons.

a Schematic of modelterm·mth that uses the neuronal responses to predict performance on a 2-AFC sound levelterm discriminationterm·mth task (100 ms broadband noise in different contrastterm environments; see Methods). b Psychometric functions produced by the modelterm·mth from A1term unitsterm (top) or awaketerm CNICterm unitsterm (bottom) in low- (20 dB, blue) and high- (40 dB, red) contrastterm conditions. c Predicted strength of contrastterm adaptationterm·dom from unitsterm recorded in awaketerm CNICterm or in CNICterm, MGBvterm, or A1term under anesthesia, compared with measured perceptualterm contrastterm adaptationterm·dom in humanterm listeners. Solid black lines connect mean values after 25 runs of the modelterm·mth (or across the eight participants in the measured humanterm adaptationterm·dom). Colored error bars denote 95% confidence intervals around the mean (individual data points are displayed next to the corresponding error bars). Source data for c are provided as a Source Data file.

The strength of perceptualterm contrastterm adaptationterm·dom predicted by the modelterm·mth (mean predicted contrastterm adaptationterm·dom: awaketerm CNICterm: 19.3%; anesthetized CNICterm: 20.2%; MGBvterm: 17.9%; A1term: 21.4%) closely resembled that measured in humanterm participants performing the contrast-dependent sound levelterm discriminationterm·mth task (28.8%, n = 8 participants; Fig. 7b, c ). No differences were found between these values (one-way ANOVAterm·stat, p = 0.22), suggesting that the gain controlterm measured at each level of the auditory pathway is sufficient to account for the perceptualterm adaptationterm·dom exhibited by humanterm listeners.

Discussion Our results demonstrate that auditory contrastterm adaptationterm·dom, which has been associated mainly with the auditory cortexterm 9 , 21 , is exhibited to a similar degree by neurons in lemniscalterm subcorticalterm structures—the CNICterm and MGBvterm. Moreover, we have shown that this subcorticalterm adaptationterm·dom is independent of cortical activity. We also found that perceptualterm thresholds in a sound levelterm discriminationterm·mth task compensate for contrastterm in a similar way, and that the strength of perceptualterm contrastterm adaptationterm·dom can be predicted from the gain controlterm exhibited by auditory neurons.

Previous work in the ferretterm has shown that contrastterm adaptationterm·dom is weaker and less consistent in the CNICterm than in A1term 6 , and does not consistently compensate for stimulus contrastterm·var. In contrastterm, the results of this study show that compensatory contrast gain controlterm·var in mice is not purely a cortical computation, but is present to a comparable degree in both the lemniscalterm auditory midbrainterm and the thalamusterm. Although the contraststerm used by Rabinowitz et al. 6 were different from those used in the present study, it is possible that this reflects a difference in subcorticalterm computations between mouseterm and ferretterm. In both species, however, the data suggest a hierarchyterm of contrastterm adaptationterm·dom, wherein subcorticalterm structures exhibit contrast gain controlterm·var but in cortexterm this becomes more consistent across neurons (in ferretsterm) or more temporally stable (in ferretsterm and mice). In the visual system, a hierarchyterm of contrastterm normalizationterm·dom is present at multiple processing levels from the retina upwards 41 . Similarly, predictionterm error signals increase along the auditory pathway 42 . Altogether, these findings suggest that some aspects of adaptationterm·dom to stimulus statisticsterm are organized in a serial fashion in the brain.

It is possible that contrast gain controlterm·var is also exhibited by neurons in more peripheral structures, particularly as adaptationterm·dom to mean sound levelterm takes place in the auditory nerve 8 . However, modeling studies suggest that contrast gain controlterm·var is present to a very limited degree in the auditory nerve 6 . In any case, our results show that auditory subcorticalterm neurons can execute contrast gain controlterm·var without the involvement of cortical activity. A full understanding of contrast gain controlterm·var will therefore require new hypotheses to be developed about the subcorticalterm neural circuitry and mechanisms that underlie this fundamental property of auditory neurons.

Although we found that the overall strength of contrast gain controlterm·var is similar in CNICterm, MGBvterm and A1term, adaptationterm·dom is not the same at each level of the processing hierarchyterm. A reduction in baselineterm firing rateterm during high-contrast stimulation, which may provide an additional mechanism for making overall firing ratesterm invariant to contrastterm, is found only in the MGBvterm and A1term. Furthermore, the temporal dynamics of contrast gain controlterm·var change as we ascend the auditory pathway, suggesting that additional contrast-dependent processing happens at each level. In keeping with Rabinowitz et al. 6 , we found that the time constantsterm for auditory contrast gain controlterm·var become longer at higher levels of the processing hierarchyterm. This mirrors previous results showing that the temporal integration window for auditory inputs becomes longer from the CNICterm, through MGBvterm, to A1term 43 , 44 .

The changes we observe in adaptationterm·dom time constantterm cannot be accounted for by temporal bandwidthterm changes in neuronal STRFsterm·mth. This suggests that neurons at each processing level may actively adapt to the recent history of stimulus contrastterm·var over a range of time scales, rather than merely acting as relaysterm for the transmission of auditory contrastterm. The progressive increase in the time constantterm of adaptationterm·dom along the auditory hierarchyterm is likely to result in an increasingly stable representation of the auditory environment in the cortexterm relative to subcorticalterm nuclei. Furthermore, the presence of multiple time scales of adaptationterm·dom at different levels of the auditory pathway may provide an effective means for representing sounds presented in different acoustical environments or tasks. Such diversity of dynamics among different cells also exists for visual contrastterm adaptationterm·dom in the retina 45 and adaptationterm·dom to mean level in the CNICterm 26 , suggesting that this may be a widespread property of sensory systems.

Contrast gain controlterm·var in the auditory cortexterm appears to be a specific case of neuronal normalizationterm·dom wherein the sensitivityterm of neurons adjusts to compensate for stimulus contrastterm·var 41 , 46 . It has been suggested that normalizationterm·dom is a canonicalterm computation in sensory systems and is present at multiple processing levels 41 . The results presented in this study expand on this idea by demonstrating that contrast gain controlterm·var is not only a property of neurons in auditory cortexterm, where it has been studied most extensively 6 , 9 , 21 , 47 , but, at least in mice, is equally robust in the CNICterm and MGBvterm. Contrast gain controlterm·var is therefore established at a relatively early processing level in the auditory pathway.

Our results demonstrate for the first time an important role for the thalamusterm in contrastterm adaptationterm·dom, by both increasing the duration of the adaptationterm·dom time constantsterm and introducing a subtractive component (y-offsetterm adaptationterm·dom) that is subsequently inherited by cortexterm. Neither the contrast gain controlterm·var nor the subtractive component in contrastterm adaptationterm·dom found in the thalamusterm is dependent on auditory cortical activity. Thus, the thalamusterm is an active contributor to contrastterm adaptationterm·dom in the ascending auditory pathway, and not merely a relayterm from the midbrainterm to the cortexterm.

The longer adaptationterm·dom time constantsterm we observe in the cortexterm suggest that further contrast-related processing happens there. As the representation of sound features changes along the ascending auditory pathway 48 , corresponding changes in contrastterm adaptationterm·dom may be required at each successive stage. If that is the case, an important question for future research will be whether contrast gain controlterm·var is implemented via different neural architectures, as has been shown for other neuromodulatory computations 49 , 50 . Thus, although auditory contrastterm normalizationterm·dom can be viewed as a canonicalterm computation in the brain, it is unlikely to be implemented by a canonicalterm neural circuit.

Corticofugalterm projections have previously been shown to have modulatory effects on the excitabilityterm and tuning properties of neurons in subcorticalterm nuclei in the auditory 33 – 37 , 51 , 52 , visual 53 , and somatosensory 54 , 55 systems. Several studies have reported a net excitatory effect of corticothalamic feedback, which can contribute to changes in receptive field shape 33 , 35 , 53 , 54 , 56 . In the auditory system, corticofugalterm modulation has mostly been assessed by measuring spontaneous activity and responses to tones and noise 33 – 35 , 56 , and evidence for how complex sound processing in the thalamusterm is affected is sparse. However, recent work suggests that corticothalamic feedback from layer VI of A1term to MGBvterm contributes to auditory scene analysis 57 . Furthermore, in the somatosensory system, in vitroterm recordings have demonstrated that the effects of corticothalamic feedback are dynamic, changing from suppressive to facilitatory depending on stimulation frequency 58 , suggesting that corticothalamic input may contribute to context-dependent processing of sensory information.

Our cortical silencingterm·mth results indicate that while corticofugalterm inputs have a strong effect on the overall excitabilityterm of thalamic neurons (and a weaker effect on the CNICterm), the receptive field properties of neurons in these subcorticalterm structures remain unchanged. The reduction in excitabilityterm induced by transient optogeneticterm·mth silencingterm·mth of the auditory cortexterm, and the difference in corticofugalterm effects on CNICterm and MGBvterm, are in accordance with what would be expected from previous studies of the effects of widespread inactivation of A1term on subcorticalterm responses to simple stimuli 59 . However, focal silencingterm·mth or activation of auditory cortical areas can shift the BF of neurons in both the MGBvterm and CNICterm 34 , 35 . Manipulating the activity of frequency-specific regions of auditory cortexterm may therefore have similar effects on the structure of the STRFsterm·mth acquired from complex sounds, which would be consistent with a potential role for corticofugalterm feedback in the task-dependent STRFterm·mth plasticity of auditory midbrainterm neurons 29 .

It has been proposed on the basis of in vitroterm investigations that corticothalamic feedback provides synaptic noise, which helps thalamic neurons to integrate synaptic inputs more linearly 60 , 61 . However, by isolating the corticofugalterm contribution to the representation of ongoing stimuli in vivo, our results suggest that corticofugalterm activity decreases the linear input–output relationship and the reliabilityterm of neuronal responses in the CNICterm and MGBvterm. In the CNICterm, this effect of cortical silencingterm·mth on the transfer function of the neurons appears to depend on wakefulness, but in the MGBvterm was present even under anesthesia, implying that it is not simply a result of trial-to-trial variability in corticofugalterm synaptic transmission due to changes in cognitive state.

Although cortical silencingterm·mth alters the excitabilityterm, reliabilityterm, and linearity of MGBvterm and CNICterm responses, we found no effect on the strength of contrast gain controlterm·var. This is consistent with the lack of effect of widespread cortical cooling on adaptationterm·dom to mean level by IC neurons 27 . However, cortical deactivation does prevent the change in the rate of adaptationterm·dom by IC neurons following repeated exposure to stimuli with different sound levelterm distributions 27 . It is therefore possible that descending corticofugalterm inputs might play a role in contrastterm adaptationterm·dom in rapidly changing acoustic environments.

The role of corticofugalterm inputs could be clarified by investigating contrastterm adaptationterm·dom in non-lemniscal subdivisions of the auditory thalamusterm and midbrainterm. In this study, we aimed to investigate the evolution of contrastterm adaptationterm·dom across the primary inputs to A1term. However, neurons in non-lemniscal regions of the MGB and IC display stronger stimulus-specific adaptationterm·dom than their lemniscalterm counterparts 25 , 38 and receive stronger inputs from the auditory cortexterm 38 , 62 ; these neurons may also show correspondingly stronger contrastterm adaptationterm·dom, which is under the influence of cortical activity.

The behavioral consequences of adaptationterm·dom to stimulus statisticsterm in the auditory system have received very little attention. Presenting sounds with interaural level differences 63 or interaural time differences 64 that follow specific statistical distributions results in comparable adaptive changes in the sensitivityterm of binaural neurons in the brain and in the perceptualterm sensitivityterm of humanterm listeners. Furthermore, adaptationterm·dom to mean level and contrastterm can improve the decoding of complex sounds from population neuronal activity, potentially providing a mechanism for establishing noise invariance 6 .

Our results show for the first time that humanterm auditory perception is subject to a behavioral form of contrast gain controlterm·var. We also show that the strength of perceptualterm contrastterm adaptationterm·dom in humansterm is predictable from physiological contrastterm adaptationterm·dom in midbrainterm, thalamic, and cortical auditory neurons in mice. This highlights the importance of adaptationterm·dom in regulating both neuronal and perceptualterm sensitivityterm according to the ongoing statisticsterm of the sensory environment. Furthermore, there is evidence that contrast gain controlterm·var may mediate the effects of attention on neural processing 65 , 66 . It would therefore be interesting to determine whether contrast gain controlterm·var at different levels of the auditory system can be differentially modulated depending on the sensory and behavioral contexts in which sounds occur.

The demonstration in this paper of the widespread and robust nature of auditory contrastterm adaptationterm·dom at both physiological and perceptualterm levels highlights the importance of this adaptive mechanism, and shows that a complex computation with strong implications for behavior can be implemented in subcorticalterm circuitry without the need of cortexterm.

Methods Mice All animal experiments conformed to ethical standards approved by the Committee on Animal Care and Ethical Review at the University of Oxford and were licensed by the UK Home Office (Animal Scientific Procedures Act, 1986, amended in 2012). A total of 39 mice were used in this study. Four strains of male and female mice were used in the electrophysiological experiments: C57BL6/J (Envigo, UK), GAD2-IRES-cre (Jackson Laboratories, USA), VGAT-ChR2-YFP (Jackson Laboratories, USA), and C57BL6/NTac.Cdh23 67 . C57BL6/J , GAD2-IRES-cre , and VGAT-ChR2-YFP were 7–12-weeks old at the time of data collection, and C57BL6/NTac.Cdh23 were 10–20-weeks old at the time of data collection. All experiments were carried out in a sound-attenuated chamber.

Humanterm subjects All procedures conformed to ethical standards approved by the Inter-divisional Research Ethics Committee at the University of Oxford ( R52936 /RE001). Eight (four male, four female) (plus two additional participants (both male) for the level control experiment) humanterm participants (18–30-years-old) with normal audiometry provided informed consent and participated in the contrast-dependent sound levelterm discriminationterm·mth study. All experiments were carried out in a sound-attenuated chamber.

Electrophysiologyterm·mth stimuli Stimuli were presented with a Tucker-Davis Technologies (TDT) RX6 Multifunction processor at ∼200 kHz. Sounds were amplified by a TDT SA1 stereo amplifier and delivered via a modified Avisoft ultrasonic electrostatic loudspeaker (Vifa) positioned ~1 mm from the ear canal. The sound presentation system was calibrated to a flat (±1 dB) frequency-level response between 500 and 64,000 Hz.

Stimuli consisted of dynamic random chords (DRCsterm) with individual chords having a duration of 25 ms (including 5 ms on and off ramps) and comprising 25 superposed frequencies, logarithmically spaced between 1000 and 64,000 Hz (1/4 octave intervals). The tones of the DRCterm were played at sound levelsterm that were randomly drawn from one of two uniform distributions: 30–50 dB sound pressure level (SPL) (low contrastterm) or 20–60 dB SPL (high contrastterm). The mean of the distribution was therefore constant, at 40 dB SPL. The logarithmic statisticsterm of the decibel scale have been found to better match the statisticsterm of natural sounds 44 , 68 . The overall sound levelterm of the DRCsterm was calibrated to be 79–83 dB SPL. A DRCterm for any given trial was played for either 40 s or 5 s (5-s trial duration in optogeneticterm·mth experiments), with inter-trial intervals of 2–10 s. DRCsterm have previously been used to assess contrastterm adaptationterm·dom in the auditory system of ferretsterm and mice 6 , 9 , 21 , 47 .

The overall sound levelterm of high-contrast stimuli was slightly (∼3 dB) higher than that of the low-contrast stimuli, due to the nonlinearityterm·mth inherent in the logarithmic scale. An additional experiment was therefore carried out in which the overall sound levels of DRCsterm·var were matched in low- and high-contrast stimuli, at the expense of equality of sound levelsterm of individual tones in the DRCsterm, to control for possible effects of this small difference in overall sound amplitude (see Supplementary Fig. 5 ).

In vivo extracellular recordings We carried out extracellular recordings using 32- or 64-channel silicon probes (NeuroNexus Technologies Inc.), in a 4 × 8, 8 × 8, or 2 × 32 electrode configuration. Electrophysiological data were acquired on a Tucker-Davis technologies (TDT) RZ2 BioAmp processor and collected and saved using custom-written Matlab code ( https://github.com/beniamino38/benware ).

For experiments carried out under anesthesia, mice were anesthetized with an intraperitoneal injection of ketamine (100 mg kg −1 ) and medetomidine (0.14 mg kg −1 ). We also administered intraperitoneal injections of atropine (Atrocare, 1 mg kg −1 ) to prevent bradycardia and reduce bronchial secretions, and dexamethasone (Dexadreson, 4 mg kg −1 ) to prevent brain edema. Prior to initial surgery, bupivacain was administered as an analgesic under the scalp. The depth of anesthesia was monitored via the pedal reflex and small additional doses of the ketamine/medetomidine mix were given subcutaneously approximately every 15 min once the recordings started (∼1–1.5 h post induction of anesthesia). The dosage of individual top-ups depended on the depth of anesthesia at the time, but corresponded to ∼50 mg kg −1 h −1 of ketamine and ∼0.07 mg kg −1 h −1 of medetomidine. All recordings were performed in the right hemisphere. A silver reference wire was positioned in visual cortexterm of the contralateral hemisphere, and a grounding wire was attached under the skin on the neck. The head was fixed in position with a metal bar acutely attached with bone cement to the skull over the left hemisphere. We then made 2-mm diameter circular craniotomies above the IC (centered ∼5 mm posterior from bregma and ∼1 mm lateral from midline), over the visual cortexterm for auditory thalamic recordings (centered ∼3 mm posterior from bregma and ∼2.1 mm lateral from midline), and/or over the auditory cortexterm (centered ∼2.5 mm posterior from bregma and ∼4.5 mm lateral from midline). Following exposure of the brain, the exposed dura mater was kept moist with saline. The silicon probe was then inserted carefully into the recording site of interest.

In the mouseterm, the dorsal surface of the IC is not covered by the cortexterm, and is very distinct 62 . The craniotomies over the IC were always large enough to see the entire exposed IC surface, so we could visually target the probes. We inserted the probe in the center of the IC, and therefore above the CNICterm, where the overlying dorsal cortexterm is relatively thin 62 . We confirmed this by checking for a clear dorso-ventral tonotopic gradient in the STRFsterm·mth that is indicative of this nucleus 69 , 70 . The tuning widths of the STRFsterm·mth measured in CNICterm are shown in Supplementary Fig. 2 . We also estimated frequency response areas using tones, which confirmed the presence of dorso-ventral tonotopic gradients with narrow tuning (data not shown). When we were positioning the electrode array, we observed tightly locked multiunit responses to noise stimuli, characteristic of CNICterm neurons, and post-mortem inspection of the midbrainterm confirmed that the probe had indeed been located in the CNICterm.

Prior to insertion into auditory thalamusterm, the probe was coated in DiI (Sigma-Aldrich) for subsequent histological verification of the recording site. Recording sites were confirmed as being located in auditory thalamusterm if multiunit activity responded to broadband noise and was frequency tuned when the tip of the probe was ∼2.5–3.5 mm below the brain surface. Auditory thalamic recordings were subsequently attributed to MGBvterm by histological investigation of recording sites and by analysis of physiological responses. Based on an immunohistochemical study by Lu et al. 71 on the shape and size of subdivisions of the mouseterm auditory thalamusterm, we allocated recording sites to the MGBvterm if they responded reliably to DRCterm stimulation on electrode channels < 500 μm from the lateral border of the MGB (see data inclusion criteriaterm·mth).

Finally, A1term was identified by robust neuronal responses to broadband noise bursts, and a caudo-rostral tonotopic axis. Cortical tonotopy was assessed in 4/10 mice by estimating frequency response areas from responses to pure tones on four recording shanks spaced 200 µm apart spanning 600 µm along a rostro-caudal gradient.

For awaketerm recordings in the IC, we chronically implanted a recording chamber under isoflurane (1.5–2% in O 2 ) general anesthesia. The recording chamber consisted of a metal cylinder positioned over a craniotomy, with a lightly attached circular window in order to close the recording chamber. We placed the recording chamber above the IC, together with a head bar and a reference (silver wire) in the contralateral hemisphere. We then fixed the implant to the skull using a dental adhesive resin cement (Super Bond C&B). Following full recovery, on a subsequent day the mouseterm was head-fixed, the recording chamber was opened, and a sterile recording probe was acutely inserted into the brain via the recording chamber.

Optogenetic silencing of auditory cortexterm·var To transiently silence the activity of auditory cortical excitatory neurons, we employed either a transgenic or a viral approach to express ChR2 in auditory cortical inhibitory neurons. VGAT-ChR2-YFP mice express ChR2-YFP in GABAergic neurons throughout the adult brain. Optogeneticterm·mth activation of cortical inhibitory neurons is the most effective available method for inhibiting cortical activity at sub-second time resolution 72 over the time window required for this study, and has been used extensively to transiently silence excitatory activity (including corticofugalterm outputs) in cortical areas in mice 18 , 73 – 75 . Viral injection surgeries were performed under isoflurane (∼1.5%) anesthesia, with the animal positioned in a stereotaxic frame (Kopf instruments, USA). For viral transfection, we injected a floxed AAV5-DIO-ChR2-eYFP (UNC gene therapy vector core) into auditory cortexterm of GAD2-IRES-cre mice. We injected ∼400 nl of virus, spread over three locations (spaced caudal-rostrally ∼400 μm apart) at three depths (700, 500, and 300 μm from cortical surface), to ensure widespread expression in auditory cortexterm (Supplementary Fig. 5 ). Mice were used for electrophysiological recordings > 4 weeks post injection of virus. This ensured strong expression of ChR2-eYFP in the auditory cortexterm.

For optogeneticterm·mth silencingterm·mth, we exposed the auditory cortexterm to blue (470 nm) LED light. This was achieved by placement of a 200 μm ( VGAT-ChR2-YFP experiments) or 1 mm optical fiber ( GAD2-cre + viral ChR2 experiments) immediately above the dura mater over the auditory cortexterm to allow for blue light exposure to ChR2-expressing cells. For silencingterm·mth of auditory cortical activity during recordings in MGBvterm or CNICterm, we stimulated with blue light at 40 Hz frequency using sinusoidal waves or 15 ms pulses (10 ms gaps). When recording from auditory cortexterm, we stimulated with blue light at 40 Hz using either sinusoidal waves or 15 ms pulses (10 ms gaps) or constant light stimulation. Light power was ∼5–7 mW mm −2 at the tip of the fiber. We found that light stimulation (40 Hz (sinusoid or pulsed) or constant light) effectively silenced activity in auditory cortical neurons by driving inhibitory neurons for the duration of the DRCterm stimulation (5 s) (Supplementary Fig. 6 ).

Humanterm psychoacoustic experiments Stimulus presentation and response collection were performed using PsychoPy 1.85.6 76 . Sounds were presented using a MOTU 828 mkII soundcard and delivered via Sennheiser 650HD headphones in a sound-attenuated chamber. The headphones were calibrated to a flat (±1 dB) frequency-level response between 125 and 19,500 Hz.

Stimuli consisted of broadband noise bursts (100 ms) and dynamic random chords (DRCsterm) comprising 25-ms duration chords with 29 frequencies logarithmically spaced between 150 and 19,200 Hz. DRCsterm were constructed with each tone of the DRCterm being played at levels randomly assigned from a uniform distribution, ranging from 35 to 45 dB SPL (low contrastterm) or 25 to 55 dB SPL (high contrastterm) around a fixed mean amplitude of 40 dB SPL. The total sound amplitude of the DRCsterm was measured to be 64–69 dB SPL. The stimulus for each trial was 1,950 ms long, consisting of 1000 ms of DRCterm, followed by 100 ms broadband noise (reference: 60 dB SPL), 250 ms of DRCterm, 100 ms of broadband noise (Target: 52–68 dB SPL), and ending with 500 ms of DRCterm. The overall sound levelterm of high-contrast stimuli was slightly higher relative to low-contrast stimuli (∼4 dB).

A control experiment was also carried out, where the overall sound levels of DRCsterm·var were matched in low- and high-contrast stimuli, at the expense of the equality of levels of individual tone levels in the DRCsterm, to determine whether the small difference in overall sound amplitude between the high- and low-contrast stimuli could account for the JNDterm change with contrastterm (Supplementary Fig. 1 ).

Spike sorting We clustered potential neuronal spikes using KiloSort 77 ( https://github.com/cortex-lab/KiloSort ). Following this automatic clustering step, we manually inspected the clusters in Phy ( https://github.com/kwikteam/phy ), and removed noise (movement artefacts, optogeneticterm·mth light artefacts etc.). We assessed clusters according to suggested guidelines published by Stephen Lenzi and Nick Steinmetz ( https://phy-contrib.readthedocs.io/en/latest/template-gui/#user-guide ).

Signal powerterm and noise power In order to identify unitsterm that were continuously responsive to DRCterm stimulation, we measured the signal powerterm (SP) and noise power (NP) of the neural responses 39 . For all results, unless otherwise specified, we excluded unitsterm for which the ratio NP/SP > 60, indicating that these unitsterm did not respond reliably to the DRCsterm on repeated trials.

Where relevant, we also tested how well a linear modelterm·mth described the data, using cross-validationterm·stat. We fitted spectro-temporal linear filters to 80–90% of the data (training dataset) and tested how well the modelterm·mth predicted the responses on the remaining data (test dataset). Unitsterm were excluded if the correlation coefficient (Pearson’s r )term·stat between predicted and real responses in the test dataset was <0.1. These cross-validated predictionterm values are referred to as cc predterm·def , indicating cross-validated correlation between the predicted response and the actual response.

Spectro-temporal receptive fields Neuronal response rates were binned to produce peri-stimulus time histograms (PSTHs) at the same temporal resolution (25 ms) as the chords in the DRCsterm. To exclude transient onsetterm responses, we excluded the first 500 ms of each stimulus and response. Linear spectro-temporal receptive fields (STRFsterm·mth, k fh ) were then estimated to describe the relationship between the PSTHs and the sound levelsterm (in dB SPL) of the tones in the DRCsterm. The STRFsterm·mth were constrained to be separableterm·mth in frequency ( f ) and time history ( h ), i.e. k f h = k f ⊗ k h , and were fitted using maximum likelihood 78 . The separability constraint was used because it reduces the number of parameters that need to be estimated, and can give good STRFsterm·mth when experimental data are limited 11 . We found that this approach produced acceptable STRFsterm·mth in all three areas that we recorded from.

For each unitterm, STRFsterm·mth were first fitted to data from individual contrastterm conditions separately, in order to assess contrast-dependent changes in spectro-temporal structure. Subsequently, a single overall STRFterm·mth was fitted to data from both contraststerm, for estimation of contrast-dependent output nonlinearities.

Contrast-dependent output nonlinearities For each contrastterm condition, we fitted a sigmoid function to the relationship between the actual firing rateterm of each neuron, y t , and the responses, z t , predicted by the unitterm’s overall STRFterm·mth 79 , 80 , producing a modelled firing rateterm ŷ t : ŷ t = a + b 1 + e − z t − c ∕ d 1 by estimating the parameters of the sigmoids in different contrastterm conditions ( a = y-offsetterm, b = y-range, c = x-offset, b/ (4 d ) = gain), we were then able to estimate contrast-dependent changes in the response properties of each unitterm.

Contrast gain controlterm·var was measured as percentage compensationterm·def in response to a doubling of contrastterm, where complete (100%) compensationterm is defined as a halving of gain, and no compensationterm is defined as no change in gain: % compensationterm = C low ( G low − G high ) G high ( C high − C low ) × 100 2 For the other variables, we report the percentage change between the values in high ( V high ) and low conditions ( V low ), relative to the low-contrast value: % change = V high − V low V low × 100 3 Contrast-dependent LN modelterm·mth with adaptationterm·dom time constantsterm In order to estimate adaptationterm·dom dynamics during changing contrastterm, we used a variation of the contrast-dependent LN modelterm·mth. In this modelterm·mth, the STRFterm·mth was fixed across conditions, but the parameters of the sigmoid output nonlinearityterm·mth were fitted separately to high- and low-contrast data, as described above. To estimate the time course of adaptationterm·dom of the output nonlinearityterm·mth, we allowed the sigmoid parameters to vary smoothly between their low- and high-contrast values, depending on the exponentially weighted history of recent stimulus contrastterm·var. For example: a = a low + ( a high − a low ) ∑ C t n t exp − t ∕ τ ′ 4 where a low and a high are the values of a in the low- and high-contrast conditions, respectively, C t is 0 for low contrastterm and 1 for high contrastterm, t indexes the time bins, n t is the number of time bins, and τ ′ is the time constantterm of the exponential in bins, corresponding to a time constantterm τ in ms. The dataset used to estimate the adaptationterm·dom time course switched between high (40 dB) and low (20 dB) contrastterm every 2 s. Contrast-dependent LN parameters were estimated from the last second of each contrastterm presentation. We allowed a maximum τ of 700 ms, which is the longest value that could be reliably estimated from 2-s epochs. All parameters of the LN modelterm·mth were contrastterm dependent, and the full modelterm·mth containing LN modelterm·mth parameters from both contraststerm along with the estimation of τ were optimized by gradient decent to minimize the square error between predicted firing rateterm and the actual firing rateterm.

In addition to the inclusion criteriaterm·mth used in the LN modelsterm·mth for contrastterm adaptationterm·dom estimation (see below), we further restricted analysis of time constantsterm to unitsterm whose activity was better described by a contrast-dependent LN modelterm·mth than a single contrast-independent modelterm·mth. Consequently, we estimated contrast adaptation time constantsterm·var only from unitsterm that underwent contrastterm adaptationterm·dom.

Psychophysics We fitted psychometric functions ( https://github.com/wichmann-lab/psignifit ) to the probability of participants indicating that the target sound was louder than the reference sound. The JNDterm was estimated as the dB difference between the 25% and 75% points on the psychometric curveterm·stat. As each listener’s sensitivityterm is inversely proportional to their JNDterm, we assume that the effective gain of the level discriminationterm·mth process is also inversely proportional to JNDterm, and therefore % compensationterm can be calculated similarly to the % compensationterm of contrast gain controlterm·var above.

Neurometric behavioral predictionterm modelterm·mth We predicted perceptualterm contrastterm adaptationterm·dom using contrast-dependent LN modelterm·mth simulated responses. We simulated responses to novel broadband noise stimuli of different sound levelsterm (reference: 70 dB SPL, target: 62–78 dB SPL) embedded in low- or high-contrast DRCsterm (similar to the stimuli used in the psychophysics experiment). This was achieved using response predictionsterm to these novel stimuli from the contrast-dependent LN modelterm·mth estimated from recorded unitsterm in the CNICterm, MGBvterm and A1term. This was done for every unitterm included in the analyses of physiological contrastterm adaptationterm·dom (separately for each processing level/anesthetic stateterm·var). For each simulated trial, the simulated response to the broadband noise for each unitterm was discretized according to a Poisson process, and the simulated onsetterm responses across unitsterm were added together. We then asked which noise stimulus elicited most spikes in the simulated trial. If the reference noise elicited fewer spikes than the target noise stimulus, we predicted a “louder” response (Fig. 7a ). This process was repeated 500 times for each sound levelterm, in each contrastterm condition, for estimation of a predicted contrast-dependent psychometric curveterm·stat from simulated neuronal responses from unitsterm in the CNICterm (awaketerm or anesthetized), MGBvterm or A1term (Fig. 7b ). We estimated predicted psychometric curvesterm·stat 25 times for each processing level/anesthetic stateterm·var.

Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Supplementary information Supplementary Information (7MB, pdf) Peer Review File (253.5KB, pdf) Reporting Summary (1.4MB, pdf) Acknowledgements The research was funded by a Wellcome Trust studentship (WT105241/Z/14/Z) to M.L., and a Wellcome Trust Principal Research Fellowship (WT108369/Z/2015/Z) to A.J.K.

Source data Source data (105.6KB, xlsx) Author contributions M.L., B.D.B.W., and A.J.K. conceived and designed the research. M.L. performed and visualized research. M.L. and B.D.B.W. analyzed the data. M.L. and A.J.K. acquired funding for the research. M.L., B.D.B.W., and A.J.K. interpreted the research. M.L., B.D.B.W., V.M.B., and A.J.K. wrote the manuscript.

Data availability The source data underlying Figs. 1 c, 3 e, f, 4 a, b, d–i, 5 c, d, 6b, c and 7c are provided as a Source Data File. All relevant data are available on request to, and will be fulfilled by, the lead contact (michael.lohsepersName@dpag.ox.ac.uk).

Code availability Matlab code for executing linear–nonlinear modelsterm·mth used in this paper can be found on https://github.com/beniamino38/benlib .

Competing interests The authors declare no competing interests.

Footnotes Peer review information Nature Communications thanks Jordi Costa-Faidella and Daniel Duque for their contribution to the peer review of this work. Peer reviewer reports are available.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Andrew J. KingpersName, Ben D. B. WillmorepersName Contributor Information Michael LohsepersName, Email: michael.lohsepersName@dpag.ox.ac.uk.

Andrew J. KingpersName, Email: andrew.kingpersName@dpag.ox.ac.uk Supplementary information Supplementary information is available for this paper at 10.1038/s41467-019-14163-5.