Corticofugal regulation of predictive coding

  1. Alexandria MH Lesicko
  2. Christopher F Angeloni
  3. Jennifer M Blackwell
  4. Mariella De Biasi
  5. Maria N Geffen  Is a corresponding author
  1. Department of Otorhinolaryngology, University of Pennsylvania, United States
  2. Department of Psychology, University of Pennsylvania, United States
  3. Department of Neurobiology and Behavior, Stony Brook University, United States
  4. Department of Psychiatry, University of Pennsylvania, United States
  5. Department of Systems Pharmacology and Experimental Therapeutics, University of Pennsylvania, United States
  6. Department of Neuroscience, University of Pennsylvania, United States
  7. Department of Neurology, University of Pennsylvania, United States

Abstract

Sensory systems must account for both contextual factors and prior experience to adaptively engage with the dynamic external environment. In the central auditory system, neurons modulate their responses to sounds based on statistical context. These response modulations can be understood through a hierarchical predictive coding lens: responses to repeated stimuli are progressively decreased, in a process known as repetition suppression, whereas unexpected stimuli produce a prediction error signal. Prediction error incrementally increases along the auditory hierarchy from the inferior colliculus (IC) to the auditory cortex (AC), suggesting that these regions may engage in hierarchical predictive coding. A potential substrate for top-down predictive cues is the massive set of descending projections from the AC to subcortical structures, although the role of this system in predictive processing has never been directly assessed. We tested the effect of optogenetic inactivation of the auditory cortico-collicular feedback in awake mice on responses of IC neurons to stimuli designed to test prediction error and repetition suppression. Inactivation of the cortico-collicular pathway led to a decrease in prediction error in IC. Repetition suppression was unaffected by cortico-collicular inactivation, suggesting that this metric may reflect fatigue of bottom-up sensory inputs rather than predictive processing. We also discovered populations of IC units that exhibit repetition enhancement, a sequential increase in firing with stimulus repetition. Cortico-collicular inactivation led to a decrease in repetition enhancement in the central nucleus of IC, suggesting that it is a top-down phenomenon. Negative prediction error, a stronger response to a tone in a predictable rather than unpredictable sequence, was suppressed in shell IC units during cortico-collicular inactivation. These changes in predictive coding metrics arose from bidirectional modulations in the response to the standard and deviant contexts, such that the units in IC responded more similarly to each context in the absence of cortical input. We also investigated how these metrics compare between the anesthetized and awake states by recording from the same units under both conditions. We found that metrics of predictive coding and deviance detection differ depending on the anesthetic state of the animal, with negative prediction error emerging in the central IC and repetition enhancement and prediction error being more prevalent in the absence of anesthesia. Overall, our results demonstrate that the AC provides cues about the statistical context of sound to subcortical brain regions via direct feedback, regulating processing of both prediction and repetition.

Editor's evaluation

This study concerns the neural representation of prediction in the central auditory pathway. The authors report that top-down inputs from the auditory cortex carry contextual cues that enable subcortical neurons to distinguish between predictable and unexpected sounds. This work provides important insights into how feedback pathways in the auditory system modulate feedforward signals in a context-dependent fashion.

https://doi.org/10.7554/eLife.73289.sa0

Introduction

Sensory systems differentially encode environmental stimuli depending on the context in which they are encountered (De Franceschi and Barkat, 2020; Herrmann et al., 2015; Jaramillo et al., 2014; Pakan et al., 2016; Takesian et al., 2018; Zhai et al., 2020). The same physical stimulus can elicit distinct neuronal responses depending on whether it is predictable or unexpected in a given sensory stream (Weissbart et al., 2020; Yaron et al., 2012). Neurons in select regions of the central auditory system are sensitive to statistical context, responding more strongly to a tone when it is presented rarely (a ‘deviant’) than when it is commonplace (a ‘standard’) (Ulanovsky et al., 2003). This phenomenon, known as stimulus-specific adaptation (SSA), is prevalent in the auditory cortex (AC) (Natan et al., 2015; Ulanovsky et al., 2003). Weaker SSA is present in regions peripheral to the AC, including the auditory midbrain, or inferior colliculus (IC), and the auditory thalamus, or medial geniculate body (MGB) (Anderson et al., 2009; Antunes et al., 2010; Duque and Malmierca, 2015; Malmierca et al., 2009; Taaseh et al., 2011; Ulanovsky et al., 2003). Subdivisions in IC and MGB that receive descending projections from AC exhibit relatively higher SSA levels than their lemniscal counterparts (Antunes et al., 2010; Duque et al., 2012), suggesting that SSA may be generated de novo in AC and subsequently broadcast to subcortical structures via corticofugal projections (Nelken and Ulanovsky, 2007). Silencing of AC through cooling, however, has been shown to modulate, but not abolish, SSA in IC and MGB of anesthetized rats (Anderson and Malmierca, 2013; Antunes and Malmierca, 2011).

Recent studies have implemented additional control tone sequences to further decompose the traditional SSA index into two distinct underlying processes: repetition suppression and prediction error (Harms et al., 2014; Parras et al., 2017; Ruhnau et al., 2012). Repetition suppression is characterized by a decrease in firing rate to each subsequent presentation of a standard tone, whereas prediction error signals an enhanced response to a deviant tone (Auksztulewicz and Friston, 2016; Parras et al., 2017). Hierarchical predictive coding posits that prediction errors signal the mismatch between predictions, formed based on prior experience with repeated presentations of the standard, and actual sensory input in the presence of a deviant (Friston, 2009; Friston and Kiebel, 2009). These predictions are generated at higher levels of the sensory hierarchy and broadcast to lower stations to minimize processing of redundant input and maximize coding efficiency (Friston, 2009; Friston and Kiebel, 2009). Prediction error has been proposed to underlie true deviance detection, while repetition suppression is thought to potentially reflect synaptic depression (Parras et al., 2017; Taaseh et al., 2011). Prediction error increases along the auditory hierarchy and is more prevalent in regions of IC and MGB that receive cortical feedback (Parras et al., 2017), suggesting that these subcortical regions may engage in hierarchical predictive coding, with AC potentially providing predictive cues to IC and MGB. However, how feedback projections from AC shape predictive processing in subcortical targets has never been directly assessed. In fact, virtually all models of hierarchical predictive coding to date have focused on intracortical connections, with the massive system of descending corticofugal projections remaining unexplored (Asilador and Llano, 2020; Bastos et al., 2012).

Here, we investigated how inputs from AC to IC, the first station in the auditory system in which prediction error is found, shape metrics associated with predictive coding and deviance detection (Parras et al., 2017). To test this, we optogenetically inactivated cortico-collicular feedback while recording neuronal responses in IC and found that prediction error, negative prediction error, and repetition enhancement in IC are altered in the absence of cortical input. Our results suggest that the cortico-collicular pathway sends cues from AC to IC regarding the statistical context of auditory stimuli.

Results

Experimental design

We used a Cre/FLEX viral injection strategy to selectively express the inhibitory opsin, ArchT, in cortico-collicular neurons of four mice by injecting a retroAAV-Cre-GFP construct into IC and an AAV9-FLEX-ArchT-tdTomato construct into AC (Figure 1A, left). The retroAAV-Cre-GFP construct is transported in a retrograde fashion and expressed in neurons that project to IC (Blackwell et al., 2020). The genes encoded in the AAV9-FLEX-ArchT-tdTomato construct can only be expressed in neurons containing the Cre construct, thereby limiting ArchT expression to neurons in AC that project to IC. In the presence of green light, ArchT, a light-driven outward proton pump, mediates rapid, reversible inactivation of the neurons in which it is expressed (Han et al., 2011).

Figure 1 with 2 supplements see all
Experimental design.

(A) Cre/FLEX dual injections for selective ArchT expression in cortico-collicular neurons. Recordings were performed in the inferior colliculus (IC) while inactivation was mediated by a 532 nm laser connected to cannulas implanted over the auditory cortex (AC). (B) Oddball stimuli consisted of pairs of pure tones separated by 0.5 octave with a 90:10 standard-to-deviant ratio. Two sequences were constructed such that each frequency is represented as both the standard and the deviant. (C) Cascade sequences consisted of 10 evenly spaced tones separated by 0.5 octaves, with both frequencies from the oddball sequence included in the sequence. Responses to tones in the cascade context were compared to responses in the standard and deviant context to analyze repetition and prediction effects, respectively. (D) A positive index of neuronal mismatch (iMM) (top diagram) indicates a stronger response to the deviant than the standard (adaptation), while a negative iMM (bottom diagram) indicates a stronger response to the standard than to the deviant (facilitation). The iMM can be further decomposed into an index of prediction error (iPE) and an index of repetition suppression (iRS). Positive iPE values represent prediction error, and negative values convey negative prediction error. Positive iRS indices indicate repetition suppression, while repetition enhancement is represented by negative values.

We implanted cannulas over AC in mice injected with the Cre/FLEX constructs and a 532 nm laser was used to provide green light illumination to the region, allowing for inactivation of cortico-collicular neurons (Figure 1A, right). The mice were head-fixed and a 32-channel probe was lowered into IC to perform awake extracellular recordings (Figure 1A). Auditory stimuli consisted of oddball sequences of two repeated pure tones, presented at a 90:10 standard-to-deviant ratio and half-octave frequency separation (Figure 1B). On a subset of trials, presentations of either the deviant or the last standard prior to the deviant were coupled with activation of the green laser (Figure 1B, right).

Units that displayed a significantly higher response to the deviant than the standard were designated as ‘adapting’ units, while those that exhibited a significantly higher response to the standard than the deviant were categorized as ‘facilitating’ units (Figure 1D). The difference in firing rate to the standard and deviant was quantified with an index of neuronal mismatch (iMM), which is equivalent to the SSA index used in previous studies (Parras et al., 2017).

A cascade stimulus consisting of 10 evenly spaced tones, including the tone pair from the oddball sequence, was presented to further decompose the neuronal mismatch between the responses to the standard and deviant (Figure 1C and D). This stimulus is unique in that each tone occurs with the same likelihood as the deviant tone in the oddball stimulus (10%), but it contains no true statistical deviants: each tone has the same likelihood of presentation, and the tone sequence overall follows a regular and predictable pattern (Parras et al., 2017). Therefore, the response to a given tone when it is embedded in the cascade can be compared to the response when it is a deviant in order to isolate prediction error effects (Figure 1C and D, top). A neuron exhibits prediction error if it fires more strongly to a tone when it is a deviant than when it is presented in the cascade sequence (Figure 1D, top). Conversely, if a neuron responds more strongly to a tone presented in the cascade sequence than when it is a deviant, the neuron encodes negative prediction error (Figure 1D, bottom). This phenomenon is quantified using an index of prediction error (iPE), with positive indices indicating prediction error and negative indices representing negative prediction error (Figure 1D).

The cascade sequence is also free from repetition effects since adjacent tone presentations never include a tone of the same frequency (Figure 1C). Therefore, the response to a given tone embedded in the cascade sequence can be compared to the response generated when that tone is a standard. The difference in response indicates either repetition suppression (stronger response to the tone in the cascade) (Figure 1D, top) or repetition enhancement (stronger response to the tone as a standard) (Figure 1D, bottom). These contrasting processes are quantified by the index of repetition suppression (iRS), with a positive index indicating repetition suppression and a negative index representing repetition enhancement (Figure 1D).

Cre/FLEX viral injection strategy enables selective inactivation of cortico-collicular neurons

Examination of fixed tissue from injected mice revealed that expression of the retroAAV-Cre-GFP construct was restricted to IC (Figure 1—figure supplement 1A, top left). Somatic expression of GFP (indicating the presence of Cre) was restricted to layer 5 and deep layer 6 of AC, which contain cortico-collicular cell bodies, and was broadly distributed throughout the rostro-caudal extent of AC (Figure 1—figure supplement 1A, right) (Bajo et al., 2007; Schofield, 2009; Yudintsev et al., 2019). Expression of tdTomato was found in the soma and processes of neurons in layers 5 and 6, with additional apical dendritic labeling observed in the upper cortical layers (Figure 1—figure supplement 1A, right). The laminar expression of tdTomato is consistent with previous studies and suggests that AAV9-FLEX-ArchT-tdTomato expression is Cre-dependent and not due to nonspecific labeling (Blackwell et al., 2020). Axons and terminals labeled with tdTomato were distributed in IC in a manner matching the known projection pattern of this pathway, with dense, ‘patchy’ labeling in shell regions of IC (Figure 1—figure supplement 1A, bottom left) (Herbert et al., 1991; Lesicko et al., 2016; Saldaña et al., 1996; Torii et al., 2013). These data confirm that our viral injection strategy leads to selective transfection of cortico-collicular neurons.

Extracellular recordings in AC of injected mice revealed a reduction in firing rate during the duration of the laser stimulus in several units (Figure 1—figure supplements 1B and 2C). In these putative cortico-collicular units, laser-induced inactivation led to a mean ~60% reduction in firing rate at baseline (Figure 1—figure supplement 1C, left; Figure 1—figure supplement 2D, top; Table 1; p=1.9e-06, Wilcoxon signed-rank test) and an average ~45% reduction in firing during presentation of pure tone stimuli (Figure 1—figure supplement 1C, right; Figure 1—figure supplement 2D, bottom; Table 1; p=1.9e-06, Wilcoxon signed-rank test). These results indicate that our optogenetic parameters significantly suppress cortico-collicular units.

Table 1
Statistical comparisons for experimental data.
ComparisonFigureMeanMedianSDSEMCI (±)TestTest statisticNdfpEffect size
Response of putative cortico-collicular units in silence (laser OFF vs. ON)Figure 1—figure supplement 1D (top)OFF: 11ON: 4.1OFF: 9.0ON: 3.5OFF: 8.9ON: 3.5OFF: 2.0ON: 0.78OFF: 4.2ON: 1.6Wilcoxon signed-rank testV = 020NA1.9e-060.88
Response of putative cortico-collicular units to pure tones (laser OFF vs. ON)Figure 1—figure supplement 1D (bottom)OFF: 18ON: 9.6OFF: 8.8ON: 4.3OFF: 24ON: 12OFF: 5.4ON: 2.7OFF: 11ON: 5.6Wilcoxon signed-rank testV = 020NA1.9e-060.88
iMM central (awake vs. anesthetized)Figure 2BAw: 0.050An: 0.25Aw: 0.045An: 0.28Aw: 0.21An: 0.49Aw: 0.024An: 0.074Aw: 0.047An: 0.15Wilcoxon rank-sum testW = 952.5Aw: 78An: 43NA8.8e-050.36
iPE central (awake vs. anesthetized)Figure 2CAw: –0.13An: 0.077Aw: –0.11An: 0.098Aw: 0.17An: 0.53Aw: 0.019An: 0.081Aw: 0.038An: 0.16Student’s t-testt = –2.5Aw: 78An: 43380.0170.52
iRS central (awake vs. anesthetized)Figure 2DAw: 0.18An: 0.18Aw: 0.17An: 0.30Aw: 0.17An: 0.56Aw: 0.019An: 0.085Aw: 0.039An: 0.17Wilcoxon rank-sum testW = 1444Aw: 78An: 43NA0.210.12
iMM shell (awake vs. anesthetized)Figure 2EAw: 0.095An: 0.27Aw: 0.090An: 0.27Aw: 0.31An: 0.35Aw: 0.025An: 0.022Aw: 0.050An: 0.043Wilcoxon rank-sum testW = 12,502Aw: 147An: 254NA3.5e-080.28
iPE shell (awake vs. anesthetized)Figure 2FAw: 0.15An: 0.018Aw: 0.15An: –0.0075Aw: 0.33An: 0.39Aw: 0.027An: 0.025Aw: 0.053An: 0.049Wilcoxon rank-sum testW = 23,368Aw: 147An: 254NA2.6e-050.21
iRS shell (awake vs. anesthetized)Figure 2GAw: –0.056An: 0.25Aw: –0.085An: 0.29Aw: 0.36An: 0.33Aw: 0.029An: 0.020Aw: 0.058An: 0.040Wilcoxon rank-sum testW = 9501.5Aw: 147An: 254NA2.5e-160.41
iMM central adapting (laser OFF vs. ON)Figure 3D (top)OFF: 0.26ON: 0.21OFF: 0.24ON: 0.19OFF: 0.096ON: 0.13OFF: 0.013ON: 0.019OFF: 0.027ON: 0.037Wilcoxon signed-rank testV = 108352NA0.000340.50
iPE central adapting (laser OFF vs. ON)Figure 3D (middle)OFF: 0.0077ON: –0.029OFF: 0.036ON: 0.0041OFF: 0.16ON: 0.16OFF: 0.022ON: 0.022OFF: 0.043ON: 0.044Wilcoxon signed-rank testV = 90752NA0.0480.28
iRS central adapting (laser OFF vs. ON)Figure 3D (bottom)OFF: 0.25ON: 0.24OFF: 0.24ON: 0.24OFF: 0.16ON: 0.16OFF: 0.023ON: 0.022OFF: 0.046ON: 0.045Wilcoxon signed-rank testV = 83252NA0.190.18
iMM shell adapting (laser OFF vs. ON)Figure 3E (top)OFF: 0.34ON: 0.31OFF: 0.32ON: 0.28OFF: 0.19ON: 0.20OFF: 0.017ON: 0.019OFF: 0.035ON: 0.037Wilcoxon signed-rank testV = 4283113NA0.00230.29
iPE shell adapting (laser OFF vs. ON)Figure 3E (middle)OFF: 0.15ON: 0.14OFF: 0.12ON: 0.10OFF: 0.30ON: 0.30OFF: 0.028ON: 0.028OFF: 0.056ON: 0.056Wilcoxon signed-rank testV = 3963113NA0.0340.20
iRS shell adapting (laser OFF vs. ON)Figure 3E (bottom)OFF: 0.19ON: 0.17OFF: 0.19ON: 0.16OFF: 0.24ON: 0.24OFF: 0.023ON: 0.023OFF: 0.045ON: 0.045Paired t-testt = 1.61131120.110.15
iMM central facilitating (laser OFF vs. ON)Figure 3G (top)OFF: –0.32ON: –0.13OFF: –0.31ON: –0.11OFF: 0.16ON: 0.19OFF: 0.042ON: 0.050OFF: 0.090ON: 0.11Paired t-testt = –3.514130.00360.95
iPE central facilitating (laser OFF vs. ON)Figure 3G (middle)OFF: –0.20ON: –0.17OFF: –0.24ON: –0.20OFF: 0.20ON: 0.17OFF: 0.054ON: 0.044OFF: 0.12ON: 0.095Paired t-testt = –1.214130.250.32
iRS central facilitating (laser OFF vs. ON)Figure 3G (bottom)OFF: –0.12ON: 0.036OFF: –0.092ON: 0.069OFF: 0.18ON: 0.24OFF: 0.049ON: 0.064OFF: 0.11ON: 0.14Paired t-testt = –3.714130.00261.0
iMM shell facilitating (laser OFF vs. ON)Figure 3H (top)OFF: –0.29ON: –0.19OFF: –0.24ON: –0.15OFF: 0.15ON: 0.16OFF: 0.024ON: 0.026OFF: 0.048ON: 0.052Wilcoxon signed-rank testV = 15938NA0.00160.50
iPE shell facilitating (laser OFF vs. ON)Figure 3H (middle)OFF: –0.026ON: 0.033OFF: 0.011ON: 0.023OFF: 0.26ON: 0.29OFF: 0.042ON: 0.047OFF: 0.085ON: 0.096Wilcoxon signed-rank testV = 22738NA0.0370.34
iRS shell facilitating (laser OFF vs. ON)Figure 3H (bottom)OFF: –0.26ON: –0.23OFF: –0.29ON: –0.23OFF: 0.32ON: 0.33OFF: 0.052ON: 0.054OFF: 0.11ON: 0.11Wilcoxon signed-rank testV = 25438NA0.0930.27
iMM central nonadapting (laser OFF vs. ON)Figure 4C (top)OFF: 0.022ON: 0.072OFF: 0.023ON: 0.065OFF: 0.12ON: 0.14OFF: 0.0094ON: 0.011OFF: 0.019ON: 0.022Wilcoxon signed-rank testV = 3419155NA2.7e-060.38
iPE central nonadapting (laser OFF vs. ON)Figure 4C (middle top)OFF: –0.096ON: –0.081OFF: –0.098ON: –0.093OFF: 0.19ON: 0.19OFF: 0.015ON: 0.015OFF: 0.030ON: 0.030Wilcoxon signed-rank testV = 5327155NA0.200.10
iRS central nonadapting (laser OFF vs. ON)Figure 4C (middle bottom)OFF: 0.12ON: 0.15OFF: 0.12ON: 0.15OFF: 0.15ON: 0.17OFF: 0.012ON: 0.013OFF: 0.024ON: 0.027Wilcoxon signed-rank testV = 4224155NA0.00110.26
iRS > 0 central nonadapting (laser OFF vs. ON)Figure 4C (bottom)OFF: 0.17ON: 0.19OFF: 0.16ON: 0.18OFF: 0.10ON: 0.15OFF: 9.1e-03ON: 0.013OFF: 1.8e-02ON: 0.026Wilcoxon signed-rank testV = 3313127NA0.0710.16
iRS < 0 central nonadapting (laser OFF vs. ON)Figure 4C (bottom)OFF: –0.13ON: –0.012OFF: –0.10ON: –0.017OFF: 0.11ON: 0.15OFF: 0.021ON: 0.029OFF: 0.044ON: 0.060Wilcoxon signed-rank testV = 3025NA0.000120.71
iMM shell nonadapting (laser OFF vs. ON)Figure 4D (top)OFF: 0.0053ON: 0.023OFF: 0.0062ON: 0.028OFF: 0.13ON: 0.16OFF: 0.0081ON: 0.010OFF: 0.016ON: 0.020Wilcoxon signed-rank testV = 12,765243NA0.0760.11
iPE shell nonadapting (laser OFF vs. ON)Figure 4D (middle)OFF: 0.053ON: 0.072OFF: 0.059ON: 0.061OFF: 0.21ON: 0.20OFF: 0.013ON: 0.013OFF: 0.026ON: 0.026Wilcoxon signed-rank testV = 13,474243NA0.220.079
iRS shell nonadapting (laser OFF vs. ON)Figure 4D (bottom)OFF: –0.048ON: –0.049OFF: –0.042ON: –0.041OFF: 0.23ON: 0.22OFF: 0.015ON: 0.014OFF: 0.029ON: 0.028Wilcoxon signed-rank testV = 14,344243NA0.660.028
FR change standard central adaptingFigure 5A2.12.05.60.781.6One-sample t-testt = 2.752510.00920.38
FR change cascade central adaptingFigure 5A–0.380.676.90.951.9One-sample t-testt = –0.4052510.690.056
FR change deviant central adaptingFigure 5A–2.3–2.25.60.781.6One-sample t-testt = –2.952510.00540.40
FR change standard shell adaptingFigure 5B0.640.895.30.500.98One-sample Wilcoxon testV = 3760113NA0.0350.20
FR change cascade shell adaptingFigure 5B0.500.447.30.681.4One-sample t-testt = 0.741131120.460.069
FR change deviant shell adaptingFigure 5B–1.8–1.37.40.691.4One-sample Wilcoxon testV = 2040113NA0.00570.26
FR change standard central facilitatingFigure 5C–6.3–7.35.81.63.4One-sample t-testt = –4.114130.00131.1
FR change cascade central facilitatingFigure 5C–0.44–0.894.11.12.4One-sample t-testt = –0.4014130.690.11
FR change deviant central facilitatingFigure 5C1.51.33.40.922.0One-sample t-testt = 1.714130.120.45
FR change standard shell facilitatingFigure 5D–2.7–3.15.40.871.8One-sample t-testt = –3.138370.00420.50
FR change cascade shell facilitatingFigure 5D0.360.445.10.841.7One-sample t-testt = 0.4338370.670.070
FR change deviant shell facilitatingFigure 5D2.62.74.50.741.5One-sample t-testt = 3.538370.00130.57
FR change standard central nonadaptingFigure 5E–2.5–2.26.20.500.99One-sample Wilcoxon testV = 2995155NA1.4e-060.38
FR change cascade central nonadaptingFigure 5E–0.68–0.446.30.511.0One-sample t-testt = –1.31551540.180.11
FR change deviant central nonadaptingFigure 5E0.570.05.80.470.93One-sample t-testt = 1.21551540.220.098
FR change standard shell nonadaptingFigure 5F–0.63–0.445.30.340.68One-sample Wilcoxon testV = 11,050243NA0.0350.14
FR change cascade shell nonadaptingFigure 5F–0.51–0.445.10.320.64One-sample Wilcoxon testV = 12,157243NA0.150.089
FR change deviant shell nonadaptingFigure 5F–0.0590.05.00.320.64One-sample t-testt = –0.182432420.860.012
FR central facilitating (first vs. last standard)Figure 6CFirst: 31Last: 36First: 29Last: 31First: 15Last: 16First: 3.9Last: 4.4First: 8.5Last: 9.5Wilcoxon signed-rank testV = 014NA0.00170.87
FR shell facilitating (first vs. last standard)Figure 6DFirst: 53Last: 57First: 38Last: 42First: 38Last: 42First: 6.2Last: 6.8First: 13Last: 14Wilcoxon signed-rank testV = 9238NA9.3e-050.64
FR central adapting (cascade vs. many standards)Figure 3—figure supplement 2B (left)Casc: 61MS: 63Casc: 50MS: 52Casc: 38MS: 40Casc: 5.2MS: 5.6Casc: 10MS: 11Wilcoxon signed-rank testV = 59552NA0.390.12
FR central facilitating (cascade vs. many standards)Figure 3—figure supplement 2B (right)Casc: 29MS: 31Casc: 26MS: 28Casc: 14MS: 16Casc: 3.8MS: 4.3Casc: 8.2MS: 9.3Wilcoxon signed-rank testV = 4114NA0.490.19
FR shell adapting (cascade vs. many standards)Figure 3—figure supplement 2C (left)Casc: 64MS: 66Casc: 43MS: 41Casc: 61MS: 68Casc: 5.7MS: 6.4Casc: 11MS: 13Wilcoxon signed-rank testV = 2653113NA0.460.064
FR shell facilitating (cascade vs. many standards)Figure 3—figure supplement 2C (right)Casc: 43MS: 45Casc: 24MS: 28Casc: 41MS: 52Casc: 6.6MS: 8.4Casc: 13MS: 17Wilcoxon signed-rank testV = 264.538NA0.410.14
Central iMM OFF (single vs. multiunit)Figure 3—figure supplement 3 (left)Single: 0.045Multi: 0.057Single: 0.048Multi: 0.064Single: 0.15Multi: 0.18Single: 0.052Multi: 0.013Single: 0.12Multi: 0.025Wilcoxon rank-sum testW = 825Single: 8Multi: 213NA0.880.010
Central iMM ON (single vs. multiunit)Figure 3—figure supplement 3 (left)Single: 0.087Multi: 0.092Single: 0.085Multi: 0.086Single: 0.17Multi: 0.16Single: 0.059Multi: 0.011Single: 0.14Multi: 0.022Student’s t-testt = –0.093Single: 8Multi: 2137.50.930.034
Shell iMM OFF (single vs. multiunit)Figure 3—figure supplement 3 (right)Single: 0.035Multi: 0.081Single: 0.028Multi: 0.055Single: 0.18Multi: 0.25Single: 0.022Multi: 0.014Single: 0.045Multi: 0.027Wilcoxon rank-sum testW = 9832Single: 67Multi: 327NA0.190.067
Shell iMM ON (single vs. multiunit)Figure 3—figure supplement 3 (right)Single: 0.046Multi: 0.091Single: 0.045Multi: 0.072Single: 0.21Multi: 0.23Single: 0.026Multi: 0.013Single: 0.051Multi: 0.025Wilcoxon rank-sum testW = 9883Single: 67Multi: 327NA0.210.064
  1. iRS: index of repetition suppression; iPE: index of prediction error; iMM: index of neuronal mismatch; Aw: awake; An: anesthetized; casc: cascading; MS: many standards.

Parsing of recording sites into central and shell locations

Shell and central regions of IC differ in their tuning, degree of adaptation, and amount of input from AC, and may also play distinct roles in predictive processing (Aitkin et al., 1975; Bajo et al., 2007; Blackwell et al., 2020; Duque et al., 2012; Herbert et al., 1991; Stebbings et al., 2014; Syka et al., 2000). We quantitatively parsed our recording sites by exploiting known differences in the sharpness of tuning and direction of frequency gradients between shell and central regions: shell IC neurons tend to have broader frequency tuning (low sparseness) than central IC neurons, and the central IC is characterized by a highly stereotyped tonotopic gradient with depth (Figure 1—figure supplement 2A; Aitkin et al., 1975; Chen et al., 2012; Malmierca et al., 2008; Stiebler and Ehret, 1985; Syka et al., 2000). Similar to previously established procedures used in human and monkey IC research, we performed clustering analysis using the mean sparsity and variation in best frequency with depth from each recording site to determine whether it was from the central nucleus or shell regions of IC (Figure 1—figure supplement 2B and C; Bulkin and Groh, 2011; Ress and Chandrasekaran, 2013). In a subset of recordings, we also marked the recording electrode with a lipophilic dye to histologically confirm the recording location (Figure 1—figure supplement 2D).

Inferior colliculus (IC) units encode different aspects of prediction and repetition in awake and anesthetized states.

(A) Experimental design for recording in the awake and isoflurane anesthetized IC in the same population of units. (B) Distribution of index of neuronal mismatch (iMM) in the awake vs. anesthetized central IC. Bar plots represent means over the population of n = 39 units. Error bars are standard error of the mean. (C) Index of prediction error (iPE) distribution in the awake vs. anesthetized central IC. (D) Index of repetition suppression (iRS) distribution in the awake vs. anesthetized central IC. (E) Distribution of iMM in the awake vs. anesthetized shell IC. Bar plots represent means over the population of n = 165 units. Error bars are standard error of the mean. (F) iPE distribution in the awake vs. anesthetized shell IC. (G) iRS distribution in the awake vs. anesthetized shell IC. Data is from four recording sessions in one mouse.

IC units in both regions exhibited multiple response types to pure tone stimuli (Figure 1—figure supplement 2E). In addition to excitatory responses (e.g., onset and sustained responses), inhibited and offset responses were common, as has previously been characterized in IC of awake animals (Figure 1—figure supplement 2E, top right, bottom middle; Duque and Malmierca, 2015). Consistent with previous findings, tuning curves from central regions were sharp and narrow, whereas units in shell regions exhibited broad frequency tuning (Figure 1—figure supplement 2F, left vs. right; Aitkin et al., 1975; Syka et al., 2000). Inhibited side bands were common in tuning curves from both regions, and some inhibited tuning curves were observed (Figure 1—figure supplement 2G). These data confirm that our experimental parameters elicit sound responses and tuning properties characteristic of central and shell regions of the awake IC (Aitkin et al., 1975; Duque and Malmierca, 2015; Syka et al., 2000).

IC units encode different aspects of prediction and repetition in awake and anesthetized states

Much of the research regarding SSA and deviance detection in IC to date has been performed in anesthetized animals, with few studies recording from awake subjects (Duque and Malmierca, 2015; Parras et al., 2017). Given that neuronal responses to sound depend on the state of anesthesia of the subject, it is possible that there are differences in predictive coding metrics between the awake and anesthetized states (Fontanini and Katz, 2008; Gaese and Ostwald, 2001; Schumacher et al., 2011). While previous studies have characterized how anesthesia affects SSA, it remains unknown whether its component repetition and prediction metrics differ with anesthetic state (Duque and Malmierca, 2015). Therefore, we first characterized how anesthesia affects these predictive coding metrics in a subset of animals. We first performed awake recordings and then repeated our experimental procedures, leaving the animal head-fixed and the probe in place, after anesthetizing the mouse with isoflurane (Figure 2A). This protocol allowed us to compare how metrics of predictive coding differ between the awake and anesthetized preparations in the same population of units.

In the central IC, the mean iMM in the anesthetized condition was positive, indicative of prevalent adaptation (Figure 2B). The iMM values under anesthesia were significantly higher than those obtained while the animal was awake (Figure 2B, Table 1; p=8.8e-05, Wilcoxon rank-sum test). To better understand what prediction or repetition effects underlie iMM in each condition, the iMM for both distributions was further decomposed into an iPE and iRS. In the anesthetized condition, the mean iPE value of 0.077 indicated the presence of modest prediction error, while a mean iPE of –0.13 indicated that negative prediction error is significantly more prevalent in the awake condition (Figure 2C, Table 1; p=0.017, Student’s t-test). Under both anesthetized and awake conditions, prominent repetition suppression was observed in the central IC (Figure 2D).

Similar to the central IC, the mean iMM was significantly more positive in shell regions during anesthesia (Figure 2E, Table 1; p=3.5e-08, Wilcoxon rank-sum test). A greater proportion of units in the awake condition had a negative iMM compared with the anesthetized distribution, indicating that facilitation (a greater response to the standard than the deviant context) is more common in the awake than the anesthetized condition (Figure 2E). The iPE values in shell IC suggest that prediction error is significantly higher in the awake compared to the anesthetized condition (Figure 2F, Table 1; p=2.6e-05, Wilcoxon rank-sum test). Although the distribution for the iRS under anesthesia had a positive mean of 0.25, indicating prevalent repetition suppression, the awake distribution exhibited a significant leftward shift by comparison (Figure 2G). Interestingly, the mean iRS for the awake condition was negative (mean = −0.056), indicating that repetition enhancement, rather than suppression, is present in the awake shell IC (Figure 2G, Table 1; p=2.5e-16, Wilcoxon rank-sum test). These results point to differences between predictive coding metrics in the awake and anesthetized states, with previously undescribed metrics such as repetition enhancement and negative prediction error more prominent in awake animals.

Adapting and facilitating units are differentially affected by cortico-collicular inactivation

We next performed recordings in IC of awake mice to determine how neuronal mismatch and its component repetition and prediction metrics were affected by cortico-collicular inactivation (Figure 3A). To inactivate cortico-collicular feedback, we shined light over AC in subjects that expressed a suppressive opsin in cortico-collicular neurons. We segregated the population of recorded units according to those that exhibited a significantly stronger response to the deviant than the standard (adapting units; Figure 3B, blue; Figure 5C), those that exhibited a significantly stronger response to the standard than the deviant (facilitating units; Figure 3B, red; Figure 5F), and those that responded equally to both stimulus contexts (nonadapting units; Figure 3B, green) for recordings in both central and shell regions of IC (Figure 3B, left vs. right).

Figure 3 with 3 supplements see all
Adapting and facilitating inferior colliculus (IC) units are differentially affected by cortico-collicular inactivation.

(A) Experimental design for recording in awake IC during laser inactivation of the cortico-collicular pathway. (B) Categorization of units according to whether they displayed significant adaptation, facilitation, or neither (nonadapting). (C) Average peristimulus time histogram for adapting units in central (top) and shell (bottom) IC. Green = during laser inactivation. (D) Index of neuronal mismatch (iMM) (top), index of prediction error (iPE) (middle), and index of repetition suppression (iRS) (bottom) for adapting units in the central nucleus. Dots represent recorded units. Bar plots represent means over the population of n = 52 units. Error bars are standard error of the mean. (E) iMM (top), iPE (middle), and iRS (bottom) for adapting units in shell regions of IC. Dots represent recorded units. Bar plots represent means over the population of n = 113 units. Error bars are standard error of the mean. (F) Average peristimulus time histogram for facilitating units in central (top) and shell (bottom) IC. Green = during laser inactivation. (G) iMM (top), iPE (middle), and iRS (bottom) for facilitating units in the central nucleus. Dots represent recorded units. Bar plots represent means over the population of n = 14 units. Error bars are standard error of the mean. (H) iMM (top), iPE (middle), and iRS (bottom) for facilitating units in shell regions of IC. Dots represent recorded units. Bar plots represent means over the population of n = 38 units. Error bars are standard error of the mean.

The iMM for adapting units in the central nucleus significantly decreased with laser inactivation of cortico-collicular neurons (Figure 3D, top; Table 1; p=0.00034, Wilcoxon signed-rank test). The iMM at baseline for adapting units predominantly represents repetition suppression (Figure 3D, bottom) and a small amount of prediction error (Figure 3D, middle). Prediction error was abolished during laser inactivation (Figure 3D, middle; Table 1; p=0.048, Wilcoxon signed-rank test), while repetition suppression remained unaffected (Figure 3D, bottom). Adapting units in shell regions of IC exhibited a similar pattern to those in the central nucleus. At baseline, these units encoded both prediction error and repetition suppression (Figure 3E, middle and bottom). A significant decrease in iMM during laser inactivation (Figure 3E, top; Table 1; p=0.0023, Wilcoxon signed-rank test) was driven by a decrease in prediction error (Figure 3E, middle; Table 1; p=0.034, Wilcoxon signed-rank test), whereas repetition suppression remained unaffected (Figure 3E, bottom). Combined, these results suggest that removing cortical feedback reduced prediction error but not repetition suppression in adapting units.

Prior studies of deviance detection in IC have focused exclusively on adapting units. However, given the relative prevalence of facilitating units discovered in the awake versus anesthetized IC (Figure 2), we further investigated this population of units to determine whether facilitation reflects prediction or repetition effects. In the central nucleus, cortico-collicular inactivation led to a significant decrease in facilitation in facilitating units (Figure 3G, top; Table 1; p=0.0036, Student’s t-test). At baseline, the iMM for facilitating units represents a combination of negative prediction error and repetition enhancement (Figure 3G, middle and bottom). During inactivation, negative prediction error remained unaffected (Figure 3G, middle), while repetition enhancement was nearly abolished (Figure 3G, bottom; Table 1; p=0.0026, Student’s t-test). Facilitating units in the shell IC were also significantly affected by cortico-collicular inactivation (Figure 3H, top; Table 1; p=0.0016, Wilcoxon signed-rank test). In this case, however, the change in iMM was driven by the near abolishment of negative prediction error (Figure 3H, middle; Table 1; p=0.037, Wilcoxon signed-rank test), while repetition enhancement was unaffected (Figure 3H, bottom).

These data suggest that adaptation and facilitation in the awake IC are composed of distinct underlying processes: adapting populations in both central and shell regions of IC exhibit prediction error and repetition suppression, while facilitating populations are characterized by negative prediction error and repetition enhancement. In adapting units in both central and shell regions, cortico-collicular inactivation significantly decreases prediction error. Facilitating units in the central IC display decreased repetition enhancement with cortico-collicular inactivation, while those in shell regions exhibit decreased negative prediction error. To ensure that the laser-induced changes described above were opsin-mediated, we performed control experiments in two mice with identical manipulations to the experimental group, but in the absence of ArchT (Figure 3—figure supplement 1A). At baseline, the control group exhibited a similar distribution of iMM values to the experimental group in both the central and shell regions of IC (Figure 3—figure supplement 1B, Table 2). Similar proportions of adapting/facilitating/nonadapting units were also found in the control (central: 23% adapting, 5% facilitating, 71% nonadapting; shell: 29% adapting, 18% facilitating, 53% nonadapting) and experimental groups (central: 24% adapting, 6% facilitating, 70% nonadapting; shell: 29% adapting, 9% facilitating, 62% nonadapting). We found no significant differences between baseline and laser trials for either adapting (Figure 3—figure supplement 1C and D, Table 2) or facilitating (Figure 3—figure supplement 1E and F) units in either region. This experiment confirmed that the observed effects of cortico-collicular inactivation were indeed due to opsin-mediated inactivation of the cortico-collicular projection neurons.

Table 2
Statistical comparisons for control data.
ComparisonFigureMeanMedianSDSEMCI (±)TestTest statisticNdfpEffect size
iMM central (control vs. experimental)Figure 3—figure supplement 1B (left)Con: 0.092Exp: 0.057Con: 0.086Exp: 0.064Con: 0.16Exp: 0.18Con: 0.011Exp: 0.012Con: 0.022Exp: 0.024Wilcoxon rank-sum testW = 791977 (control)221 (exp.)NA0.370.052
iMM shell (control vs. experimental)Figure 3—figure supplement 1B (right)Con: 0.083Exp: 0.073Con: 0.069Exp: 0.053Con: 0.23Exp: 0.24Con: 0.012Exp: 0.012Con: 0.023Exp: 0.024Wilcoxon rank-sum testW = 22,364119 (control)394 (exp.)NA0.450.034
iMM central adapting (laser OFF vs. ON)Figure 3—figure supplement 1C (top)OFF: 0.35ON: 0.33OFF: 0.35ON: 0.32OFF: 0.11ON: 0.15OFF: 0.026ON: 0.034OFF: 0.054ON: 0.072Wilcoxon signed-rank testV = 12418NA0.0990.40
iPE central adapting (laser OFF vs. ON)Figure 3—figure supplement 1C (middle)OFF: 0.16ON: 0.19OFF: 0.10ON: 0.081OFF: 0.39ON: 0.40OFF: 0.091ON: 0.094OFF: 0.19ON: 0.20Paired t-testt = –1.118170.300.25
iRS central adapting (laser OFF vs. ON)Figure 3—figure supplement 1C (bottom)OFF: 0.19ON: 0.14OFF: 0.24ON: 0.14OFF: 0.38ON: 0.37OFF: 0.090ON: 0.087OFF: 0.19ON: 0.18Paired t-testt = 1.918170.0770.44
iMM shell adapting (laser OFF vs. ON)Figure 3—figure supplement 1D (top)OFF: 0.38ON: 0.38OFF: 0.35ON: 0.38OFF: 0.19ON: 0.22OFF: 0.032ON: 0.037OFF: 0.065ON: 0.075Paired t-testt = –0.001335340.990.00022
iPE shell adapting (laser OFF vs. ON)Figure 3—figure supplement 1D (middle)OFF: 0.16ON: 0.14OFF: 0.12ON: 0.15OFF: 0.24ON: 0.26OFF: 0.041ON: 0.044OFF: 0.083ON: 0.090Paired t-testt = 0.5835340.560.099
iRS shell adapting (laser OFF vs. ON)Figure 3—figure supplement 1D (bottom)OFF: 0.22ON: 0.24OFF: 0.24ON: 0.20OFF: 0.23ON: 0.22OFF: 0.040ON: 0.038OFF: 0.081ON: 0.077Paired t-testt = –0.7835340.440.13
iMM central facilitating (laser OFF vs. ON)Figure 3—figure supplement 1E (top)OFF: –0.37ON: –0.33OFF: –0.36ON: –0.37OFF: 0.15ON: 0.18OFF: 0.077ON: 0.090OFF: 0.25ON: 0.29Paired t-testt = –1.1430.340.57
iPE central facilitating (laser OFF vs. ON)Figure 3—figure supplement 1E (middle)OFF: –0.043ON: 0.030OFF: –0.0047ON: 0.077OFF: 0.47ON: 0.45OFF: 0.24ON: 0.22OFF: 0.75ON: 0.71Paired t-testt = –0.93430.420.47
iRS central facilitating (laser OFF vs. ON)Figure 3—figure supplement 1E (bottom)OFF: –0.33ON: –0.36OFF: –0.49ON: –0.53OFF: 0.55ON: 0.60OFF: 0.27ON: 0.30OFF: 0.87ON: 0.95Paired t-testt = 0.49430.660.24
iMM shell facilitating (laser OFF vs. ON)Figure 3—figure supplement 1F (top)OFF: –0.38ON: –0.31OFF: –0.32ON: –0.30OFF: 0.22ON: 0.20OFF: 0.048ON: 0.043OFF: 0.10ON: 0.090Wilcoxon signed-rank testV = 6321NA0.0700.40
iPE shell facilitating (laser OFF vs. ON)Figure 3—figure supplement 1F (middle)OFF: –0.090ON: –0.094OFF: –0.11ON: –0.081OFF: 0.18ON: 0.20OFF: 0.040ON: 0.044OFF: 0.083ON: 0.093Wilcoxon signed-rank testV = 10921NA0.840.050
iRS shell facilitating (laser OFF vs. ON)Figure 3—figure supplement 1F (bottom)OFF: –0.29ON: –0.21OFF: –0.28ON: –0.15OFF: 0.24ON: 0.21OFF: 0.053ON: 0.047OFF: 0.11ON: 0.097Paired t-testt = –1.821200.0910.39
iMM central nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1G (top)OFF: 0.021ON: 0.060OFF: 0.014ON: 0.050OFF: 0.24ON: 0.23OFF: 0.032ON: 0.031OFF: 0.064ON: 0.063Paired t-testt = –1.855540.0750.24
iPE central nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1G (middle)OFF: 0.12ON: 0.14OFF: 0.034ON: 0.092OFF: 0.34ON: 0.35OFF: 0.046ON: 0.047OFF: 0.092ON: 0.095Paired t-testt = –1.255540.230.16
iRS central nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1G (bottom)OFF: –0.095ON: –0.083OFF: –0.064ON: –0.072OFF: 0.31ON: 0.29OFF: 0.042ON: 0.038OFF: 0.084ON: 0.077Paired t-testt = –0.5755540.570.077
iMM shell nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1H (top)OFF: 0.063ON: 0.051OFF: 0.040ON: 0.031OFF: 0.16ON: 0.22OFF: 0.021ON: 0.027OFF: 0.042ON: 0.054Wilcoxon signed-rank testV = 113363NA0.390.11
iPE shell nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1H (middle)OFF: 0.053ON: 0.027OFF: 0.0ON: 0.0OFF: 0.25ON: 0.26OFF: 0.031ON: 0.032OFF: 0.063ON: 0.065Paired t-testt = 0.8863620.380.11
iRS shell nonadapting (laser OFF vs. ON)Figure 3—figure supplement 1H (bottom)OFF: 0.011ON: 0.024OFF: 0.028ON: 0.041OFF: 0.27ON: 0.28OFF: 0.034ON: 0.035OFF: 0.068ON: 0.071Paired t-testt = –0.4363620.670.054
iRS > 0 central nonadapting (laser OFF vs. ON)N/AOFF: 0.21ON: 0.18OFF: 0.20ON: 0.16OFF: 0.12ON: 0.16OFF: 0.026ON: 0.034OFF: 0.054ON: 0.070Paired t-testt = 1.522210.160.31
iRS < 0 central nonadapting (laser OFF vs. ON)N/AOFF: –0.31ON: –0.26OFF: –0.27ON: –0.27OFF: 0.21ON: 0.21OFF: 0.036ON: 0.037OFF: 0.074ON: 0.075Paired t-testt = –1.732310.0990.30
  1. iRS: index of repetition suppression; iPE: index of prediction error; iMM: index of neuronal mismatch.

Adapting and facilitating units respond similarly to the cascade and many standards controls

Though the cascade sequence is free of repetition effects between adjacent tone pairs, it does exhibit global repetition across the entire tone sequence. To assess whether global stimulus regularity affects the response to the cascade context, we used a shuffled version of the cascade sequence, known as the ‘many standards’ sequence, as an additional control stimulus (Figure 3—figure supplement 2A). The many standards sequence contains the same 10 tones as the cascade but presented in random order (Figure 3—figure supplement 2A). This reduces the potential for adaptation across adjacent frequency channels and also eliminates the global predictability of the stimulus, both of which could lead to suppression of responses to tones in the cascade context and potentially affect the calculations of iMM, iPE, and iRS. We compared the responses of adapting and facilitating units in both central and shell regions of IC to tones in the cascade versus the many standards context (Figure 3—figure supplement 2A). We found no significant differences in firing rates to the cascade versus the many standards contexts (Figure 3—figure supplement 2B and C, Table 1), suggesting that the global structure of the cascade sequence does not significantly affect how units in IC respond to this stimulus, as has been shown in other structures (Casado-Román et al., 2020; Parras et al., 2021).

iMM distribution does not differ between single- and multiunit types

The analysis of changes in predictive coding metrics is performed on pooled single- and multiunit responses of IC units. To determine whether the expression of neuronal mismatch differs between these unit types, we plotted the iMM for laser OFF and ON conditions for each of the subgroups in the central and shell regions of the IC separated by single- (displayed in teal) and multiunits (Figure 3—figure supplement 3). We observed no differences in the distributions of these unit types in central or shell IC (Table 1; central OFF: p=0.88, Wilcoxon rank-sum test; central ON: p=0.93, Student’s t-test; shell OFF: p=0.19, Wilcoxon rank-sum test; shell ON: p=0.21, Wilcoxon rank-sum test). We therefore combined data from both single- and multiunits for the analyses of predictive coding metrics.

Nonadapting units also display top-down repetition enhancement

The majority of units in both central and shell IC do not exhibit either adaptation or facilitation but respond similarly to tones when they are presented as a standard or deviant (Figure 4A). However, since both negative and positive metrics are included in the calculation of iMM, it is still possible that these units exhibit predictive processing that may not be reflected in the overall iMM value. We further characterized these nonadapting units (Figure 4B) and tested how they are affected by cortico-collicular inactivation. Nonadapting units in the central nucleus exhibited a significant increase in iMM during inactivation (Figure 4C, top; Table 1; p=2.7e-06, Wilcoxon signed-rank test), whereas those in the shell IC were unaffected (Figure 4D, top). The change in iMM for nonadapting units in the central nucleus was driven by a significant increase in iRS (Figure 4C, bottom middle; Table 1; p=0.0011, Wilcoxon signed-rank test). To determine whether this reflected a change in repetition suppression or enhancement, we further segregated central nonadapting units according to whether their baseline iRS values were negative or positive (Figure 4C, bottom). Only those units with negative baseline iRS values (i.e., those units showing repetition enhancement) were significantly affected by cortico-collicular inactivation (Figure 4C, bottom; Table 1; p=0.00012, Wilcoxon signed-rank test). In control experiments without ArchT, no significant changes were observed in nonadapting units (Figure 3—figure supplement 1G and H, Table 2). These results indicate that, similar to central facilitating units, central nonadapting units display repetition enhancement, and that input from the cortex is critical for expression of this phenomenon.

Nonadapting units also display top-down repetition enhancement.

(A) Distribution of adapting types (adapting, facilitating, and nonadapting) for units in central (left) and shell (right) regions of the inferior colliculus (IC). (B) Average peristimulus time histogram for nonadapting units in central (top) and shell (bottom) IC. (C) Index of neuronal mismatch (iMM) (top), index of prediction error (iPE) (middle), and index of repetition suppression (iRS) (bottom) for nonadapting units in central regions of IC. Dots represent recorded units. Bar plots represent means over the population of n = 155 units. Error bars are standard error of the mean. (D) iMM (top), iPE (middle), and iRS (bottom) for nonadapting units in shell regions of IC. Dots represent recorded units. Bar plots represent means over the population of n = 243 units. Error bars are standard error of the mean.

Standard and deviant responses are bidirectionally modulated by cortico-collicular inactivation

The observed changes in repetition metrics with cortico-collicular inactivation could reflect an effect on either the standard or cascade context. Similarly, the shift in prediction metrics observed with inactivation could be due to altered responses to either the cascade or deviant contexts. We next determined whether the laser-induced changes in the iMM, iPE, and iRS for adapting units reflect changes in the firing rates to the standard, deviant, or cascade contexts. We found that adapting units in the central nucleus increased responses to the standard (Figure 5A, Table 1; p=0.0092, one-sample t-test) and decreased responses to the deviant (Figure 5A, Table 1; p=0.0054, one-sample t-test) during inactivation. These results explain the decrease in iMM for this population during the laser stimulus (Figure 3D, top): the firing rate to the cascade stimulus did not change during cortico-collicular inactivation, which means that the decrease in firing rate to the deviant alone underlies the decrease in prediction error observed for this population (Figure 3D, middle). Adapting units in the shell exhibited the same pattern of bidirectional changes to the standard (Figure 5B, Table 1; p=0.035, one-sample Wilcoxon test) and deviant (Figure 5B, Table 1; p=0.0057, one-sample Wilcoxon test), similarly accounting for their decrease in iMM and prediction error (Figure 3E), with no change in response to the cascade condition (Figure 5B). These data suggest that inactivation of the cortico-collicular pathway induces bidirectional changes in firing rates to the standard and deviant for adapting units in both central and shell regions of IC.

Standard and deviant responses are bidirectionally modulated by cortico-collicular inactivation.

(A) Responses to the standard (left), cascade (middle left), and deviant (middle right) for adapting units in central regions of the inferior colliculus (IC) under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 52 units. Error bars are standard error of the mean. (B) Responses to the standard (left), cascade (middle left), and deviant (middle right) for adapting units in shell regions of IC under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 113 units. Error bars are standard error of the mean. (C) Responses to the standard (left), cascade (middle left), and deviant (middle right) for facilitating units in central regions of IC under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 14 units. Error bars are standard error of the mean. (D) Responses to the standard (left), cascade (middle left), and deviant (middle right) for facilitating units in shell regions of IC under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 38 units. Error bars are standard error of the mean. (E) Responses to the standard (left), cascade (middle left), and deviant (middle right) for nonadapting units in central regions of IC under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 155 units. Error bars are standard error of the mean. (F) Responses to the standard (left), cascade (middle left), and deviant (middle right) for nonadapting units in shell regions of IC under baseline and laser conditions. Change in firing rate between the laser and baseline condition for each stimulus (right). Dots represent recorded units. Bar plots represent means over the population of n = 243 units. Error bars are standard error of the mean.

We also investigated how responses to each stimulus context changed with cortico-collicular inactivation for facilitating units. For central facilitating units, only the firing rate to the standard context changed during inactivation (Figure 5C, Table 1; p=0.0013, one-sample t-test), explaining the observed change in repetition enhancement for this population (Figure 3G). For shell facilitating units, a decreased response to the standard (Figure 5D, Table 1; p=0.0042, one-sample t-test) and an increased response to the deviant (Figure 5D, Table 1; p=0.0013, one-sample t-test) were elicited on laser trials, accounting for changes in the iMM and the abolishment of negative prediction error (Figure 3H). These changes are directionally opposite to the observed firing rate changes observed for adapting units under inactivation, with a decrease to the standard context for both central and shell units and an increase to the deviant context for shell units.

For nonadapting units, a significant decrease in response to the standard context was observed in both central (Figure 5E, Table 1; p=1.4e-06, one-sample Wilcoxon test) and shell (Figure 5F, Table 1; p=0.035, one-sample Wilcoxon test) regions of IC. The decrease was only significant enough to produce an effect on the iMM in central regions (Figure 4C, top), leading to an increase in repetition suppression (Figure 4C, bottom).

For adapting and facilitating units, these data exhibit that IC responses to the standard and deviant contexts in the absence of cortical input are bidirectionally modulated, such that neurons respond more similarly to both contexts rather than firing differentially to each. For nonadapting units, the response to the standard context alone is diminished during cortico-collicular inactivation, causing these units to become more adapting. These changes suggest that under normal conditions AC provides information regarding sound context to neurons in IC.

IC units have distinct combinations of iPE and iRS

To determine whether IC units exhibit particular combinations of repetition suppression/enhancement and prediction error/negative prediction error, we plotted the iPE values against the iRS values for each unit in the adapting, facilitating, and nonadapting groups. Both the adapting and nonadapting groups in the central IC contained units with significant values for both iPE and iRS, most often resulting from a combination of negative prediction error and repetition suppression (Figure 6A, maroon dots). In the shell IC, a greater variety of response combinations was observed. All three groups contained units with both significant negative prediction error and repetition suppression, as well as a separate population exhibiting significant prediction error and repetition enhancement (Figure 6B, maroon dots). Some shell adapting units also exhibited a combination of both repetition suppression and prediction error (Figure 6B, left). These results suggest that the units in IC exhibit distinct combinations of repetition suppression/enhancement and prediction error/negative prediction error.

Inferior colliculus (IC) units exhibit distinct combinations of index of prediction error (iPE) and index of repetition suppression (iRS).

(A) Distribution of both iRS and iPE in adapting (left), facilitating (middle), and nonadapting (right) units in central IC. (B) Plots of distributions of both iRS and iPE in adapting (left), facilitating (middle), and nonadapting (right) units in shell IC. (C) Response to three subsequent standards prior to or following the deviant in facilitating units in central IC. Comparison between the last standard before and the first standard after the deviant demonstrates significant repetition enhancement. Bar plots represent means over the population of n = 14 units. Error bars are standard error of the mean. (D) Response to three subsequent standards prior to or following the deviant in facilitating units in shell IC. Comparison between the last standard before and the first standard after the deviant demonstrates significant repetition enhancement. Bar plots represent means over the population of n = 38 units. Error bars are standard error of the mean.

Facilitating units exhibit true repetition enhancement

Facilitating units in both central and shell regions of IC exhibited repetition enhancement at baseline, as defined by the difference in firing rate to the last standard and the same tone embedded in the cascade sequence (Figure 3G and H). We sought to further characterize the response to the standard context to determine whether the repetition enhancement captured by the iRS indicates true repetition enhancement (an incremental increase in firing rate on subsequent presentations of the standard) or simply a net increase in firing rate to the standard versus cascade condition. We calculated the mean firing rate for each of the three standards before the deviant and each of the three standards after the deviant (Figure 6C and D). The progression of standards by position exhibited subsequent enhancements in firing rate that was plateaued by the second to last standard before the deviant for both central (Figure 6C) and shell facilitating units (Figure 6D). The firing rate to the last standard was significantly higher than the first in both regions (Figure 6C, Table 1; p=0.0017, Wilcoxon signed-rank test; Figure 6D, Table 1; p=9.3e-05, Wilcoxon signed-rank test). These data provide evidence that facilitating units in IC exhibit true repetition enhancement.

Discussion

Summary of findings

The results of this study indicate that AC is critically involved in regulating both repetition and prediction effects in the awake IC, providing evidence for the implementation of predictive coding in cortico-subcortical networks. Adapting and facilitating units were bidirectionally modulated by cortico-collicular inactivation, with adapting units becoming less adapting and facilitating units becoming less facilitating on laser trials (Figure 3). The decrease in adaptation for adapting units was driven by a decrease in prediction error for units in both central and shell regions of IC ( Figure 3D, Figure 5E, Figure 7, pink arrows). For facilitating and nonadapting units in the central nucleus, inactivation-driven changes were caused by a decrease in repetition enhancement (Figure 3G, Figure 7, gold dashed arrows). The decrease in facilitation in the shell IC, however, was caused by the abolishment of negative prediction error (Figure 3H, Figure 7, pink dashed arrows).

Corticofugal regulation of predictive coding.

Laser inactivation led to the abolishment of repetition enhancement in central facilitating units and the abolishment of negative prediction error in shell facilitating units. Prediction error decreased during inactivation for adapting units in both shell and central regions of the inferior colliculus (IC). Repetition suppression remained unaffected during cortical inactivation, suggesting that it may reflect fatigue of bottom-up sensory inputs.

In adapting units, these changes were modulated by an increased response to the standard and decreased response to the deviant, while the opposite pattern was true for facilitating units (Figure 5). Overall, these bidirectional changes indicate that, without input from AC, IC responds more similarly to tones in the standard and deviant contexts. These findings demonstrate that AC provides critical contextual cues about the statistics of the auditory environment to targets in IC under normal conditions. We further discuss these results in the context of a hierarchical predictive coding framework below.

iMM in the awake versus anesthetized IC

Our results include the first investigation of how the repetition and prediction processes that underlie deviance detection in the awake IC compare to the anesthetized condition. Our data suggest that while iMM values are higher under anesthesia, they almost entirely reflect repetition suppression, with only a small contribution of prediction error (Figure 2). In the central IC, modest prediction error is present under anesthesia, but negative prediction error becomes dominant when the animal is awake. In the shell IC, the same units exhibit drastically different iPE and iRS values for the awake versus the anesthetized condition. Prediction error is substantially higher in the awake IC, and repetition enhancement, rather than repetition suppression, is observed (Figures 2F and 4G). These findings suggest that the iMM values in the awake and anesthetized brain reflect different underlying processes, and that anesthesia induces bidirectional changes in metrics of repetition and prediction.

Facilitating units in IC

We also provide here the first analysis of facilitating units in IC. Previous studies that have investigated iMM have focused selectively on the positive side of the iMM distribution since these units display adaptation. However, facilitation seems to be enriched in the awake IC (Figures 2B and 4E) and reflects other potentially interesting parameters, such as repetition enhancement (represented as a higher response to the standard than the cascade sequence) (Figure 2G) and negative prediction error (represented as a higher response to the cascade than the deviant) (Figure 2C).

Repetition enhancement and repetition suppression in IC

Because previous studies that have applied a predictive coding framework to decompose neuronal mismatch have focused exclusively on adapting neurons, the repetition enhancement found here in facilitating units has not been previously described (Parras et al., 2017). However, it is well-documented in fMRI literature that repetition enhancement is a common phenomenon in humans, existing either alongside or in place of repetition suppression (de Gardelle et al., 2013; Müller et al., 2013; Segaert et al., 2013). Interestingly, repetition enhancement has been proposed to reflect novel network formation and consolidation of novel sensory representations (Segaert et al., 2013). Once new representations have been formed, repetition suppression is hypothesized to take over, reflecting the minimization in prediction errors that occurs when new representations give rise to accurate predictions (Auksztulewicz and Friston, 2016; de Gardelle et al., 2013; Friston and Kiebel, 2009). Though the repetition enhancement described in human studies differs drastically on spatial and temporal scales from the phenomenon described here, we find that it similarly involves a sequential enhancement in the response to subsequent presentations of the standard (Figure 6C and D). Repetition enhancement has also been observed in the MGB in response to temporally degraded stimuli that are hypothesized to engage top-down resources to compensate for bottom-up acoustic information loss (Cai et al., 2016; Kommajosyula et al., 2019). Interestingly, this enhancement is reversed when cortico-thalamic pathways are blocked, further suggesting that repetition enhancement in the auditory system reflects a top-down phenomenon (Kommajosyula et al., 2021).

While repetition suppression can be understood from a predictive coding framework, it can also be viewed from the perspective of neuronal fatigue, whereby the incremental decrease in firing rate to a repeated standard tone is simply explained by synaptic depression (Escera and Malmierca, 2014; Taaseh et al., 2011). Interestingly, we did not find any effect on repetition suppression during cortico-collicular inactivation, suggesting that it may reflect fatigue of bottom-up sensory inputs rather than an active predictive process (Figures 3D and 5E, Figure 7, gold arrows). While these data do not provide definitive proof of either perspective, they do suggest that the processes that underlie repetition suppression in IC do not involve top-down cortical signals. This notion is supported by the fact that repetition suppression was much more prevalent when animals were under anesthesia, a state in which the auditory responsiveness in the cortex is compromised (Figure 2G; Brugge and Merzenich, 1973; Katsuki et al., 1959).

Prediction error in IC

In both central and shell populations that exhibited prediction error at baseline, cortico-collicular inactivation led to a decrease, or complete abolishment, of prediction error (Figures 3D and 5E). According to models of hierarchical predictive coding, higher-order stations generate predictions that they broadcast to lower centers (Friston and Kiebel, 2009). These predictions are compared with representations of the actual sensory input, and if there is a mismatch, a prediction error is generated and forwarded up the hierarchy (Friston and Kiebel, 2009). Under this framework, the inactivation of top-down inputs would interfere with communication of predictions, leading to dysfunction in the prediction error response, as seen in our data. Another possibility is that prediction errors are directly backpropagated from AC to IC. While this contradicts canonical predictive coding models, evidence for prediction error has been found in deep layers of the cortex in which feedback neurons reside (Asilador and Llano, 2020; Rummell et al., 2016). Though the precise mechanism underlying the generation of prediction error in IC remains unclear, our data show that feedback from AC plays a critical role in this process.

Negative prediction error in IC

In addition to units with prediction error, we found that units in IC that responded more strongly to the cascade than the deviant context (Figure 3G and H), consistent with previous reports (Parras et al., 2017). A stronger response to a tone in the cascade sequence compared to the context in which it is a deviant could simply reflect a relative lack of cross-frequency adaptation; the oddball stimulus consists of repeated tone presentations of two neighboring frequencies, making it more likely to generate cross-frequency effects than the cascade stimulus, which cycles through repetitions of 10 evenly spaced frequencies (Parras et al., 2017; Taaseh et al., 2011). Previous studies that have investigated the effective bandwidth for cross-frequency adaptation, however, have found that it occurs between channels with a frequency separation of a third of an octave or less (Taaseh et al., 2011). The stimuli used in this study had a half-octave frequency separation, indicating that cross-frequency effects should be minimized. Therefore, it is unlikely that the negative prediction error responses observed in this study simply reflect cross-frequency adaptation to the oddball stimulus.

A stronger response to a tone when it is embedded in a completely predictable sequence, such as the cascade sequence, than when it is a deviant could also signify that a neuron encodes predictions, rather than prediction errors. In hierarchical predictive coding, both predictions and prediction errors are generated at every level of the hierarchy, with prediction errors being forwarded to ascending sensory centers and predictions being backpropagated (Friston and Kiebel, 2009). In the shell IC, the region that receives the vast majority of descending cortical input, evidence for negative prediction error was abolished during cortico-collicular inactivation (Figure 3H), consistent with the notion that feedback from the cortex may carry predictions to IC (Bajo et al., 2007; Herbert et al., 1991; Saldaña et al., 1996; Stebbings et al., 2014). Interestingly, negative prediction error in the central nucleus remained unperturbed during inactivation of cortical feedback (Figure 3G). Given that only a small fraction of cortico-collicular fibers terminate in the central nucleus, it is likely that it receives predictions from another source (Bajo et al., 2007; Herbert et al., 1991; Saldaña et al., 1996; Stebbings et al., 2014). An intriguing potential candidate for this source of predictions could be the shell IC, given the extensive network of intracollicular connections in IC (Lesicko et al., 2020; Saldaña and Merchán, 1992; Saldaña and Merchán, 2005). Future studies will be required to determine whether the negative prediction error metric described here captures the type of top-down predictions described in canonical predictive coding models.

Technical considerations

One limitation of this study is that laser inactivation achieved only partial and not complete inactivation of the cortico-collicular pathway. Given that light itself can have a modulatory or toxic effect on neurons, these types of optogenetic experiments require a careful titration between using enough power to substantially affect the population of interest without causing nonspecific light or heat-based perturbations (Tyssowski and Gray, 2019). Though other techniques, such as chemogenetic approaches or cooling, provide more complete inactivation, they do not allow for rapid and reversible inactivation (English and Roth, 2015). With our laser power parameters, we found a mean 60% reduction in firing in putative cortico-collicular neurons at baseline and a 45% reduction during presentation of pure tone stimuli (Figure 1—figure supplement 1D). This reduction produced clear effects on repetition and prediction processing in IC, in several cases with the severe reduction or complete abolishment of certain metrics of deviance detection, such as prediction error and repetition enhancement in the central nucleus and negative prediction error in the shell IC (Figure 3). The interpretation of these results should bear in mind that they reflect only partial and not complete inactivation.

The analyses in this study were performed on pooled single- and multiunit data. Although we observed no differences in the iMM distribution between single- and multiunits (Figure 3—figure supplement 3), the results of this study should be interpreted with this limitation in mind, namely, photosuppression-induced changes in these units may not reflect changes in single neurons.

Whereas this study focuses on changes specific to the cortico-collicular pathway, it should be noted that cortico-collicular neurons are known to branch to additional subcortical targets besides the IC, including the MGB, caudal regions of the dorsal striatum, and the lateral amygdala (Asokan et al., 2018). The fact that our photo-suppression experiments produce short-latency effects in the IC (Figure 3C and F) indicates that the observed changes are likely due to direct, monosynaptic AC to IC pathways, and that multisynaptic effects from other collateral sites are unlikely. Nevertheless, the potential contribution from these additional downstream targets cannot be definitely ruled out and should be factored into the interpretation of the results.

Conclusions

Our findings indicate that deviance detection and predictive coding in IC involve additional complexity than has been previously described. We provide here the first description of facilitating units in IC, as well as evidence for the existence of repetition enhancement and negative prediction error in these units. We show that AC regulates these metrics and is also involved in the generation of prediction error in IC. Repetition suppression is unaffected by inactivation of cortical input to IC, providing evidence that this process may reflect bottom-up fatigue rather than top-down predictive processing. These results demonstrate the role of AC in providing contextual cues about the auditory stream to targets in IC.

Materials and methods

Key resources table
Reagent type (species) or resourceDesignationSource or referenceIdentifiersAdditional information
Strain, strain background (Mus musculus)Cdh23 miceJackson LaboratoriesCdh23tm2.1Kjn/J;RRID:IMSR_JAX:018399
Recombinant DNA reagentAAV9-CAG-FLEX-ArchT-tdTomatoUNC Vector CoreAddgene_28305
Recombinant DNA reagentRetroAAV2 hSyn Cre-GFPIn-houseVector generated and maintained in the di Biasi lab
Software, algorithmKilosort2Marius Pachitariuhttps://github.com/MouseLand/Kilosort; RRID:SCR_016422
Software, algorithmMATLABMathWorkshttps://www.mathworks.com/; RRID:SCR_001622
Software, algorithmImageJNIHRRID:SCR_003070

Animals

We performed experiments in six adult Cdh23 mice (Cdh23tm2.1Kjn/J, RRID:IMSR_JAX:018399; four males and two females, age 3–8 months). This mouse line has a targeted point reversion in the Cdh23 gene that protects against the age-related hearing loss common to C57BL/6 strains (Johnson et al., 2017). Animals were housed on a reversed 12 hr light–dark cycle with water and food available ad libitum. All procedures were approved by the University of Pennsylvania IACUC (protocol number 803266) and the AALAC Guide on Animal Research. We made every attempt to minimize the number of animals used and reduce pain or discomfort.

Virus injection

Request a detailed protocol

Mice were continuously anesthetized with isoflurane and mounted in a stereotaxic frame. Buprenex (0.1 mg/kg), meloxicam (5 mg/kg), and bupivicane (2 mg/kg) were injected subcutaneously for preoperative analgesia. We performed small craniotomies bilaterally over AC (−2.6 mm caudal to bregma, ±4.3 mm lateral, +1 mm ventral) and IC (−4.96 mm caudal to bregma, ±0.5 mm lateral, +0.5 mm ventral and −4.96 mm caudal to bregma, ±1.25 mm lateral, +1.0 mm ventral). A glass syringe (30–50 μm diameter) connected to a pump (Pump 11 Elite, Harvard Apparatus) was used to inject modified viral vectors (AAV9-CAG-FLEX-ArchT-tdTomato or AAV9-CAG-FLEX-tdTomato; 750 nL/site; UNC Vector Core) into AC and a retroAAV construct (retro AAV-hSyn-Cre-GFP; 250 nL/site) into IC (Figures 1A and 2A, Figure 3—figure supplement 1A). Large viral injections were performed to broadly target cortico-collicular neurons throughout all regions of the AC. We implanted fiber-optic cannulas (Thorlabs, Ø200 μm Core, 0.22 NA) bilaterally over AC injection sites (0.4 mm ventral to brain surface) and secured them in place with dental cement (C and B Metabond) and acrylic (Lang Dental). IC injection sites were covered with a removable silicone plug (Kwik-Sil). A custom-built headplate was secured to the skull at the midline and a ground-pin was lowered into a small craniotomy over bregma. We injected an antibiotic (5 mg/kg Baytril) subcutaneously for 4 days postoperatively. Virus injection sites were confirmed postmortem for all animals included in the study.

Extracellular recordings

Request a detailed protocol

We performed recordings a minimum of 21 days after virus injection surgeries to allow adequate travel time for the viral constructs (Figure 1A). Recordings were carried out inside a double-walled acoustic isolation booth (Industrial Acoustics) or a custom-built table-mounted acoustic isolation booth. For IC recordings, mice were briefly anesthetized to remove the silicone plug over IC virus injection sites. Following recovery from anesthesia, the headplate was clamped within a custom base to provide head-fixation. We lowered a 32-channel silicon probe (Neuronexus) vertically into IC during presentation of broadband noise clicks and monitored sound responses online to confirm localization within IC (Figure 1A). In a subset of animals (seven recording sites in two mice), the probe was first coated in a lipophilic dye (DiD or DiA; Invitrogen) to aid in post hoc reconstruction of recording sites. In each animal, two recordings were performed per IC (four total recording sessions bilaterally). We attempted to target both shell and central IC regions in each animal, and our post hoc analysis of recording sites (see details in ‘Analysis’ section) revealed that all but one animal was recorded from in both regions. Recordings that did not show significant sound responsiveness were removed from the analysis. Following completion of all IC recording sessions, we recorded the activity of neurons in AC using the same procedure (Figure 1—figure supplement 1B). We performed a square craniotomy (2 mm × 2 mm) over AC and oriented the probe vertically to the cortical surface (35° angle of the stereotaxic arm). Electrophysiological data were filtered between 600 and 6000 Hz to isolate spike responses and then digitized at 32 kHz and stored for offline analysis (Neuralynx). For a subset of recordings, the experimental procedures were repeated while recording from the same units after the animal had been anesthetized with isoflurane (Figure 2A). We performed spike sorting using Kilosort2 software (https://github.com/MouseLand/Kilosort; RRID:SCR_016422, version 2). Both single and multiunits were included for all analyses (experimental IC: 50 single units, 354 multiunits; control IC: 17 single units; 111 multiunits; anesthetized: 10 single units, 129 multiunits; AC: 95 single units, 300 multiunits; putative cortico-collicular: 9 single units; 11 multiunits).

Laser inactivation

Request a detailed protocol

We inactivated cortico-collicular neurons using a 532 nm DPSS laser (GL532T3-300, Slocs lasers, 3 mW power at cannula tip or OptoEngine, MGL-III-532, 15 mW power at cannula tip) connected via optical fibers to the implanted cannulas (Figures 1A, 2C and D). Data collected using either laser was pooled together as no significant differences were observed in the strength of inactivation in AC during silence (p=0.054, Wilcoxon rank-sum test) or the presentation of pure tone stimuli (p=0.072, Wilcoxon rank-sum test) between the two lasers. Square laser pulses were timed to coincide with tone onset and lasted for 100 ms. Evidence of inactivation in putative cortico-collicular units (infragranular AC units with a minimum 30% reduction in both baseline and sound-evoked neuronal activity) was confirmed for all animals included in the study.

Stimuli

Request a detailed protocol

We generated an initial frequency response function from a sequence of 50 pure tones, 1–70 kHz, repeated 20 times at 70 dB SPL in pseudo-random order. This response function was generated online to select suitable frequencies for the oddball stimuli, that is, frequencies that would fall into the average response area for units in a given recording. Each tone was 50 ms duration (1 ms cosine squared ramps) with an inter-stimulus interval of 200 ms and presentation rate of 4 Hz. A similar tuning curve stimulus, with eight amplitude levels (35–70 dB, 5 dB increments) and five repetitions, was used to further characterize the tuning properties of each unit (Figure 1—figure supplement 2E, F).

Oddball tone pairs were chosen to fit within the average response area for units from a given recording. Given the prevalence of inhibited regions in the tuning curves, and the fact that this often led to differences in the response profile of the unit to each frequency in the oddball tone pair, the responses to each frequency were analyzed separately (Figure 1—figure supplement 2G). Oddball stimuli consisted of a frozen sequence of two pure tones (with the same tone parameters as those used in the initial frequency response functions) with a 90:10 standard-to-deviant ratio and half-octave frequency separation. The number of standards interleaved between two deviants was counterbalanced and varied between 3 and 17 standards. The stimuli were divided into blocks (with the end of a block defined by the presentation of a deviant), and tone type and laser pairings were alternated on subsequent blocks. For example, on the first block the laser stimulus was paired with the deviant, on the second block it was paired with the last standard, and the corresponding tones in the third block served as baseline controls, with no laser stimulus. The number of preceding standards in the blocks was balanced for all three laser conditions (deviant, last standard, and baseline). Each block type (laser + standard, laser + deviant, no laser) was presented 45 times, and the total number of tones in each sequence was 1250. Two oddball sequences were created, both with the same frozen pattern, but with the frequencies of the standard and the deviant switched.

Cascade sequences consisted of either an ascending or descending set of 10 evenly log-spaced (half-octave separation) pure tones (same tone parameters as described above) (Figure 1C). The two tones used in the oddball sequences were always included as adjacent tones in the cascade sequences, though their position within the cascade was varied. To generate the many standards control sequence, we shuffled the cascade sequences using an algorithm that does not allow for repetition of tones of the same frequency on subsequent presentations.

Analysis

To distinguish between shell and central IC recording locations, we plotted the best frequency for each unit from a given recording against its depth and fit the data with a robust linear regression model (Figure 1—figure supplement 2B). Additionally, we computed the mean sparseness for all units from a given recording site to quantify the sharpness of tuning. The R2 metric from the linear fit and the mean sparseness from each recording were used to perform k-means clustering with two groups. Each recording was assigned to a location (either central or shell) according to the k-means output, with central sites typically having high sparseness and high R2 values and shell sites having low sparseness and low R2 metrics (Figure 1—figure supplement 2C).

Sound response profiles were categorized quantitatively from analysis of the combined responses to the standard and deviant tones using MATLAB’s ‘findpeaks’ function with a minimum peak height set to the mean of the baseline period (50 ms before tone onset) ± 3 SDs. Units that did not display maxima or minima during the tone duration period (0–50 ms) or in the 50 ms after (the ‘offset window’) were labeled as sound unresponsive and were removed from the analysis. Units that showed only a single minimum (‘inhibited’ units) or only a response in the offset window were similarly removed from the analysis. Units that showed at least one maxima during the tone duration period were included in the analysis and further categorized as either onset (single maxima in the first 10 ms after tone onset), sustained (single maximum after the first 10 ms after tone onset), E-I or I-E (units that displayed both a maximum and minimum during the tone duration period), biphasic (units that displayed two maxima during the tone duration period), or mixed (units with greater than two maxima and/or minima during the tone response period). It was common for units to display a response both during the tone duration window and the offset window, and in these cases a combined response profile was assigned (e.g., onset/offset, sustained/inhibited offset). Units with only inhibited or offset responses were removed from the dataset.

Significant adaptation or facilitation for each unit was assessed with a Wilcoxon rank-sum test between the trial-by-trial firing rates to the standard and deviant on the 45 baseline trials. The iMM, identical to the traditional SSA index, was further deconstructed into an iPE and an iRS such that iMM = iPE + iRS. The raw firing rates to the standard, cascade, and deviant conditions were normalized by dividing by the Euclidean norm, N = FRDev2+ FRCasc2+ FRStan2 . The iPE was calculated as the difference in normalized firing rate to the deviant and cascade conditions (iPE = FRDevN - FRCascN), while the iRS was calculated as the difference in normalized firing rate to the cascade and standard conditions (iRS = FRCascN - FRStanN). Predictive coding metrics for the laser condition were calculated similarly, but using trials from laser + standard, laser + cascade, and laser + deviant pairings.

Statistical analysis

Request a detailed protocol

Shapiro–Wilk tests were used to assess normality. For normally distributed data, Student’s t-tests were performed. When the assumption of normality was violated, Wilcoxon rank-sum tests were used for nonpaired data and Wilcoxon signed-rank tests were used for paired data. Cohen’s d was calculated as a measure of effect size for t-tests. For Wilcoxon tests, the effect size r was calculated as the z statistic divided by the square root of the sample size.

Data availability

The data is available for review on the dryad depository, https://doi.org/10.5061/dryad.m905qfv13.

The following data sets were generated

References

    1. Friston K
    2. Kiebel S
    (2009) Predictive coding under the free-energy principle
    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 364:1211–1221.
    https://doi.org/10.1098/rstb.2008.0300
  1. Book
    1. Paxinos G
    2. Franklin KBJ
    (2019)
    Paxinos and Franklin’s the Mouse Brain in Stereotaxic Coordinates
    Academic press.
  2. Book
    1. Saldaña E
    2. Merchán MA
    (2005) Intrinsic and commissural connections of the inferior colliculus
    In: Driscoll ME, editors. In The Inferior Colliculus. Springer. pp. 155–181.
    https://doi.org/10.1007/b138578

Decision letter

  1. Jennifer M Groh
    Reviewing Editor; Duke University, United States
  2. Barbara G Shinn-Cunningham
    Senior Editor; Carnegie Mellon University, United States
  3. Jennifer M Groh
    Reviewer; Duke University, United States

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Decision letter after peer review:

[Editors’ note: the authors submitted for reconsideration following the decision after peer review. What follows is the decision letter after the first round of review.]

Thank you for submitting the paper "Cortico-fugal regulation of predictive coding" for consideration at eLife. Your submission has been reviewed by three peer reviewers, including Jennifer Groh as the Reviewing Editor and Reviewer #3, and the evaluation has been overseen by a Senior Editor. Although the work is of interest, we are not convinced that the findings presented have the potential significance that we require for publication in eLife.

Specifically, after discussion among the reviewers, the most important consensus concerns that emerged were (a) whether the findings are novel in comparison to similar existing studies, including those that involved cooling in auditory cortex and the impact of such cooling on the IC, and (b) the suggestion that some of the reported effects could be due to regression to the mean. This is a potentially addressable problem via the suggestion of Reviewer 2 point 2b. Finally (c) it was noted that the central/shell distinction is critical to the novelty of the findings. Histological confirmation of the assignment of sites to these subdivisions would strengthen the paper. While there were some differences of opinion regarding the clarity of the manuscript, we hope you find even the more critical comments useful.

Reviewer #1:

In this study Lesicko and colleagues have studied the effect of AC inactivation using the optogenetic technique to analyze neuronal mismatch in the interior colliculus of the awake mouse.

The study is interesting and is potentially beneficial for the people working on predictive coding. However, the manuscript is long and could be substantially trimmed and focused to make it more useful. While I was originally excited about the manuscript, my enthusiasm decreased after I read it. I had to read several times to make sense and get a general idea of results and still I am not totally convinced about the data and the presentation. In its present form I cannot recommend acceptance as in my humbly view this manuscript needs to be substantially revised.

As opposed to what the authors claims, and after close inspection of previous studies by Parras´ and by their team, it seems to me much of the basic/general results are similar to Parras and colleagues or even Duque and Malmierca. These authors also studied SSA and iMM in awake mouse. The main results are similar. Low levels of iMM and iPE in the IC. This is a major issue for me. Given the low levels of SSA/iMM, I wonder how authors consider if a neuron shows significant iMM…(bootstrapping??) as AC may have mostly subtle changes on the IC responses.

Another major issue is that authors claim that have recorded separately neurons from the central nucleus and the shell. However, they don´t show any histological probe of the electrode recording. The central nucleus in mouse is very small and although it shows a distinct tonotopic organization and response are areas are usually V shaped, these responses can also occur in the shell. So, my concern is that most of the neurons may be actually recorded in the shell. I would like to draw the attention to the authors that previous studies on subcortical SSA/iMM have demonstrated the lack of adapting responses in the central nucleus. Of course, all previous studies may have missed this, but I dare to suggest that much of the authors central nucleus data may actually be from the shell. In any case, this needs to be unambiguously demonstrated here with some histological data; it is not adequate to rely solely on the frequency response areas, Even if this is shown, I would like to see a convincing conceptual framework to understand it.

Also, the conclusion that the cascade and many-standard controls yield similar results has also been reported in Parras and Casado-Roman recently.

The most original part of this study is the in-depth analysis of the repetition enhancement responses, but I also would like to draw the authors´ attention to the previous study by Parras where they also reported negative iPE (cf. Figures3 rat and Figure 7 awake mouse). This is mentioned on page 29, lines 694-696, but it should be more clearly stated in abstract etc. so that previous studies get a fairer recognition. Also, Duque and Malmierca already have reported that SSA (which reflects iMM) is lower in awake than in anesthetized mouse. This is mostly due to the high rate of spontaneous activity, which is not mentioned in this study.

The most interesting part of the paper is the section related to the effect of cortical deactivation. However, as the authors themselves note, the technique has some limitations. I would like a more detailed discussion on how such limitations may have affected the results and hence the conclusions. Another important issue not addressed is where in AC the injections were made (A1, AAF, A2, etc ???) and how large the injection sites are. These technicalities may have affected the results and should also be considered and different fields may have different projections to IC.

Reviewer #2:

Lesicko et al. studied the role of corticofugal feedback in predictive coding in the auditory system, building on previous studies of stimulus-specific adaption. With a focus on the IC, the authors measured contextual effects on tone responses with and without Arch-mediated inactivation of AC neurons that project to the IC. Using a tone cascade stimulus as a baseline, they divided SSA effects into adaptation to a regular, repeating stimulus (repetition suppression) and enhanced responses to an oddball stimulus (prediction error). This work nicely replicates some previous findings, including the relatively low rates of SSA in IC, especially central areas, the decrease in SSA magnitude for awake versus anesthetized animals, and the decrease in SSA magnitude following inactivation of auditory cortex. By breaking SSA effects into its components, they are able to argue that feedback from AC primarily signals prediction error (rather than suppression). In addition, they identify a group of neurons that shows an opposite pattern to standard SSA, with enhanced responses to repetition and decreased responses to oddball stimuli.

Several new observations help refine understanding of the role of cortical feedback in sound processing. Perhaps most substantial is the observation of enhanced repetition responses in the auditory midbrain. While these results may be important, there are some methodological/analytical concerns that should be addressed, especially given that the pattern of repetition enhancement has not been reported previously.

Concerns:

1. The novelty of the current result could be spelled out more clearly. The authors cite previous work from the Malmierca group reporting diminished SSA in awake animals and when cortex was silenced, but don't provide a more detailed comparison. Is the novelty of the current study that the cortical effect was linked specifically to prediction error? But why then is the major effect of Arch on non-adapting cells to increase repetition suppression? While there are some overarching themes, the results currently seem a bit scattered and would be strengthened if linked more directly to the previous work.

2. The report of facilitating neurons is quite interesting, but some important details are not clear.

a. (L. 213-214) It took some close reading, but it appears that neurons with suppressive responses were not analyzed for SSA effects. How were neurons with suppressive responses defined? If a neuron showed a transient response followed by suppression (eg, left panel of Figure 3D), did that count as suppressive? It appears in this example, that a tone actually evokes a sustained response that is suppressive, relative to the spontaneous rate. If repetition suppression is computed from raw firing rate, a decrease in suppression in this case could actually appear as an enhancement. The example in Figure 5F, of course, provides a clear demonstration of facilitation, but it is important to know what the typical response profile is for the adapting versus suppressive neurons and if they are qualitatively different. Perhaps the average PSTH responses could be compared?

b. (L. 386-391) How are neurons defined as adapting vs. facilitating? Based only on non-laser trials? One worries that some of the reported effects may reflect a regression to the mean. That is, if there was experimental noise that made responses slightly adapting or facilitating in the laser-off conditions, then the absence of adaptation or facilitation in the laser-on condition may be that the noise was absent. A potentially less biased approach would be to define adapting versus facilitating neurons based on responses averaged across both laser-on and -off trials. Given that the adaptation effects are relatively infrequent and small, this additional control seems important.

L. 47-48. "suppression … suppression" Not a concern really, but a request. The language is technically correct, but the word suppression refers both to a neural computation (suppression of error signals) and optogenetic manipulation (Arch-mediated suppression). The authors might consider an alternative term for one of these elements of the paper, e.g., "optogenetic inactivation"?

L. 102. "repetition and prediction" unclear. "repetition suppression and prediction error"?

L. 127. What volume of virus was injected in AC and IC?

L. 152. "experimental procedures were repeated" please clarify, were the same units recorded in both conditions?

L. 155. "experimental" does this refer to recordings from IC? Also, please clarify if there were any differences in the results for SUA vs. MUA. This is particularly important for central IC, where single unit isolation is typically difficult.

L. 193. "evenly spaced" Please confirm, "evenly log-spaced"?

L. 217. "iMM, equivalent to…" Does "equivalent to" mean "identical to" the traditional SSA index?

L 218. "iMM = iPE and iRS" Should "and" be "+"?

L 220. "FR" please clarify if spontaneous rate was subtracted or considered as part of the analysis.

L. 262. The term "error suppression" is a bit confusing. Any difference between standard and deviant response (positive or negative) seems like an error signal. Here the term appears to indicate weaker response to the deviant. It's fine if this is a standard term used elsewhere, but the authors might otherwise consider an alternative.

L. 289. What volume of AC was labeled with virus? Was it limited to a tonotopic region? If so, were effects of laser inactivation frequency-specific in IC? This is understandably a difficult question to answer definitively, but some information about the extent of transduction would be helpful.

L. 411. (Figure 5 legend) Were the examples in C and F recorded in central or shell IC?

L. 497. "non-adapting neurons … increase repetition suppression" this result is confusing. Is the idea that two factors cancel each other out, and the optogenetic inactivation unmasks an adaptation process? Some help is needed for interpretation.

L. 681. "According to hierarchical predictive coding" Confusing, maybe "According to models of hierarchical…"?

L. 707. "… stronger response … in a completely predictable sequence" this is an interesting point. If this is the case, should the response to a tone in a cascade differ from the many standards condition?

Reviewer 3:

This is an excellent study testing the influence of auditory cortical connections to the inferior colliculus on context-dependent aspects of response patterns in the IC. The authors deploy a paradigm that permits dissociating two different forms of context-dependence, the predictability of a sound and the overall statistics of the sounds, and they use optogenetic methods to investigate the contributions of auditory cortex to how IC neurons alter their responses in these two different types of contexts – this is the particularly novel finding in the paper as the basics of IC neuron response characteristics as a function of context had been previously explored. The manuscript also provides a direct comparison of awake vs anesthetized mouse results, which is a very nice addition and service to the literature as it helps the community incorporate both awake and anesthetized results together and understand their similarities and differences. The manuscript is very well written and the figures are carefully constructed.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Cortico-fugal regulation of predictive coding" for further consideration by eLife. Your revised article has been evaluated by Barbara Shinn-Cunningham (Senior Editor) and Jennifer Groh (Reviewing Editor, Reviewer 1)

The manuscript has been improved, and the reviewers agreed that this work is important, but there are some remaining issues involving the analyses that need to be addressed as outlined in greater detail below (Reviewers 2 and 3).

Reviewer #1:

This is a well designed and well written study to investigate the role of descending inputs from auditory cortex to the inferior colliculus in modulating responses to sound. The behavioral task permits a distinction between various forms contextual interpretation of sound, and distinct roles for descending inputs in these contexts are identified. The work replicates and extends prior work in this area, and is likely to be highly impactful regarding our understanding of how the brain's "backwards" connections govern processing in sensory pathways.

Reviewer #2:

This study measures the impact of feedback from auditory cortex (AC) to the auditory midbrain (inferior colliculus, IC) on neural encoding of predicable versus unpredictable sounds. Consistent with previous work, the authors find that inactivation of AC feedback reduces differential adaptation to high- versus low-probability stimuli. By introducing a new stimulus condition to their analysis, they provide evidence that the specific effect of AC feedback is to enhance prediction errors and that other forms of adaptation occur independent of AC.

The authors have done a good job addressing concerns raised during the initial review. In particular they have more clearly contrasted their new results on prediction error with previous corticofugal/SSA work that did not distinguish between adaptation and prediction error. This provides a substantial advance on previous work.

However, some concerns do persist around the validity of statistical methods used.

L. 323-324: "To ensure that the laser-induced changes described above were opsin-mediated, we performed control experiments in two mice with identical manipulations to the experimental group, but in the absence of ArchT…"

It feels like nagging, but the circumstances described in this study--where fairly symmetric tails of a distribution both shift toward zero--are exactly the case where regression to the mean is a concern. The control cited above provides evidence that activation of ArchT has some impact on adaptation, but it is not clear that this single control is adequate to support all the subsequent findings in the manuscript. It would be more convincing if the authors could categorize units based on responses averaged across laser-off and -on conditions. If this is not feasible, then the authors should address the following:

1. Are the experimental groups in Figure 3S1B the same as in Figure 3B? Perhaps some details are missing or there is a labeling issue with the x axis? The distributions in 3S1B appear narrower than in 3B. A narrower distribution might indicate less noise, which would reduce the possibility of regression to the mean. Please clarify if in fact the experimental data should be the same and or if differences in the width of the iMM distributions might impact the validity of the control. Are the fractions of adapting/facilitating/non-adapting neurons similar for the control and experimental groups?

2. While the control experiment/analysis supports conclusions reported in Figure 3, it is not clear that it is adequate to support the conclusions in the subsequent analyses, where the data are further processed (e.g., Figure 4C, IRS<0 only) or different quantities are analyzed (e.g., firing rate in Figure 5). One solution would be to run the same analysis on control data in each case. This option does seem cumbersome, and authors might have a better idea for how to address this concern.

L. 88. "However, it remains unknown whether these modulations in the SSA index with cortical deactivation reflect changes in predictive processing." Minor. This sentence seems out of place, as the relationship between SSA and prediction error is not laid out until the next paragraph.

L. 95. "Prediction error…" It might help to rephrase this sentence to provide a definition of prediction error as a component of SSA, in the same way that the previous sentence defines repetition suppression.

L. 253 "… while an iPE value…" Should this be "… while the mean iPE value"? Also, since the same neurons were recorded in both conditions could a paired test be used here? Students T usually treats the two distributions as independent.

Reviewer #3:

This manuscript will be of interest to the broad sensory neuroscience field. Prediction error signals, reflecting a mismatch between expected and actual sensory inputs, have been described across sensory modalities and the contributions of bottom-up and top-down processing are still unknown. This work reveals the contributions of descending cortico-collicular inputs to predictive coding of neurons within the inferior colliculus.

In the present manuscript, Lesicko and colleagues studied the contributions of cortico-collicular neurons to the components underlying stimulus specific adaptation (SSA) of neurons within the inferior colliculus (IC). This is a subject of interest for the broad sensory neuroscience field, as SSA has been described across modalities and the contributions of bottom-up and top-down processing are still unknown. Overall, the experimental design and results are convincing, straight-forward and well presented. Although some of these results simply corroborate previous findings, this study does provide the following conceptual advances from prior work:

1. The contribution of the decomposed processes underlying SSA in subdivisions of the inferior colliculus (IC) has been studied previously in awake mice (Parras et al. 2017) and cortical manipulations have demonstrated the influence of cortical feedback on SSA (Anderson and Malmierca 2013). However, this work combined these approaches to evaluate the contribution of cortico-collicular projections to the distinct SSA processes (repetition suppression and prediction error) in the IC of awake mice.

2. By recording from the same units in the IC of mice in awake and anesthetized states, the authors provide evidence for changes in the repetition and prediction processes underlying SSA across behavioral states. This result is significant, as much of the previous work has been performed in anesthetized rodents.

3. The authors focused on previously ignored facilitating and non-adapting neurons. They discover populations of IC neurons that show repetition enhancement and negative prediction error, which were suppressed during cortico-collicular inactivation. This novel finding has important implications for top-down cortical regulation of predictive coding in IC.

Together, the results from this study will contribute to our understanding of predictive coding in the central auditory system. The strength of this study is the rigorous experiments and comprehensive analysis: the authors examined SSA in neurons from distinct IC subdivisions in both awake and anesthetized mice, assessed the specific effects of corticocollicular projections and analyzed previously-ignored facilitation. Addressing the following concerns will greatly increase the significance of the findings:

1. The authors pool together data from single and multi-unit recordings. Although the inclusion of multi-units might not necessarily affect the described results, this limitation should be clearly stated in the "technical considerations" section of the discussion. For instance, the inclusion of multi-unit recordings may underlie the finding of "mixed" firing types in the IC, as shown in Figure 1-S2C. Importantly, the figure comparing responses between single- and multi-unit responses (provided in the response to the reviewers) should be included as a supplementary figure. Moreover, the author's interpretation of Figure 6 is that 'individual neurons exhibit distinct combinations of iPE and iRS'. Unless these data are only representing single units (which would not be consistent with the total number of single-unit recordings shown in the plot given to the reviewers), Figure 6 (panels A,B) fails to demonstrate that single neurons are exhibiting distinct iPE and iRS. This limitation should be clear in the Results section and included in the 'Technical Considerations' section of the Discussion.

2. The data shown in Figure 1-S2 are not convincing that the recordings were obtained from either the central or shell IC. The authors should provide more histological evidence if the data are available. For example, the panel showing the recording site within the central IC (Figure 1-S2D) also shows some DiA signal in the shell. The authors claim to have performed histological reconstruction by DiD/A probe coating in a subset of animals (lines 708-709). The authors should state how many recording sites used in the study were evaluated histologically, and include all data in a supplemental figure. In addition to the histology, the authors also use the response properties of the IC neurons to distinguish between central and shell recordings. This is a nice complementary method. However, while there are sites clearly distinguished as central or shell recordings based on sparseness and correlations between BF and depth, some sites are borderline (Figure 1-S2C). For example, several sites with high mean sparseness (characteristic of central IC) are categorized as shell recordings based on low correlations between BF vs depth. Could these sites instead be recordings in central IC with an electrode penetration angle slightly off the tonotopic axis? The differences between central and shell IC neurons in predictive coding, effects of anesthesia, and effects of cortico-collicular silencing are interesting findings. If the authors could provide additional data to give us more confidence in their ability to distinguish these sites, these results would be more compelling.

3. Figure 1-S1C shows an off/rebound response for cortico-collicular neurons when the laser is turned off. The iMM (index of neuronal mismatch) is calculated by comparing the firing in response to a standard and a deviant tone that occurs 200ms after the standard. Thus, rebound activation of the cortico-collicular neurons after the last standard stimulus may alter the response to the subsequent deviant stimulus, impacting iMM. This potential confound should be discussed within the 'technical limitations' section of the Discussion.

4. Figure 1-S1A. Although the study emphasizes the specific manipulation of the cortico-collicular projections, it should be noted that auditory corticofugal neurons that innervate the IC have other widespread targets including the medial geniculate body, striatum and lateral amygdala (Asokan et al., 2018, Nature Communications). Did the authors also see expression of axons in widespread brain regions with their viral strategy? If so, this could be shown in this panel or a supplemental figure. The involvement of these other regions in the inactivation studies should be addressed in the Discussion.

https://doi.org/10.7554/eLife.73289.sa1

Author response

[Editors’ note: the authors resubmitted a revised version of the paper for consideration. What follows is the authors’ response to the first round of review.]

Reviewer #1:

1. “the manuscript is long and could be substantially trimmed and focused to make it more useful”.

We have revised the present draft of the manuscript to make it more focused. Specifically, we have moved Figure 2 and 3 to the supplementary material, as these figures are meant to provide validation of the experimental methods and not new findings.

2. “it seems to me much of the basic/general results are similar to Parras and colleagues or even Duque and Malmierca. These authors also studied SSA and iMM in awake mouse. The main results are similar. Low levels of iMM and iPE in the IC. This is a major issue for me.”

We have extensively studied Parras et al., 2017 and Duque and Malmierca, 2015 as they provide much of the groundwork and inspiration for the present study (Duque and Malmierca, 2015; Parras et al., 2017). We have emphasized in the abstract and discussion several novel findings. Specifically:

– Like Parras et al., 2017, we decomposed stimulus specific adaptation into two distinct processes: prediction error and repetition suppression. A novel contribution is that we specifically determined how corticofugal inputs affect prediction error and other metrics of deviance detection in the inferior colliculus.

– While we identify “low levels of iMM and iPE in the IC”, it is not the main result of the study. Our main findings are that top-down inputs from the auditory cortex regulate prediction error, as well as other metrics such as repetition enhancement, in the IC, and that the cortex routes contextual information subcortically.

– As mentioned by reviewer #2, Anderson et al. 2013 previously investigated how cortico-collicular inputs affect SSA in the inferior colliculus through cortical deactivation in anesthetized rats (Anderson and Malmierca, 2013). In the present study we also deactivate cortico-collicular inputs and measure SSA. However, we furthermore determine how the index of prediction error, and the index of repetition suppression are affected by deactivation, not just SSA as a whole. This distinction is critical, as these two metrics reflect different underlying processes, whereas previously reported results could have captured changes in either and/or both. It is also worth noting that these findings are derived from awake animals, while the previous study used anesthetized animals. Our study presents novel data that show that the prediction and repetition processes that are reflected in the SSA index differ substantially between the awake and anesthetized condition, with prediction error and repetition enhancement being much more prevalent in the absence of anesthesia. It is likely that previously reported findings reflect different underlying processes than those studied here in the awake animal, further emphasizing the novelty of the current study.

– Duque and Malmierca, 2015 reported whether and how anesthesia and spontaneous activity affect SSA in the inferior colliculus. The authors conclude that SSA is “similar, but not identical, in the awake and anesthetized preparations” and show that the differences are “mostly due to the higher spontaneous activity observed in the awake animals.” We also compare how anesthesia affects SSA, but we (1) record from the same units while the animal is awake and under anesthesia to directly compare the two conditions and (2) further assess how the index of prediction error and the index of repetition suppression are affected. We find that SSA in awake animals reflects entirely different underlying processes than those in anesthetized animals. Specifically, prediction error and repetition enhancement are significantly more prevalent in shell IC units when animals are awake, while low levels of prediction error and high repetition suppression dominate under anesthesia. In the central nucleus, negative prediction error becomes dominate when the animal is awake. Given that much of the prior research on SSA in the auditory midbrain has been conducted in anesthetized animals, we believe this is a finding of major significance for the field and suggests that the state of anesthesia must be taken into consideration when interpreting SSA studies.

3. “Given the low levels of SSA/iMM, I wonder how authors consider if a neuron shows significant iMM…(bootstrapping??) as AC may have mostly subtle changes on the IC responses”.

We performed this analysis. This information is provided in lines 791-792 of the present manuscript: “Significant adaptation or facilitation for each neuron was assessed with a Wilcoxon rank sum test between the trial-by-trial firing rates to the standard and deviant on the 45 baseline trials.”

We want to emphasize that the fact that there are relatively low levels of SSA/iMM (in comparison to the cortex) is not a major point or conclusion of the present study; rather, we are interested in how inputs from the cortex affect the indices of neuronal mismatch, prediction error, and repetition suppression in IC units. A “low level” of SSA/iMM does not necessarily mean that a unit is absent of prediction error, etc. For example, we find that a high level of prediction error and a negative index of repetition suppression (indicating repetition enhancement) can result in a SSA index/iMM close to zero (see Figure 2E-G; Figure 6).

4. “Another major issue is that authors claim that have recorded separately from the central nucleus and the shell. However, they don’t show any histological probe of the electrode recording.”

We have added histological data to the manuscript from a recording that was categorized as a central nucleus site and a recording that was categorized as a shell site (Figure 1 —figure supplement 2D) using the analytic methods detailed in lines 768-775 of the present manuscript. In both instances the recording electrode was submerged in a lipophilic dye prior to tissue insertion to mark the recording location. An atlas image overlay was used to define the locations of the central (denote as “CIC” here) and shell (denoted as “ECIC” and “DCIC” here) regions of the IC (Paxinos and Franklin, 2019). Notably, our histological data provide the same categorization of electrode sites as the analytical methods used in the present manuscript.

5. “The central nucleus in mouse is very small and although it shows a distinct tonotopic organization and response are areas are usually V shaped, these responses can also occur in the shell. So, my concern is that most of the neurons may be actually recorded in the shell.”

Indeed, responses in the shell can be V-shaped and regionally tonotopic (Barnstedt et al., 2015; Wong and Borst, 2019). However, the shell does not exhibit the same stereotyped tonotopic gradient with depth that is highly characteristic of the central nucleus of the IC (see Figure 3B) (Aitkin et al., 1975; Malmierca et al., 2008; Stiebler and Ehret, 1985; Syka et al., 2000). Our method for parsing the recording sites considers the patterns in best frequency and sparsity in all units across the entire depth of the electrode – whereas it is possible to encounter some V-shaped responses or local regions of tonotopy in the shell, it is unlikely that these characteristics will produce the same highly linear fits and consistent high sparsity (reflecting V-shaped tuning) that we find with our central nucleus recordings. Because histological assessment alone can render inconclusive subdivision assignments, specifically in instances when the recording location is near a boundary between subdivisions, these methods allow for an unbiased analysis of site location.

6. “I would like to draw the attention to the authors that previous studies on subcortical SSA/iMM have demonstrated the lack of adapting responses in the central nucleus. Of course, all previous studies may have missed this, but I dare to suggest that much of the authors central nucleus data may actually be from the shell.”

This is an important concern, and we compared our results more directly to those reported in the literature. We found comparable mean baseline SSA/iMM values for the adapting units in the central nucleus (mean = 0.26, Figure 4A and Table 1) to those in the literature for awake mouse (mean = 0.24, Parras et al., 2017 Table 2). We also find that there are fewer adapting units in the central nucleus compared to the shell (Figure 3B,), and those adapting units have lower SSA/iMM indices than their counterparts in the shell (Figure 3B), also in line with previous studies (Duque et al., 2012; Parras et al., 2017). Furthermore, if most of the recordings categorized as central were actually recorded in the shell, we would likely see similar trends in predictive coding metrics from both regions. In fact, we find that both the baseline values and the laser-induced changes in these regions are distinct.

7. “In any case, this needs to be unambiguously demonstrated here with some histological data; it is not adequate to rely solely on the frequency response areas, even if this is shown, I would like to see a convincing conceptual framework to understand it.

We agree and we have added histological data as detailed in response to comment 4. We exploit known differences in the tonotopic organization across the depth of the tissue between shell and central regions to further parse the recording sites. This pattern of increasing best frequency with depth has been very well categorized (see response to point #5) and has been similarly employed by other researchers to categorize IC recordings sites, as mentioned by reviewer #3 (Bulkin and Groh, 2011; Ress and Chandrasekaran, 2013). We have edited the manuscript to include these references and expand the conceptual framework that justifies the use of these analytical techniques for sorting recording sites in lines 202-222 of the present manuscript.

8. “Also the conclusion that the cascade and many-standard controls yield similar results has also been reported in Parras and Casado-Roman recently.”

We thank the reviewer for bringing this study to our attention and have cited it in lines 375-376 of the edited manuscript (Casado-Román et al., 2020). The fact that the cascade and many-standards sequences yield similar results is not a major conclusion of this study (Figure 3 —figure supplement 2), but rather an important control to validate the use of the cascade sequence in further decomposing the iMM into an index of prediction error and an index of repetition suppression. We find it reassuring that Parras and Casado-Roman have also found little difference between the use of these two control sequences.

9. “The most original part of this study is the in-depth analysis of the repetition enhancement responses, but I also would like to draw the authors’ attention to the previous study by Parras where they also reported negative iPE (cf. Figures3 rat and Figure 7 awake mouse). This is mentioned on page 29, lines 694-696, but it should be more clearly stated in abstract etc. so that previous studies get a fairer recognition.

This is a great suggestion. Repetition enhancement reflects a higher response to the standard context than to the cascade context, resulting in a negative iRS value. This is different than the negative iPE value mentioned in the Parras et al. study, which we have termed “negative prediction error” in the present manuscript. We are aware that Parras et al. also found units with negative iPE and have contrasted their interpretation of these results with our own in the Discussion (lines 608-638). Whereas Parras et al. do find units with negative iRS indices, they do not show that negative iRS reflects repetition enhancement (Figure 6C,D), which we report for the first time here. Further, one of our main results is that repetition enhancement decreases during cortical suppression, suggesting that it is a top-down phenomenon, which is a novel finding.

10. “Also, Duque and Malmierca already have reported that SSA (which reflects iMM) is lower in awake than in anesthetized mouse. This is mostly due to the high rate of spontaneous activity, which is not mentioned in this study.”

This is an insightful comment. We also find that SSA index/iMM is lower in awake vs. anesthetized animals (Figure 2). However, this is not a major conclusion of the present study. Rather, we sought to determine what underlying repetition/prediction processes the SSA index/iMM reflects in the awake vs. anesthetized condition, as these processes have drastically different functional implications for predictive processing. We show that prediction error and repetition enhancement are significantly more prevalent in the awake animal in shell IC units, and that negative prediction error dominates the awake central nucleus.

11. “The most interesting part of the paper is the section related to the effect of cortical deactivation. However, as the authors themselves note, the technique has some limitations. I would like a more detailed discussion on how such limitations may have affected the results and hence the conclusions.”

We appreciate the reviewer’s recognition of the merits of the study. We discuss in the manuscript that the main drawback of using laser photosuppression to mediate cortico-collicular deactivation is that it does not achieve full inactivation (lines 641-654). Indeed, this is a concern to us, but we believe the inactivation is robust. We found a mean 60% reduction in firing in putative cortico-collicular neurons at baseline and a 45% reduction during presentation of pure tone stimuli with our laser parameters and observed clear effects on repetition and prediction processing in IC.

12. “Another important issue not addressed is where in AC the injections were made (A1, AAF, A2, etc ???) and how large the injection sites are. These technicalities may have affected the results and should also be considered, and different fields may have different projections to IC.”

We have provided these details in lines 687-692 of the manuscript: “A glass syringe (30-50 µm diameter) connected to a pump (Pump 11 Elite, Harvard Apparatus) was used to inject modified viral vectors (AAV9-CAGFLEX-ArchT-tdTomato or AAV9-CAG-FLEX-tdTomato; 750 nL/site; UNC Vector Core) into AC and a retroAAV construct (retro AAV-hSyn-Cre-GFP; 250 nL/site) into IC (Figure 1A, 2A, Figure 3 —figure supplement 1A). Large viral injections were performed to broadly target cortico-collicular neurons throughout all regions of the auditory cortex.” We agree that it could be informative to consider whether projections from specific sub-fields differentially affect metrics of predictive coding and deviance detection in the IC, and that would be an interesting future direction.

13. “Figures are difficult to read. I am not sure if this is due to the pdf that the system generates, but the dot raters of responses etc. e.g., Figure 2, Figure 5, 6. etc. (dots in scatter plots) are almost impossible to make out and I have to simply rely on the text. The figure quality needs to be significantly improved in order to see the data.”

Thank you for pointing this out. We have increased the dot size and the panel size for the figures mentioned and several others in order to improve the visibility of the data.

Reviewer #2:

1. “The novelty of the current result could be spelled out more clearly.”

We have edited the manuscript in several places to address this issue (lines 40-41; 86-90; 103-104; 238244; 404-405) and have included a summary diagram in Figure 7 to further elucidate the main findings. Please see responses to reviewer 1 comments 2, 8, 9 and 10.

We also provide below a summary of the major findings from the present study:

– We show for the first time that cortical input is critical for generating prediction error in IC units, suggesting that the cortex regulates predictive coding subcortically. To our knowledge, this is the first demonstration of predictive coding in a cortico-subcortical network, as virtually all prior studies have focused on cortico-cortical interactions.

– We show for the first time that repetition suppression is unaffected by cortical inactivation, suggesting that this process may reflect fatigue of bottom-up sensory inputs rather than deviance detection.

– We also show for the first time that a subset of IC neurons exhibit repetition enhancement. Further, we show that repetition enhancement is abolished in the absence of cortical input for central IC units, suggesting that it is a top-down phenomenon.

– We show that cortico-collicular inactivation leads to bidirectional changes in the response to the standard vs. deviant tone contexts, such that IC cells respond more similarly to both contexts in the absence of cortical input. These findings suggest that under normal conditions the cortex routes contextual information to the IC.

– We provide the first direct comparison of SSA (i.e., in the same units) in awake vs. anesthetized conditions. We show that SSA reflects drastically different repetition and prediction processes in awake vs. anesthetized animals: in the central IC, negative prediction error rather than prediction error dominates when the animal is awake, and in the shell IC, prediction error and repetition enhancement are significantly more prominent. These findings have important implications for the field, as the vast majority of previous SSA studies have been conducted in anesthetized animals.

2. “The authors cite previous work from the Malmierca group reporting diminished SSA in awake animals and when cortex was silenced, but don’t provide a more detailed comparison.”

Please see response to points 2, 8, 9 and 10 of reviewer #1.

3. “Is the novelty of the current study that the cortical effect was linked specifically to prediction error? But why then is the major effect of Arch on non-adapting cells to increase repetition suppression? While there are some overarching themes, the results currently seem a bit scattered and would be strengthened if linked to the previous work.”

One main finding is that cortico-collicular deactivation decreases prediction error in IC units. However, we also show that the cortex plays a role in repetition processing; specifically, we find that repetition enhancement is abolished in facilitating units in the central IC, suggesting that it may be a top-down phenomenon (Figure 3G). As the reviewer mentions, we additionally find an effect on repetition processing in central non-adapting cells with cortical deactivation, leading to an enhanced index of repetition suppression. This change also reflects a decrease in repetition enhancement. We have added data to better illustrate this finding (Figure 4C, bottom) showing that when the iRS is further parsed for non-adapting units, it is those with negative indices (i.e. those that show repetition enhancement) that show a significant laser effect. We believe that this finding further solidifies the notion that repetition enhancement is a top-down phenomenon.

4. “How were the neurons with suppressive responses defined? If a neuron showed a transient response followed by suppression (eg, left panel of Figure 3D), did that count as suppressive? It appears in this example, that a tone actually evokes a sustained response that is suppressive, relative to the spontaneous rate. If repetition suppression is computed from raw firing rate, a decrease in suppression in this case could actually appear as an enhancement.”

“Suppressed units”, henceforth referred to as “inhibited” units were defined quantitatively from analysis of the combined responses to the standard and deviant tones using MATLAB’s “findpeaks” function with a minimum peak height set to the mean of the baseline period (50 ms before tone onset) +/- 3 standard deviations. In brief, units that displayed a single minimum peak in the PSTH during the 0-50 ms tone duration were termed either “inhibited” units or “inhibited onset” units (if the peak of the minimum occurred in the first 10 ms after tone onset). These units were removed from the analysis. Units that showed a combination of a minimum and maximum during tone duration were labeled as either E-I or I-E (excited-inhibited or inhibited-excited) units, depending on whether the minimum or maximum occurred first. These units were included in the analysis such that their excitatory sound responses could be analyzed (see lines 776-789 for further details of categorizing sound response profiles).

5. “The example in Figure 5F, of course, provides a clear demonstration of facilitation, but it is important to know what the typical response profile is for the adapting versus suppressive neurons and if they are qualitatively different. Perhaps the average PSTH responses could be compared?”

We thank the reviewer for this excellent suggestion and have replaced the single examples with the average PSTH responses for adapting and facilitating units in the revision. The average PSTH for both adapting and facilitating units show excitatory, rather than inhibited responses (Figure 3C,F).

6. “How are the neurons defined as adapting vs. facilitating? Based only on non-laser trials? One worried that some of the reported effects may reflect a regression to the mean. That is, if there was experimental noise that made responses slightly adapting or facilitating in the laser-off conditions, then the absence of adaptation or facilitation in the laser-on condition may be that the noise was absent. A potentially less biased approach would be to define adapting versus facilitating neurons based on responses averaged across both laser-on and -off trials. Given that adaptation effects are relatively infrequent and small, this additional control seems important.”

We edited the methods to state: “Significant adaptation or facilitation for each neuron was assessed with a Wilcoxon rank sum test between the trial-by-trial firing rates to the standard and deviant on the 45 baseline trials.” (lines 791-792).

We agree that it is critical to rule out that the observed effects are simply due to regression to the mean. We took multiple steps at the study design stage to control for regression to the mean, namely using multiple baseline measurements and including a control group to provide an estimate of the change caused by regression to the mean, both standard practices that “can be combined to give even greater protection against regression to the mean” (Barnett et al., 2005). Units were defined as adapting/facilitating/non-adapting based on a statistical comparison of the trial-by-trial responses to the standard and deviant on 45 separate baseline trials. Although it is possible to have experimental noise on a given baseline trial that is not present in a laser trial, the use of multiple baseline trials protects against the possibility that a spurious extreme value will affect the overall categorization of the units as adapting/facilitating/non-adapting. We also included a control group that underwent identical manipulations to our experimental group except ArchT was not present in the viral construct that was injected in the auditory cortex (Figure 3 —figure supplement 1). For this group, we observed no significant differences in iMM, iPE, or iRS for any of the adapting, facilitating, or non-adapting cells in both the central and shell regions. Given that this control group provides an estimate of the change caused by regression to the mean, we conclude that the significant effects in the experimental group were not caused by this phenomenon.

7. L. 47-48. "suppression … suppression" Not a concern really, but a request. The language is technically correct, but the word suppression refers both to a neural computation (suppression of error signals) and optogenetic manipulation (Arch-mediated suppression). The authors might consider an alternative term for one of these elements of the paper, e.g., "optogenetic inactivation"?

Per the reviewer’s suggestion, we have edited the manuscript to avoid using “suppression” to describe different processes:

– Photosuppression/cortical suppression refers to Arch-mediated suppression of cortico-collicular neurons.

This term will be changed to “inactivation” or “optogenetic inactivation” throughout the manuscript.

– Error suppression refers to a greater response to the cascade context (a tone embedded in a predictable sequence) than to the deviant context (an unpredictable tone). We will be changing this term to “prediction signals”.

– Suppressed units (see Figure 3) refer to those units whose firing rate during tone presentation are lower than the preceding baseline period. We will be re-naming these units as “inhibited” units.

– Repetition suppression refers to a greater response to the standard context than the cascade context. Following the established convention in the literature, we will leave this term as is.

8. L. 155 "experimental" does this refer to recordings from IC? Also, please clarify if there were any differences in the results for SUA vs. MUA. This is particularly important for central IC, where single unit isolation is typically difficult.

We performed this analysis and believe that we can pool the units together in the manuscript. Figure 3—figure supplement 3 includes plots of the index of neuronal mismatch in laser off and on conditions for each of the subgroups in the central and shell regions of the IC separated by single (displayed in teal) and multi units (similar to Figure 3D,E,G,H, top panel and Figure 4C,E, top panel). No major differences exist in the distributions of these two groups, further justifying the decision to pool data from both for the analysis.

9. L. 707. "… stronger response … in a completely predictable sequence" this is an interesting point. If this is the case, should the response to a tone in a cascade differ from the many standards condition?

This is an interesting observation. The many standards condition, similar to the deviant condition, is completely unpredictable. However, this sequence does not contain the establishment or violation of a prediction, as is seen in the oddball sequence. It is possible that cells with negative prediction error are suppressed by prediction violation, which would explain their higher firing rates to the cascade than the deviant and equal responses to the cascade and many standards sequences.

All other reviewer suggestions have been addressed in the manuscript.

Reviewer #3:

1. "additional response profiles were assigned for each tone in the oddball pair based on the number, timing, duration, and direction of peaks in the peristimulus time histogram. Neurons with only suppressive or offset responses were removed from the data set." These two sentences are unclear – what is a response profile? If this is a quantitative judgment, what were the criteria used? If qualitative, I'd suggest rephrasing to say that neurons were excluded by visual inspection of the PSTH to ensure that only neurons with excitatory peaks within a reasonable onset latency were included, and give parameters for what was considered a reasonable onset latency and duration.”

As mentioned in response #4 to reviewer #2, the categorization of sound responses (i.e., “response profiles”) was done quantitatively. We have edited lines 776-789 of the methods to include the following:

“Sound response profiles were categorized quantitatively from analysis of the combined responses to the standard and deviant tones using MATLAB’s “findpeaks” function with a minimum peak height set to the mean of the baseline period (50 ms before tone onset) +/- 3 standard deviations. Units that did not display maxima or minima during the tone duration period (0-50 ms) or in the 50 ms after (the “offset window”) were labeled as sound unresponsive and were removed from the analysis. Units that showed only a single minimum (“inhibited” units) or only a response in the offset window were similarly removed from the analysis. Units that showed at least one maxima during the tone duration period were included in the analysis and further categorized as either onset (single maxima in the first 10 ms after tone onset), sustained (single maximum after the first 10 ms after tone onset), E-I or I-E (units that displayed both a maximum and minimum during the tone duration period), biphasic (units that displayed two maxima during the tone duration period), or mixed (units with greater than 2 peaks during the tone response period). It was common for units to display a response both during the tone duration window and the offset window, and in these cases a combined response profile was assigned (e.g., onset/offset, sustained/inhibited-offset).

2. “Anesthetized vs. awake comparison – this is an important part of the study and I’d suggest mentioning it in the abstract (anything not mentioned in the abstract might as well not have happened….).”

We thank the reviewer for pointing out this omission and have edited lines 53-57 of the abstract to reflect the anesthetized vs. awake findings: “We also investigated how these metrics compare between the anesthetized and awake states by recording from the same neurons under both conditions. We found that metrics of predictive coding and deviance detection differ depending on the anesthetic state of the animal, with negative prediction error emerging in the central IC and repetition enhancement and prediction error being more prevalent in shell regions in the absence of anesthesia.”

3. “Central vs shell quantification by regression of best frequency vs depth – there is precedent for this method from the monkey and human literature; it might be worth citing this work to strengthen the justification for this choice.”

We thank the reviewer for bringing these important papers to our attention and have cited them in an updated version of the manuscript. We believe that this precedence in conjunction with our added histological data further strengthens the justification for using analytical methods to sort recording sites.

[Editors’ note: what follows is the authors’ response to the second round of review.]

The manuscript has been improved, and the reviewers agreed that this work is important, but there are some remaining issues involving the analyses that need to be addressed as outlined in greater detail below (Reviewers 2 and 3).

Reviewer #2:

This study measures the impact of feedback from auditory cortex (AC) to the auditory midbrain (inferior colliculus, IC) on neural encoding of predicable versus unpredictable sounds. Consistent with previous work, the authors find that inactivation of AC feedback reduces differential adaptation to high- versus low-probability stimuli. By introducing a new stimulus condition to their analysis, they provide evidence that the specific effect of AC feedback is to enhance prediction errors and that other forms of adaptation occur independent of AC.

The authors have done a good job addressing concerns raised during the initial review. In particular they have more clearly contrasted their new results on prediction error with previous corticofugal/SSA work that did not distinguish between adaptation and prediction error. This provides a substantial advance on previous work.

However, some concerns do persist around the validity of statistical methods used.

L. 323-324: "To ensure that the laser-induced changes described above were opsin-mediated, we performed control experiments in two mice with identical manipulations to the experimental group, but in the absence of ArchT…"

It feels like nagging, but the circumstances described in this study--where fairly symmetric tails of a distribution both shift toward zero--are exactly the case where regression to the mean is a concern. The control cited above provides evidence that activation of ArchT has some impact on adaptation, but it is not clear that this single control is adequate to support all the subsequent findings in the manuscript. It would be more convincing if the authors could categorize units based on responses averaged across laser-off and -on conditions. If this is not feasible, then the authors should address the following:

We thank the reviewer for this suggestion. We used well-established methods for estimating regression to the mean including using multiple baseline measurements and a control group, as we have done here. Notably, we do not observe any significant changes in the control group for any of the predictive coding metrics between laseroff and laser-on trials, suggesting that regression to the mean does not drive the effects in our experimental group. Categorizing the units on responses averaged across laser-off and laser-off conditions, as reviewer suggests, would not adequately parse out an effect due to regression to the mean from a true photo-suppression effect. In the instance that a real photo-suppression effect is present in the data, categorizing units in this way would lead to the biased selection of only those units that maintain similar responses to the standard and deviant on both laser-on and -off trials, i.e., those units that do not show photo-suppression effects.

1. Are the experimental groups in Figure 3S1B the same as in Figure 3B? Perhaps some details are missing or there is a labeling issue with the x axis? The distributions in 3S1B appear narrower than in 3B. A narrower distribution might indicate less noise, which would reduce the possibility of regression to the mean. Please clarify if in fact the experimental data should be the same and or if differences in the width of the iMM distributions might impact the validity of the control. Are the fractions of adapting/facilitating/non-adapting neurons similar for the control and experimental groups?

The data from the experimental groups in Figure 3S1B and Figure 3B are the same. These plots depict the SSA index for laser-off trials, computed from the mean firing rates over 45 trials in response to a tone in either the standard or the deviant context. Whereas indeed, a narrower distribution of trial-by-trial firing rates for an individual unit could signify less noise in the responses, therefore reducing the possibility of regression to the mean, we do not believe the same interpretation applies for these plots, as they do not depict measures across multiple trials (i.e., test/re-test data). A narrower distribution in this case would indicate that the population shows less extreme adaptation/facilitation and that units respond more similarly to tones regardless of statistical context. However, this is not what we observe here. Rather, we find that the width of the distributions is relatively matched between the control and experimental groups in Figure 3S1B, which suggests that the range of SSA indices in the control sample is similar to the experimental group, thus enhancing the validity of the control.

We also find similar proportions of adapting/facilitating/non-adapting neurons in the control groups (central: 23% adapting, 5% facilitating, 71% non-adapting; shell: 29% adapting, 18% facilitating, 53% non-adapting) compared to the experimental groups (see Figure 4A; central: 24% adapting, 6% facilitating, 70% non-adapting; shell: 29% adapting, 9% facilitating, 62% non-adapting), and have added these comparisons to the edited manuscript (lines 332-336).

2. While the control experiment/analysis supports conclusions reported in Figure 3, it is not clear that it is adequate to support the conclusions in the subsequent analyses, where the data are further processed (e.g., Figure 4C, IRS<0 only) or different quantities are analyzed (e.g., firing rate in Figure 5). One solution would be to run the same analysis on control data in each case. This option does seem cumbersome, and authors might have a better idea for how to address this concern.

The data were analyzed, as depicted in Figure 4C, in order to determine whether changes in repetition suppression (iRS > 0) or repetition enhancement (iRS < 0) drive the statistically significant change in the overall iRS metric in central non-adapting units (see Figure 4C, third row). We have run this comparison in the control data as well and did not observe a laser effect for the overall iRS metric (Figure 4C) , repetition suppression (p=0.16, paired t-test) or enhancement (p=0.099, paired t-test). This analysis has been added to Table 2 of the edited manuscript.

The firing rate analysis in Figure 5 was performed to further characterize whether the changes in predictive coding metrics with laser photo-suppression arise from changes to the standard, cascade, and/or deviant contexts. Given that no significant changes were seen in these metrics in the control group, we did not perform an additional characterization.

L. 88. "However, it remains unknown whether these modulations in the SSA index with cortical deactivation reflect changes in predictive processing." Minor. This sentence seems out of place, as the relationship between SSA and prediction error is not laid out until the next paragraph.

We have removed this sentence and further edited the Introduction section per the suggestions of Reviewer 3.

L. 95. "Prediction error…" It might help to rephrase this sentence to provide a definition of prediction error as a component of SSA, in the same way that the previous sentence defines repetition suppression.

We have edited this sentence to the following: “Repetition suppression is characterized by a decrease in firing rate to each subsequent presentation of a standard tone while prediction error signals an enhanced response to a deviant tone”.

L. 253 "… while an iPE value…" Should this be "… while the mean iPE value"? Also, since the same neurons were recorded in both conditions could a paired test be used here? Students T usually treats the two distributions as independent.

We thank the reviewer for this correction and have changed it to “mean iPE” in the text. Though the units were recorded in both the awake and anesthetized conditions, some units showed sound-evoked responses in only one of the two conditions that were either inhibited, offset, or non-responsive, and were subsequently removed from the analysis. Therefore, paired data for both conditions was not available for all units, and an unpaired test was used instead.

Reviewer #3:

[…]

Together, the results from this study will contribute to our understanding of predictive coding in the central auditory system. The strength of this study is the rigorous experiments and comprehensive analysis: the authors examined SSA in neurons from distinct IC subdivisions in both awake and anesthetized mice, assessed the specific effects of corticocollicular projections and analyzed previously-ignored facilitation. Addressing the following concerns will greatly increase the significance of the findings:

1. The authors pool together data from single and multi-unit recordings. Although the inclusion of multi-units might not necessarily affect the described results, this limitation should be clearly stated in the "technical considerations" section of the discussion.

We thank the reviewer for this suggestion, and have added the following to lines 662-666 of the Discussion:

“The analyses in the present manuscript were performed on pooled single- and multi-unit data. Although we observed no differences in the iMM distribution between single- and multi-units (Figure 3 —figure supplement 3), the results of the present study should be interpreted with this limitation in mind. Namely, photosuppression-induced changes at the level of individual units may not reflect changes in single neurons.

For instance, the inclusion of multi-unit recordings may underlie the finding of "mixed" firing types in the IC, as shown in Figure 1-S2C.

The example of a “mixed” firing type shown in Figure 1-S2C is from a single unit. We have added the unit type (single or multi-unit) for each example depicted in Figure 1-S2 to the figure legend.

Importantly, the figure comparing responses between single- and multi-unit responses (provided in the response to the reviewers) should be included as a supplementary figure.

We thank the reviewer for this suggestion and have included the figure as a supplement (Figure 3—Figure Supplement 3).

Moreover, the author's interpretation of Figure 6 is that 'individual neurons exhibit distinct combinations of iPE and iRS'. Unless these data are only representing single units (which would not be consistent with the total number of single-unit recordings shown in the plot given to the reviewers), Figure 6 (panels A,B) fails to demonstrate that single neurons are exhibiting distinct iPE and iRS. This limitation should be clear in the Results section and included in the 'Technical Considerations' section of the Discussion.

We apologize for this error—this statement was meant to read “individual units exhibit distinct combinations of iPE and iRS”. The reviewer is correct that the data in Figure 6 are from both single and multi-units, and thus that we cannot conclude that individual neurons are exhibiting distinct iPE and iRS. We have edited this sentence and further discussed these limitations in the Discussion.

2. The data shown in Figure 1-S2 are not convincing that the recordings were obtained from either the central or shell IC. The authors should provide more histological evidence if the data are available. For example, the panel showing the recording site within the central IC (Figure 1-S2D) also shows some DiA signal in the shell. The authors claim to have performed histological reconstruction by DiD/A probe coating in a subset of animals (lines 708-709). The authors should state how many recording sites used in the study were evaluated histologically, and include all data in a supplemental figure.

Given that the shell IC surrounds the central IC dorsally, our electrode insertion sites (orthogonal to the dorsal surface of the IC) will always pass through the shell IC, even if the recording site is in the central IC. The source of dye in the shell IC in Figure 1-S2D is likely due to this initial insertion. The fact that histological assessment can render inconclusive subdivision assignments motivated us to pursue further categorization of the electrode sites based on the response properties of IC neurons, a method that has been established previously. We have added further details regarding the number of recording sites used for histological comparison in line 730.

In addition to the histology, the authors also use the response properties of the IC neurons to distinguish between central and shell recordings. This is a nice complementary method. However, while there are sites clearly distinguished as central or shell recordings based on sparseness and correlations between BF and depth, some sites are borderline (Figure 1-S2C). For example, several sites with high mean sparseness (characteristic of central IC) are categorized as shell recordings based on low correlations between BF vs depth. Could these sites instead be recordings in central IC with an electrode penetration angle slightly off the tonotopic axis? The differences between central and shell IC neurons in predictive coding, effects of anesthesia, and effects of cortico-collicular silencing are interesting findings. If the authors could provide additional data to give us more confidence in their ability to distinguish these sites, these results would be more compelling.

Given that our electrodes are inserted orthogonal to the dorsal surface of the IC and that the tonotopic gradient in the central nucleus runs dorso-lateral to ventro-medial, a slightly off angle penetration would likely still capture the increasing BF with depth relationship in this nucleus. It is not surprising to have some shell sites with higher sparseness values, as this metric can be affected by multiple tuning characteristics, such as inhibited regions of the tuning curve. Despite these few borderline cases, the BF vs. depth correlation and sparseness metrics show clear trends for parsing each IC sub-region.

3. Figure 1-S1C shows an off/rebound response for cortico-collicular neurons when the laser is turned off. The iMM (index of neuronal mismatch) is calculated by comparing the firing in response to a standard and a deviant tone that occurs 200ms after the standard. Thus, rebound activation of the cortico-collicular neurons after the last standard stimulus may alter the response to the subsequent deviant stimulus, impacting iMM. This potential confound should be discussed within the 'technical limitations' section of the Discussion.

In order to avoid potential laser rebound effects, the iMM was not computed by comparing immediately subsequent standard and deviant tones on laser trials. Rather, for a given last standard and subsequent deviant pair, only one of the tones was paired with the laser, and the pairing was switched on the next last standard/deviant presentation. This “block” approach to stimulus presentation and calculation of the iMM is described in lines 775782 of the text.

4. Figure 1-S1A. Although the study emphasizes the specific manipulation of the cortico-collicular projections, it should be noted that auditory corticofugal neurons that innervate the IC have other widespread targets including the medial geniculate body, striatum and lateral amygdala (Asokan et al., 2018, Nature Communications). Did the authors also see expression of axons in widespread brain regions with their viral strategy? If so, this could be shown in this panel or a supplemental figure. The involvement of these other regions in the inactivation studies should be addressed in the Discussion.

Consistent with Asokan et al. 2018, our viral strategy does label cortico-collicular collaterals in the medial geniculate body (see Blackwell et al. 2020) and additional downstream targets. The laser-induced changes seen in the present study are likely specific to the cortico-collicular pathway because they produce short-latency effects in the IC (Figure 3C, 3F), making muti-synaptic effects from other collateral sites unlikely. We have discussed this potential limitation in the Discussion in lines 667-674.

https://doi.org/10.7554/eLife.73289.sa2

Article and author information

Author details

  1. Alexandria MH Lesicko

    Department of Otorhinolaryngology, University of Pennsylvania, Philadelphia, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4766-2258
  2. Christopher F Angeloni

    Department of Psychology, University of Pennsylvania, Philadelphia, United States
    Contribution
    Data curation, Investigation, Resources, Software
    Competing interests
    No competing interests declared
  3. Jennifer M Blackwell

    1. Department of Otorhinolaryngology, University of Pennsylvania, Philadelphia, United States
    2. Department of Neurobiology and Behavior, Stony Brook University, Stony Brook, United States
    Contribution
    Data curation, Software
    Competing interests
    No competing interests declared
  4. Mariella De Biasi

    1. Department of Psychiatry, University of Pennsylvania, Philadelphia, United States
    2. Department of Systems Pharmacology and Experimental Therapeutics, University of Pennsylvania, Philadelphia, United States
    3. Department of Neuroscience, University of Pennsylvania, Philadelphia, United States
    Contribution
    Resources, Validation
    Competing interests
    No competing interests declared
  5. Maria N Geffen

    1. Department of Otorhinolaryngology, University of Pennsylvania, Philadelphia, United States
    2. Department of Neuroscience, University of Pennsylvania, Philadelphia, United States
    3. Department of Neurology, University of Pennsylvania, Philadelphia, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing – original draft, Writing – review and editing
    For correspondence
    mgeffen@pennmedicine.upenn.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-3022-2993

Funding

National Institute on Deafness and Other Communication Disorders (F32MH120890)

  • Alexandria MH Lesicko

National Institute on Deafness and Other Communication Disorders (R01DC015527)

  • Maria N Geffen

National Institute on Deafness and Other Communication Disorders (R01DC014479)

  • Maria N Geffen

National Institute of Neurological Disorders and Stroke (R01NS113241)

  • Maria N Geffen

National Institute on Deafness and Other Communication Disorders (F31DC016524)

  • Christopher F Angeloni

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank the members of the Geffen laboratory for helpful discussions. This work was supported by funding from the National Institute of Health grants F32MH120890 to AMHL, R01DC015527, R01DC014479 and R01NS113241 to MNG, F31DC016524 to CA, and R01DA044205, R01DA049545 and U01AA025931 to MDB.

Ethics

Animals were housed on a reversed 12-hour light-dark cycle with water and food available ad libitum. All procedures were approved by the University of Pennsylvania IACUC (protocol number 803266) and the AALAC Guide on Animal Research. We made every attempt to minimize the number of animals used and to reduce pain or discomfort.

Senior Editor

  1. Barbara G Shinn-Cunningham, Carnegie Mellon University, United States

Reviewing Editor

  1. Jennifer M Groh, Duke University, United States

Reviewer

  1. Jennifer M Groh, Duke University, United States

Publication history

  1. Preprint posted: April 14, 2021 (view preprint)
  2. Received: August 24, 2021
  3. Accepted: March 12, 2022
  4. Accepted Manuscript published: March 15, 2022 (version 1)
  5. Version of Record published: April 5, 2022 (version 2)

Copyright

© 2022, Lesicko et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 929
    Page views
  • 194
    Downloads
  • 2
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Alexandria MH Lesicko
  2. Christopher F Angeloni
  3. Jennifer M Blackwell
  4. Mariella De Biasi
  5. Maria N Geffen
(2022)
Corticofugal regulation of predictive coding
eLife 11:e73289.
https://doi.org/10.7554/eLife.73289

Further reading

    1. Developmental Biology
    2. Neuroscience
    Jennifer M Lin, Tyler A Mitchell ... Paolo Emanuele Forni
    Research Article Updated

    Neuronal identity dictates the position in an epithelium, and the ability to detect, process, and transmit specific signals to specified targets. Transcription factors (TFs) determine cellular identity via direct modulation of genetic transcription and recruiting chromatin modifiers. However, our understanding of the mechanisms that define neuronal identity and their magnitude remain a critical barrier to elucidate the etiology of congenital and neurodegenerative disorders. The rodent vomeronasal organ provides a unique system to examine in detail the molecular mechanisms underlying the differentiation and maturation of chemosensory neurons. Here, we demonstrated that the identity of postmitotic/maturing vomeronasal sensory neurons (VSNs), and vomeronasal-dependent behaviors can be reprogrammed through the rescue of Tfap2e/AP-2ε expression in the Tfap2eNull mice, and partially reprogrammed by inducing ectopic Tfap2e expression in mature apical VSNs. We suggest that the TF Tfap2e can reprogram VSNs bypassing cellular plasticity restrictions, and that it directly controls the expression of batteries of vomeronasal genes.

    1. Neuroscience
    David S Jacobs, Madeleine C Allen ... Bita Moghaddam
    Research Advance Updated

    Previously, we developed a novel model for anxiety during motivated behavior by training rats to perform a task where actions executed to obtain a reward were probabilistically punished and observed that after learning, neuronal activity in the ventral tegmental area (VTA) and dorsomedial prefrontal cortex (dmPFC) represent the relationship between action and punishment risk (Park and Moghaddam, 2017). Here, we used male and female rats to expand on the previous work by focusing on neural changes in the dmPFC and VTA that were associated with the learning of probabilistic punishment, and anxiolytic treatment with diazepam after learning. We find that adaptive neural responses of dmPFC and VTA during the learning of anxiogenic contingencies are independent from the punisher experience and occur primarily during the peri-action and reward period. Our results also identify peri-action ramping of VTA neural calcium activity, and VTA-dmPFC correlated activity, as potential markers for the anxiolytic properties of diazepam.