Introduction

Classically, perception is considered to rely on the flow of information from the sensory periphery via a sequence of hierarchically-organized brain structures up to the cortex. The ascending sensory pathways connecting these structures have been studied extensively and much has been learned about how signals are relayed, how features are extracted, and how information is integrated to produce increasingly abstract representations of the sensory environment. These pathways are paralleled by descending pathways that can feed information back to lower-order sensory structures. The fact that descending projections often outnumber their feedforward counterparts (Sherman, 2007) attests to their likely importance for brain function. This may include turning an otherwise passive, stimulus-driven device into an active and adaptive brain that is capable of processing sensory input within its behavioral context and, therefore, able to learn and create meaning (Engel et al., 2001; Kraus and White-Schwoch, 2015; Malmierca, Anderson and Antunes, 2015).

The descending projections of the auditory cortex target all major subcortical stations of the auditory pathway and are among the largest pathways of the brain (Winer, 2006; Bajo and King, 2013; Antunes and Malmierca, 2021), making them a particularly suitable system for investigating the behavioral and physiological consequences of corticofugal processing. One of their main targets is the inferior colliculus (IC), an obligatory midbrain relay for nearly all ascending auditory input. The corticocollicular projection primarily terminates in the non-lemniscal shell of the IC. The shell encapsulates and is extensively connected with the central nucleus of the IC, which forms part of the tonotopically organized core or lemniscal auditory pathway to the primary auditory cortex. The projection from the auditory cortex to the midbrain was identified almost a century ago (Mettler 1935) and decades of research have since demonstrated that manipulating the activity of descending projection neurons can alter the collicular representations of multiple sound features, influence adaptive plasticity and perceptual learning, and even trigger an innate flight response (Suga 2008; Nakamoto et al 2008; Bajo et al 2010; Xion et al 2015; Blackwell et al 2020). However, experimental evidence, especially from behaving animals, that could help explain what information the auditory midbrain and other subcortical sensory structures rely on their cortical input for is still very limited.

Interactions between different sensory pathways occur at multiple processing levels and they are also closely linked with the brain’s motor centers and neuromodulatory regions. Indeed, recordings in awake animals have shown that behavior, cognition and brain state can strongly influence activity in the sensory pathways (Schneider and Mooney, 2018; McCormick et al., 2020; Parker et al., 2020). Consistent with a hierarchical view of sensory processing in which neurons at higher levels carry progressively more complex representations of the world, such contextual influences appear particularly strong in the cortex (Stringer et al,. 2019; Musall et al., 2019) and may to a large extent be the result of intracortical processing (Noudoost et al., 2010; Schneider et al., 2014; Song et al., 2017). Nevertheless, non-acoustic and contextual variables can also alter sensory processing at subcortical levels, including the IC and particularly its shell (Metzger et al., 2006; Gruters and Groh, 2012; Chen and Song, 2019; Yang et al., 2020; Parras et al., 2017; Saderi et al., 2021). This raises the possibility that these context-dependent effects may be inherited from the auditory cortex.

To test whether auditory midbrain neurons convey behaviorally-relevant signals that depend on descending cortical inputs, we imaged corticorecipient IC shell neurons in mice engaged in a sound detection task. We found that the activity of most neurons contained information beyond the physical attributes of the sound and that this information could be used to decode the animals’ behavior with a high degree of accuracy. Surprisingly, this was the case both in mice with an intact cortex and those in which the auditory cortex had been lesioned. These findings suggest that subcortical auditory structures have access to a wealth of non-auditory information independently of descending inputs from the auditory cortex. Consequently, the contextually-enriched representations that are characteristic of sensory cortices can arise from subcortical processing.

Results

Transient suppression of the auditory cortex impairs sound detection

Our aim was to characterize the activity of neurons in the shell of the IC in animals engaged in sound-guided behavior and assess how this activity is influenced by the input from the auditory cortex. To this end, we trained water-regulated mice on a sound detection task (Figure 1A) in which they were rewarded with a drop of water for licking in response to a click sound. Transient pharmacological silencing of auditory cortex using the GABA-A agonist muscimol has been shown to abolish the ability of rodents (Talwar et al., 2001), including head-fixed mice (Li et al., 2017), to perform a sound detection task, making this approach unsuitable for our aim of exploring the role of IC during behavior. We found that optogenetic suppression of cortical activity by photoactivating ChR2-expressing inhibitory neurons in GAD2-IRES-cre mice (Lohse et al., 2020) also significantly impaired sound detection performance (d’: 1.51 + 0.71 vs 1.10 + 1.18 (median + interquartile range); p = 0.0009; Wilcoxon signed rank test; Figure 1B,C), albeit not to the same degree as pharmacological silencing. Optogenetic suppression of cortical activity is thus also unsuitable for this study as it would leave us unable to determine whether any changes in the activity of IC neurons arise from removal of their auditory cortical input or are a consequence of alterations in the animals’ behavior.

Figure 1.

Auditory cortex lesions leave detection ability intact

Several recent studies have shown that in contrast to the disruptive effects of transient silencing, cortical lesions leave performance in some sensory tasks intact (Hong et al., 2018; Ceballo et al., 2019; O’Sullivan et al., 2019). In order to compare how auditory cortex lesions impact sound detection performance, we therefore lesioned the auditory cortex bilaterally in 7 out of 16 animals.

Most corticocollicular neurons project ipsilaterally, with a substantial proportion also sending axons to the contralateral midbrain (Stebbings et al., 2014). The majority of corticocollicular neurons are found in the temporal cortex, and overwhelmingly in the auditory fields, while a small fraction populates adjacent areas, such as the temporal association area (Figure 2 - figure supplement 1). After the experiments, we injected a retrogradely-transported viral tracer (rAAV2-retro-tdTomato) into the right IC to determine whether any corticocollicular neurons remained after the AC lesions (Figure 2).

Figure 2.

Lesioned mice were able to learn and perform the click detection task. In fact, there was no significant difference between them and the non-lesioned mice in their learning speed (Figure 3A, p > 0.05, mixed-design ANOVAs) and psychometric functions (Figure 3B, p > 0.05, mixed-design ANOVA). Cortical lesioning thus leaves behavioral sensitivity to clicks intact and therefore provides a means of examining the effects of removing corticocollicular input, albeit non-reversibly, without directly affecting sound detection performance.

Figure 3.

Transsynaptic labeling and two-photon calcium imaging of auditory corticorecipient IC neurons

Manipulations of auditory cortical activity can influence the activity of neurons throughout the IC, including the central nucleus (Suga 2008, Nakamoto et al., 2008), where corticocollicular axons are relatively sparse (Stebbings et al 2014). The strongest effects, however, tend to be observed in the shell, where cortical input is densest (Nakamoto et al., 2008; Vila et al 2019; Blackwell et al., 2020). But even here, effects can be subtle (Vila et al., 2019) or undetectable (Blackwell et al., 2020) especially for cortical silencing. It is also unclear whether the IC neurons recorded in these studies receive cortical input or not. Therefore, we took a projection-specific approach to record the activity of IC neurons that receive direct input from the auditory cortex. More specifically, we injected AAV1.hSyn.Cre.WPRE, a virus with anterograde transsynaptic spread properties (Zingg et al., 2017), into the right auditory cortex of, initially, a tdTomato (Ai9) reporter mouse. This resulted in the expression of cre-recombinase and the reporter gene in neurons that receive input from the auditory cortex, including the corticorecipient neurons of the IC (Figure 4A). By employing this approach in GCaMP6f (Ai95D) reporter mice, we could target the expression of a calcium indicator to corticorecipient IC neurons and then proceed to record the activity of those neurons using two-photon microscopy (Figure 4B).

Figure 4.

Corticorecipient IC neurons display heterogeneous response profiles

The activity of individual corticorecipient IC neurons showed distinct response profiles across neurons and trial outcomes (hit vs miss) (Figure 4C). To capture the heterogeneity of response patterns across all recorded neurons, we used an unsupervised clustering algorithm (Namboodiri et al. 2019) to group the average responses of hit and miss trials for each neuron. This yielded 10 clusters that displayed different response patterns over the course of the trial (Figure 5A, B). Remarkably, most of the clusters exhibited distinct activity for hit vs miss trials. Some hit trial profiles were characterized by increases or decreases in activity, with a very sharp, short-latency onset, as in clusters 4 and 10 (see Figure 5 - figure supplement 1 for a scaled version of cluster 10), and others by much more gradual changes in which a peak occurred seconds after the trial onset, as in clusters 5 and 9. Cluster 3, which contained the smallest number of neurons, was an exception in that it showed a transient, short latency response to the stimulus for both trial outcomes. The response profiles of some other clusters, especially clusters 6 and 8, were also qualitatively similar across hit and miss trials and/or only weakly modulated across both trial types.

Fig. 5

This suggests that the activity of the majority of neurons in the recorded population contained information beyond the physical properties of the stimulus. Given that licking causes self-generated sounds, IC neurons could, in principle, respond to the sound of licking. However, given how quiet these are - estimated to be just 12 dB SPL (Singla et al., 2017) - and that much of the response to such lick-related sounds is already canceled out at the level of the cochlear nucleus (Singla et al., 2017), it is highly unlikely that lick-related sounds play a major role in driving activity in the IC.

To assess whether certain response profiles depended on auditory cortical input, we compared the ratio of neurons from lesioned vs non-lesioned mice in each cluster to that of the overall recorded population. The number of recorded neurons was unequal for lesioned and non-lesioned mice (952 vs 1697, respectively), reflecting the fact that a greater proportion of imaging sessions in non-lesioned animals were carried out using a larger field of view, which contained larger numbers of neurons (Figure 5 - figure supplement 2). To account for this, the percentages shown on the pie charts were normalized to the ratio in the overall population (Figure 5C). Neurons from both groups were well represented across all 10 clusters and while a significant difference in the lesioned/non-lesioned ratio was found for four clusters, only one deviated from a 50/50 split by more than 10%. This suggests that in mice engaged in a sound detection task, the activity of the neurons was broadly consistent in lesioned and non-lesioned animals and therefore that the IC shell can produce very similar output regardless of whether auditory cortical input is available or not.

Behavior can be accurately decoded from neural activity in lesioned and non-lesioned mice

The average responses of individual neurons in the IC shell exhibited a variety of activity patterns associated with both the stimulus and the trial outcome (Figure 5A,B). To gain insight into how these activity patterns can be read out collectively on a trial-by-trial basis, we assessed the relationship between the trial-by-trial network activity and the trial outcome. We trained logistic regression models to classify hit vs miss trials on a trial-by-trial, frame-by-frame basis. As different populations of neurons were recorded in different imaging sessions, the models were trained separately for each session. “Dummy models”, which randomly classified trials while taking into account the probability of hit vs miss trials in a given session, were used as the baseline model performance. If the population activity of the IC shell contained information about the trial outcome, the performance of the models would be significantly above baseline.

In both lesioned and non-lesioned mice, the average model performance was significantly above baseline in classifying hit vs miss trials (p < 0.05, one-sided Wilcoxon signed-rank test or paired t-test with Bonferroni correction, Figure 6A) and showed a temporal profile that is consistent with the dynamics of the activity profiles of some of the clusters, in particular clusters 1, 2, 4, 5, 9, 10 (Figure 5A,B). Additionally, the model performance in non-lesioned mice was significantly better than that in lesioned mice (p < 0.05, one-sided Mann-Whitney U test or t-test with Bonferroni correction, Figure 6A). The difference in the decoding performance was not the result of the difference in the number of neurons between non-lesioned and lesioned mice (Figure 6 - figure supplement 1).

Figure 6.

By examining the corticocollicular labeling and referencing the histological sections against a mouse brain atlas (Paxinos and Franklin, 2001), we categorized the mice according to lesion size. Four of the seven lesioned animals had “(near-)complete” lesions, meaning that all (Figure 2) or an estimated ∼95% (Figure 2 - figure supplement 2) of the auditory cortex had been lesioned, while the remaining mice had “partial” lesions, with an estimated 15% - 25% of the auditory cortex left intact. To assess whether the size of the lesions impacted the decoding performance, we compared the model performance between mice that had (near-)complete lesions and mice that had partial lesions. This revealed that the average decoding performance for mice with (near-)complete lesions was significantly better than that measured for mice with partial lesions. While this pattern of results may be unexpected, it is consistent with work showing smaller lesions being associated with greater somatosensory processing deficits (Hong et al., 2018). Additionally, the decoding performance in mice with (near-)complete lesions was largely indistinguishable from that in mice with an intact auditory cortex, implying that the population activity of IC shell neurons can contain similar amounts of information about the animal’s behavior regardless of whether descending input from the cortex is available or not (Figure 6B).

Pre-stimulus activity is predictive of the upcoming trial outcome. Remarkably, decoding accuracy was better than baseline even before stimulus onset. This could reflect changes in the network state that led or contributed to the upcoming trial outcome. For instance, changes in arousal or motivation can alter both the probability that an upcoming stimulus is detected and the activity of neurons in the network (Lee and Dan, 2012, McGinley et al., 2015). The decoding models might detect such changes in activity, resulting in higher decoding accuracy prior to stimulus onset. Additionally, pre-stimulus differences in hit and miss trial activity could also reflect the anticipation of an upcoming stimulus (Ruth et al., 1974; Nienhuis and Olds, 1978; Metzger et al., 2006) and the resulting change in attentional state. Inter-trial intervals in our experiments were randomly drawn from a normal distribution with a mean and standard deviation of 8 s and 2 s, respectively, and a lower bound of 3 s. Nevertheless, spontaneous licks did not occur at random times during the peri-catch trial periods following hit trials. Instead, average lick rates approximated the inter-trial interval distribution (Figure 6 - figure supplement 2A-D), suggesting that mice learned to adapt their behavior to this distribution and anticipate the timing of upcoming stimuli (Figure 6 - figure supplement 2E,F). Assuming that successfully anticipating the timing of an upcoming stimulus confers a greater chance of detecting the stimulus, neurons whose activity reflects that anticipation might be expected to show differences in pre-trial activity between hit and miss trials that could be detected by a decoding model. Note that for the analysis illustrated in Figures 5 and 6, hit trials were excluded if there were any licks between -500 ms and +120 ms (the latter number representing the lower bound of the animals’ lick-latency) relative to stimulus onset, suggesting that changes in pre-stimulus activity cannot be directly related to licking.

Discussion

Imaging auditory corticorecipient neurons in the dorsal shell of the IC in mice trained to perform a sound detection task revealed that the majority of neurons exhibited distinct activity profiles for hit and miss trials, implying that they encode information beyond just the physical attributes of the stimulus. Indeed, using logistic regression models to classify hit vs miss trials, we found that the animals’ behavioral choice can be read out from these neurons with a high degree of accuracy. Surprisingly, neural activity profiles and the decoding performance were similar in mice in which the auditory cortex had been lesioned bilaterally, suggesting that the midbrain has, independently of the auditory cortex, access to a wealth of non-acoustic information, which may be sufficient to support sound detection behavior.

Auditory corticocollicular axons terminate predominantly in the shell of the IC (Stebbings et al., 2014; Bajo and King, 2013) and the strongest effects of cortical manipulations have been reported in this region (Nakamoto et al., 2008; Vila et al 2019; Blackwell et al., 2020). However, these effects can be subtle (Cruces-Solis et al., 2018; Vila et al., 2019) or undetectable, especially when optogenetic silencing is used (Blackwell et al., 2020). Because of this and uncertainties over exactly what proportion of neurons in the shell of the IC is innervated by the auditory cortex and even where the border lies with the underlying central nucleus (Barnstedt et al., 2015), we used an anterograde transsynaptic tagging approach (Zingg et al., 2017) to identify corticorecipient neurons. This therefore maximized the chances of revealing the contribution of descending cortical input to the response properties of these midbrain neurons.

Inferior colliculus neurons exhibit task-related activity

Our recordings from corticorecipient neurons in the IC are consistent with previous studies demonstrating that neural representations of behavioral variables can be found in the auditory midbrain (Ruth et al., 1974; Nienhuis and Olds, 1978; Metzger et al., 2006; Gruters and Groh, 2012; Chen and Song, 2019; Yang et al., 2020; Saderi et al., 2021, Franceschi and Barkat, 2021). In keeping with responses recorded in the auditory cortex (Francis et al., 2018; Franceschi et al., 2021) and IC (Chen and Song 2019; Yang et al., 2020; Franceschi et al., 2021) of behaving mice, we found that the activity of most neurons was facilitated and about a third were suppressed during the sound detection task. Overall, only a small minority of clusters (mostly cluster 3) in our dataset showed what could be characterized as largely behavior-invariant response profiles to the auditory stimulus. In contrast, a large number of neurons were clearly driven by variables other than the stimulus itself. Their activity may represent the choice (to lick or not to lick) that an animal made, preparatory motor activity, corollary discharge or the reward and the somatosensory or gustatory feedback associated with its consumption, as well as modulation by the animal’s cognitive and behavioral state. Due to the task structure used, for the most part, it was not possible to unambiguously assign activity profiles to a particular variable. Nevertheless, we can speculate that neurons with late transients, such as in cluster 5, are more likely to represent corollary discharge and signals associated with the consumption of the reward, while those with very short latency peaks, as in clusters 4 and 10, may represent the animals’ choice and/or preparatory motor activity.

When engaged in the detection task, an animal’s arousal or motivational state may vary spontaneously or as a result of changes in, for instance, thirst, time of day or time into a session. In addition, cognitive factors, such as expectations about the timing of an upcoming trial (Ruth et al., 1974; Nienhuis and Olds, 1978; Metzger et al., 2006), which mice may have derived by learning the shape of the inter-trial interval distribution, may lead to variations in arousal or attentional state. Pre-trial differences in activity as well as the above-chance decoding performance before trial onset likely reflect the joint impact of those state changes on the activity of IC corticorecipient neurons and detection sensitivity (McCormick et al., 2020).

Contribution of the auditory cortex to task-related activity in the midbrain

Given the massive corticofugal projections that exist within the auditory system (Bajo and King, 2013), we hypothesized that task-related activity in the IC might depend on descending inputs from the auditory cortex. To address this, we imaged corticorecipient IC neurons during the same sound detection task after removing the cortical input. Consistent with previous work in the auditory (O’Sullivan et al., 2019) and somatosensory systems (Hong et al., 2018), we found that transient optogenetic silencing of the auditory cortex impaired sound detection, whereas cortical lesions had no effect on detection behavior, with lesioned mice learning the task as quickly as non-lesioned animals and achieving the same level of performance. In order to determine whether the absence of auditory cortical input alters the activity of IC neurons during sound detection behavior, we therefore focused on mice with bilateral cortical lesions to avoid the potentially confounding effects that reduced detection sensitivity produced by transient cortical silencing might have on the activity of IC neurons. For the same reason, we opted against the more targeted approach of optogenetic silencing of corticocollicular axons. Furthermore, it would have been difficult to silence the entire corticocollicular projection and the higher light powers required for manipulating axons compared to somata would have risked transmitting light to the cortex or other corticofugal targets, potentially causing behavioral changes and/or sacrificing specificity. Locally silencing corticocollicular axons would also have left indirect transmission via the thalamus between the auditory cortex and IC intact and would have been very challenging to verify. Finally, it has been reported that using optogenetic silencing tools in axons can have unintended consequences (Wiegert et al., 2017).

In keeping with our findings, numerous studies (reviewed in e.g. Pickles, 1988; Buser and Imbert, 1992) have shown that simple auditory skills, including the ability of freely moving rats to detect sounds (Kelly, 1970), are unaffected by the removal of the auditory cortex. However, transient pharmacological silencing of the auditory cortex in freely moving rats (Talwar et al., 2001), as well as head-fixed mice (Li et al., 2017), completely abolishes sound detection. The time course of the effects produced by muscimol application (Talwar et al 2001) suggests that there is a relationship between the size of the behavioral deficit and the degree of cortical inactivation. Consequently, milder impairments may be produced by the optogenetic approaches employed by us and others (Kato et al 2015; O’Sullivan et al., 2019) because of incomplete suppression of cortical activity. Alternatively, the larger behavioral effects reported following muscimol application may be due to diffusion of the drug to other brain structures, potentially including the IC. Although our results cannot speak directly to the question of whether the preservation of sound detection without auditory cortex reflects a rewiring or repurposing of circuits in the brain, this seems unlikely given that other studies have shown that trained mice achieve pre-lesion performance levels on simple auditory discrimination (Ceballo et al., 2019; O’Sullivan et al., 2019) or somatosensory detection (Hong et al., 2018) tasks suddenly and within 48 hours following cortical ablation.

Why then does transient inactivation produce behavioral deficits? One possibility is that disabling the auditory cortex impacts behavior not because it contributes necessary computations or information, but because of the sudden and disruptive removal of tonic excitation (Oberle et al., 2021) to downstream targets (Otchy et al., 2015) that are indispensable for successful sound detection. In this scenario, normal operation would resume once synaptic scaling (Keck et al., 2013) had homeostatically restored normal activity in these structures, a process that has been suggested to take up to 48 hours and is consistent with the time course of recovery after lesions (Ceballo et al., 2019; Hong et al., 2018). Alternatively, several circuits may redundantly support sound detection. Silencing the auditory cortex might then transiently impede sound detection until the relevant downstream decision and motor structures have updated their synaptic weights and/or processing has shifted to the other circuits. Two observations, however, argue against this possibility. First, removing one of several redundant structures should leave some residual function intact and not have the devastating effect that pharmacological cortical silencing achieves (Talwar et al., 2001, Li et a., 2017). Second, other circuits mediating the acousticomotor transformation required for successful sound detection behavior very likely incorporate subcortical auditory structures, including the auditory midbrain. Activity in the IC may trigger actions (Cassedey and Covey, 1996), such as licking, via its direct projections to the superior colliculus, pontine nuclei, the periaqueductal gray (Hufman and Henson, 1990, Wenstrup et al., 1994, Casseday and Covey, 1996; Xiong et al., 2015) and the striatum (Allen Mouse Brain Connectivity experiment 100141596) or indirectly via its projections to the auditory thalamus. If cortical lesioning results in a greater weight being placed on the activity in spared subcortical circuits for perceptual judgements, we would expect the accuracy with which trial-by-trial outcomes could be read out from IC neurons to be greater in mice without auditory cortex. However, that was not the case.

A difference was observed, however, for decoding the behavior of mice from IC activity in mice with (near-)complete vs only partial lesions of the auditory cortex. Somewhat counterintuitively, trial outcomes could be read out with lower accuracy from the mice with smaller lesions. That smaller lesions can have a more disruptive effect than larger, more complete, lesions has been observed before, in that the time it takes mice to learn a whisker-dependent sensory detection task is anticorrelated with the size of their somatosensory cortex lesion (Hong et al., 2018). While the complete destruction of a cortical area severs all its communication with downstream structures, a partial lesion may actually be more disruptive by eradicating normal local processing while at the same time leaving intact some tissue, especially in the deeper output layers, which continue to transmit what are now aberrant activity patterns. The difference in decoding accuracy that we observed in the IC could thus be a consequence of residual and now disruptive cortical input.

Our results show that behavioral variables are encoded by corticorecipient neurons in the dorsal shell of the IC independently of their main source of descending input, the auditory cortex. It therefore seems likely that this region of the auditory midbrain is part of the circuit that supports sound detection behavior in the absence of the auditory cortex. Nevertheless, except for the regions immediately bordering the auditory cortex, corticocollicular neurons located in other areas were left intact. These relatively sparse descending projections to the IC, such as those originating from somatosensory cortical areas (Lohse et al., 2021; Lesicko et al., 2016) may have contributed to the response profiles that we observed. Additional non-acoustic sensory input can reach the IC via brainstem nuclei (Lesicko et al., 2016; Shore and Zhou, 2006) and the superior colliculus (Chen et al., 2020; Coleman and Clerici, 1987). The latter, together with input from the substantia nigra (Olazabal and Moore, 1989) and the globus pallidus (Morizumi and Hattori, 1991) may also be a source of motor signals, while state changes may impact on the IC via inputs from neuromodulatory structures, including the locus coeruleus and the subparafascicular, dorsal raphe and tegmental nuclei (Chen et al., 2020, Liu et al., 2023).

Conclusion

Behavior is a major determinant of activity in the non-lemniscal auditory midbrain and thus key to understanding how it contributes to hearing. The anatomical feature that defines this structure more than any others is its connection with the auditory cortex. While modulation of IC activity by this descending projection has been implicated in various functions, most notably in the plasticity of auditory processing, we have shown that the auditory cortex is not required for the task-related activity recorded in IC neurons of mice performing a sound detection task. These results therefore emphasize more than ever the need to factor in subcortical processing when considering how the cortex contributes to sound-guided behavior.

Materials and methods

Animals

All experiments were approved by the Committee on Animal Care and Ethical Review at the University of Oxford and were licensed by the UK Home Office (Animal Scientific Procedures Act, 1986, amended in 2012). We used 22 (3 female, 19 male) Ai95 (RCL-GCaMP6f)-D (JAX 024105, Jackson Laboratories, USA), three (one female, two male) Gad2-IRES-Cre (JAX 010802, Jackson Laboratories, USA), six female Ai9 (RCL-tdT) (JAX 007909, Jackson Laboratories, USA), two female Ai95 (RCL-GCaMP6f)-D X VGAT-cre (JAX 016962, Jackson Laboratories, USA), three female Ai95 (RCL-GCaMP6f)-D X T29-1 (Camk2a-cre, JAX 005359, Jackson Laboratories, USA) and three (one male, two female) C57BL6/NTac.Cdh23 (MRC Harwell, UK) mice. All mice were 9–15 weeks old during data collection. They were maintained on a 12-h light/dark cycle and were housed at 20–24°C with a relative humidity of 45–65%.

Surgeries

For all surgical procedures, mice were premedicated with intraperitoneal injections of dexamethasone (Dexadreson, 4 mg), atropine (Atrocare, 1 mg) and carprofen (Rimadyl, 0.15 mg) before being anesthetized with isoflurane (1.5-2%) and administered with buprenorphine (Vetergesic, 1 ml/kg) postoperatively. Mice were then placed in a stereotaxic frame (Model 900LS, David Kopf Instruments, CA, USA) and their body temperature was kept constant at 37°C by the use of a heating mat and a DC temperature controller in conjunction with a temperature probe (FHC, ME, USA).

For injections in the auditory cortex of AAV1.hSyn.Cre.WPRE (Penn Vector Core), the skin over this part of the brain was shaved and an incision was made, after which three small holes were drilled (Foredom K.1070, Blackstone Industries, CT, USA) into the skull with a 0.4 mm drill bit and the virus injected using a pulled glass pipette and a custom pressure injection system. In order to express GCaMP6f or tdTomato in IC neurons that receive auditory cortical inputs, a total of 150-200 nl of AAV1.hSyn.Cre.WPRE was injected at three sites in the right auditory cortex of GCaMP6f (Ai95D) or tdTomato (Ai9) reporter mice, respectively, at depths of 450-550 μm below the brain surface. Given the anterograde transsynaptic spread properties of AAV1 (Zingg et al. 2017), this caused the expression of the desired fluorescent protein in structures that the auditory cortex projects to, including the shell of the IC (Figure 4A,B).

In order to prepare GAD2-Ires-Cre mice for the optogenetics experiments, we removed a large flap of skin over the parietal and temporal bones, partially removed the temporal muscles and performed a circular craniotomy of 3 mm diameter over each auditory cortex. We then injected a total of 500 nl of AAV5-EF1a-DIO-hChR2-EYFP, UNC Vector Core) bilaterally across 4 sites and two depths (200 and 600 μm) into the auditory cortex. Each craniotomy was covered with a circular 3 mm glass window that was attached to the edges of the skull with cyanoacrylate glue (Pattex Ultra Gel, Henkel), and the exposed skull was sealed with dental acrylic (C&B Superbond, Sun Medical, Japan) into which a custom steel bar was embedded for head fixation. Experiments commenced approximately three weeks afterwards.

The IC window implantation and cortical lesioning in the Ai95D mice were performed at least three weeks after the injections. The window implantation involved removing a flap of skin over the (inter-)parietal and occipital bone and making a circular 3 mm craniotomy over the midbrain. A 3-mm diameter glass coverslip that had been glued to a ∼1 mm tall steel cylinder with 0.5 mm wall thickness was inserted into this craniotomy. The cylinder allowed us to press the glass window gently onto the brain (in order to minimize brain movement during experiments) and was then glued to the edges of the skull. For head fixation, we embedded a custom steel plate in the dental acrylic used to seal the exposed bone.

Lesions were performed as part of the cranial window implantation surgery. In those mice undergoing lesions, we removed a slightly larger flap of skin on both sides in order to expose the temporal bone, detached and deflected and/or partly removed the temporal muscle and then made, on both sides, an elliptical craniotomy over the auditory cortex of ∼3 mm (dorsoventral) by 4 mm (rostrocaudal). The exposed tissue was then aspirated (Hong et al ., 2018) with a blunted 19 G needle connected to a suction pump (Eschmann Vp25, UK) or destroyed by thermocoagulation (Ceballo et al., 2019) with a cauterizer (Small Vessel Cauterizer Kit, FST, Germany) and the piece of skull that had been removed for the craniotomy was glued (Pattex Ultra Gel) back in place. In some of the lesioned mice, after completion of the imaging, 150 nl of a retrograde viral construct (rAAV2-CAG-tdTomato, UNC Vector Core) was injected into the dorsal IC across three sites at depths of 100-400 μm below the brain surface in order to visualize the remaining IC-projecting cortical neurons. The extent of the lesions was estimated from the histological sections and by referencing them against sections from a mouse brain atlas (Paxinos and Franklin, 2001).

In order to visualize the distribution of IC-projecting neurons in mice without cortical lesions, 150 nl of the retrograde rAAV2-CAG-cre (UNC Vector Core) construct was injected into the dorsal IC of one Ai9 mouse with an intact cortex across three sites at depths of 100-400 μm below the brain surface.

Histology

For histological processing, mice were perfused transcardially, first with phosphate buffered saline (PBS) and then with 4% paraformaldehyde in (PBS), and their brains were sectioned coronally (100 μm thick) with a vibratome (Leica). Images were taken manually using Leica DMR microscope, a confocal laser scanning microscope (Olympus FV1000) or with an automated slide scanner (Zeiss Axioscan Z1). The brain of one mouse (Figure 2 - figure supplement 1) was sectioned and imaged on a custom-built two-photon whole brain tomograph.

Click detection task

Starting 2-3 days before training commenced, the mice were habituated to head fixation in the experimental setup and their access to water was restricted to about ∼1 ml per day, bringing their body weight down to about ∼85% of the pre-restriction values. During the training phase, the mice were required to report a 0.5 ms broadband click stimulus of 80 dB SPL by licking a water spout positioned in front of them. Licking within a 1.5-second response window (occasionally this was reduced in duration to discourage excessive licking) triggered an immediate water reward (∼2 μl). Stimulus trials and catch (no stimulus) trials were randomly interleaved with an inter-trial interval drawn from a normal distribution with a mean and standard deviation of 8 s and 2 s, respectively, and a lower bound of 3 s. Successful reporting of the sound within the response window was scored a ‘hit’, while failure to respond was scored a ‘miss’. During catch trials, neither licking (‘false alarm’) nor withholding licking (‘correct rejection’) triggered a reward. To help the mice form an association between sound and reward, they received occasional ‘free’ rewards in stimulus trials during the initial training even when no licking occurred.

Once the mice had achieved a stable level of performance (typically two days with d’ > 1.5), quieter stimuli (41-71 dB SPL) were introduced. The range of sound levels was adjusted to each animal’s behavioral performance to avoid floor and ceiling effects and could, therefore, differ from mouse to mouse. The behavioral experiments were run using custom MATLAB (MathWorks) scripts interfacing with a National Instruments board (NI USB-6501) for reward delivery and lick registration. The stimuli were presented using Psychtoolbox through a free-field speaker (Vifa, Avisoft Bioacoustics, Germany), positioned about ∼15 cm from the snout of the mouse. Stimuli were calibrated using a Pettersson M500 microphone, which was itself referenced to a sound-level calibrator (Iso-Tech SLC-1356). Stimulus levels were calibrated by integrating the recorded RMS of clicks over the mouse hearing range (1-100 kHz) and comparing this to the RMS of stimuli from the reference sound-level calibrator.

In the optogenetics experiments, the behavioral task was identical except that a single sound level (80 dB SPL) was used and on 50% of the trials bilateral photostimulation (20 Hz, 10 ms pulses, 0.2 mW/mm²) was performed via two 470 nm LEDs (CREE-XP-E2, LED-Tech, Germany) positioned above the cranial windows. LED-on and LED-off trials were randomly interleaved and stimulation lasted for 700 ms starting 50 ms before trial onset. Furthermore, masking flashes were presented on all trials from two bright LEDs (60 mW) positioned a few cm in front of the animals’ eyes.

Two-photon calcium imaging

Imaging was performed at a depth of 50 μm – 150 μm from the IC surface using a commercial two-photon laser-scanning microscope (B-Scope, ThorLabs, VA, USA), a SpectraPhysics Mai-Tai eHP laser (Spectra-Physics, CA, USA) tuned to 930 nm, and a Nikon 16x 0.8 NA objective. Images were acquired with a resolution of 512 by 512 pixels at a rate of ∼28 Hz. The size of the field of view was either 500 µm by 500 µm or 666 µm by 666 µm, which allowed us to, typically, image dozens of corticorecipient IC neurons simultaneously. Each imaging session lasted around 1-2 hours.

Image processing

Rigid and non-rigid image registration, segmentation, neuropil correction and signal extraction were performed using the Python version of suite2p (Pachitariu et al. 2017). To assess the extent of image displacement in the z-axis, we compared the average of the top and the bottom 500 frames of each spatial principal component (PC) of the registered images for every 8-16 minutes of the recordings. Any region of interest (ROI) with substantial z-axis movement was excluded from further analysis. Sessions in which the majority of ROIs had to be excluded were discarded entirely. Furthermore, in order to specifically assess brain motion caused by the motor component of the task, i.e. the animal’s licking, lick-triggered movies of the imaging frames were created for every 8-16 minutes of the recordings. The rationale here is that if licking causes a stereotypical displacement of the imaging plane, this will become apparent when image sequences are averaged across lick events. Specifically, non-registered image sequences surrounding (from 2 s before to 2 s after) lick events were used to produce averaged lick-triggered movies. These lick-triggered movies, as well as non-averaged sequences, were then visually inspected and ROIs were excluded from subsequent analysis if they were affected by substantial z-motion.

Clustering analysis

To identify sub-populations of neurons with distinct response profiles, a clustering analysis was performed using an approach very similar to that described in Namboodiri et al. (2019). For each neuron, the trial-averaged activity was obtained by averaging across all the sound levels presented in a given session separately for hit and miss trials (given the small number of catch trials, approximately one tenth of all trials, this analysis was restricted to stimulus trials only). Differences in the field of view size between sessions resulted in slight differences in frame rate and thus frame duration. Therefore, the activity traces were linearly interpolated to have the same number of data points (193 frames). For each neuron, the trial-averaged activity for miss trials was appended to that for hit trials, producing 386 data points per neuron for a total of 2649 neurons (n = 1697 neurons from 38 sessions with 9 non-lesioned mice; n = 952 neurons from 37 sessions with 7 lesioned mice). To reduce the dimensionality of this dataset before applying the clustering algorithm, we performed principal components analysis (PCA) along the time axis to capture the temporal response profile for each neuron. Guided by the ‘elbow’ point in a scree plot visualizing the fraction of variance explained by each PC, we decided to project the dataset to the lower dimensional subspace formed by the first 9 PCs.

Spectral clustering was used to cluster the resulting data. The affinity matrix was constructed by computing a graph of nearest neighbors. The hyperparameters of the clustering algorithm, including the number of nearest neighbors and the number of clusters, were optimized by a grid search to maximize the mean Silhouette Score for all samples. The Silhouette Score is a measure of the compactness of individual clusters (intra-cluster distance) and the separation amongst clusters (inter-cluster distance). For a given sample i that belongs to cluster C_I, the Silhouette Score is defined as:

where a_i is the mean distance between sample i and all the other samples in the same cluster, and b_i is the mean distance of sample i to the nearest cluster that sample i is not part of. Let |C_I| and |C_J| be the number of samples belonging to clusters C_I and C_J, and d(i, j) be the distance between samples i and j; a_i and b_i are defined as:

The resulting clusters from the hyperparameter search were further examined by plotting clusters in pairs against each other with t-distributed Stochastic Neighbor Embedding, a statistical method for visualizing high-dimensional data that involves giving each data point a location in a two or three-dimensional space (van der Maaten and Hinton 2008), which showed that the clusters were generally separable.

Population decoding

Logistic regression models were trained on the network activity of each session to classify hit vs miss trials on a frame-by-frame basis, i.e., each time point (frame) of each session was trained separately. Sessions were only included if there were at least 15 instances for both hit and miss trials. The models were trained with L2 regularization, which gave similar contributions to correlated features (i.e., individual neuronal activity) instead of discarding some of the correlated features that were also related to behaviorally-relevant information. The strength of the regularization for each model was hyperparameter-tuned and the reported results were cross validated. Specifically, neuronal data in each session was split into 5 stratified folds, and each fold preserved the percentage of hit and miss trials in a given session. Four folds were used for cross-validated hyperparameter search (randomized search drawn from the log-uniform distribution between 1 × 10^-4 and 1 × 10²), and the remaining 1 fold was used for evaluating the model after the best hyperparameters were refitted on the 4 folds of data. To more reliably estimate the model results, the evaluation was done for each of the 5 folds for each session and the average of these 5 results was taken as each session’s model performance at each timepoint.

The percentage of hit and miss trials was different in each session, and the number of hit trials often exceeded the number of miss trials. To include as many trials as possible while preventing the models from taking advantage of class imbalances, balancing procedures were performed at both the model-level and the metrics-level. First, logistic regression was trained with the class weights adjusted inversely proportional to the frequency of each trial type in the training data, giving higher weights to the minority class and lower weights to the majority class. Given the total number of trials in the training data N_T, the number of classes N_C, and the number of trials for a given class N_i, the weight for a given class W_i was defined as follows:

These weights were then applied to the cost function during the training process to increase the penalty for minority class misclassifications and reduce the penalty for majority class misclassifications. Second, to avoid the estimated model performance being inflated due to class imbalance, balanced accuracy (Brodersen et al. 2010) was used to report the model performance. Balanced accuracy was defined as the arithmetic mean of the true positive rate and the true negative rate. For a model performing equally well on either class, the balanced accuracy is the same as the conventional accuracy (i.e., the number of correct predictions divided by the total number of predictions). However, for a model scoring above chance only because the model takes advantage of the class imbalance (i.e. consistently predicts the majority class), the balanced accuracy is at chance level.

Additionally, dummy models were used as baseline models to compare against the performance of the logistic regression models. Dummy models predicted the class labels (i.e., hit or miss trials) randomly while taking into account the probability of each class.

To assess whether the model performance was correlated with the number of ROIs recorded in a session, Spearman’s correlation coefficient was computed between the number of ROIs in a session and the mean model performance over different 1-second time periods relative to stimulus onset (from 2 seconds before to 5 seconds after stimulus onset).

Statistical tests were conducted to compare the model performance between lesioned and non-lesioned mice, as well as between the trained models and dummy models. Since the frame rate varied slightly with the size of the field of view, the numbers of frames (193 - 197 frames) per 7-s trial could be different across sessions. Thus, model performance was linearly interpolated to make all sessions contain the same number of frames before statistical tests were performed at each timepoint. The model performance of each session was cross-validated and averaged across folds, and the statistical tests were performed on the distributions of the sessions’ model performance. The Shapiro–Wilk test was used to determine whether a parametric or nonparametric test should be used, using p < 0.05 as a criterion. A one-sided Wilcoxon signed-rank test or paired t-test was performed for comparing trained vs dummy models, while a one-sided Mann-Whitney U test or t-test was performed for comparing trained models for different groups of mice. Because of the smaller sample sizes, the statistical tests in Figure 6B were carried out after binning the scores for every two timepoints. Statistical significance was defined as p < 0.05 after Bonferroni correction.

Data availability

Data will be made available in a public repository.

Acknowledgements

We are grateful to Christopher Breen and Robert Campbell for helping with the histology and for the financial assistance provided by an Oxford-Taiwan Graduate Scholarship from the University of Oxford and the Taiwan Ministry of Education to TYL, a Wellcome 4 year PhD Studentship to YW (102372/Z/13/Z), and by a Wellcome Principal Research Fellowship to AJK (WT108369/Z/2015/Z).

Figure supplements

Figure 2 - figure supplement 1.

Figure 2 - figure supplement 2.

Figure 5 - figure supplement 1.

Figure 5 - figure supplement 2.

Figure 6 - figure supplement 1.

Figure 6 - figure supplement 2.