Balance or imbalance: inhibitory circuits for direction selectivity in the auditory system

Abstract

The auditory system detects and processes dynamic sound information transmitted in the environment. Other than the basic acoustic parameters, such as frequency, amplitude and phase, the time-varying changes of these parameters must also be encoded in our brain. Frequency-modulated (FM) sound is socially and environmentally significant, and the direction of FM sweeps is essential for animal communication and human speech. Many auditory neurons selectively respond to the directional change of such FM signals. In the past half century, our knowledge of auditory representation and processing has been updated frequently, due to technological advancement. Recently, in vivo whole-cell voltage clamp recordings have been applied to different brain regions in sensory systems. These recordings illustrate the synaptic mechanisms underlying basic sensory information processing and provide profound insights toward our understanding of neural circuits for complex signal analysis. In this review, we summarize the major findings of direction selectivity at several key auditory regions and emphasize on the recent discoveries on the synaptic mechanisms for direction selectivity in the auditory system. We conclude this review by describing promising technical developments in dissecting neural circuits and future directions in the study of complex sound analysis.

Introduction

In the sensory world, external signals are never constant. Our brain has adapted to processing these changing cues, especially those encoding socially significant information. Such processing has been exemplified by the extensive studies in the visual system and the somatosensory system, in which the directions of moving objects and the deflection of the whiskers are well represented [1–15]. In parallel to the visual system or the somatosensory system, which can detect spatially or spectrally moving stimuli, the auditory system can exploit and represent time-varying frequency-modulated (FM) acoustic signals. FM sounds consist of several fundamental parameters: modulation range, e.g., the starting frequency and the ending frequency; modulation direction, e.g., from low frequency to high frequency (upward) or from high frequency to low frequency (downward); and modulation speed or rate, e.g., how fast is frequency changed. More complex signals could be reduced as a combination of upward FM and downward FM components as shown in Fig. 1 [16, 17].

Fig. 1.

Ultrasonic vocalization evoked by male mice (Lin and Wu, unpublished data). Sonogram of a 1.4 s vocalization recording consists of upward and downward FM sweeps. Colored arrow indicates the direction of FM. Orange upward; green downward

The direction of FM sweeps is an important acoustic cue in animal communication and human speech [18–21]. Thus, the representation of FM information has been a long-lasting focus of auditory, speech and hearing studies. Although the spectrotemporal signals in communication vary among different species, the general principles for processing time-varying frequency such as FM sweeps might be shared among them. This conclusion has been reinforced by the discovery of direction-selective (DS) neurons in rats, mice, cats, bats and other animals [22–29]. In the auditory system of these animals, many neurons respond robustly by firing action potentials to a preferred direction of FM sweeps, while few spikes are evoked by the opposite or null direction [22, 30, 31]. Such neurons have been identified at many stages of central auditory processing such as in the cochlear nuclei (CN) of the brainstem [25, 32], the inferior colliculus (IC) of the midbrain [33–35], the medial geniculate body (MGB) of the thalamus [36] and primary auditory cortex (A1) [37, 38], although the evidence of direction-selective responses in the peripheral auditory system is scarce [32].

Direction selectivity becomes enhanced, and the number of direction-selective neurons increases progressively at higher stages of central auditory processing [23, 36, 39]. To quantify the level of direction selectivity, a direction selectivity index (DSI) was widely used and calculated for the responses to the pairs of opposing directional sweeps with the same speed and intensity [23, 36, 39]. The DSI for spike, membrane potential or synaptic input responses evoked by opposing directions can be calculated as (Ru − Rd)/(Ru + Rd), where Ru is the response amplitude to upward FM sweeps and Rd is that to downward sweeps. A DSI with a positive value indicates upward direction selectivity, while a DSI with a negative value indicates downward direction selectivity. The correlation of neurons’ characteristic frequency (CF) and their DSI is also progressively better in higher levels [23, 37] (Fig. 2; Table 1). From these studies, it is reasonable to assume that direction selectivity emerges in the nuclei of the central auditory pathway, rather than in the periphery. Because experimental design and results in the study of rate/speed selectivity vary greatly in many studies, we will focus our discussion on the direction selectivity; for more systematic review of potential mechanisms of the rate/speed selectivity, other reviews could be referred [40].

Fig. 2.

Schematic drawing of the location of auditory brainstem (CN), midbrain (IC), thalamus (MGB) and auditory cortex of rats. A sagittal section of the rat brain with electrodes indicating the recording areas. FM direction selectivity (DSI) was topographically correlated with CF in the central nucleus of inferior colliculus (CNIC), the ventral area of medial geniculate body (MGBv) and the primary auditory cortex (A1), while such correlation was not obvious in the CN. Numbers in black indicate correlation coefficients and numbers in red indicate mean absolute value of DSI, which is progressively increased along the ascending central auditory pathway. DSI data of CN, CNIC and MGBv extracted from ref. 23. Asterisk for cortical data, see ref. [37]

Table 1.

Proposed mechanisms of direction selectivity

Location	Level of DS	Mechanisms
Hair cells/spiral ganglion cells/auditory nerves	Lack of or weak in ultrafast modulation	Neural adaptation; systems nonlinearity caused by the mechanical and physiological properties of the cochlea [25, 32]
Brainstem	Weak	Asymmetrical Inhibition; coincidental synaptic inputs, but no strong evidence yet [23–29]
Midbrain	Strong	Inherited from presynaptic neurons for some neurons; generated DS locally by asymmetrical inhibition [22, 23, 33–35, 40–42]
Thalamus	Strong	Unknown, lack of evidence, maybe also shaped by inhibition as IC inhibitory neurons could project to MGB [36, 39, 108–112]
Primary auditory cortex	Strong	Inherited from presynaptic neurons, shaped by inhibition, and spike thresholding [37, 57]

To reveal the synaptic circuitry mechanisms that generate direction-selective responses, several criteria need to be met. First, neurons with DS responses should be identified and targeted; second, DS neurons should not receive direction-selective inputs from presynaptic neurons. Thus, both neurons’ output responses and input response to FM sweeps should be clearly dissected and examined in sufficient detail. Not until recently, in vivo whole-cell voltage clamp recordings have targeted DS neurons and revealed their synaptic responses to FM signals. Several studies suggest in the IC and the A1 that some neurons’ direction selectivity is inherited from presynaptic neurons and sharpened by local inhibition [37, 41]. Most recent computational simulation and electrophysiological experiments on rodent IC demonstrated DS could be constructed by the synaptic receptive field properties [23, 42]. They shed light on the function of inhibitory circuits on creating feature selectivity in the sensory system.

In this review, we will begin with two hypothetical models for creating direction selectivity in the auditory system [Box 1]. Then, evidence to support or undermine the proposed models will be discussed for different auditory processing centers along the central auditory pathway. We will conclude by discussing questions that remain to be answered and future directions for experimental analysis. For the sake of conciseness, we limit this review largely to auditory studies in rodents, although a very significant amount of work has been done in other animal models such as cats and bats (reviewed in [40, 43–45]).

Box 1: theoretical models for direction selectivity in the auditory system.

A progressive increase of direction-selective neurons has been observed along the central auditory pathway (as summarized in Fig. 2). Neurons might acquire their direction selectivity from their presynaptic neurons as demonstrated in the auditory cortex [37]. However, the essential question is how neurons receiving non-selective inputs become direction selective at the subcortical auditory nuclei. Two principal mechanisms have been hypothesized to explain the generation of FM direction selectivity in the auditory system [30, 41] (Fig. 3). Both hypotheses are based on two common assumptions that (1) the presynaptic neurons need not to show direction selectivity; (2) the timing of synaptic inputs plays a crucial role in generating direction selectivity. To differentiate the two models, the first one is referred to as ‘asymmetrical inhibition model’ and the second one as ‘coincidental excitation model’.

Fig. 3.

Schematic drawing of asymmetrical inhibition model for direction selectivity. a A proposed frequency–intensity receptive field of synaptic inputs for an upward direction-selective neuron. Red, receptive field of excitatory inputs; blue, that of inhibitory inputs. Arrows indicate the direction of FM sweeps. Orange upward; green downward. b Synaptic receptive fields for a downward direction-selective neuron. c Synaptic inputs and membrane potential outputs of an upward direction-selective neuron evoked by FM sweeps. Upper panel EPSC (red) and IPSC (blue) evoked by FM sweeps. Dashed lines onset latency of EPSC and IPSC; lower panel membrane potential changes evoked by FM sweeps. Dashed line spiking threshold. d Synaptic inputs and membrane potential outputs of a downward direction-selective neuron evoked by FM sweeps

In the ‘asymmetrical inhibition model’, inhibitory inputs and synaptic frequency–intensity receptive fields are taken into account for the determining factors. A lateral inhibitory sideband could appear next to the excitatory synaptic receptive field either in the high-frequency domain or in the low-frequency domain (Fig. 3a, b) [46–52]. As FM sweeps can be regarded as a decomposition with individual frequency components presented in a continuous fashion, in the preferred direction (upward sweep, shown in Fig. 3a; or downward sweep in Fig. 3b), the early-frequency components in the FM sweeps will enter into the excitatory synaptic input receptive field first and initiate large depolarization (Fig. 3c, d, left panels). Although the later frequency components enter into the inhibitory synaptic input receptive field and generate large hyperpolarization, it lags behind the depolarization. Thus, the early depolarization could not be suppressed and eventually suprathreshold responses could occur. In the null direction (downward sweep, Fig. 3a; or upward sweep in Fig. 3b), the early-frequency components enter into the inhibitory receptive field first and initiate an early hyperpolarization. Although the later-frequency components enter into the excitatory synaptic receptive field and recruit more excitatory inputs, they cannot depolarize the cell membrane sufficiently to counteract the early hyperpolarization (Fig. 3c, d, right panels). Thus, suprathreshold responses could not be achieved. In this model, both excitatory postsynaptic currents (EPSC) and inhibitory postsynaptic currents (IPSC) are not selective to either direction, but their latency creates a temporal asymmetry to allow integration with opposing effects: in the preferred direction, EPSC leads, while IPSC leads in the null direction. Such temporal asymmetry could result from the non-overlapping and side-by-side excitatory and inhibitory synaptic receptive fields as shown in Fig. 3a, b. Recently, in vivo whole-cell voltage clamp recording has been applied to reveal the patterns of both excitatory and inhibitory synaptic inputs and construct their synaptic frequency–intensity receptive fields. These studies suggest that most auditory neurons receive approximately balanced excitatory and inhibitory synaptic inputs [37, 53, 54]. Even those neurons with unbalanced excitatory and inhibitory synaptic inputs have large overlapped excitatory and inhibitory synaptic receptive fields [23, 55]. Thus, Fig. 3a, b only represents an extreme case of the configuration of synaptic receptive fields as the excitatory and inhibitory synaptic receptive fields are never completely separated in the frequency domain.

In the ‘coincidental input model’, primary consideration is given to excitatory inputs rather than inhibitory inputs. In this model, the frequency components of FM sweeps sequentially evoke individual inputs. An internal delay line mechanism compensates the arrival time of each input and allows the coincidental occurrence and the summation of these inputs evoked by different frequency channels. For excitatory synaptic inputs evoked by preferred direction, their net arrival time is coincidental (Fig. 4a, b). Under these circumstances, the preferred direction evokes a large excitatory postsynaptic potential (EPSP) and results in action potential. In the null direction, the internal delay line amplifies the difference of the arrival time of excitatory inputs (Fig. 4c, d). Thus, the summation of excitatory inputs over the time could not generate depolarization large enough to evoke spikes. Alternatively, a delay line could be brought up from inhibitory connections. In the preferred direction, such mechanism amplifies the timing difference of inhibitory inputs; in the null direction, it compensates the arrival time to create large hyperpolarization to suppress the responses of the neuron. This model is based on temporal summation of excitation/inhibition and dendritic integration of synaptic inputs; however, experimental data to support this hypothesis is scarce. There are several difficulties in testing the validity of this model: first, the individual inputs evoked by individual frequency components need to be isolated in vivo; second, the stimulation has to be well controlled in temporal domain, as each frequency component only represents a fairly short time point in FM sweeps, especially when the speed or rate of modulation is high; third, as the frequency changes sequentially in a relatively fast time window, neural adaptations could introduce inevitable nonlinearity to neural responses to nearby frequency components.

Fig. 4.

Schematic drawing of coincidental excitation model. a Synaptic inputs and membrane potential outputs evoked by FM sweeps in preferred directions. f ₀, f ₁,…f _n indicate individual frequency components in the FM sweeps. Each of them evokes excitatory inputs with varying latencies. Blue blocks represent internal delay lines that compensate the latencies. b Individual excitatory inputs (red) after timing compensation arrive coincidentally at the postsynaptic neurons, which evoke large depolarization of membrane potential (blue trace). c Synaptic inputs and membrane potential outputs evoked by FM sweeps in null directions. d Individual excitatory inputs (red) after further delayed timing, which evoke smaller depolarization of membrane potential (blue trace)

Recently, computational modeling has been applied to predict neuronal responses to FM sweeps from their responses to pure tones of different frequencies [23, 42]. Interestingly, the onset latency of synaptic inputs evoked by pure tones does not follow a systematic delay that could be correlated with a sequential change of the frequency as suggested by a direct measurement of synaptic inputs. Moreover, the profiles of synaptic inputs evoked by FM sweeps in both directions are similar. If this model stands out, a large synaptic input and scattered inputs should be observed respectively when a preferred and a null direction of FM sweeps are presented. Although such data undermine the coincidental model, there remains the possibility that only a subset of frequency components might actually evoke synaptic inputs and the arrival time of these inputs could still be compensated, especially when we consider the nonlinear neuronal responses evoked by nearby similar frequencies, such as two-tone facilitation/inhibition. Other than these two classical models emphasizing the synaptic inputs, the examination of spike outputs suggests a larger DSI than the DSI of EPSP or EPSC [23, 41, 56, 57]. It could be related to the contribution of the spike threshold to the direction selectivity. Spike threshold has sharpening effects on the direction selectivity of neurons in the IC and the A1. However, such effects could be considered more as an ‘amplifier’ rather than a ‘generator’ of feature selectivity.

Experimental methods dissecting synaptic inputs

To illustrate the neural circuitry mechanisms for direction selectivity to FM sweeps, especially for examining the validity of both aforementioned models, it is crucial to acquire synaptic inputs to the identified DS neurons that receive non-selective inputs and directly examine the spectrotemporal patterns of both their excitatory and inhibitory inputs in sufficient detail. Earlier studies using extracellular recordings have obtained invaluable information about the spiking responses of the DS neurons and incorporate reverse correlation techniques to understand the transfer function underlying FM sweeps [58–60]. As neurons’ responses are often nonlinear and the stimuli are often limited, it is not practical to probe the operational rules of neural circuits. An alternative way is to investigate subthreshold responses, which may improve the understanding of neural connectivity and function. To measure subthreshold membrane responses, sharp electrodes with high impedances can record depolarization or hyperpolarization of the cell’s membrane potential (EPSP or IPSP, inhibitory postsynaptic potential) evoked by sound [61–64]. Such recordings presumably reflect the overall effects of excitatory or inhibitory inputs to membrane potential change. This method using sharp electrodes, however, cannot fully assess the strength of the inputs. In DS neurons, membrane potential is not only determined by synaptic inputs, but also affected by many voltage-dependent conductances within the cell. Recordings acquired from in vivo whole-cell current clamp recording mode, based on the assumptions of an isopotential cell and a linear relationship between membrane potential change and synaptic input currents, can be used to derive excitatory and inhibitory input conductances [4, 65–68]. However, the lack of control of voltage-gated channels using current clamp results in nonlinearity of the neurons associated with instantaneous voltage-dependent varying conductances and the contamination of synaptic conductances. Not until recently have voltage clamp recordings been conducted in sensory systems including various auditory neurons along the auditory pathway in vivo [37, 53, 54, 69]. By clamping the cell’s membrane potential at different voltages, voltage-dependent conductances are minimized. In voltage clamp mode of in vivo whole-cell recording, excitatory inputs are measured by clamping the neuron’s membrane potential at −70 mV, which is close to the reversal potential for GABA_A receptors, whereas inhibitory inputs are recorded at 0 mV holding potential, which is the reversal potential for glutamate receptor-mediated currents. With this technique, excitatory and inhibitory synaptic inputs could be isolated, which increased our knowledge on neural circuits underlying sound processing.

These studies provide considerable evidence to the asymmetrical inhibition model for how direction selectivity is generated and subsequently modulated or sharpened. However, under these somatic voltage clamp recordings, the deviation of real neurons from isopotentiality due to space clamp error and cable attenuation for synaptic inputs at the distal dendrites [70, 71] should be cautioned for data interpretation, as extensively discussed in several recent studies [53, 54]. In reality, distal dendrites may not be clamped at the same potential as the cell body, thus excitatory and inhibitory synaptic inputs on these dendrites can be underestimated and distorted [71]. It is difficult to determine the effectiveness of space clamping in the distal dendrites of a neuron recorded in vivo. A thorough examination of this issue will depend on the knowledge of the fine dendritic organization of the recorded neuron, the detailed location of synaptic inputs on the dendrites and the combination of these electrophysiological data with subcellular imaging and computational modeling [53, 72]. As space clamp or cable effects could lead to quantitative errors for estimating the real parameters of synaptic conductances [71], the conclusions drawn from these data should not be heavily based on the absolute values. For qualitative measurements such as synaptic receptive fields from the same cortical or subcortical neurons, as well as relative onset latency between excitatory and inhibitory inputs, they are less susceptible to be affected [53, 55]. However, to ensure optimal assessments, several strategies still need to be implemented to reduce the errors to certain extent [70, 71, 73]: (1) obtaining high-quality recordings with compensated and comparably low effective series resistance (10–20 MΩ); (2) examining the derived reversal potential of evoked excitatory currents around 0 mV (no early-phase inward components should be observed); (3) assuring that the linearity of synaptic IV curves that could also suggest synaptic conductances are not strongly affected by nonlinearities of neurons; (4) including cesium, TEA, QX-314 in intracellular solutions and inducing anesthesia by ketamine to block most voltage-dependent currents [70, 71]. Together, although somatic voltage clamp recording provided essentially new and valuable information on synaptic mechanisms underlying functional responses of auditory neurons, quantitative measurements of synaptic conductances in vivo can only be achieved by incorporating other techniques and detailed analysis. A more detailed review of techniques, potential limitations and future directions for in vivo voltage clamp approaches in the auditory system can be found elsewhere [74].

Basic frequency processing in the peripheral auditory system

Frequency analysis is the fundamental function of the auditory system. In the cochlea, inner hair cells transduce mechanical sound waves into the release of neurotransmitters, which in turn activate the spiral ganglia neurons. Cochlear nerve consisting of the axons of the spiral ganglia neurons transmits the auditory information from the peripheral auditory system to the central pathway. The cochlea is far from a passive frequency detector and analyzer, as it receives efferent innervation from central auditory structures. Whether the efferent pathway has any contribution to FM processing of inner hair cells is unclear, as no in vivo cellular-level recording data of hair cells have been achieved so far. Each Type I spiral ganglion neuron only innervates one single inner hair cell; therefore, the firing pattern of individual auditory nerve fibers is generally used to represent the overall function of the peripheral auditory system. Discharge patterns of auditory nerve fibers in response to FM sweeps have been recorded from cats [25, 32]. Upward direction generally evoked similar firing patterns in comparison with the downward direction. This suggested the lack of direction selectivity of peripheral auditory system as all the recorded units have DSI less than ±0.33, a criterion for direction selectivity. Nevertheless, most of the opposing FM-evoked responses were ‘mirror symmetrical’, and a subtle difference was observed when varying certain parameters, such as speed (or rate) or intensity. At lower intensity or slow rate, FM-evoked responses were ‘perfectly’ symmetrical to both up and down directions. At higher intensity or fast rate, however, such responses seemed moderately asymmetrical. This could be a reflection of the neural adaptation and system nonlinearity, or the mechanical and physiological properties of the cochlea [32].

Such scenarios evoking imperfect symmetrical auditory nerve responses might seldom occur in natural environment around the animals or rarely carry significant social signals. To some extent, responses of auditory nerve fibers to FM sweeps can be understood by considering linear filtering, which could be manifested by the tuning curves of frequency–threshold or frequency–intensity receptive fields [25, 32]. However, such consideration of linear modeling cannot explain nonlinear processing such as demodulation of the FM signals [75]. Inhibition plays a crucial role in shaping neuronal responses in the central auditory pathway [76–79]. Lack of inhibitory neurons in the peripheral system limits the possibility of neural circuitry mechanisms for creating feature selectivity for more complex signals. Conservatively, the evidence obtained from peripheral auditory pathway so far implies that DS neurons have to be constructed somewhere centrally [25, 35, 80–82]. We will now focus on three major stages of auditory processing to illustrate the direction selectivity in these central nuclei.

Brainstem, debut of inhibition and heterogeneous responses

Primary-like neuronal responses are relayed to the dorsal and ventral cochlear nuclei (DCN and VCN) in the brainstem, the first station of central auditory processing. Although the input to the CN is generally considered not sensitive to direction of FM sweeps, the anatomical and physiological diversity of neuronal types are thought to reflect substantial frequency processing in the cochlear nuclei, including FM [25, 26, 83, 84]. However, whether CN is the primary location constructing DS is somewhat controversial: neurons with asymmetrical discharge patterns to ascending and descending portions of FM signals were found in the CN in cats [25, 26, 85], but they were not prominent in the CN of bats or rats [23, 24, 27–29]. Such contradicting observations in different species are not yet fully understood. For example, the testing frequency range of FM sweeps, the linear vs. logarithmic profile of FM and the intensity and speed of stimulation could all compromise the consistency of measurement across species [40, 42]. Also, the neural circuits in the brainstem could be different ethologically, such as the proposed circuits for sound localization in mammals versus avians [86]. To fully address the discrepancy, a series of carefully designed experiments with a focus on the same species might bring a more systematic view toward FM processing, the fundamental function of the auditory system to some extent.

A prominent feature in the CN is that inhibitory neurons are abundant [87, 88]. These inhibitory neurons are thought to play an important role in shaping neuronal responses [76–79]. Moreover, the heterogeneous cell types in the CN suggest that extensive transformation of response could happen [83, 84]. In cat studies, primary-like neurons mimicked the response of auditory nerve fibers and did not show selectivity to the directions of FM sweeps. Build-up, onset and pauser neurons show symmetrical responses to both directions except the highest modulation speed [25], while nonmonotonic onset cells show unidirectional response. As the auditory nerve fibers themselves could show a minimal DS in some extreme cases, the weak DS observed in CN might be merely inherited from the peripheral system. Recordings in rats’ CN only demonstrated a negligible DS well below the criteria for strong direction selectivity [23]. Due to the difficulty in accessing the CN, in vivo intracellular recordings were somehow limited in contrast to the many studies in the IC and the A1. A recent study showed that the excitatory and inhibitory inputs of nonmonotonic cells were not balanced very well in the CN, which resulted in the generation of intensity selectivity [79]. Although no excitatory and inhibitory inputs and their synaptic receptive fields of there CN neurons with weak direction selectivity were presented in previous studies, it is possible that the imbalanced inhibition could also exist and sharpen or create direction selectivity at some level. Considering the diversity of neuronal types in the CN, it is also worth noting that the frequency range of FM sweeps should also be wide enough to cover the whole receptive fields of synaptic inputs of the tested neurons, as diversity of receptive field size and excitatory/inhibitory configuration exist among different neurons. Further examination of synaptic receptive fields of CN neurons and synaptic inputs and spike outputs of CN neurons will be needed to convince the existence of direction-selective neurons in the CN.

Is direction selectivity generated in the midbrain?

The inferior colliculus (IC), a key auditory processing center in the midbrain, is likely to be essential for transforming and representing behaviorally significant vocalization. As FM is an important component of animal communication, the IC has become a focus for elucidating critical mechanisms for processing FM sweeps. Previous studies suggest that the IC is the major processing center where direction selectivity is constructed, because most of the cells in the “lower” brainstem auditory nuclei are not direction selective, especially in rats [24, 82]. The direction selectivity of FM sweeps has been studied in the IC of various species [22, 23, 33–35, 40, 42]. Such knowledge is enriched by investigations in bats, as a higher proportion of direction-selective neurons have been found in the IC and the cortex of bats [28, 41, 89–91]. This prominence of DS neurons could be due to the evolutional and functional significance of using FMs in their echolocation system and communication system [40, 92]. Neurons are biased with downward FM selectivity in bats [30, 91, 93, 94], while the ratio of upward direction-selective neurons to downward direction-selective neurons is different in other species. In the IC or the auditory cortex of ferrets or rats, neurons selective to upward or downward direction are roughly equally abundant [35, 37, 95, 96]. Even for the same species, some results are not always comparable. For example, the percentage of DS neurons in the rat’s IC has been addressed by several studies that used different FM stimuli and yielded different results, ranging from 10 to 80 % [34, 82, 95, 97]. The functional significance behind such differences across and within various species is still unclear, as it could be due to the varying strengths of projections from different nuclei in the brainstem to the IC among species, the experimental method in designing the acoustic stimulation or recording techniques [98].

Recently, in vivo whole-cell recordings, especially voltage clamp, have been applied to the IC of rats, bats and mice [23, 41, 42]. It allows a more careful dissection of synaptic inputs or membrane potentials. In the study of Gittelman and colleagues, they derived synaptic conductances from the membrane potential changes recorded under the current clamp mode [41]. Although such analysis could estimate synaptic inputs to some extent, the nonlinearity of the biophysical properties of IC neurons might generate uncertain inaccuracy at some level [99]. Based on their experimental procedures and presented data, a larger number of second-order neurons inheriting direction selectivity were encountered in the bat’s IC, which suggested that direction selectivity presented in EPSP or IPSP. This could reflect an upstream processing that is inherited by IC neurons or a nonlinear integration of synaptic inputs. If DS is inherited from upstream processing stages, the likely candidates worthy of examining are neurons in the medial superior olive (MSO) and the lateral superior olive (LSO) in the brainstem or other CNIC neurons. Evidence suggests that contralateral DCN and VCN are major excitatory sources of projection to IC with minimal DS, while interestingly there have been few studies targeting LSO and MSO for their FM responses [25, 26, 85]. Although most of the studies were conducted through contralateral stimulation, a recent study on mice demonstrated that only a limited number of neurons were direction selective to ipsilateral FM sweeps [42]. These data somehow undermine the significant contribution of MSO and LSO to the DS observed in the IC.

Inhibition has been suggested to play a significant role in the direction selectivity and tuning curves of IC neurons: when inhibition is blocked by GABA and glycine receptor antagonists bicuculline and strychnine, respectively, the tuning curve is expanded, direction selectivity is abolished and the selectivity to vocalization is reduced [81, 93, 100–102]. Interestingly, by manipulating the acoustic stimulation of band-pass FM sweeps with various starting frequencies, FM sweeps excluding the frequency domain covering the inhibitory sidebands can evoke neuronal responses in the null direction and eliminate direction selectivity [40]. These results indirectly demonstrate asymmetrical inhibition’s role in DS generation. In recent studies, the FM stimulation covering most of the hearing range of the animal was used, as demonstrated by the receptive fields for each recorded DS neuron, which was more efficient to search for DS neurons in different brain areas [23, 37, 42]. In rat CNIC, revealed by in vivo voltage clamp recordings, most of the recorded neurons showed similar synaptic inputs evoked by the opposing directions of FM sweeps, with minimal DSI, while these neuron’s membrane potential change showed a large DSI [23]. These neurons are less likely to inherit DS from their upstream sources and are fully capable of converting non-selective inputs into selective outputs to the direction of FM sweeps. When looking into the mechanisms, it is noted that the coincidental EPSC or IPSC is not observed; otherwise, the EPSC will be integrated in a smaller time window in the preferred direction, which will result in different temporal profile of synaptic input in response to opposing directions. Instead, FM sweep-evoked excitatory and inhibitory synaptic inputs differ in their relative latency that favors the asymmetrical inhibition model. In the same study, at optimal speeds, excitatory inputs were led in the preferred direction, while inhibitory inputs were led in the null direction [23]. Interestingly, it also showed that at non-optimal speeds, the excitatory inputs were similar for both sweep directions, but the inhibitory inputs were more scattered or less coincidental and not able to strongly suppress excitation evoked by both directions, resulting in a decreased DSI. Several mechanisms have been proposed for rate/speed selectivity including duration tuning, delayed high-frequency inhibition and asymmetrical facilitation [40]. Such observation might provide another insight toward our understanding of the mechanisms underlying rate/speed selectivity of auditory neurons, as non-optimal speed desynchronizes the individual inhibitory inputs.

Patterns of synaptic inputs can be largely reflected by their frequency–intensity tonal receptive fields. These patterns represent basic structural properties of neural circuitry underlying the function of individual neurons. As the temporal relationship between excitation and inhibition evoked by FM sweeps was hypothesized to link with their receptive fields, tuning curves or receptive fields were constantly obtained in the studies of DS neurons [34, 57, 91, 103–105]. Most of the studies so far only estimate such patterns from spiking responses obtained from two-tone paradigm or from evaluating membrane potentials. In vivo whole-cell voltage clamp recording provides a powerful way to dissect the excitatory and inhibitory inputs. Together with acoustic signals varying in frequency and intensity, a much clearer synaptic input map for both excitation and inhibition could be acquired [23]. For the DS neurons in the IC, the frequency–intensity synaptic receptive fields of excitatory and inhibitory inputs do not overlap as observed in the cortex and other areas [106]. Instead, much broader inhibitory input receptive fields with extended frequency domains were observed, e.g., for upward selective and low CF neurons, inhibitory sidebands flank in the higher frequency domain to the excitatory receptive fields, and vice versa for the downward selective and high CF neurons. The correlation between DSI and CF is also shown in Fig. 2, which is enhanced along the ascending pathway [23]. When constructing the FM responses from the inputs evoked by individual frequency components in the receptive fields, the predicted latencies of excitation and inhibition well correlated with the experimental data evoked by actual FM sweeps [23, 42]. These data provide strong supporting evidence for the asymmetrical inhibition model for the generation of DS. A recent study by Geis and Borst focused on the dorsal cortex of mouse IC to study FM responses; their electrophysiology experiment and computational modeling suggest that multiple mechanisms could be involved in generating DS in the dorsal cortex of IC, although most of neurons’ DS selectivity could be generated by asymmetric inhibition [42]. Overall, the inhibitory inputs with broader frequency-intensity receptive fields and asymmetrical temporal patterns evoked by FM sweeps indicate that imbalanced inhibition is crucial for the emergence of feature selectivity and functional topography. Although lower auditory stages showed a minimal number of DS units, they might share the same mechanisms to create direction selectivity as in the IC neurons, because they should receive direction-non-selective inputs from the auditory nerve fibers, and inhibitory neurons in the cochlear nuclei are abundant [32, 107].

Thalamus and auditory cortex: further sharpening

Although only limited studies have contributed to our understanding of thalamic processing of FM sweeps, evidences support neural sensitivity for FM stimuli in the cat, bat and rat MGB [36, 39, 108–112]. The strength or level of direction selectivity in the MGB is somehow between that of IC and A1 (Fig. 2). Many inhibitory neurons in the IC can also be projection neurons and send inhibitory inputs to MGBv [113–115]. This suggests that the observed higher level of direction selectivity in MGB compared to IC could be due to the inhibitory projection circuits in the MGB, which may further sharpen the direction selectivity. However, due to technical difficulties, there has been no in vivo whole-cell voltage clamp recording that directly measures synaptic inputs in the MGB neurons so far. It is not affirmative to say MGB neurons only inherit direction selectivity from the IC, or that MGB neurons could construct DS again. Thus, the frequency–intensity receptive fields of synaptic inputs may or may not be the determinant of DS in the MGB. Studies in the A1 have shown the direction selectivity among different species. Zhang et al. showed that the direction selectivity of cortical neurons was inherited from their excitatory inputs and shaped by cortical inhibition, and its topography highly correlated with the tonotopic map [37]. How neural circuits shape or strengthen direction selectivity in the cortex drew a large interest. In their studies, FM sweeps evoked temporally reversed excitatory and inhibitory inputs to the opposing directions, and excitatory and inhibitory synaptic receptive fields were overlapped and marked by covaried tone-evoked excitatory and inhibitory synaptic inputs [37, 53]. However, both the excitatory and inhibitory synaptic receptive fields were skewed: for low-frequency neurons, weaker excitatory and inhibitory inputs occupied a large high-frequency region within the receptive field and vice versa for the high-frequency neurons. For these primary auditory cortical neurons, the sharpening of direction selectivity can be attributed to such temporal asymmetry and skewed pattern of their synaptic frequency-intensity receptive fields [37]. This balanced excitation and inhibition suggests a feedforward inhibitory circuit: the presynaptic GABAergic neurons may be innervated by the same set of thalamocortical afferents as the recorded A1 cell, which is similar to previously proposed circuitry for other sensory cortices [116]. Until recently, imbalanced inhibition had not been observed for normal sensory processing [23, 55]. Recordings of cortical intensity-selective neurons demonstrated that temporally imbalanced inhibition sharpened the intensity selectivity that was inherited from afferent inputs, although the excitatory and inhibitory synaptic TRFs frequency-intensity receptive fields were still overlapped [55]. These studies revealed that imbalanced inhibition is prominent in both cortical and subcortical nuclei to a much larger extent. It is reasonable to posit that to create direction selectivity, instead of only sharpening the inherited direction selectivity, a much stronger inhibition must be entrained.

Future directions

The proposed mechanisms for generating or shaping direction selectivity to FM sweeps are summarized in Table 1. In the past several decades, with the advancement of emerging technology, we have witnessed tremendous progress in our understanding of the communication and auditory processing. Extracellular recordings allowed extensive characterization of neurons’ responses to direction and rate of FM sweeps along the entire auditory pathway. Intracellular recordings shed light on the synaptic mechanisms for the generation and sharpening of FM in the auditory system. Pharmacological application demonstrated that inhibition was the leading factor for FM selectivity. However, there are still many mysteries unknown to us.

1. Source of inhibition. Although evidence has suggested that the IC might be the primary location for creating direction selectivity of FM sweeps by inhibitory circuits, next to nothing is known about the structure and source of these unbalanced inhibitions [23, 40, 42, 43]. IC neurons receive inhibition from heterogeneous sources, including local interneurons and inhibitory neurons from the ventral and dorsal nuclei of the lateral lemniscus (VNLL and DNLL) [98]. In contrast to the balanced inhibition in the auditory cortex, which shows that a feedforward circuit consists of local interneurons, afferent inhibitory projections to IC neurons could be in concert with such local inhibitory circuits to provide inhibition in the same frequency domain; or, their inhibitory outputs might not occur in the same frequency domain as the excitatory outputs from CN neurons. IC is also the major recipient structure for corticotectal efferent pathways [117, 118]. Inhibitory neurons could be activated by such efferent innervation from layer 5 neurons of the auditory cortex [98, 119]. Whether such efferent innervation is also in concert with afferent innervation in the frequency domain is unknown. A realistic investigation will rely on accurately isolating the inhibitory sources when performing in vivo voltage clamp recording. Pharmacological application might not the best resort, due to its lack of specificity. The burgeoning optogenetic tools available for activating and inactivating neural circuits provide a promising direction to look at the contribution from different inhibitory sources to the synaptic receptive fields of DS neurons and their responses to FM sweeps [120, 121].

2. Cell types. Most of the recorded neurons by in vivo voltage clamp whole-cell recording demonstrated a bias toward the large flat/disc-shaped neurons; they are presumably excitatory projection neurons to MGB [23, 122]. However, at least three major types of neurons exist in the IC, large excitatory projection neurons, inhibitory projection neurons and smaller inhibitory local neurons [115]. Lack of anatomical studies combined with electrophysiological examination of DS neurons also hindered our understanding toward the specific type of neurons and their DS. The other morphologically distinct neurons, stellate or multipolar cells, are proposed to integrate frequency from multiple isofrequency laminar in the IC and might also be good candidates to provide neuronal responses over an extended frequency domain [123]. To obtain a complete understanding of the neural circuits for FM processing, it is essential to know the functional and structural properties of inhibitory neurons themselves. In several recent studies, in vivo whole-cell recordings have been combined with post hoc histology to identify inhibitory neurons in the cortex [124, 125]. However, it is very difficult to monitor inhibitory neurons specifically in blind recordings. Recent development in two-photon imaging of Ca²⁺ responses [126–128] and two-photon imaging guided targeted recording [129–132] allows efficient examinations of functional properties of inhibitory neurons in transgenic mouse cortex. Considering the easier access of mouse IC, we believe the same techniques could also be applied to the investigation of IC neurons in vivo. Moreover, such technique allows an examination of a population of neurons simultaneously. In this case, different neurons’ responses to FM sweeps could be dynamically monitored. It will shed light on the basic principles of how neuronal ensemble functions to complex sound or in the natural acoustic environment. Thus, our understanding toward both excitatory and inhibitory neurons in the IC will be largely enriched by the development of imaging and genetic tools. To construct a realistic model of neural circuits for FM responses, we will need to examine how excitatory and inhibitory neurons project to a single neuron. The recent development of trans-synaptic labeling [133] will allow neurons presynaptic to a transfected neuron to be labeled. Ca²⁺ imaging or targeted recording can then be performed on these neurons to understand their functional relationship.

3. Neural circuits underlying FM processing. Finally, a realistic model of neural circuits that could generate feature selectivity including DS will be beneficial to many areas of application and research, e.g., artificial intelligence, human–machine interface, and hearing aid. From our efforts toward the neural representation of FM sweeps, especially the DS of auditory neurons, we could propose a schematic model circuits to mimic the possible mechanisms that create DS as shown in Fig. 5, based on the anatomical, histological and functional organization of the IC. One possible circuit is that IC neurons receive excitatory inputs from a group of CN neurons with similar CF; however, it receives inhibitory inputs from a larger pool of VNLL neurons with different CF’s, which provide inhibition in a much broader frequency domain. Alternatively, local inhibitory neurons in the IC integrate a larger pool of excitatory innervation and send feedforward inhibition to the IC projection neurons (Fig. 5). As mentioned in the previous session, to fully uncover the fine structure of neural circuits for FM selectivity, a variety of innovative techniques will have to be in place to examine the function and anatomy of neurons in detail in the next decade. Over the past several decades, the mechanisms behind it seem clearer and clearer. When we look forward, the renovations of electrophysiology, imaging and genetic manipulations will allow a reverse engineering of neural circuits that could function to extract external information efficiently.

Fig. 5.

A schematic drawing of potential circuits for direction selectivity. IC is characterized by isofrequency laminae with low-frequency laminar located dorsolaterally and high frequency ventromedially. Among flat cells in blue, Cell I represents upward direction-selective neuron; Cell II, non-selective neuron; Cell III, downward direction-selective neuron. Cell I has excitatory synaptic receptive fields similar to the laminar it resides in and receives inhibitory inputs from inhibitory neurons with higher-frequency responding domain (multipolar cells in orange). Cell II has excitatory synaptic receptive fields similar to the laminar as well, but receives inhibitory inputs either confined within the laminar or with both higher- and lower-frequency responding domain (symmetrical inhibition). Cell III has excitatory synaptic receptive fields similar to the laminar and receives inhibitory inputs from multipolar cells with lower-frequency responding domain