From elementary synaptic circuits to information processing in primary auditory cortex

Abstract

A key for understanding how information is processed in the cortex is to unravel the dauntingly complex cortical neural circuitry. Recent technical innovations, in particular the in vivo whole-cell voltage-clamp recording techniques, make it possible to directly dissect the excitatory and inhibitory inputs underlying an individual cortical neuron’s processing function. This method provides an essential complement to conventional approaches, with which the transfer functions of the neural system are derived by correlating neuronal spike outputs to sensory inputs. Here, we intend to introduce a potentially systematic strategy for resolving the structure of functional synaptic circuits. As complex circuits can be built upon elementary modules, the primary focus of this strategy is to identify elementary synaptic circuits and determine how these circuit units contribute to specific processing functions. This review will summarize recent studies on functional synaptic circuits in the primary auditory cortex, comment on existing experimental techniques for in vivo circuitry studies, and provide a perspective on immediate future directions.

1. Introduction

Sound is a time-varying signal, specified by frequency and intensity distribution in a time domain. How cortical neurons represent these physical features of sound has been intensively investigated (e.g. Aitkin, 1990; Merzenich and Schriner, 1992; Woolsey, 1960; Goldstein and Knight, 1980; Brugge and Reale, 1985; Clarey et al., 1992; Winer, 1992; de Ribaupierre, 1997; Rouiller, 1997). Neuronal functional properties in auditory cortex such as frequency tuning, intensity tuning and tonotopic organization are highly preserved among different species (mouse: Stiebler et al., 1997; Linden et al., 2003; Bandyopadhyay, et al., 2010; Rothschild, et al., 2010; rat: Horikawa et al., 1988; Polley et al., 2007; cat: Merzenich et al., 1975; Reale and Imig, 1980; monkey: Imig et al., 1977; Merzenich and Brugge, 1973; Kaas and Hackett, 2000; human: Formisano et al., 2003; Humphries et al., 2010), suggesting that the underlying neural circuits in different species are organized under common principles. In this review, we will focus on the primary auditory cortex (A1) of rats to discuss the recent progress on the synaptic circuitry basis for auditory cortical processing.

2. A reductive approach to unraveling neural circuitry

2.1. Realistic neuronal circuits

Cortical neurons are well organized into ensembles or circuits to process auditory information (Schreiner et al., 2000; Douglas and Martin, 2004; Read et al., 2002; Winer and Lee, 2007). Since neurons communicate with each other through synaptic connections, the ensemble of synaptic strengths can define the connectivity within a neural circuit (Fig. 1A). However, the neural circuitry in the brain is dauntingly complex. The human brain is estimated to contain about 10¹¹ neurons (Azevedo et al., 2009). Each neuron makes thousands to tens of thousands of synaptic connections with other neurons, which brings the total number of synaptic connections to a level of 10¹⁵. In addition, the intricately intermingled axonal and dendritic processes of neurons make it extremely difficult to trace individual connections. The diagnostic method used in dissecting electronic circuits by removing specific components cannot be easily applied to neural circuits. Finally, the heterogeneity of neuronal populations further adds to the complexity of neural circuits. To overcome these obstacles, a well-defined systematic approach has to be developed to reduce the complexity of neural circuits, which will facilitate our understanding of their function.

Fig. 1. Schematic diagrams of cortical circuits.

A. Upper: a general neural network. Blue and triangle, principal cell; red and circle, inhibitory cell. Thin line, axon; thick line, dendrite; dot, synapse. Lower: the connectivity of this network can be described by a matrix with synaptic strength of synapses between each pair of neurons as individual elements.

B. Columnar and laminar processing in auditory cortex. Layer 4 (L4) and Layer 6 (L6) are recipient layers of thalamic inputs. Layer 2/3 (L2/3) and Layer 5 (L5) send outputs to other cortical regions and subcortical nuclei respectively. L6 sends corticothalamic feedback.

C. A simplified input circuit for a L4 neuron.

2.2. Cortical columnar and laminar organization

A prominent feature of the neocortex is its columnar and laminar organization, which provides a platform for parallel and serial processing, respectively (Schreiner et al., 2000; Winer et al., 2005). Within a cortical column neurons are thought to exhibit similar functional properties, e.g. they are tuned to similar tonal frequencies. As other sensory cortices, the auditory cortex is stratified into six layers (Fig. 1B). Among them, Layer 4 is the major recipient of inputs from the auditory thalamus, while Layer 6 is weakly innervated by the thalamus (Cruikshank et al., 2002; Kimura et al., 2003, Romanski and LeDoux, 1993; Winer and Lee, 2007; Barbour and Callaway, 2008). The non-pyramidal neurons in Layer 1, an integration layer, have also been reported to receive direct thalamic input (Mitani and Shimokouchi, 1985; Huang and Winer, 2000; Zhu and Zhu, 2004; Theyel et al., 2010), although the functional significance of this direct input remains to be determined. Layer 2/3 neurons are largely innervated by Layer 4 projections and they send outputs to other cortical regions for intercortical processing, while Layer 5 neurons receive relayed information from Layer 2/3 and project to various subcortical nuclei to modulate subcortical responses (Winer and Prieto, 2001; Winer, 2005; Barbour and Callaway, 2008). Layer 6 sends cortical feedback specifically to the thalamus (Ojima, 1994; Prieto and Winer, 1999; Winer, 2005). In addition to the direct feedforward input, principal neurons receive prominent intracortical excitatory input. The highest density intracortical excitatory input comes from neurons within a 500μm radius (Douglas et al., 1995; Roerig and Chen, 2002; Marino et al., 2005). Besides the excitatory input, inhibition from cortical inhibitory interneurons plays an important role in shaping response properties of principal neurons (Sillito, 1977, 1979; Kyriazi et al., 1996; Wang et al., 2002). Based on the columnar and laminar organization of the cortex, and the belief that even the most complex circuits are built upon repeated circuit modules, we propose that cortical circuits can be reduced by identifying columnar circuit modules at each cortical layer. It is worth noting that within the module, the circuit may involve both intralaminar and interlaminar interactions. The layer-specific circuit modules can then be integrated and repeated across columns to construct a more complete cortical circuit.

2.3. Simplified elementary circuit module

For each cortical layer, we assume an elementary circuit module lie at the heart of the processing capability. Within the module, the excitatory neurons can be regarded as the core component. No matter how complex the modular circuit is, the spiking response properties of individual excitatory neurons are determined by all the sound-driven synaptic inputs received by the neuron, including both excitatory and inhibitory inputs, as well as the cell’s intrinsic membrane properties. In auditory cortex, the ascending synaptic inputs are topographically organized to form a gradient of frequency representation (Schreiner, et al., 2000; Winer et al., 2005). Due to the correspondence between spatial and spectral locations, stimulation with single tones of different frequencies can reveal the information on the spatial distribution of synaptic inputs. In simplification, each tone-evoked input can be viewed as an individual input, although it may not result from a single synaptic connection. Thus, the spectral distribution of synaptic inputs, both excitatory and inhibitory, provides fundamental information on the structure of a functional synaptic input circuit. From the spectrotemporal integration of functional synaptic inputs evoked by simple tones, it would be possible to predict the neuron’s other response properties (e.g. direction selectivity) based on this simplified model. A starting place for examining functional synaptic input circuits is the thalamo-recipient layers, since cortical processing starts here. For neurons in these input layers, in principle they receive synaptic input from three major sources: thalamocortical, intracortical excitatory and intracortical inhibitory inputs (Schreiner, et al., 2000; Winer, et al., 2005). The network in Fig. 1A can then be simplified into a more concise form with a single excitatory neuron receiving three sources of input (Fig. 1C). We will discuss recent progresses on dissecting the contribution of these three types of input to auditory processing in the rat A1. It is worth noting that the functional architecture of the rat auditory cortex and the general response properties of rat auditory cortical neurons are similar to those in cats and mice (e.g. Schreiner, et al., 2000; Wu, et al., 2006)

3. Experimental approaches to dissecting neural circuits

The representation and processing of individual neurons is a fundamental basis for cortical function. Many previous studies using the reverse correlation technique have focused on the spike responses of individual neurons as to understand their transfer function under sound stimuli (e.g. Kim and Young, 1994; Klein et al., 2000; Miller et al., 2002). Such “black-box” method only deals with the properties of sensory input and neuronal spike output, without knowing what occurs in the box. Because of the nonlinearity of neurons and a finite number of stimuli that can be applied, it is difficult to obtain a complete understanding of the operations performed by neural circuits. Alternatively, an investigation into the synaptic mechanisms underlying the spike responses can provide insights into the connectivity of neural circuits and the way the connectivity determines the processing function of the neuron.

To access the cell interior, traditional sharp electrodes with high impedances are used to impale the neuron and record sound-driven depolarization or hyperpolarization of the cell’s membrane potential (Coombs et al., 1955; Fatt et al., 1951; de Ribaupierre et al., 1972; Volkov and Galazjuk, 1991; Ojima and Murakami, 2002). Although these changes in membrane potential suggest the presence of excitatory or inhibitory inputs, the absolute strength of the inputs cannot be determined since the membrane potential response is an integrated result of the interaction between both excitatory and inhibitory inputs as well as voltage-dependent conductances. Whole-cell current-clamp recordings have been more widely used to measure subthreshold and suprathreshold membrane potential responses evoked by sensory stimuli (Anderson et al., 2000; Moore and Nelson, 1998; Margrie et al., 2002). Due to a better electronic access to the cell interior, different levels of currents can be injected into the cell while membrane voltages are monitored. Excitatory and inhibitory conductances are derived based on a linear neuron model, which assumes that the recorded neuron is isopotential and that there is a linear relationship between membrane potential and synaptic current (Anderson et al., 2000; Monier et al., 2003; Marino et al., 2005). However, because the membrane potential is not controlled in current-clamp recording, any voltage-gated conductances that have not been blocked by appropriate ion-channel blockers can vary instantaneously in a voltage-dependent manner and contaminate synaptic conductances, sabotaging linearity of the neuron. Not until recently has voltage-clamp recording technique been applied in functional studies of sensory cortices in live animals (Borg-Graham et al., 1998; Zhang et al., 2003; Wehr and Zador, 2003; Tan et al., 2004). By clamping the cell’s membrane potential, fluctuation of unblocked voltage-dependent conductances is minimized. With this technique, our knowledge on neural circuits underlying sound processing has been increased. We review here the current progresses, potential limitations, and future directions along this promising research line.

3.1. Voltage-clamp recording to dissect synaptic inputs

A key for dissecting excitatory and inhibitory inputs is to achieve low access-resistance whole-cell recording, which requires using patch pipettes with relatively large tip openings (1–2μm). In a typical blind recording experiment, the electrode resistance is continuously monitored as the pipette advances in small steps. When the tip contacts cell membrane, the pipette resistance increases. The positive pressure within the pipette is then released. In a lucky case, a tight seal between the pipette and cell membrane (a few giga-ohms in seal resistance) is formed, which is essential for high-quality experiments. Pulses of negative pressure are then applied to break in the cell’s membrane and bring the recording to whole-cell mode. With sufficient compensations of whole-cell capacitance and series resistance, the cell’s membrane potential is clamped at −70mV (the reversal potential for GABA_A receptor-mediated Cl⁻ currents) to record excitatory currents, or 0mV (the reversal potential for glutamate receptor-mediated excitatory currents) to record inhibitory currents (Fig. 2). From the recorded synaptic currents under different levels of membrane potentials, excitatory and inhibitory conductances can be derived, assuming that the neuron is isopotential and linear. The large pipette opening and the low effective series resistance after compensation (~ 10–20 mega-ohms) allows a better control of the cell’s membrane voltage. However, the space-clamp issue (Hille, 1992) cannot be omitted when estimating synaptic inputs, which will be discussed below.

Fig. 2.

In vivo voltage-clamp recording to dissect different components of input. Excitatory and inhibitory components are separated by clamping the membrane potential of the recorded neuron at two different voltages. Thalamocortical inputs are further isolated by silencing the cortex. Cell morphology is recovered by post hoc histological staining.

3.2. Reconstruction of the output response by integrating synaptic inputs

The functional properties of the neuron are usually defined by its spike responses. By comparing the evoked synaptic inputs and spike outputs of the same cell, we would gain an understanding of how the synaptic input circuit contributes to the function. However, for high-quality voltage-clamp recording, various blockers of voltage-gated conductances, including those blocking Na⁺ spikes, are infused into the cell through the patch pipette. Under this circumstance, the bona fide spike responses cannot be obtained. Nonetheless, the membrane potential response resulting from the integration of excitatory and inhibitory inputs can be estimated by feeding the experimentally determined synaptic conductances into an integrate-and-fire neuron model (Wehr and Zador, 2003; Wu et al., 2008; Sun et al., 2010; Zhou et al., 2010). Without knowing the detailed dendritic distribution of individual inputs, the model is a single-compartment neuron with inputs all arriving near the soma. Thus this model is better applied to inputs that are known to be close to the soma, e.g. thalamocortical inputs in layer 4 (Richardson et al., 2005). Spike rate can be calculated from the derived membrane potential based on either a “threshold-linear” or a power-law model (Carandini and Ferster, 2000; Priebe et al., 2004). One advantage of the computational simulation of membrane potential responses is that the inhibitory component can be easily removed to examine its contribution to the spike response, while the experimental feasibility of removing intracellular inhibition specifically is still unclear.

To estimate the accuracy of the computational simulation of output responses, the simulated membrane potential responses can be compared to those recorded under current-clamp mode. It should be noted that as spikes are blocked, the recorded membrane potentials can be considered as bona fide because they have not been disturbed by spike generation. Alternatively, a technically more challenging experiment can be performed: cell-attached recording followed by whole-cell recording (Sun et al., 2010). In the cell-attached recording, the cell membrane seals the pipette tip, allowing a superb isolation of a single unit, i.e. spikes only from the patched cell can be recorded (DeWeese et al., 2003; Hromadka et al., 2008; Wu et al., 2008). The subsequent tight-seal whole-cell voltage-clamp recording determines the evoked excitatory and inhibitory conductances. It is worth noting that under cell-attached mode the cell’s intracellular milieu is intact, while voltage-dependent channels may be blocked after forming the whole-cell. If the functional property (e.g. frequency tuning) as exhibited in simulated spike responses is close to that in recorded spike responses, it may be concluded that the experimentally measured excitatory and inhibitory conductances contribute largely to the observed functional property of the neuron.

3.3. Quality control for in vivo whole-cell recording

When deriving excitatory and inhibitory synaptic conductances, it is assumed that the neuron is isopotential (like a small spherical cell) and linear (Zhang et al., 2003; Wehr and Zador, 2003; Tan et al., 2004). However, the real neuron extends dendritic processes and is far from spherical. Although the soma may be well clamped at the set holding potential, local membrane potentials in distal dendrites are not controlled as well due to an imperfect space clamp. As synaptic inputs arrive at the distal dendrites, there may be a large deviation of the local potential from the set holding potential. Under this circumstance, when the holding potential (e.g. 0mV) differs from the resting membrane potential substantially, there can be large errors in the measured synaptic conductances (Spruston et al., 1993). Such space-clamp error will be less an issue if all inputs are close to the soma, and can be alleviated by using cesium and tetraethylammonium (TEA, a blocker of potassium channels) in the intracellular solution, which reduces membrane permeability and cable attenuation (Spruston et al., 1993). To estimate how well the evoked synaptic current as measured from the soma is clamped, the current-voltage (I–V) relationship can be examined for the synaptic current at different time points. The linear I–V relationship supports linearity of the neuron, which may have been improved by the blockade of various voltage-dependent conductances including NMDA-receptor conductances. In addition, the instantaneous reversal potential of synaptic current can be derived from the I–V curve. Since inhibitory input usually follows the excitatory input by a temporal delay (Zhang et al., 2003; Wehr and Zador, 2003), within a small time window right after the response onset, the current is almost exclusively contributed by excitatory conductance. Space-clamp error can be detected when the reversal potential within this small time window deviates largely from the actual reversal potential for excitatory currents (0mV). In our high-quality recordings from layer 4 neurons, we consistently observed a close proximity of the reversal potential for early currents to 0mV (Wu et al., 2006, 2008; Liu et al., 2007; Sun et al., 2010). This indicates that the synaptic inputs under examination are mostly in close proximity to the cell body.

3.4. Criteria for studies under anesthesia

Because of the level of mechanical stability required for high-quality voltage-clamp recordings, up to now the recording technique has only been applied in anesthetized animals. A few studies have explored the feasibility of whole-cell recordings in awake animals (Margrie et al., 2002; Lee et al., 2006; Gentet et al., 2010; Okun et al., 2010), but none of them have attempted to dissect synaptic inputs or conductances. Notably, anesthetized animals can only be considered as a reduced model, as anesthesia reduces activity level and may affect the temporal pattern of sound evoked responses even in the primary sensory processing pathway (e.g. Young and Brownnell, 1976; Evans and Nelson, 1973; Kuwada et al., 1987; Gaese and Ostwald, 2001; Wang et al., 2005). Among commonly used anesthetics, pentobarbital prolongs the opening time of GABA-activated chloride channels (MacDonald et al., 1989; Nicoll et al., 1975), and ketamine blocks NMDA receptors and reduces the general activity level (Thomson et al., 1985; Rennaker et al., 2007). One important criterion for using anesthetized animal models is that the studied response property can be observed in both anesthetized and awake animals. Various fundamental processing properties mostly involving ascending sensory pathways (e.g. tonal receptive field and frequency tuning) are likely preserved in anesthetized animals. The temporal relationship between sensory evoked excitatory and inhibitory inputs, especially at the early phase of response, is also likely preserved. Thus, the results on the spectrotemporal distribution of excitatory and inhibitory responses to simple tones obtained from anesthetized animals would still be valuable in addressing functional synaptic circuitry in awake models. Nonetheless, the limitation of anesthetized models should be recognized and experimental data should be interpreted with appropriate caution. It is recommended that results be further verified by using a different anesthesia method.

4. Functional Neural Circuitry underlying feature selectivity in the primary auditory cortex

4.1. Recurrent and feedforward circuits in the recipient layer 4

4.1.1. Approximately balanced excitation and inhibition underlying tonal receptive field

Neurons in the auditory system are tuned to specific frequencies (and intensities). The distribution pattern of synaptic inputs within the frequency-intensity tonal receptive field (TRF) reflects the basic synaptic input circuitry, and is a basis for processing properties such as frequency/intensity selectivity and direction selectivity in response to frequency modulated (FM) sound sweeps. Neurons in the recipient layer 4 (L4) of primary sensory cortices receive synaptic input from three major sources: thalamocortical inputs which are excitatory, and intracortical excitatory and inhibitory inputs (Liu et al., 2007; Zhou et al., 2010; Happel et al., 2010). With the application of whole-cell voltage-clamp recordings, tone evoked excitatory and inhibitory synaptic inputs underlying the TRF have been analyzed. As observed in several studies, receptive field composed of inhibitory inputs matches that of excitatory inputs (Zhang et al., 2003; Wehr and Zador, 2003; Tan et al., 2004; Wu et al., 2006, 2008). Inhibition is tuned to the same frequency as excitation, with essentially the same (or sometimes slightly narrower) total frequency responding range. The amplitudes of inhibitory and excitatory conductances evoked by the same tone exhibit a strong linear correlation (Zhang et al., 2003; Wehr and Zador, 2003; Tan et al., 2004; Tan and Wehr, 2009; Dorrn et al., 2010). In addition, inhibition always follows excitation with a short temporal delay (about 2–4ms), which is essentially constant across different frequencies (Wehr and Zador, 2003). These observations led to the conclusion that excitation and inhibition are balanced (Wehr and Zador, 2003). The balanced inhibition is easily achieved by a feedforward circuit with the cortical inhibitory neurons driven by the same set of thalamocortical projections as the recorded L4 neuron (Fig. 3A; a potential inhibitory relay from layer 1 remains to be investigated -see Mitani and Shimokouchi 1985; Huang and Winer, 2000). Under balanced excitation and inhibition, inhibition sharpens frequency tuning of spike responses through an “ice-berg” effect, suppressing membrane potential responses to off-optimal frequencies below the spike threshold. In addition, the shortly delayed inhibition enhances the temporal precision of the evoked spike (Wehr and Zador, 2003). A more detailed analysis, however, reveals that the inhibitory tuning is in fact broader than the excitatory tuning (in terms of tuning shape) within a frequency range flanking the best frequency (Wu et al., 2008). Comparing the normalized tuning curves, the frequency range of inhibitory input is about 27% broader than that of excitatory input at a level of 60% maximum. Through an analogous lateral-inhibition effect, the broader inhibition within the putative suprathreshold response range further sharpens the frequency tuning of membrane potential response (Wu et al., 2008). Thus, excitation and inhibition can only be viewed as approximately balanced. The disparity between inhibitory and excitatory tunings can be explained by the tuning properties of inhibitory neurons and intracortical excitatory inputs (see below).

Fig. 3. Diverse cortical functional circuits.

A. A schematic diagram of a canonical circuit in Layer 4 of the primary auditory cortex (A1). Triangle, excitatory neuron; circle, inhibitory neuron. The frequency tuning of each connection is also presented along the projections. Gray curve depicts the frequency tuning curve. The thalamic and feedforward inhibitory inputs have a broad tuning, while the recurrent excitatory input has a sharp tuning. Together, the summed input exhibits a sharp frequency representation.

B. A “silent” circuit in Layer 6 of A1. The recorded neuron is not directly innervated by thalamic neurons. Instead, it receives disynaptic excitatory input.

C. A circuitry mechanism for cortical FM direction selectivity. The recorded neuron has a low CF and prefers upward FM sweeps. The spectral distribution of its excitatory inputs is skewed to low frequency side, and that of inhibition inputs is relatively broader and less skewed.

D. A hypothesized circuit for intensity-tuned neurons in Layer 4 of non-monotonic zone (NM), which is adjacent to A1. The recorded neuron receives intensity-tuned thalamic input (green), and non-intensity-tuned inhibitory input from interneurons which are innervated by non-intensity-tuned thalamic neurons. The amplitude-intensity tuning curves for each type of input and the summed input are shown. Te and Ti represent the latencies of thalamic and inhibitory inputs. Their difference reduces as intensity increases (left), which further contributes to the sharpening of intensity selectivity of the recorded neuron.

4.1.2. Contribution of intracortical excitation to frequency tuning

Whether the functional properties of L4 neurons are determined by the convergence of thalamocortical inputs or are attributed mainly to intracortical inputs has been an issue of intense debate (Reid and Alonso, 1995; Ferster et al., 1996; Douglas et al., 1995; Somers et al., 1995). It is a great challenge to determine the respective functional roles of thalamocortical and intracortical inputs. Previously three methods have been used to isolate thalamocortical inputs by preventing spiking of cortical neurons: a) pharmacologically silencing the cortex with muscimol (Fox et al., 2003; Kaur et al., 2004; Zhang and Suga, 1997); b) cooling the cortex (e.g. Ferster et al., 1996; Volgushev et al., 2000; Villa et al., 1991); c) local electrical stimulation of the cortex (Ferster et al., 1996). In these methods, the thalamocortical transmission can potentially be affected significantly, confounding the interpretation of results. For example, a widely used method is to apply muscimol through microinjection, iontophoresis or surface perfusion, assuming that muscimol specifically activates GABA_A receptors. However, it was found recently that muscimol can activate GABA_B receptors at relatively low concentrations and can thus reduce transmission through presynaptic GABA_B receptors (Yamauchi et al., 2000). To prevent muscimol’s effect on release from thalamic axons (Porter and Nieves, 2004), we recently developed a new pharmacological method for cortical silencing: co-applying muscimol and a specific antagonist of GABA_B receptors which prevents activation of GABA_B receptors by muscimol (Liu et al., 2007; Khibnik, et al., 2010). Using this method, the frequency tunings of excitatory input before and after cortical silencing have been compared (Liu et al., 2007). The total frequency range is not reduced after cortical silencing, indicating that thalamocortical input determines the area of synaptic receptive field. However, the tuning of thalamocortical input exhibits a broad and flat tuning peak, and is significantly weaker compared to that of total excitatory input. In contrast, the tuning of derived intracortical input is much sharper, and closely resembles that of total excitatory input within the putative frequency range of suprathreshold response. Thus, it is the intracortical excitatory input that largely defines the frequency tuning of the neuron by selectively amplifying responses at the optimal and near-optimal frequencies. Using a similar silencing method, a current source density (CSD) study in gerbil A1 (Happel et al., 2010) confirms that short-range recurrent input amplifies the thalamocortical afferent signal. Different from our conclusion, the study suggests that the subthreshold responses to off-optimal tones are attributed to long-range horizontal inputs between cortical sites in layer 2/3, while the sensitivity of CSD for detecting small responses (Mitzdorf, 1985) and the specificity of perfusion of drug in cortical surface need to be further examined. It is worth noting that a recent study in the mouse barrel cortex indicates that the range of horizontal connections within layer 2/3 is within 300 μm (Adesnik and Scanziani, 2010).

4.1.3. Broader inhibitory tuning

As described above, high-resolution mapping of excitatory and inhibitory inputs has revealed a broader tuning of inhibition than excitation within the putative frequency range of suprathreshold response (Wu et al., 2008). This observation is consistent with extracellular studies (e.g. Sadagopan and Wang, 2010), and can be sufficiently explained by the known circuit in layer 4. Inhibitory input is likely attributed to a feedforward circuit, with the local inhibitory neurons providing the input directly driven by thalamic axons (Fig. 3A). Therefore, the receptive field property of L4 inhibitory neurons is key to understanding the tuning of inhibitory input. Because fast-spiking inhibitory neurons are the major source of cortical inhibition in layer 4 (Agmon and Connors, 1992; Gil and Amitai, 1996; Gibson et al., 1999; Inoue and Imoto, 2006; Sun et al., 2006; Atencio and Schreiner, 2008), we developed a method of cell-attached recording combined with subsequent juxtacellular labeling or intracellular recording to selectively target fast-spiking inhibitory neurons (Wu et al., 2008). Compared to nearby excitatory neurons, fast-spiking inhibitory neurons exhibit a lower threshold and a broader bandwidth of spike TRFs, despite that the two types of neurons receive similar ranges of excitatory input and are organized by the same tonotopic map. The broader spike TRF of inhibitory neurons is thus due to their higher efficiency in converting synaptic input to spike output, which may be attributed to a stronger thalamocortical input to L4 inhibitory neurons than L4 excitatory neurons (Gibson et al., 1999; Beierlein et al., 2003). Also as discussed above, the tuning of thalamocortical input is broad while that of intracortical excitatory input is sharp, which indicates that intracortical excitatory input derives from cortical neurons with similar frequency tuning properties. The summation of thalamocortical and intracortical excitatory inputs is analogous to adding a pyramid on top of a flat base. Since L4 inhibitory neurons have a broader spiking response tuning than excitatory neurons, inhibition would exhibit a broader tuning than excitation near the apex of the pyramid. Such intracortical inhibition with higher sensitivity and lower selectivity than intracortical excitation can laterally sharpen the frequency tuning of neurons, ensuring their highly selective representation (Fig. 3A).

4.2. A “silent” circuit in layer 6

It has been widely observed that sensory stimulation or stimulation of the thalamus results in a stereotypic temporal relationship between evoked excitation and inhibition, with inhibition occurring shortly after excitation (Douglas and Martin, 1991; Zhang, et al., 2003; Wehr and Zador, 2003; Tan, et al., 2004). The later arriving inhibition can curtail the evoked spiking responses, preventing runaway excitation within the cortical circuit. Besides layer 4, layer 6 of the A1 also receives direct thalamocortical input (Winer et al., 2001; Winer et al., 2005; Llano and Sherman, 2008; Kaur et al., 2005; Lakatos et al., 2007; Wallace and Palmer, 2008). Conversely, layer 6 sends feedback projections predominantly to the first-order thalamic nucleus (i.e. ventral nucleus of the medial geniculate body, MGBv) (Ojima, 1994; Prieto and Winer, 1999; Winer, 2005; Takayanagi and Ojima, 2006; Rouilller and Welker, 2000; Llano and Sherman 2008). It has been proposed that the corticothalamic feedback from layer 6 modulates thalamic responses (Villa et al. 1991; Zhang and Suga, 1997; Yan and Ehret, 2002), and plays a role in mediating the induction of sound-specific plasticity in the auditory thalamus (Zhang and Suga, 2000; Suga and Ma, 2003; Zhang and Yan, 2008). Surprisingly, sensory stimuli do not drive spike responses in a large proportion of L6 excitatory neurons (Zhou et al., 2010, Tsumoto and Suda, 1980; Sirota et al., 2005), but suppress their spontaneous firings within the expected tonal receptive field (Zhou et al., 2010). Whole-cell recordings reveal that the suppression of evoked spike responses results from a novel synaptic integration pattern with a strong inhibitory input preceding the co-activated excitatory input. The reversed temporal relationship between excitatory and inhibitory inputs in many L6 neurons can be attributed to a parallel feedforward circuit with both the excitatory and inhibitory inputs disynaptically relayed (Fig. 3B). Since the first-order L6 inhibitory neurons spike earlier than the first-order excitatory neurons (Zhou et al., 2010), the second-order excitatory neuron would receive inhibition preceding the disynaptic excitatory input. Interestingly, anatomical evidence suggests that the silenced L6 neurons may project back to the thalamus, leading to a hypothesis that the corticothalamic feedback is only activated under specific circumstances such as conditioning during which the suppression by preceding inhibition can be relieved (Zhou et al., 2010). While this hypothesis awaits future investigations, the specific “silent” circuit in layer 6 highlights an essential role of inhibition in creating a diversity of response properties, through its specific spectral and temporal patterns inherited from the local circuitry.

4.3. From elementary circuits to more complex auditory processing

As discussed above, the spatial distributions and temporal patterns of excitatory and inhibitory synaptic inputs underlying the TRF can reveal the basic cyto-architecture of neuronal connectivity. Receptive field structures have been shown to be major determinants of cortical responses to complex sound stimuli (Kowalski, et al., 1996; deCharms, et al., 1998; Schnupp, et al., 2001; Sen, et al., 2001; Linden, et al., 2003; Zhang, et al., 2003), although dynamic properties of synaptic inputs may also be important in generating responses to complex sound. The elementary circuits derived from synaptic TRFs will allow us to better understand how cortical neurons process complex acoustic signals.

4.3.1. FM direction selectivity

Frequency-modulated (FM) sweeps are important cues in animal vocalization and human communication (Winter et al., 1966; Kanwal et al., 1994; Wang, 2000; Lindblom and Studdert-Kennedy, 1967; Gold and Morgan, 2000; Zeng et al., 2005). Neurons selective for direction of FM sweeps are found in the A1 (Suga, 1965; Mendelson and Cynader, 1985; Zhang et al., 2003), the auditory thalamus (O’Neill and Brimijoin, 2002) as well as other subcortical nuclei such as the inferior colliculus (Nelson et al., 1966; Gordon and O’Neill, 1998; Fuzessery et al., 2006). Mapping studies suggest that in the auditory cortex, direction selectivity is topographically ordered in parallel with characteristic frequency (CF): low CF neurons prefer upward sweeps, whereas high CF neurons prefer downward sweeps (Heil, et al., 1992; Zhang, et al., 2003; Godey, et al., 2005). Both the topographic organization of direction selectivity and the selectivity per se can be well explained by the spectral distributions of excitation and inhibitory inputs and their temporal interaction. Firstly, the spectral distribution of input strength is asymmetric or skewed in direction-selective neurons. The skewness co-varied with CF, as well as of the measured direction selectivity, suggesting a synaptic circuitry basis for the topographic ordering of the cortical representation of direction selectivity (Zhang et al., 2003). Secondly, the slightly delayed inhibitory input further strengthens the directional selectivity inherited from excitatory input by suppressing synaptic excitation more effectively when stimulus is in the non-preferred direction than in the preferred direction (Zhang et al., 2003). Thirdly, there appears to be a spectral offset between excitation and inhibition in direction-selective neurons (Suga, 1965; Shamma et al., 1993; Nelken and Versnel, 2000; Zhang et al., 2003; Razak and Fuzessery, 2006; Ye et al., 2010). For example, “inhibitory sidebands”, the suppressive regions flanking the TRF of spiking response exhibit asymmetry, and the asymmetry co-varied with CF (Zhang, et al., 2003). In addition, relatively stronger inhibition on one side of excitatory spectral tuning curve has been observed for direction selective A1 neurons (Ye et al., 2010). In fact, the inhibitory sidebands revealed by extracellular recording experiments with two-tone suppression, and the spectral offset between excitation and inhibition can be attributed, at least partially, to a broader tuning of inhibition than excitation (Wu et al., 2008). A schematic diagram of synaptic inputs to a direction selective neuron is shown in Fig. 3C.

4.3.2. Intensity selectivity

Intensity tuned auditory neurons are characterized by their nonmonotonic responses to sound intensities (Greenwood and Maruyama, 1965). Such neurons (also named nonmonotonic neurons) have been observed along the central auditory pathway, including the cochlear nucleus (Greenwood and Maruyama, 1965; Young and Brownell, 1976), inferior colliculus (Aitkin, 1991; Kuwabara and Suga, 1993), medial geniculate body (Aitkin and Webster, 1972; Rouiller et al., 1983), and auditory cortex (Davies et al., 1956; Evans and Whitfield, 1964; Brugge et al., 1969; Schreiner et al., 1992; Phillips et al., 1995). In the rat auditory cortex, the response properties of intensity-tuned neurons (Phillips et al., 1995; Heil and Irvine, 1998) and their susceptibility to specific changes after training animals with a sound magnitude discrimination task (Polley et al., 2004, 2006) suggest that these neurons may play important roles in encoding sound loudness and envelop transients. Studies using extracellular recordings with two-tone masking paradigms (Suga and Manabe, 1982; Calford and Semple, 1995; Sutter and Loftus, 2003) and GABA receptor blockade (Faingold et al., 1991; Pollak and Park, 1993; Wang et al., 2002; Sivaramakrishnan et al., 2004), as well as those using intracellular recordings (Ojima and Murakami, 2002) suggest that intensity tuning may be created by spectral and temporal interactions between excitation and inhibition. Recently, voltage-clamp recordings from intensity-tuned and non-intensity-tuned neurons indicate that cortical intensity tuning is determined by the interplay between imbalanced excitatory and inhibitory synaptic inputs (Wu et al., 2006; Tan et al., 2007). In intensity-tuned neurons, excitatory input already exhibits intensity tuning. On the other hand, the strength of inhibitory input increases monotonically with intensity increments and quickly saturates. Interestingly, the temporal delay of inhibitory input relative to excitatory input is reduced as intensity increases, resulting in an enhanced suppression of excitation at high intensities and a significant sharpening of intensity selectivity. These findings imply a unit circuit controlling the relative timing of excitation and inhibition, which may also be capable of creating intensity selectivity at the first step (Fig. 3D).

5. Conclusions and future directions

5.1. Imbalanced excitation and inhibition underlying functional diversity

The initial whole-cell recording studies in the rat A1 have led to the conclusion that balanced excitation and inhibition underlie the frequency-intensity tonal receptive field (Zhang et al., 2003; Wehr and Zador, 2003; Tan et al., 2004). The balance is defined as matched excitatory and inhibitory tunings together with a constant temporal interval between excitation and inhibition. However, a series of more recent studies with high-resolution mapping and more detailed analysis indicate that balanced excitation and inhibition cannot be taken as a universal rule (Wu et al., 2006, 2008; Tan et al., 2007; Zhou et al., 2010; Ye et al., 2010). First, the tunings of excitation and inhibition do not match perfectly. For frequency tuning, inhibition is broader than excitation within the frequency range that generates spike responses. For intensity tuning, in a subset of neurons excitation exhibits nonmonotonic amplitude-intensity function, whereas inhibition displays monotonic intensity tuning. Second, the temporal interaction can vary with stimulus parameter. For intensity-tuned neurons, the relative delay of inhibition reduces with the intensity increment. In a subset of L6 neurons, the temporal sequence of excitation-inhibition is completely reversed. Through various spectral and temporal interactions between excitation and inhibition, a wide diversity of feature-selective response properties is created. Thus, deviations from perfectly balanced excitation and inhibition endow the cortex with much increased computational capacity.

5.2. Elementary circuit modules

The canonical microcircuit (Douglas and Martin, 1991) was initially proposed to account for the response profile of cortical neurons to thalamic stimulation, which shows a stereotypic temporal sequence of activation of inhibition shortly after that of excitation. The core of the canonical microcircuit is a tripartite feedforward circuit involving disynaptic inhibitory inputs from cortical inhibitory neurons driven by the same set of afferents as the excitatory neuron of interest (Fig. 3A). Such circuit module is prevalent in layer 4 of the A1, and is also observed for a subpopulation of L6 neurons (Zhou et al., 2010). Other circuit modules cannot be described as canonical. In layer 4 of the “nonmonotonic” zone, i.e. an area posterior-ventrally adjacent to the A1 where neurons mostly exhibit nonmonotonic intensity tunings (Wu et al., 2006), the disynaptic inhibitory input is from inhibitory neurons driven by a different set of thalamic axons (Fig. 3D). In a second subpopulation of L6 neurons in the A1, both excitatory and inhibitory inputs are disynaptic (Fig. 3B). While it remains to be tested whether these elementary circuits also exist in subcortical nuclei, it is already remarkable that cortical circuits, by exploiting such simple connectivity strategies, can create a wide variety of functional properties. It is likely that we will be more amazed by the cortex during further explorations of other layers as well as other functional auditory cortical areas.

5.3. Inhibitory neurons in cortical circuits

Most of the in vivo recordings so far are obtained from excitatory neurons, a dominant cell type in the cortex. Although cortical neurons receive prominent inhibitory inputs through feedforward or feedback circuits, these inputs only come from a small population (15–20%) of cortical neurons (Peters and Kara, 1985; Hendry et al., 1987; Priet et al., 1994). To obtain a complete understanding of the structure of functional circuits, it is essential to know the functional properties of inhibitory neurons themselves. The difficulty in studying inhibitory neurons lies not only in the small number of these neurons, but also in their morphological and neurochemical diversity (Kawaguchi and Kubota, 1997; Gonchar and Burkhalter, 1997; Gupta et al., 2000) (Fig. 4), which suggests that inhibitory neurons are also functionally diverse. Although in a few cases, in vivo whole-cell recordings combined with post hoc histology have identified inhibitory neurons in the sensory cortex (Azouz et al., 1997; Hirsch et al., 2003), in general it is extremely difficult to encounter inhibitory neurons in blind recordings, especially those minor types of inhibitory neurons. Recent development in two-photon imaging of Ca²⁺ responses (Sohya et al., 2007; Kerlin et al., 2010; Runyan et al., 2010) and two-photon imaging guided targeted recording (Margrie et al., 2003; Liu et al., 2009; Gentet et al., 2010; Ma et al., 2010) allows efficient examinations of functional properties of inhibitory neurons in transgenic mice in which a specific type of inhibitory neurons is genetically labeled. It is expected that our knowledge on inhibitory neurons in auditory cortex will grow rapidly as better imaging and genetic tools are being developed.

Fig. 4.

A cortical excitatory neuron receives inhibitory input from a diversity of interneurons, which have differential subcellular targeting preferences.

Finally, for a realistic model of cortical circuitry, further understanding of the distribution pattern of excitatory and inhibitory neurons projecting to a single neuron will be needed. The recent development of trans-synaptic labeling (Marshel et al., 2010) 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. Taken together, the renovations of electrophysiology, imaging and genetic techniques will greatly facilitate a quantitative analysis of elementary neuronal circuits with a bottom-up approach. The results will enhance our understanding of circuitry mechanisms underlying both normal and abnormal brain functions.

Highlights.

• A systematic reductive approach is developed to resolve the cortical circuitry.

• In vivo whole-cell voltage-clamp recordings represent a key to probe circuitry.

• Excitatory-inhibitory interplay largely determines functional response properties.

• Elementary circuits can be derived from spatiotemporal patterns of synaptic inputs.