INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Many neurons in the inferior colliculus (IC) exhibit displaced tuning to auditory stimuli presented under static or dynamic conditions. Recent studies have shown that tuning to interaural delay (McAlpine et al. 2000; Spitzer and Semple 1991a, 1993, 1998), interaural level difference (Sanes et al. 1998), simulated free-field motion (Wilson and O'Neill 1998), and monaural frequency (Malone and Semple 2001) may be different when the tuning curve is generated by continuously varying the stimulus parameter rather than by testing a range of discrete parameter values.

These dynamic conditioning effects have been particularly well studied with manipulation of the interaural phase difference (IPD)a binaural cue for low-frequency sound localization. Temporal variation of IPD simulates the binaural cue produced by a moving sound source, and the variable may be manipulated across its entire range (i.e., 360°). Although IC responses to IPD modulation depend on the recent stimulus-response history (McAlpine et al. 2000; Spitzer and Semple 1991a, 1993, 1998), responses in the medial superior olive (MSO, where IPD is initially encoded) reflect a coincidence-detection process that effectively tracks the instantaneous value of IPD (Spitzer and Semple 1995, 1998; Yin and Chan 1990).

One explanation for this transformation of response properties between MSO and IC is that it reflects "adaptation of excitation" (firing rate adaptation) (Cai et al. 1998b; McAlpine et al. 2000; Spitzer and Semple 1998). Alternatively (Spitzer and Semple 1998), motion sensitivity could be the result of interaction between excitatory and inhibitory inputs, aided by firing rate adaptation and postinhibitory rebound. However, McAlpine and Palmer (2002) argued that leaving the key role to inhibition contradicts their data. They showed that sensitivity to the apparent motion cues is decreased in the presence of the inhibitory transmitter, GABA, and increased in the presence of bicuculline.

In this study, we propose a framework that unifies these different ideas. We explore how intrinsic cellular mechanisms, such as adaptation and rebound, together with the interaction of excitatory and inhibitory inputs, can contribute to the shaping of dynamic conditioning. We suggest that each of these mechanisms has a specific influence on response features. To examine the role of various mechanisms in shaping IC response properties, we developed a mathematical model of an IC cell that receives IPD-tuned excitatory and inhibitory inputs and possesses the properties of adaptation and postinhibitory rebound. In earlier modeling studies of the role of adaptation and inhibition in IC, Cai et al. (1998a,b) developed detailed cell-based spike-generating models, involving multiple stages of auditory processing. In contrast, our model is minimal in that it only involves the components whose role we want to test and does not explicitly include spikes (firing-rate-type models). This approach (extending the prototype version of Borisyuk et al. 2001a) greatly facilitates examination of computations that can be performed with rate-coded inputs (spike-timing precision on the order of 10 ms), consistent with the type of input information available to neurons at higher levels of processing (such as the inferior colliculus compared with the superior olive). In addition, our model is computationally efficient, can be easily implemented, and requires minimal assumptions about the underlying neural system.

First, we explore the extent to which our model, using the experiment-based hypotheses of firing rate adaptation and IPD tuned excitatory and inhibitory inputs, satisfactorily accounts for experimental findings. Next, the novel feature of a postinhibitory rebound mechanism (modeled as a transient membrane current activated by release from prolonged hyperpolarization) allows us to explain the strong excitatory dynamic response to sweeps in the silent portion of the static tuning curve ("rise-from-nowhere"). Preliminary reports of some aspects of these results have been presented in abstract form (Borisyuk et al. 2001b).

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Cellular properties

Our model cell represents a low-frequency-tuned IC neuron that responds to interaural phase cues. Since phase-locking to the carrier frequency among such cells is infrequent and not tight (Kuwada et al. 1984), our model assumes that binaural input is rate-encoded. We adopt the idealization of our earlier model (Borisyuk et al. 2001a), i.e., we eliminate spikes by implicitly averaging over a short time scale (say, 10 ms). Here, we take as the cell's observable state variable its averaged voltage relative to rest (V). Our formulation is in the spirit of previous rate models (e.g., Grossberg 1973), but extended to include intrinsic cellular mechanisms such as adaptation and rebound. Parameter values are not optimized, most of them are chosen to have representative values within physiological constraints, as specified in this section.

The current-balance equation for our IC cell is

(1)

The terms on the right-hand side of the equation represent (in this order) the leakage, adaptation, synaptic, post-inhibitory rebound, and applied currents. The applied current I is zero except in Fig. 8B, and all other terms are defined below.

Leakage current has reversal potential V_L = 0 mV and conductance g_L = 0.2 mS/cm². Thus with C = µF/cm², the membrane time constant _V = C/g_L is 5 ms.

We model adaptation by a slowly-activating voltage-gated potassium current _aa(V  V_a). Its gating variable a satisfies

(2)

Here a(V) = 1/[1 + exp((V  _a)/k_a)], time constant _a = 150 ms, k_a = 5 mV, _a = 30 mV, maximal conductance _a = 0.4 mS/cm², and reversal potential V_a = 30 mV. The parameters k_a and _a were chosen so that little adaptation occurs below firing threshold, and _a so that when fully adapted for large inputs, the firing rate is reduced by 67%. The value of _a, assumed to be voltage-independent, matches the adaptation time scale range seen in spike trains (Semple, unpublished observations), and it leads to dynamic effects that provide good comparison with results over the range of stimulus rates used in experiments.

The model's readout variable, r, represents firing rate, normalized by an arbitrary constant. It is an instantaneous threshold-linear function of V: r = 0 for V < V_th and r = K · (V  V_th) for V  V_th, where V_th = 10 mV. We set (arbitrarily) K = 0.04075 mV¹. Its value does not influence the cell's responsiveness, only setting the scale for r.

Synaptic inputs

As proposed in Spitzer and Semple (1998) and implemented in the spike-based model of Cai et al. (1998a,b), our model IC cell receives binaural excitatory input from neurons in ipsilateral MSO (e.g., Adams 1979; Oliver et al. 1995) and indirect (via dorsal nucleus of lateral lemniscus (DNLL); Shneiderman and Oliver 1989) binaural inhibitory input from the contralateral MSO (Fig. 1). Thus DNLL, although it is not specifically included in the model, is assumed (as in Cai et al. 1998a) to serve as an instantaneous converter of excitation to inhibition.

View larger version (13K): [in this window] [in a new window]   Fig. 1. Schematic of the model. A neuron in the inferior colliculus (IC), receiving excitatory input from ipsilateral medial superior olive (MSO); contralateral input is inhibitory. Each of the MSO-labeled units represents an average of several MSO cells with similar response properties. Dorsal nucleus of lateral lemniscus (DNLL) is a relay nucleus and is not included in the model. Interaural phase difference (IPD) is the parameter whose value at any time defines the stimulus. The tuning curves show the stimulus (IPD) tuning of the MSO populations. They also represent dependence of synaptic conductances (gE and gI) on IPD.

The afferent input from MSO is assumed to be tonotopically organized and IPD-tuned. Typical tuning curves of MSO neurons are approximately sinusoidal with preferred (maximum) phases in contralateral space (Spitzer and Semple 1995; Yin and Chan 1990). We use a sign convention that the IPD is positive if the phase is leading in the ear contralateral to the IC cell, to choose the preferred phases _E = 40° for excitation and _I = 100° for inhibition. Preferred phases of MSO cells are broadly distributed (Malone et al. 2002; McAlpine et al. 2001). Similar ranges work in the model, and in that sense, the particular values that we picked are arbitrary.

While individual MSO cells are phase-locked to the carrier frequency (Spitzer and Semple 1995; Yin and Chan 1990), we assume that the effective input to an IC neuron is rate-encoded. That is, the convergent afferents from MSO represent a distribution of preferred phases and phase-locking properties so that, either through filtering by the IC cell or dispersion among the inputs, spike-timing information is not of critical importance to these IC cells, for the issues under consideration here. Under these assumptions, the synaptic conductance transients from incoming MSO spikes are smeared into a smooth time course that traces the tuning curve of the respective presynaptic population (Fig. 1). We define g_E and g_I to be these smoothed synaptic conductances, averaged over input lines and short time scales. They are proportional to the firing rates of the respective MSO populations and they depend on t if IPD is dynamic

(3)

(4)

In some cases we consider IPD-insensitive inhibition; and then g_I,tonic  0. Reversal potentials are V_E = 100 mV and V_I = 30 mV. For these maximum conductance values: _E = 0.3 mS/cm², _I = 0.43 mS/cm², total membrane conductance doubles or triples when the stimulus is maximal. Similar conductance changes have been shown in other parts of the CNS (e.g., Anderson et al. 2000). All of our qualitative results are robust with respect to the input parameters.

Rebound mechanism

The postinhibitory rebound (PIR) mechanism (motivated and described in the corresponding subsection of RESULTS) is implemented as a transient inward current (I_PIR in the Eq.1)

(5)

The current's gating variables, fast (instantaneous) activation, m, and slow inactivation, h satisfy

(6)

(7)

where m(V) = 1/[1 + exp((V  _m)/k_m)], h(V) = 1/[1 + exp((V  _h)k_h)]. The parameter values k_m = 4.55 mV, _m = 9 mV, k_h = 0.11 mV, and _h = 11 mV are chosen to provide only small current at steady state for any constant V and maximum responsiveness for voltages near rest. We set _h = 150 ms and V_PIR = 100 mV. Maximum conductance g_PIR equals 0.35 mS/cm² in the subsection on PIR and zero elsewhere. We treat _h, V_PIR, and g_PIR as parameters and discuss their values in Fig. 10 and the accompanying text.

Stimuli

The only sound input parameter that we vary is IPD as represented in Eqs. 3 and 4; the sound's carrier frequency and pressure level are fixed. The static tuning curve is generated by presentation of constant-IPD stimuli for 500 ms. The dynamic stimuli are from two classes. First, the binaural beat stimulus is generated as: IPD(t) = 360° · f_b · t (mod 360°), with beat frequency f_b that can be negative or positive. Second, the partial range sweep is generated by IPD(t) = P_c+ P_d triang(P_rt/360). Here, triang(·) is a periodic function (period 1) defined by triang(x) = 4x  1 for 0 < x < 1/2 and triang(x)= 4x + 3 for 1/2  x < 1. The stimulus parameters are the sweep's center P_c; sweep depth P_d (usually 45°, which we refer to as a sweep of ±45° depth); and sweep rate P_r (in degrees per second, usually 360°/s). We call the sweep's half cycle where IPD increases the "up-sweep," and where IPD decreases the "down-sweep." A dynamic stimulus was usually presented for 1 s or 4 stimulus cycles, whichever is longer.

Simulation data analysis

The response's relative phase (Figs. 5 and 11) is the mean phase of the response to a binaural beat stimulus minus the mean phase of the static tuning curve. The mean phase is the direction of the vector that determines the response's vector strength (Goldberg and Brown 1969). To compute the mean phase, we collect a number of response values (r_j) and corresponding stimulus values (IPD_j in radians). The average phase  is such that tan = [r_j sin(IPD_j)]/[r_j cos(IPD_j)]. For the static tuning curve, the responses were collected at IPDs ranging from 180° to 180° in 10° intervals. For the binaural beat we used the response to the last stimulus cycle recorded with 1-ms steps in time.

We compute a hysteresis measure (Figs. 6, 7, and 11) by using the up-sweep (r_up,j) and the down-sweep (r_down,j) responses, recorded with 1-ms time increments during the final stimulus cycle. They are normalized by the maximum response over the period to yield _up,j and _down,j. The hysteresis measure is then the area of the hysteresis loop formed by the normalized responses approximated by the sum |_{up,j  down,j}| · dt · P_r/360, where dt is the time step (0.001 s) and P_r is the rate of the sweep (in degrees per second). Similar measures were used by Spitzer and Semple (1993) and McAlpine and Palmer (2002).

Transmission delays

In its essence, our model deals with information processing in the IC as represented by a transformation between rate-coded inputs and rate-coded output. (The rate-coding means that the timing information is only represented with accuracy on the order of 10 ms). Therefore most of our qualitative results do not depend on the presence or absence of the short (5-10 ms) delays that exist in the transmission of signal from ear to IC (transmission delay). Moreover, in most of the paper (as in the experiments), we use periodic stimuli with frequencies of only 1 or 2 Hz. Inclusion of the transmission delay would shift the observed response, relative to the stimulus, only by a small fraction of the cycle, resulting in a very slight modification of the results. Therefore for clarity of presentation, we only consider transmission delay in the binaural beats section, where higher frequency stimuli are used. Of course, the discussion on the role of the delays would be more complicated had we varied the carrier frequency.

Simulations

The model is implemented in MATLAB. The system of ordinary differential equations is integrated by the low-order stiff solver "ode23s." Numerical simulations were run under Solaris on a Sun Ultra10 workstation. To check accuracy, the error tolerance was decreased until no noticeable differences were observed.

Experimental procedures

In selected instances, model properties are compared directly with physiological response properties recorded extracellularly from single cells in the IC of anesthetized gerbils. These neural responses derive from a database that has formed the basis of prior published reports (Spitzer and Semple 1993, 1995, 1998) in which methods are described in full detail. All procedures concerning the use of animals were approved by the institutional animal care and use committee. Adult gerbils (Meriones unguiculatus) were anesthetized (pentobarbital sodium, 60 mg/kg ip) for surgical preparation, and anesthesia was maintained throughout the recording session with supplemental injections of ketamine (30 mg/kg/h, im). Single neurons were recorded at histologically confirmed sites in the IC. Digitally synthesized stimuli were presented dichotically through calibrated sealed sound-delivery systems. Sound pressure level (SPL) was calculated in dB (approximately 20 µPa). Static IPDs were generated by dichotic presentation of tone pips differing only in starting phase. Dynamic IPDs were in the form of binaural beats or IPD sweeps generated by triangular modulation of the phase at one ear.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Static tuning curve

SHAPING OF STATIC TUNING CURVE BY SYNAPTIC CONDUCTANCES AND CURRENTS. Static IPD tuning of the model IC neuron is determined by the sharpness of tuning, preferred IPDs and mutual strengths of the incoming excitatory and inhibitory inputs from MSO (Fig. 1), and by the cell's intrinsic properties. Figure 2A shows the tuning of model MSO populations as well as the resulting adapted response of the IC neuron for statically presented inputs. The peak of the model IC cell's tuning curve is close to that of the excitatory MSO population. However, because for our choice of parameters excitation and inhibition are neither in-phase nor exactly in anti-phase, inhibition shifts the peak of the IC curve closer to the minimum of inhibition.

View larger version (21K): [in this window] [in a new window]   Fig. 2. Static tuning curves of the model neurons. A: black curves represent, as functions of IPD, the synaptic conductances (gE, gI) that are delivered to an IC neuron due to the activity of the excitatory (thick) and inhibitory (thin) MSO populations. Gray line is the steady (normalized) firing rate of the IC cell (computed as the steady-state solution of Eqs. 1-4)the IC static tuning curve. B: gray line is the same static IC response as in A, in terms of voltage (V). Black curves are the synaptic currents IE,I = gE,I(V VE,I), corresponding to the conductances in A. Horizontal dotted lines are the resting potential (Vrest= 0 mV) and the threshold potential (Vth = 10 mV).

(8)

TUNED VERSUS TONIC INHIBITION. The relative strength of excitation and inhibition determines the width of the tuning curve in terms of firing rate. Given a certain tuning width for excitation, the tuned inhibition provides a shift in the position of the peak (unless inhibition is exactly in antiphase with excitation) and a mechanism for narrowing the tuning curve (Fig. 3A). Narrower tuning (at the given level of excitation) can also be achieved by application of the IPD-insensitive inhibition or by raising the threshold of the cell (Fig. 3B), at the expense of substantially decreasing firing rates. Comparison between Fig. 3, A and B, leads to a prediction that tuning properties of major inhibitory input in IC may be determined by observation of phase shifts in inhibition blocking experiments (see DISCUSSION).

View larger version (23K): [in this window] [in a new window]   Fig. 3. A: static (fully adapted) tuning curve at different levels of IPD-tuned inhibition (gI = 0, 0.2, 0.43 mS/cm2, gI,tonic = 0 mS/cm2). B: static (fully adapted) tuning curve at different levels of IPD-insensitive inhibition (gI = 0 mS/cm2, gI,tonic = 0, 0.1, 0.2 mS/cm2).

Binaural beats

The binaural beat stimulus has been used widely in experimental studies (e.g., Kuwada et al. 1979; McAlpine et al. 1998; Spitzer and Semple 1993; Wernick and Starr 1968) to create interaural phase modulation: a linear sweep through a full range of phases (Fig. 4A). Positive beat frequency means IPD increases during each cycle; negative beat frequency means IPD decreases (Fig. 4A). Using the linear form of IPD(t) to represent a binaural beat stimulus in the model, we obtain responses whose time courses are shown in Fig. 4A (bottom).

View larger version (29K): [in this window] [in a new window]   Fig. 4. Responses to binaural beat stimuli. A: time courses of IPD stimuli (top) and corresponding model IC responses (bottom). Beat frequencies, 2 Hz (solid lines) and 2 Hz (dashed lines). B-D: black lines show responses to binaural beats of positive (solid lines) and negative (dashed lines) frequencies. Gray line is the static tuning curve. B: example of experimentally recorded response. Carrier frequency, 400 Hz; presented at binaural SPL 70 dB. Beat frequency, 5 Hz. C and D: model responses over the final cycle of stimulation vs. corresponding IPD. Beat frequencies are equal to 2 and 20 Hz, respectively.

To compare responses between different beat stimuli and also with the static response, we plot the instantaneous firing rate versus the corresponding IPD. Figure 4B shows an example that is typical of many experimentally recorded responses (e.g., Kuwada et al. 1979; Spitzer and Semple 1991a, 1993). Replotting our computed response from Fig. 4A in Fig. 4C reveals features in accord with the experimental data (compare Fig. 4, B and C): dynamic IPD modulation increases the sharpness of tuning, preferred phases of responses to the two directions of the beat are shifted in opposite directions from the static tuning curve, and maximum dynamic responses are significantly higher than maximum static responses. These features, especially the phase shift, depend on beat frequency (Fig. 4, C and D).

DEPENDENCE OF PHASE ON THE BEAT FREQUENCY. In Fig. 4D, the direction of phase shift is what one would expect from a lag in response. In Fig. 4C, it is what would be expected in the presence of adaptation (e.g., Smith et al. 2001; as the firing rate rises, the response is "chopped off" by adaptation at the later part of the cyclethis shifts the average of the response to earlier phases). In particular, if we interpret the IPD change as an acoustic motion, then for a stimulus moving in a given direction (e.g., solid curve in Fig. 4, C and D), there can be a phase advance or phase-lag, depending on the speed of the simulated motion (beat frequency).

View larger version (38K): [in this window] [in a new window]   Fig. 5. A: dependence of phase shift on beat frequency in the model. Phase shift is relative to the static tuning curve. B: experimentally recorded phase (relative to the phase at 0.1-Hz beat frequency) vs. beat frequency. Six different cells are shown. Carrier frequencies and binaural SPLs are as follows: 1,400 Hz, 80 dB (*); 1,250 Hz, 80 dB (); 200 Hz, 70 dB (); 800 Hz, 80 dB (); 900 Hz, 70 dB (); 1,000 Hz, 80 dB (). Inset: zoom-in at low beat frequencies. C: phase of the model response, recomputed after accounting for transmission delay from outer ear to IC, for various values of delay (d), as marked. Inset: computation of delay-adjusted phase. D and E: average firing rate as a function of beat frequency: experimental (D, same cells with the same markers as in B) and in the model (E). Multiple curves in Efor different values of gI (gI = 0.1, 0.2, 0.3, 0.43, 0.5, 0.6, 0.7, 0.8 mS/cm2). Thick line, parameter values given in METHODS.

ROLE OF TRANSMISSION DELAY. We propose that the absence of phase advance and high values of phase lag should be attributed to the transmission delay (5 to tens of milliseconds) that exists in transmission of the stimulus to the IC. This delay was not included in our model up to this point, as outlined in METHODS. In fact, we can find its contribution analytically (Fig. 5C, inset). Consider the response without transmission delay to the beat of positive frequency f_b. In the phase computation, the response at each of the chosen times (each bin of poststimulus time histogram) is considered as a vector from the origin with length proportional to the recorded firing rate and pointing at the corresponding value of IPD. In polar coordinates r_j = (r_j,IPD_j), r_j is the jth recorded data point (normalized firing rate) and IPD_j is the corresponding IPD. The average phase of response ('_f) can be found as a phase of the sum of these vectors. If the response is delayed, then the vectors should be r_j= (r_j,IPD_j + d · f_b · 360°), because the same firing rates would correspond to different IPDs (see Fig. 5C, inset). The vectors are the same as without transmission delay, just rotated by d · f_b fractions of the cycle. Therefore the phase with transmission delay _f = '_f + d · f_b. If we plot the relative phase with the transmission delay, it will be _f  0 = ('_f  0) + d · f_b. The phase of the static tuning curve (₀) does not depend on the transmission delay. Figure 5C shows the graph from Fig. 5A (right), modified by various transmission delays (thick curve is without delaysame as in Fig. 5A, right). With transmission delay included in the model, the range of observed phase shifts increases dramatically, and the phase-advance is masked. In practice, neurons are expected to have various transmission delays, which would result in a variety of phase-frequency curves, even if the cellular parameters were similar.

MEAN FIRING RATE. Another interesting property of response to binaural beats is that the mean firing rate stays approximately constant over a wide range of beat frequencies (Fig. 5D and Spitzer and Semple 1998). The particular value, however, is different for different cells. The same is valid for the model system (Fig. 5E). The thick curve is the firing rate for the standard parameter values. It is nearly constant across all tested beat frequencies (0.1-100 Hz). The constant value can be modified by change in other parameters of the system e.g., _I (shown).

DEPENDENCE ON MODEL PARAMETERS. We have illustrated in Fig. 5C that variability in phase-frequency curves can be created by introduction of the transmission delay. Additional variability arises from changes in cellular and input parameters of the system. For example, decreased level of inhibition will increase phase-advance (shift the phase curve downwards) by increasing the effect of adaptation through the increase in firing rates; slower adaptation processes (larger _a) would narrow the range of beat frequencies where phase advance is observed, due to the slowing in the system's response. These parameter dependencies are not shown here.

Partial range sweeps

Spitzer and Semple (1991a) pointed out that the physiologically realistic range of IPDs for small mammals (such as cats and gerbils) at low carrier frequencies is restricted. Therefore they elected to use restricted range IPD modulations as an alternative to binaural beats (Spitzer and Semple 1991a, 1993). We call it partial range sweep stimulus (see METHODS). This stimulation can be crudely interpreted as a motion of a sound source back and forth on an arc in the azimuthal plane around the head of the animal (Spitzer and Semple 1993). An example of IPD time course is shown in Fig. 7A (top).

As in the case of binaural beats, we plot the instantaneous firing rate (averaged over all periods) versus IPD. Figure 6A shows three examples of responses of the model neuron to sweep stimuli with different sweep centers. Figure 6B shows an example of experimental recording. Both in the model and in the data, dynamic responses deviate from the static tuning curve and responses to two different directions of motion in IPD space form hysteresis loops. These are the properties of IPD-sensitive cells in inferior colliculus, as previously reported (e.g., McAlpine and Palmer 2002; McAlpine et al. 2000; Spitzer and Semple 1993, 1998). Such properties are not usually observed for cells at lower levels of processing, such as MSO (Spitzer and Semple 1995, 1998) but are even more pronounced in auditory cortex (Malone et al. 2002).

View larger version (22K): [in this window] [in a new window]   Fig. 6. Responses to periodic triangular sweeps over different ranges of IPD. For illustrative stimulus time course see Fig. 7A, top. Sweep width ±45° at a rate of 360°/s. A: examples of model responses to sweeps with centers at 35°, 55°, and 145°. Black solid lines show the part of responses to the increasing IPD and dashed lines show responses to decreasing IPD (direction is shown with arrows at the top part of the graph). Gray line is the static tuning curve. Shaded area is a measure of hysteresis in the response. B: example of in vivo recording. Notations are the same as in A. Carrier frequency, 1,200 Hz; binaural SPL, 70 dB. C: thick curve shows the measure of hysteresis vs. sweep's center for the model neuron from A. , 3 sweeps in A. Hysteresis is generally reduced as tonic inhibition is increased (successive curves for gI,tonic value from 0 to 0.5 mS/cm2 in 0.1-mS/cm2 increments).

Using the shaded areas (Fig. 6A) to quantify the degree of hysteresis (see METHODS), we plot the hysteresis measure versus sweep center in Fig. 6C. The hysteresis measure curve or, for short, hysteresis curve, readily demonstrates multiple properties of dynamic responses, some of which have been described experimentally. First, there is a local minimum at the neuron's best phase, consistent with the observation that dynamic response properties are less evident near the peak of the static curve. Next, the dynamic effects increase as the static tuning curve steepens. Notice that there is a local maximum near 0° IPD (midline). The physiological range of IPDs lies around that point and the dynamic effect is strong there. The hysteresis measure falls to zero at some range of IPD as there are no static or dynamic responses at those IPDs.

An important feature of the hysteresis curve is its asymmetry. It has been noted before (Spitzer and Semple 1991b) that dynamic effects are stronger on one side from the best phase of the neuron than on the other (asymmetric). Our model reveals that the asymmetry can be due to IPD-tuned inhibition that is neither in-phase nor in anti-phase with excitation. If the tuned inhibition is blocked or replaced with nontuned inhibition, the asymmetry in the model disappears. This leads to another model prediction distinguishing effects of tuned inhibition (see DISCUSSION).

McAlpine and Palmer (2002) studied the effect of blockade and induction of inhibition on hysteresis. The thin curves in Fig. 6C illustrate qualitatively similar effects in the model, because IPD insensitive inhibition is varied (as modeled by changing g_I,tonic see METHODS). As more inhibition is added (which corresponds to an increase in concentration of local GABA application in experiments), the overall hysteresis level is decreased, less asymmetry is observed, the region of nonzero hysteresis measure becomes narrower, and the dip near the best phase disappears.

VARYING SWEEP PARAMETERS. McAlpine et al. (2000) and Spitzer and Semple (1993) reported that variation of sweep parameters does not dramatically affect the response dynamics. We explored the changes using the hysteresis measure curve and varying in the model modulation depth and rate.

View larger version (23K): [in this window] [in a new window]   Fig. 7. Variation of sweep parameters. A: example of stimulus used in partial range sweep simulations. Parameters: 80° center, ±45° depth, and 360°/s rate (top). Modifications of the stimulus (bottom; a single cycle's up-sweep only is shown in each case; the stimulus is continued periodically). B and C: hysteresis of the responses to triangular stimuli from A vs. center of the sweep. Thick curve is the same in each panel (same as in Fig. 6C). B: sweep rate is varied (depth is constant, see A, bottom left). C: sweep depth is varied (rate is constant, see A, bottom right).

PIR

In in vivo IC recordings, the sweep stimuli sometimes evoke a strong excitatory response over the zero-response portion of the static tuning curve (Spitzer and Semple 1998; Fig. 9A of this paper). We call this type of response a "rise-from-nowhere." We set out to test the hypothesis that the "rise-from-nowhere" can be explained by the presence of an intrinsic PIR mechanism. To do this, we have developed and included a voltage-gated postinhibitory rebound (PIR) current in our model (I_PIR, see METHODS). It is meant to give a qualitative/semi-quantitative match with the experimental observations. We do not consider it as a biophysically identified current. We also investigate how the features of dynamic responses, as described in the previous two sections, are modified by PIR.

IMPLEMENTATION AND PROPERTIES OF PIR. In accordance with our hypothesis that dynamic response properties of IC cells primarily depend on intrinsic mechanisms, we implement the rebound mechanism as a transient inward current (I_PIR; see METHODS). The steady-state PIR current is very small at any constant value of voltage, V, as evident from the lack of overlap in the steady state gating functions (Fig. 8A). This means that I_PIR does not contribute significantly to the cell's static response properties, such as the static tuning curve. On the other hand, when the membrane is hyperpolarized for a sufficient duration, the current deinactivates (h grows). If the membrane is abruptly released from hyperpolarization, then the cell can begin to depolarize. I_PIR activates instantaneously and it contributes to regeneratively depolarizing the membrane until h decreases again, inactivating I_PIR on the time scale _h. This current behaves in a manner analogous to the Hodgkin-Huxley Na current (Hodgkin and Huxley 1952), but on different time scales and at different voltage ranges, resembling more closely a transient Ca current (Jahnsen and Llinas 1984; Zhan et al. 1999). Figure 8B shows how the model cell with I_PIR responds to release from hyperpolarizing current steps of different amplitudes. From modest hyperpolarization V returns to rest soon after release. If the hyperpolarizing current is strong enough, there is a large, but saturating, rebound response.

View larger version (17K): [in this window] [in a new window]   Fig. 8. Implementation of postinhibitory rebound. A: steady-state levels of gating variables: activation, m, and inactivation, h, vs. V. B: response to release from hyperpolarizing steps of current. In each case, the current is applied for 200 ms, and then it is turned off. If hyperpolarization is large enough, there is a rebound response.

"RISE-FROM-NOWHERE", DEPENDENCE ON PARAMETERS. For IC cells, the inhibitory input presumably dominates at the lowest portion of the static IPD tuning curve. The membrane, therefore might be hyperpolarized. As the stimulus moves to adjacent regions of IPD (toward the peak), the cell is released from inhibition. This suggests that if the cell possesses a mechanism of postinhibitory rebound, it can be engaged to produce a dynamic response at the zero portion of the static tuning curve ("rise-from-nowhere," Fig. 9A, arrow). Essentially, due to PIR, dynamic stimuli can produce responses beyond the statically determined excitatory receptive field limits. The above scenario is robustly realized in the model with parameter values given in METHODS and the sweep rate of 360°/s. It is illustrated in Fig. 9B (the "rise-from-nowhere" is marked with an arrow). A less dramatic rebound occurs on the "left side" of the tuning curve.

View larger version (32K): [in this window] [in a new window]   Fig. 9. "Rise-from-nowhere." A: experimentally observed response arising in silent region of the static tuning curve (marked with an arrow). Carrier frequency, 1,050 Hz; binaural SPL, 40 Hz. Notations are the same as in Fig. 7A. B: model responses to various sweeps. The rightmost sweep produces "rise-from-nowhere" (arrow). C: model variables during stimulation marked with an arrow in B. C1: stimulus (IPD). Two periods are shown. C2: voltage time course, showing 4 main phases, marked with numbers: 1, hyperpolarization; 2, rebound; 3, removal of inhibition; 4, relaxation. C3: time course of IPIR current (solid line) and its inactivation gating variable h (dotted line). C4: firing rates of excitatory (MSOE) and inhibitory (MSOI) afferent populations, proportional to corresponding synaptic conductances. C5: excitatory, inhibitory, and total synaptic currents.

View larger version (14K): [in this window] [in a new window]   Fig. 10. Dependence of "rise-from-nowhere" on parameter values. A: rebound diminishes for faster h. Shown are responses for h = 60, 90, and 150 ms. B: static tuning curve (gray line). Responses to ±45° sweep, centered at 190°, with a rate of 360°/s. Only one-half of response period is shown (direction indicated by arrow at the top). Different curves, different values of PIR parameters gPIR, VPIR, and h. They are for curve 1, 0.3 mS/cm2, 100 mV, 150 ms; curve 2, 0.35 mS/cm2, 100 mV, 150 ms; curve 3, 0.4 mS/cm2, 120 mV, 150 ms; and curve 4, 0.4 mS/cm2, 120 mV, 250 ms, respectively.

PIR EFFECT ON DYNAMIC RESPONSE: PHASE AND HYSTERESIS. We have shown that the presence of PIR can explain the "rise-from-nowhere" phenomenon. In this section, we show how it modifies the dynamic properties of responses described earlier in the paper.

View larger version (18K): [in this window] [in a new window]   Fig. 11. Change of dynamic response properties in presence of PIR. A: increase of phase advance in response to binaural beats. Thick curve, same as in Fig. 5A, right. B: increase in hysteresis level in responses to sweeps. Thick curve is the same as in Fig. 6C. C: hysteresis level decreases with added inhibition, but less so in presence of PIR. Thick curve is the same as the thin curve in B; thin curves correspond to successively increased levels of IPD-insensitive inhibition: gI,tonic is 0, 0.1, 0.2, 0.3 mS/cm2.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

We have developed an idealized firing-rate-type model for interaural phase sensitive IC cells. This minimal model reproduces many features of physiological responses to simulated acoustic motion and suggests underlying mechanisms. For example, adaptation, PIR and transmission delay shape phase advance and phase lag for beat stimuli; adaptation or PIR shape hysteresis loops (including "rise from nowhere") at different sweep center locations; tuned inhibition provides asymmetry of the hysteresis curve; and overall level of inhibition determines the average hysteresis value. We also propose additional experiments, both in vitro and in vivo, to test our predictions.

Advantages and limitations of averaged voltage model

In contrast to typical firing rate models (e.g., Grossberg 1973; Wilson and Cowan 1972), our model includes intrinsic mechanisms. On one hand, their presence allows us to study a broader repertoire of dynamic behaviors, in which these mechanisms are directly involved. On the other hand, the model is simple enough so that we can discriminate the mechanistic contributions to individual properties. Also, our results show that many of the dynamic response properties can be reproduced without consideration of spike timing or individual synaptic events.

Even though we find the use of an activity model appropriate in this study, it does impose some limitations on the types of questions that we can address. For example, we cannot explore the dependence of response on fast synaptic time scales (say, 10 ms or less) or emergence of loosely phase-locked IC response from convergent tightly phase-locked MSO inputs. Also, the simulation of responses to brief stimuli or responses that are manifested with only a few spikes occurring over a short period of time, is beyond the scope of this study.

Inhibitory inputs to the IC

Still under debate (McAlpine et al. 2000) are the properties of inhibitory inputs to IPD-sensitive IC neurons and the role of inhibition in shaping IC cells' dynamic responses. If the principal inhibition comes from DNLL (Li and Kelly 1992; Shneiderman and Oliver 1989), it could be IPD-tuned (Brugge et al. 1970), as in our model. It is also possible that the inhibition is IPD-insensitive (tonic). Our results suggest experiments that can help to characterize the nature of inhibitory inputs.

It was shown by McAlpine and Palmer (2002) and Cai et al. (1998b) that the overall level of simulated motion sensitivity is reduced by the local application of GABA and increased by the application of bicuculline (McAlpine and Palmer 2002). From these observations on the effect of the pharmacological agents, McAlpine et al. concluded that inhibition is not important in shaping dynamic responses of the IC cells (McAlpine and Palmer 2002; McAlpine et al. 2000). Contrary to this conclusion, the inhibition does play a crucial role in our model, whereas our results on the effect of blockade and induction of inhibition are consistent with those in McAlpine and Palmer (2002). The reason for this apparent contradiction is that McAlpine et al. concentrated in their work on dynamic properties of IC that, in our model, can be explained without engagement of inhibition or PIR.

Adaptation in the IC

It was previously suggested (Cai et al. 1998b; McAlpine et al. 2000; Spitzer and Semple 1998) that firing rate adaptation (firing rate decrease with time) may be responsible for the dynamic response properties in IC. Several different sources can contribute to the firing rate decrease, e.g., adapting excitatory or facilitating inhibitory input, synaptic depression, or intrinsic cellular mechanisms. Recent data show the following: first, the major source for creati