Diverse Synaptic Mechanisms Generate Direction Selectivity in the Rabbit Retina

Directional tuning

The preferred–null axis of each cell was established from extracellular recordings. A stimulus bar was swept across the receptive field in 12 directions, spaced at 30° intervals (Fig.1). The stimulus direction that generated the most action potentials was designated as the preferred direction, and the opposite direction was designated as the null direction. This preferred–null axis was used for all further stimuli during subsequent patch-clamp recordings. The preferred direction was calculated more precisely, during later analysis, from the vector sum of the leading- and trailing-edge spikes (Fig. 1, arrow). In a sample of 10 cells, the average absolute difference between the estimated preferred–null axis and the calculated axis was 12 ± 8°, with the largest difference being 28°. The calculated preferred directions for the on- and off-responses agreed very closely, differing by only 0.3 ± 11° in a group of 11 cells. Directional tuning, evaluated as defined by Equation 5 in Materials and Methods, was the same for the on- and off-responses (Table 1).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Determination of the preferred–null axis. The distance from the origin represents the total number of spikes produced by the leading edge (●) or trailing edge (■) of a single stimulus. The surrounding traces show the extracellular responses for the corresponding directions. The first burst of spikes, produced by the leading edge of the stimulus bar, is the off-response, because the moving bar was darker than the background (0.7 contrast). The second burst of spikes, produced by the trailing edge, is the on-response. The preferred directions for the on- and off-responses, calculated from the vector sum of the data points, were indistinguishable in this cell, and therefore they are marked by asingle arrow, the length of which has been truncated; the lengths of the vector sums were 117 spikes for the on-responses and 120 spikes for the off-responses.

Synaptic current–voltage relations

After the DSGC was patched, the membrane potential was stepped to a range of levels, and the visual stimulus was swept in opposite directions along the preferred–null axis. At negative potentials, the peak inward currents in the preferred and null directions were coincident both for the on- and off-responses, suggesting that the receptive field of the DSGC was well centered with respect to the stimulus (Fig.2A,B). Therefore, comparisons of response timings in preferred and null directions are made directly from the records.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Analysis of the membrane conductance during motion along the preferred–null axis for a single stimulus.Blue data represent preferred direction;red data represent null direction. A,B, Synaptic currents elicited by a bright bar moving in the preferred and null directions at a series of holding potentials starting at −100 mV and incrementing by 15 mV to +20 mV. Thebroken vertical lines delineate the start of the leading- and trailing-edge responses. Because the stimulus bar was brighter than the background, the leading-edge response is an on-response, and the trailing-edge response is an off-response. Thesymbols above the records show the time points for measuring the current–voltage relations shown in E. Thesolid lines show examples of linear fits used to determine the synaptic reversal potentialV_r and the synaptic conductanceg_T . C, D,I–V relations were measured at 10 msec intervals to plot out the time dependence of these parameters.F, V_r in another cell showing the relative timing of extracellular action potentials during preferred-direction motion and recorded before applying the patch-clamp electrode. G, H,V_r andg_T recorded during negative-contrast stimuli showed characteristics similar to those for positive contrasts.

The sets of current records for the preferred and null directions (Fig.2A,B) show systematic changes in the time course of the responses as a function of the holding potential, reflecting the changing balance of excitation and inhibition that occurs as the stimulus sweeps across the receptive field. In particular, at positive holding potentials, where the inhibitory inputs are more evident, the pronounced differences in the current wave-forms for preferred and null directions support the hypothesis that postsynaptic inhibition is important for generating direction selectivity (Taylor et al., 2000).

A quantitative analysis of these responses was performed by generatingI–V relations of the net light-evoked current every 10 msec for the duration of the visual responses. TheI–V relations were essentially linear over much of the voltage range (Fig. 2E), and the conductance was estimated by fitting a straight line to eachI–V plot. The fitted lines provided two parameters at each time point: the intercept, giving an estimate of the synaptic reversal potential V_r (Eq. 1; Fig. 2C,F,G), and the slope, giving the size of the light-evoked conductanceg_T (Fig.2D,H).

Synaptic reversal potential

The trajectory of the synaptic reversal potential was similar in every DSGC. In the preferred direction, V_r ramped up rapidly toward 0 mV during the early phase of the response, to both the leading edge (on-response) and the trailing edge (off-response) of a positive-contrast stimulus (Fig. 2C,blue line). V_r then ramped down during the late phase of the response, typically ending around −50 mV. The occurrence of the positive reversal potential correlated precisely with the timing of extracellular spikes recorded before the intracellular recording, as shown for another cell (Fig.2F).

In the null direction, V_r generally remained negative, consistent with the dearth of spikes in the extracellular recordings. During the late phase of the off-response, however, V_r consistently ramped toward more positive values, producing a crossover in the reversal potential trajectories. When a positive-contrast stimulus was used, the crossover occurred during the late phase of the trailing-edge response (Fig.2C,F, X arrow). When a negative-contrast stimulus was used, the crossover occurred during the late phase of the leading-edge response (Fig. 2G, X arrow). The presence or absence of the crossover was robust enough to allow the sign of the contrast edge to be reliably determined by inspection of the V_r records. Such crossover was never observed in the reversal potential trajectories of the on-response, which superimposed during the late phase of the response, indicating that the synaptic currents were not direction selective at these times. These distinct characteristics suggest that there are basic differences in the synaptic mechanisms underlying direction selectivity for the on- and off-responses.

Synaptic conductance

The peak conductance was larger in the null direction than the preferred direction (Fig.2D,H). This would be compatible with a presynaptic model of direction selectivity in which inhibition was larger in the null direction, with no change in the time course of the inputs. However, a greater peak conductance in the null direction would also be compatible with a postsynaptic model in which the closer temporal correlation of inhibition and excitation in the null direction results in greater summation of the inhibitory and excitatory conductances, with no change in their relative magnitudes.

The off-response may satisfy the latter possibility. Although the peak off-conductance was smaller in the preferred direction, the waveform was consistently broader (Fig.2D,H, blue line). Thus it is possible that the same total conductance is activated, but with differing temporal dispersion. In contrast, the on-response appeared to have a very similar time course in the preferred and null directions, more in line with a presynaptic mechanism. The reversal potential trajectories support this interpretation. During the on-response, the reversal potential in the null direction was essentially constant, indicating that excitation and inhibition had very similar time courses (Eq. 2). During the off-response, the reversal potential in the null direction changed continuously, consistent with different spatiotemporal characteristics for the two inputs, as expected for a postsynaptic mechanism.

For a purely postsynaptic model, the total synaptic conductance should be equal in the preferred and null directions, whereas for presynaptic mechanisms, there should be an imbalance. We examined this quantitatively by integrating the conductance records in the two directions. When the preferred G_T is plotted against the null G_T (Fig.3), many of the points lie below the unity line, indicating that the integrated conductance was slightly larger in the null direction, averaging 118 ± 23% for 28 DSGCs. Although the difference was small and not shown by some cells, it suggested that there might be presynaptic components to the directional mechanism.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Integrated conductance. Each pointshows a single measurement of the total integrated conductance,G_T. The solid symbolsshow averages of one to three determinations in the 16 cells that were analyzed further. The open symbols show the average of one to three measurements in 12 additional cells.

Excitatory and inhibitory components

To identify whether the asymmetry lies in the excitatory or inhibitory inputs to the ganglion cells, we resolved their separate contributions to the conductance records. Similar to previous electrophysiological studies, we found that the on-response of a DSGC could be quite different in size from the off-response, with the off-responses tending to be larger. To make a meaningful comparison between the on- and off-responses, cells with markedly different responses were excluded from the analysis, leaving a subset of 16 cells with broadly comparable responses (Fig. 3, ●). For these analyzed cells, the G_T of the on-responses in both directions ranged from 21 to 131% of theG_T of the off-responses (mean = 49 ± 26%).

To separate the conductance records into excitatory and inhibitory components (Eq. 3 and 4), we made the following assumptions (see Materials and Methods). (1) The DSGC is isopotential; theoretical results from Koch et al. (1990) indicate that this approximation will result in errors in the magnitude but not the ratio of the conductances (see Discussion). (2) Synaptic and dendritic membrane conductances are linear. (3) There are only two synaptic conductances contributing to the responses: excitation and inhibition. The reversal potentials for excitation and inhibition were assigned for each cell as described in Materials and Methods. In the sample of 16 cells,V_e = +4.3 ± 5.5 mV with a range 0 to +20 mV, and V_i = −65.7 ± 2.3 mV with a range −65 to −75 mV. (4) The excitatory and inhibitory conductances are greater than or equal to zero.

The calculated excitatory and inhibitory components provide a rationale to account for the time-dependence of V_r and g_T . The qualitative features described here were similar for all cells, although there were marked quantitative differences from cell to cell. The peak amplitude of the excitatory conductance, g_e , tended to be larger in the preferred direction, but the overall time course was independent of direction for both the on- and off-responses (Fig.4A,E). By contrast, the time course of the inhibitory conductance,g_i , differed dramatically in the preferred and null directions (Fig.4B,F).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Separation of g_Tinto g_e andg_i . Each trace is the response to a single stimulus. Comparison of the time course ofg_e (A, E) andg_i (B, F) for positive (A, B) and negative (E, F) contrast stimuli. C, D,G, H, The same records replotted to compare preferred- and null-direction responses. The solid lines under the records delineate the intervals used to calculate the direction-selective integrals in Figure 5.

For the off-response, the inhibition in the preferred direction was significantly delayed compared with the null direction, consistent with an asymmetric spatial offset (Fig.4B,F). For the on-response, a distinction must be made between positive- and negative-contrast stimuli. During negative-contrast stimuli, the inhibition was smaller in the preferred direction but its time course was unchanged, suggesting that presynaptic mechanisms control this input (Fig.4B). During positive-contrast stimuli, inhibition comprised an early transient phase (Fig. 4F, delineated by the solid line beneath the trace) followed by a sustained phase (Fig. 4F, delineated by thebroken line beneath the trace). The magnitude of the early phase was strongly modulated in opposite directions, also consistent with presynaptic direction-selective mechanisms. The late phase of the on-response was nondirectional and superimposed for opposite-direction stimuli. Moreover, synaptic currents active during this late phase did not generate spikes in either direction (Fig. 2F). In contrast, a sustained phase was not present during on-responses elicited during negative-contrast stimuli (Fig. 4B), and it was clear during such responses that, unlike the off-responses, the inhibition is not delayed in the preferred direction. Although it is more difficult to discern during positive-contrast stimuli, the on-responses appeared to lack any inhibitory delay in the preferred direction because, as noted above, the late phase of the on-response was nondirectional (Fig. 4F).

The reversal potential trajectories can be understood in terms of the relative trajectories of the excitatory and inhibitory conductances in each direction. In the preferred direction, excitation dominated because the inhibition was either reduced (on-response) or delayed (off-response) (Fig. 4C,G). In the null direction, V_r was essentially constant during the on-response because the excitation and inhibition had similar time courses (Fig.4D,H). By contrast,V_r tended to increase during the off-response, leading to the characteristic crossover of theV_r trajectories. The crossover resulted from a fall in the ratio of inhibition to excitation late in the response, attributable to an earlier onset and a narrowing of the inhibitory component. In summary, on-responses appeared to be dominated by a directional inhibitory input, whereas off-responses displayed both a directional inhibitory input and a change in temporal correlation as expected for a postsynaptic mechanism.

The inhibition during the on-responses was invariably larger in the null direction, but it is not clear whether other factors contribute to the imbalance in the total conductance shown in Figure 3. Therefore, we integrated the two conductance components,g_e andg_i, separately for the on- and off-responses. The integrals for the on-responses during positive-contrast stimuli were calculated over the transient phase (see above), delineated by the solid lines beneath the records in Figure 4F, thereby avoiding the sustained nondirectional inhibition active late in the on-response. Inclusion of this nondirectional component during the analysis only reduced the estimated asymmetry of the on-inhibition.

The excitatory component was larger in the preferred direction (Fig.5A), and on average, this presynaptic asymmetry was similar for the on- and off-responses (Table1). Similarly, the inhibitory component was consistently larger in the null direction (Fig. 5B). These differences are illustrated more clearly when the preferred/nullG_e ratio is plotted against the null/preferred G_i ratio (Fig.5C). For a purely postsynaptic scheme, the points should cluster around the intersection of the unity lines. For the great majority of points, both the G_e ratio and the G_i ratio were greater than unity, indicating that the cells received both directional excitation and directional inhibition, consistent with their preferred direction. The off-responses clustered nearer the intersection than the on-responses, indicating that the presynaptic mechanisms were more potent for the on-responses. The mean values of theG_e andG_i ratios for the on- and off-responses are shown in Table 1, which also shows the preferred–null ratio of the total directional conductance,G_D, defined as the sum ofG_e andG_i evaluated over the “On” and “Off” intervals delineated in Figure 4.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 5.

Integrals of the conductance components for the on-response (A, ■) and the off-response (B, ●). Note the change in scale betweenA and B. The unity slope line shows the expectation for equal responses in the preferred and null directions. C, Conductance ratios obtained from the data in A and B; the on–off pair for an outlier is connected by a dashed line.

For the off-responses, the directional excitation and directional inhibition were of equal magnitude, and thus the total off-conductance was balanced for the preferred and null directions. For the on-responses, directional inhibition tended to be stronger, and thus the total on-conductance was greater in the null direction. The imbalance observed in Figure 3 for the integrated conductance can be attributed, therefore, mainly to the on-response, diluted by the larger off-response, and by inclusion of the nondirectional conductance active during the late phase of the on-response. The broad range of ratios in Figure 5C is noteworthy, indicating that the cells can be direction selective when they receive nondirectional excitation or nondirectional inhibition.

The extent of the spatial offset of the inhibition was estimated from the timing differences of the inhibitory peaks. The calculations assume that the receptive field is centered with respect to the stimulus (which is aligned to the soma) and that the synaptic delays are the same in the null and preferred directions. Consistent with this assumption, the excitatory conductance peaks were coincident in the preferred and null directions (Fig.4A,E). The absolute peak-conductance time, relative to the start of a stimulus trial, was measured for excitation and inhibition in the preferred and null directions. The difference, dT, preferred minus null was converted into an equivalent spatial offset; for a stimulus speed of ν μm/sec, the spatial offset was estimated as dX = ν · dT/2. The only spatial offset that diverged significantly from zero was that for the inhibition during the off-response, for which dX = 160 ± 56 μm (Fig.6, Table 1). Because the excitatory conductance peaks were coincident in the null and preferred directions, the spatial offset of the inhibition was also estimated from the timing differences between the excitation and inhibition in a given direction. This does not assume that the receptive field is centered relative to the stimulus. The derived spatial offset for the inhibition during the off-response in the null direction averaged 148 ± 48 μm, in close agreement with the figure above.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 6.

The off-inhibitory conductance is offset in the null direction. The symbols show the inferred spatial offset, dX, calculated from the shift in the peak-conductance time observed for preferred- and null-direction motion. A positive spatial offset means that the conductance was delayed during preferred-direction motion relative to null-direction motion. ■, On-responses; ●, off-responses. The on-response data are a subset of 9 of the 16 cells in which the preferredg_i displayed a clear peak early in the response. Also shown are the mean and SDs for the groups of points. Only the inhibitory off-conductance is significantly different from zero.

High chloride

Previously, we showed that the inhibition acts directly on the dendrites of the DSGC, and we argued that the direction selectivity is generated postsynaptically (Taylor et al., 2000). The key evidence was the apparent loss of directional responses observed at −30 mV when the patch electrode contained a high chloride concentration (Fig.7A,C). The conductance analysis provides further insights into the effects of high intracellular chloride (Fig.7B,D). At the commencement of patch recording, the V_r trajectory was similar to that observed in control cells (Fig. 7B). Approximately 7.5 min later,V_r(t) had shifted to more positive potentials, with little change in the conductance (Fig.7D), and the crossover was still evident during the late phase of the off-response. The main effect of the high chloride appeared to be a shift in V_i. In five cells, V_i andV_e were initially −63 ± 7 mV and +8 ± 4 mV, but 2–6 min later,V_i andV_e were set to −28 ± 5 and 18 ± 11 mV to satisfy assumption (4) above. The high chloride did not change the magnitude of the excitatory and inhibitory conductances that impinge on the DSGC; thus, the total integrated conductance did not change significantly from the initial value. Under physiological conditions, the positive reversal potential for inhibition tends to generate spikes in the null direction, thus reducing direction selectivity.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 7.

Effect of high intracellular chloride.A, B, Records in response to preferred (blue) and null (red) stimuli at the commencement of the whole-cell recording. C,D, Similar data obtained 7.5 min later. The shift inV_i is evident, but the overall time course and magnitude of the conductances are unchanged.

Locus of direction selectivity

The results reported here indicate that direction selectivity is generated by the combination of three distinct synaptic asymmetries: (1) a presynaptic mechanism producing stronger excitation in the preferred direction, (2) a presynaptic mechanism producing stronger inhibition in the null direction, and (3) postsynaptic interaction of the excitation with spatially offset inhibition. The three mechanisms were not uniformly expressed by the on- and off-responses, although the extracellularly recorded responses were very similar. The on-responses appeared to rely on presynaptic mechanisms. The off-responses were more consistent with a mix of presynaptic and postsynaptic mechanisms.

For both the on- and off-responses, the precise mix of presynaptic and postsynaptic mechanisms was quite variable, and in some cells one or another of the presynaptic components was absent (points close to the unity lines in Fig. 5). This variability raises the possibility that DSGCs are programmed to generate a specific functional phenotype during development but that there is considerable latitude in the eventual mix of synaptic mechanisms used to achieve this goal. The large variability in the contribution from each of the three mechanisms, both among different cells and between the on- and off-responses, suggests that no single mechanism is essential for the generation of direction selectivity. Moreover, the mechanistic variability argues against the idea that different mechanisms are required to process different types of visual stimuli (Grzywacz et al., 1998). Thus, the robustness of the directional trigger feature over a wide range of stimulus conditions (Barlow and Levick, 1965; Wyatt and Daw, 1975) probably reflects robustness in the individual synaptic mechanisms.

A postsynaptic mechanism reliant on shunting inhibition requires that the synapses be electrotonically remote from the soma for direction selectivity to be generated locally within the dendritic tree (Torre and Poggio, 1978; Koch et al., 1983). The linearity of theI–V relations and the clear reversal potentials observed at positive potentials seem inconsistent with an electrotonically extensive dendritic arbor. However, the dendritic arbors were probably more electrotonically compact during the recordings shown here, because the intracellular solution contained cesium ions rather than potassium ions. Further experiments will need to be performed using potassium-based intracellular solutions to determine whether the off-arbors really are electrotonically extensive. The off-dendritic arbor has a greater density of thin terminal dendrites than the on-dendritic arbor and is located more distally from the soma (Vaney, 1984; Oyster et al., 1993). Both of these factors could favor a more electrotonically extensive arbor. Moreover, the off-dendrites can arise from on-dendrites of any branching order, and thus any postsynaptic mechanism operating within the on-arbor may inappropriately shunt excitation being transmitted from the off-arbor.

The conductance analysis revealed that both the excitatory and inhibitory inputs to DSGCs may be tuned directionally and that the response bias is consistent with the preferred direction of the cells. In an earlier study we argued for a simpler scheme involving only postsynaptic mechanisms (Taylor et al., 2000). Yet the study reported here presents new evidence for presynaptic mechanisms. How can this be reconciled with the previous findings? The earlier study attempted to account for the behavior of the DSGCs by the most parsimonious interpretation, which involved testing the data against each possible mechanism rather than combinations of mechanisms. In the first experiment, depolarization of the DSGCs from −70 to −30 mV increased the direction selectivity of the cell, providing evidence against the presynaptic excitatory mechanism. In the second experiment, raising the inhibitory reversal potential with high intracellular chloride mostly abolished the direction selectivity, consistent with a postsynaptic mechanism. If the response had relied on a presynaptic mechanism, then the high chloride should have either reversed the direction selectivity (for a directional inhibitory input) or not affected the direction selectivity (for a directional excitatory input). The parsimonious conclusion of Taylor et al. (2000), that there appeared to be no asymmetry in the synaptic inputs to the DSGCs, is consistent with the present finding that the total conductance is almost equal for the preferred and null directions (Fig. 3). This balance of the conductance in the preferred and null directions is attributable partly to the complementary balance of directional excitation and directional inhibition and partly to the tendency for off-responses to be larger, a factor that was not controlled for previously.

Errors in conductance estimates in a non-isopotential neuron

To separate the conductance records into excitatory and inhibitory components, we assumed that the DSGCs were isopotential (Anderson et al., 2000; Borg-Graham, 2001). However, a simple test confirmed the expectation that this is an approximation. During a small voltage step, the capacitive transient in the DSGCs required a sum of exponentials to be fitted accurately (data not shown), demonstrating that these cells are not isopotential. In a non-isopotential cell, a point voltage clamp at the soma will not faithfully resolve synaptic currents impinging on the dendritic arbor. Because the inputs become more electrically distant, their visibility to an electrode at the soma deteriorates.

Koch et al. (1990) showed that the actual conductance at the dendritic site is always larger than the conductance measured at the soma but that this error is independent of the synaptic reversal potential, even for mixed excitatory and inhibitory inputs. Thus, even if there are nonlinear shunting interactions between inhibition and excitation at some given potential, this will not affect the estimate of the conductance ratios of the two inputs determined over a range of potentials. The theoretical approach used by Koch et al. (1990) assumed that the time-constants of the synaptic inputs are long compared with the time-constants for the decay of voltage transients within the dendritic arbor. This requirement is adequately satisfied in the present case. The capacitive transients at the onset and termination of a voltage step are barely visible in the records shown in Figure 2because they decayed rapidly compared with the time course of the light-evoked synaptic currents.

Separation of the conductance into excitatory and inhibitory components also required estimates for the reversal potentials of the two inputs. These were not measured directly and therefore were set to reasonable values: 0 mV for excitation and −65 mV for inhibition. Adjustment was sometimes required to ensure that both conductance components remained positive. Changing either V_e orV_i will change the relative magnitude of the associated conductance (Eq. 3 and 4). Thus, uncertainty in the reversal potentials implies uncertainty in the magnitude of the conductances. However, this uncertainty does not affect the conclusions from the data based on the ratios of the conductances, because the same reversal potentials were used for analysis in the preferred and null directions.

The estimation of the presynaptic components relies on the ratio of conductances measured in the preferred versus null directions. Because the total light-evoked conductance was essentially constant in the preferred and null directions (Fig. 3), the conductance errors will be independent of stimulus direction. The conductance errors for the off-responses, however, are complicated by the delay in the inhibitory conductance. The temporal offset of inhibition relative to excitation in the preferred direction will tend to reduce the conductance errors of both inputs, whereas the temporal correlation of the inputs in the null direction will tend to increase the errors. Thus, the excitatory ratio will tend to be overestimated, and the inhibitory ratio will tend to be underestimated. Without quantitative modeling, it is not possible to say how large these errors in estimation might be, but they are probably not large, because the observed conductance ratios were not correlated with the magnitude of the total conductance.

Implications for synaptic circuitry

Our results indicate that direction selectivity for both on- and off-responses is generated to some extent by a push–pull mechanism, where complementary changes in excitation and inhibition drive the cell to a greater or lesser degree with only small changes in overall membrane conductance (Watanabe and Murakami, 1984; Anderson et al., 2000). However, although this study characterized the synaptic inputs to the DSGCs in unprecedented detail, our results raise many questions. For example, the results do not reveal whether the stronger excitation in the preferred direction arises from preferred-direction facilitation or from null-direction inhibition acting presynaptically on the excitatory inputs. Nor do the results tell us whether both the glutamatergic input from cone bipolar cells and the cholinergic input from starburst amacrine cells are directional and, if so, whether similar synaptic mechanisms are responsible for their direction selectivity. For example, it has been proposed that the terminal synapses on the radial dendrites of starburst amacrine cells will be more strongly activated by centrifugal stimulation than by centripetal stimulation (Borg-Graham and Grzywacz, 1992), but such a structural mechanism is unlikely to apply to the compact axon terminal of a bipolar cell.

Many of the neuronal models of direction selectivity that have been put forward invoke a key role for starburst cells, which are two mirror-symmetric populations of amacrine cells that cofasciculate with the bistratified DSGCs (Vaney et al., 1989; Vaney and Pow, 2000). These amacrine cells contain both acetylcholine and GABA and thus could contribute to both the excitatory and inhibitory mechanisms (Vaney et al., 1989; Borg-Graham and Grzywacz, 1992; Grzywacz et al., 1997). The role of starburst cells within the direction-selective circuit is controversial, however, with one report claiming that they are critical (Yoshida et al., 2001) and another claiming that they are not essential (He and Masland, 1997). Available evidence indicates that the two populations of starburst cells are structurally and functionally equivalent (except for the sign of their responses) (Bloomfield, 1992), and therefore they may underlie the presynaptic mechanisms that are common in the on- and off-responses. In contrast, a different type of amacrine cell branching only in the off-sublamina may mediate the spatially offset inhibition that distinguishes the off-responses from the on-responses. A change in the timing of the off-inhibition in the preferred and null directions predicts a spatial offset of ∼160 μm, which is approximately half the dendritic diameter of a starburst cell in mid-peripheral retina (Tauchi and Masland, 1984; Vaney, 1984).

Our results show that as far as the spike output is concerned, the on- and off-dendritic arbors of a DSGC perform essentially the same computation but use different combinations of synaptic mechanisms. The challenge for future research is to identify the synaptic circuitry that generates the directional signals, both presynaptically in the excitatory and inhibitory neurons and postsynaptically in the DSGCs. This is the same challenge that faced Barlow et al. (1964, 1965) almost 40 years ago, when they first characterized the DSGCs in rabbit retina; however, we now know that the synaptic mechanisms underlying direction selectivity are more diverse than they had envisaged.