The ability to discriminate between lights of different spectral composition is a property common to many visual systems, both vertebrate and invertebrate. Most mammals can be classified as dichromats because they express only two spectrally distinct cone pigments, but some primates, including humans, express three different cone pigments and are trichromats (Jacobs, 1993). The tammar wallaby, a small kangaroo-like marsupial from Australia, has been shown to behave as a dichromat (Hemmi, 1999). The short and middle wavelength-sensitive cones make up about 5% of the total photoreceptor population, a typical mammalian value. The cones are expressed over the entire retina, although with marked differences in their abundance (Hemmi & Grünert, 1999). The distribution of the middle wavelength-sensitive cones follows approximately the ganglion cell distribution and shows a strong visual streak

These results are based on experiments performed on 18 adult tammar wallabies (Macropus eugenii) destined to be culled from a breeding colony maintained at the Research School of Biological Sciences at the Australian National University. The animals were bred and raised in social groups in large outdoor paddocks. Twenty-nine cells in seven animals produced adequate light responses. The color opponent cells reported here were obtained from two of these animals.

Experiments were performed on a whole mount, isolated retinal preparation (Taylor & Wässle, 1995; Peters & Masland, 1996). The relatively low yield reflects difficulties particular to the wallaby eye. The avid association between the pigment epithelium and the retina made it very difficult to isolate the retina without damaging or even losing the photoreceptor outer segments. Thus a number of retinas, which otherwise appeared healthy, failed to produce light responses. The experiments comply with the Australian Capital Territory Animal Welfare Act (1992) and were covered by the Ethical Protocol (J.NS.83.98) approved by the Animal experimentation Ethics Committee of the Australian National University.

The animals were dark-adapted for a period of at least 1 hr prior to the experiment, and all subsequent manipulations were performed under infrared illumination (>900 nm). The anesthetic procedures were performed in accordance with the relevant institutional guidelines. The wallaby was sedated by intramuscular injection of 20 mg/kg ketamine (Ketalar; Parke-Davis, West Ryde, New South Wales, Australia) and 8 mg of xylazine (Rompun; Bayer, Sydney, New South Wales, Australia). Deep anesthesia was induced by an intravenous injection of 10 mg/kg thiopentone sodium (5% solution Pentothal; Boehringer, Ingelheim, Germany) through a butterfly needle (23G) inserted into the lateral tail vein. Both eyes were removed and the animal immediately killed by an anesthetic overdose of sodium pentobarbitone (Nembutal, Boehringer). Eyecups were prepared and placed in oxygenated Ames medium (Sigma, Castle Hill, New South Wales, Australia) bubbled with 95% O₂/5% CO₂. The retinas were dissected free and cut into several pieces, which were adhered photoreceptor side down onto glass coverslips previously coated with Celltak (Becton Dickinson, North Ryde, NSW, Australia). These pieces were maintained in a light-tight holding chamber and placed into the recording chamber as required. The recording chamber was continuously perfused (3 ml/min) with fresh oxygenated Ames medium.

Cells were visualized under infrared differential interference contrast optics (820 nm) on an Olympus
BX-50 upright microscope. Healthy retinas were completely transparent and while the vitreal surface and fibre tracts were clearly seen, the cell somas beneath the surface could not be identified. For this reason, we injected 3 μL of a 3% Acridine orange (Sigma-Aldrich, Sydney, NSW, Australia) at the inflow end of the recording chamber. Using very brief exposure to dim epifluorescent light, we were able to locate cell somas. The overlying vitreal membrane was microdissected clear of the cell somas using a broken patch electrode. Recordings were made in whole-cell mode using patch clamp electrodes (5-8 MΩ resistance) filled with the following internal solution: K-gluc 110 mM, NaCl 10 mM, EGTA 10 mM, Na-ATP 5 mM, Na-GTP 0.1 mM, HEPES 5 mM and pH 7.4. The current signal was sampled at 10kHz and filtered at 2.5kHz through the 4-pole Bessel filter in the Axoclamp 200B patch clamp amplifier. In an effort to recover the morphology, the cells were routinely filled with neurobiotin during the recording period. Unfortunately, none of the color opponent cells were recovered, probably due to cell damage upon removal of the patch electrode resulting in loss of the neurobiotin (Taylor & Wässle, 1995).

The responses of the ganglion cells are described using system identification techniques, which characterize response properties in terms of weighting functions called kernels. Kernels generalize the impulse response in the classical linear time-invariant system formulation (see James, 1992, for a more complete description). In each case, the first order, or linear kernel, represents the weighting function, which, when convolved with the stimulus, describes the linear component of the response. The more familiar impulse response is related to this linear kernel by a constant scalar, but only for stimuli within the linear range. This constraint upon the stimulus is not necessary for an accurate estimation of the linear kernel from the systems analysis because static nonlinearities are accounted for by the higher order kernels. The responses to stimuli well outside the linear range, however, will be dominated by the higher order terms, and, thus, the relatively smaller linear kernel may be poorly resolved. This was not an issue in the current series of experiments, because the linear kernel dominated the responses in all the cells tested.

We used high-bandwidth pseudo-random “white noise” stimuli, which efficiently probed the system behavior at a given level of light adaptation in the limited recording time available. This approach has been successfully adopted in neuroscience and has been described in a number of publications to which the reader may refer for more information (e.g., Sakuranaga, Sato, Hida, & Naka, 1986; James, 1992; Reid & Shapley, 1992).

Least squares estimates were obtained for the first- and second-order kernels by performing multilinear regression of the vector of response values on a set of regression component vectors. The regression component vectors comprise delayed versions of the stimulus signals, corresponding to the first-order kernel values at the range of delays under consideration, and pointwise products of delayed versions of the stimulus signals, corresponding to the second-order kernel values at the delay-pairs considered. This procedure can be contrasted with the cross-correlation procedure, which converges in the long run to the desired kernel values, but which does not correct for nonorthogonality of the regression components in a finite-length dataset. It also contrasts with the M-sequence technique sometimes used, which has (almost) perfectly orthogonal regression components but which must come from a very specific set of signal sequences. The method used here allows great generality in the choice of test signals used.

For second-order kernels, only temporal delays with a significant amount of power were included in the final fit. Significance was judged based on a cross-validation procedure, using different repeats of the same stimulus. When inclusion of a component in the model does not lead to an improvement of the predicted fit to the validation data, or leads to a worsening, then that component is excluded from the model. Third-order kernels never contained significant information and, therefore, were not included in the final fit. The kernels calculated for a number of repeats were averaged. The number of repeats varied between one and four for the different cells and the recording time per repeat varied between 17 s and 2 min.

Light stimuli were generated on a Barco CRT monitor (model CCID 7551, Barco, Kortrjk, Belgium), imaged onto the photoreceptors through a 35 mm camera objective and the 40x/0.8NA microscope objective. The refresh rate of the monitor was 75 Hz with a maximum luminance of 110 cd/m², which produced a maximum illuminance at the photoreceptor outer segments of 101 lm/m². Light intensity was further attenuated using calibrated neutral density filters (Melles Griot, Japan) so that the final illuminance was about 10 lm/m². Images were generated using programs developed in the lab and incorporating routines from the VideoToolbox (Pelli 1997). The stimulus routines were run from within the Igor program (Wavemetrics, Lake Oswego, OR, USA). The spectral characteristics of the monitor were measured using a spectrophotometer (s1000; Ocean Optics, Dunedin, FL, USA), calibrated against a secondary standard light source (LS-65-8D Rev-B; Hoffman Engineering, Stamford, CT, USA).

Three types of light stimuli were presented. (1) A color full-field stimulus was used to explore the spectral characteristics of the cells. The three phosphors (red, green, and blue) of the CRT were modulated independently of each other according to a cyclically delayed version of the same pseudo-random noise sequence with contrasts of -1, 0, and 1 for each phosphor. The probabilities for the three contrasts were 0.25, 0.5, and 0.25, respectively. (2) An achromatic spatio-temporal white-noise stimulus consisted of an 11 x 11 checkerboard, where each check was 43 x 43 μm. The contrast of each check was set to -1, 0, and 1 with the same probabilities as above. The different checks were again stimulated by cyclically delayed versions of the same random sequence. This stimulus allows a description of the spatial properties of the receptive fields. (3) A chromatic center surround stimulus, where a central disk and a surrounding annulus were each modulated in the same way as the full-field stimulus, allowed us to describe the spectral responses of the center and the surround independently. The size and position of each stimulus was based on the results of the spatio-temporal stimulus. Unmodulated regions of the monitor were kept at mean luminance at all times, for all conditions.

The first and the last second (75 stimulus frames) of each response were discarded prior to analysis. The final run length was either 1,024 or 2,048 frames for the full-field stimulus, and 8,192 for the spatio-temporal and center-surround stimuli.

The red, green, and blue phosphor kernels, which best reproduced the responses from the stimulus, represent the sum of contributions originating from the short and middle wavelength-sensitive cones, designated S cones and M cones, respectively. This is because both the S cones and M cones absorb light emitted by all three phosphors and therefore contribute to all three kernels, albeit with very different weightings. For each cone type, however, we can calculate its sensitivity toward each of the three phosphors. This is purely a function of its own spectral sensitivity and the spectral characteristics of the phosphors. The spectral sensitivities of the cones were modeled based on the reported peak sensitivities of 539 nm for the M cones and 420 nm for the S cones (Hemmi, 1999; Hemmi et al., 2000) and using the spectral sensitivity templates of Stavenga et al., 1993. The resulting sensitivities for the M cones were (red:green:blue [RGB]) 0.27:1:0.30 and for the S cones (RGB) 0.025:0.078:1.

These sensitivities were used to derive the corresponding linear and nonlinear second-order kernels for the S-cone pathway and M-cone pathway. These pigment kernels were calculated according to Equations 1 and 2 shown below. Note that the peak in spectral sensitivity of the S cones is so low (420 nm) that the S cones contribution to the green phosphor kernel is less than 8% of its contribution to the blue phosphor kernel. This means that the green phosphor kernel is almost identical to the M (cone) kernel. Nonetheless, it is important to keep in mind the distinction between the S and M kernels, which provide an estimate of the inputs provided by the S and M cones, and the RGB kernels, which are actually measured and show the system response to the three phosphors.

The relationship at any point in time between the phosphor intensities and cone excitations can be expressed by multiplication of the appropriate vectors by a matrix, A, consisting of weights obtained by integrating over wavelength the product of phosphor spectral emission with the cone spectral sensitivity. If p(t) is a 3 x 1 vector of RGB intensities at time t and e(t), the photoreceptor excitations, then

The inverse relation is obtained by inverting the 2 x 3 matrix, to give

The first-order model for mapping the three channel phosphor stimulus signal to the response r(t) is

Substitution for p reveals the following:

The second-order response is modeled as the two-dimensional weighted integral,

A full-field flickering noise stimulus produced strong modulation of the membrane current measured by the patch electrode applied to the cell soma (red solid line, Figure 1). At the termination of the stimulus sequence, there is a clear reduction in the current variance. As described in “Methods,” kernels were calculated that provided the best least squares estimate of the data based on the stimulus (dotted line, left half, Figure 1). As a test, the kernels could then be used to predict the current response over unfitted regions directly from the stimulus (dotted line, right half, Figure 1). The fit provides an accurate prediction of the measured response, and accounts for 79% of the observed variance.

We were able to record from three color opponent cells. All three cells were found in the dorsal part of the retina. First-order kernels were calculated for the red, green, and blue phosphors (Figure 2A). In all three cells, the green and blue phosphors elicited inward currents (excitatory responses) and outward currents (inhibitory responses), respectively. Therefore, they can be classified as M-on/S-off cells. The temporal characteristics of the kernels were very similar in the three cells.

The action spectra of the two wallaby cone types (peak for M cones at 539 nm and for S cones at 420 nm) overlap with the spectra of the three stimulus phosphors, and, therefore, each kernel in Figure 2A represents contributions from both M- and S-cone pathways, albeit at different ratios. Using the known spectral tuning curves for the cone pigments, and the measured spectral output of the phosphors, we calculated optimal (least squares) S- and M-pathway kernels that will reproduce the RGB kernels (see “Methods”). The M-pathway kernel is very similar to the green kernel. This is because the green phosphor stimulates the S cones very weakly. The S-pathway kernel, however, is slightly faster and has lost the initial negative dip as compared with the blue kernel. This is because the blue phosphor stimulates the M cones significantly, and, thus, the M pathway will contribute to the blue kernel. Because the M pathway is an on-response, correction for this effect removes the small inward dip at the onset of the blue kernel. Qualitatively, the basic result can be derived from either set of kernels. The signals arising from M cones excite the ganglion cell, while the S-cone pathway produces an opponent inhibitory input that is significantly slower than that from the M pathway. In all cases, the S-pathway kernels were delayed by 15 ms compared to the M-pathway kernels.

We obviated the rod-photoreceptors as a possible basis for the color opponency observed, because the fits to the RGB-kernels, assuming a short wavelength pigment with a peak at the rod maximum (500nm), were much worse than those obtained using the S-cone pigment sensitivity. This was evident as a seven-fold increase in the sum of the squared residuals for the fits. The absence of a rod input could be explained by rod saturation at the background intensities used.

How do these responses compare to S- and M-cone pathways in spectrally nonopponent ganglion cells? We calculated S- and M-pathway kernels for a number of spectrally nonopponent ganglion cells (solid lines, Figure 3), and compared them to the mean of the S- and M-pathway kernels of the opponent cells (broken lines, Figure 3). Each S- and M-pathway kernel pair has been normalized with respect to its M-pathway member, therefore preserving their amplitude ratios. Nonopponent ganglion cells were either M-on or M-off, and had very little or no input from the S pathway. The time-to-peak varies considerably between the different M-pathway kernels, but these differences can probably be explained by the amount of surround inhibition contributing to the response. The delays before the response onset, however, are all very similar and considerably shorter (15 ms) than that for the opponent S-pathway response. It is important to keep in mind that the transformation to pigment kernels has removed the negative going dip at the beginning of the blue phosphor kernel. However, it has actually shortened and not delayed both the time-to-peak and the onset of the positive response component. It therefore cannot explain the difference in delay between the S- and M-cone pathways.

A significantly better fit to the data is obtained by including second-order components of response. Second-order components are fitted for each input channel (self-second-order), and for each pair of input channels (cross-second-order). For example, for the M-cone input channel, the second-order component is of the form

To aid interpretation of the kernels, a further property is that the second-order contribution to response due to unit pulses of stimulation to M-cone and S-cone channels, at times t₁, t₂ respectively, would be as t increases, that is, a slice running diagonally upward through the kernel H_ms.

Estimations of the second-order kernels were made for the red, green, and blue phosphor stimuli, and these were transformed to correspond to cone channels for medium and short wavelength as described in “Methods.” Figure 4 shows the first- and second-order kernels, averaged over the three S-off cells. The first-order kernels are plotted at top left, along with second-order kernels H_mm, H_ss and H_ms. These are shown as contour plots on t₁, t₂ axes, representing the second-order contribution to response at a given time of stimulation preceding by t₁ on the first channel and by t₂ on the second channel. The second-order kernels account for about 25% of mean squared power of the fitted response.

First, consider H_ss; recall that negative responses represent excitation to an on-stimulus. The first-order kernel, H_s, has a single positive lobe, and hence represents excitation in response to blue-off. The second-order kernel, H_ss, consists of a single negative lobe, which hence augments the excitatory response to blue-off. It thus represents an accelerating nonlinearity in response to blue-off. The medium wavelength self-quadratic kernel, H_mm, also has a major negative peak, which will augment the response to green-on stimulation. It has positive off-diagonal lobes and a second negative on-diagonal lobe with peaks at latencies corresponding with the two phases of the first-order kernel, H_m. It is thus also consistent with an accelerating nonlinearity following a biphasic linear filter. The cross-kernel, H_ms, has a major positive lobe, and a lesser negative lobe. It describes the second-order interaction of input through the short and medium wavelength channels. Because the major lobe is positive, a simultaneous blue-off and green-on stimulation leads to a negative contribution (H_m) x a positive (H_s) x the positive kernel value (H_ms), resulting in another negative, excitatory contribution. Latencies of the two peaks of H_ms correspond to the single peak of H_s and the two peaks of H_m. James (1992) and James and Osorio (1996) provide further examples of this type of decomposition.

In summary, the second-order analysis allows the decomposition of system transfer characteristics into noncommutative processing steps. It suggests there is initial linear filtering of the two input channels, followed by the opponent summation, and then an accelerating nonlinear transformation. This type of model is broadly similar to one recently proposed to account for the firing properties of a wide range of neurons in the cat, rabbit, and salamander retinas (Keat, Reinagel, Reid, & Meister, 2001). The accelerating nonlinearity is of the same nature as the harmonic distortion, or partial rectification, which is seen when recording extracellularly, in response to drifting sinusoidal grating stimulation.

We were able to hold one of the opponent cells for long enough to explore the spatial aspects of its receptive field (Figure 5). The left-most column of Figure 5 shows the results for the color opponent cell, whereas the middle and the right columns show two examples of spectrally nonopponent cells from the same animal. The middle column shows an on-cell and the right column an off-cell. Clearly, both spectrally nonopponent cells are spatially opponent. In Figure 5A, two repeats of the same center-surround stimulus are superimposed to give an indication of the reliability of the recordings for all three cells. For the color opponent cell (right column), the responses to both the center and the surround stimulus show clearly the same M-on/S-off pattern as previously found with the full-field stimulus (Figure 2). This strongly suggests that the cell is not spatially opponent. In contrast, both spectrally nonopponent cells show a reversal of response polarity from the center to the surround, indicating spatial opponency (Figure 5A, middle right). To ensure that for the color opponent cell we had not missed the surround altogether, we examined the kernels derived from a pseudo-random spatio-temporal checkerboard stimulus (Figure 5B, left). Each kernel is plotted in the center of its respective check. The three rings plotted over the checkerboard delineate the boundaries of the center (central ring) and surround stimuli (middle and outer ring) of Figure 5A. The center check of the spatio-temporal stimulus covers approximately the same area as the disk of the center-surround stimulus. Its first order kernel shows the clear biphasic profile we would expect if we added the blue and green first-order kernels together (solid line, Figure 5C, left). Moreover, the sum of the kernels for the surrounding region (medium grey, Figure 5B) shows a very similar time course (dotted line, Figure 5C) indicating that this cell was not spatially opponent. Further out, the responses become very weak and the sum of all first-order kernels becomes very noisy, but the sign of the response does not reverse (dashed line, Figure 5C, left). This indicates that our center-surround stimulus did indeed cover the entire receptive field of the cell. In contrast, the two spectrally nonopponent cells show a clear reversal of the polarity of the response from the central to more peripheral checks (Figure 5B and 5C, middle right). In both cells, the surround response is slightly delayed, which leads to the biphasic response for the medium grey area, where both the center and the surround contribute. In fact, we were able to perform the necessary experiments to determine whether a cell is spatially opponent in a total of 12 cells. In only one of these cells, the color opponent cell shown in Figure 5, we did not find clear evidence for opponency. The remaining 11 cells, all spectrally nonopponent, showed clear signs of spatial opponency.

Previous behavioral studies have indicated that wallabies have good dichromatic color vision (Hemmi, 1999). In this study, we demonstrate the existence of color opponent retinal ganglion cells that could form the neural basis for this perceptual ability. As expected, all the color opponent ganglion cells we recorded from came from the S-cone rich dorsal part of the retina (Hemmi & Grünert, 1999). Similar to color opponent channels in other mammals, these cells receive antagonistic signals

The synaptic mechanisms for color opponency have been most extensively studied in primate systems. Trichromatic primates also possess a much more recent L/M color opponent system, but we will restrict our consideration to the more ancient S/M+L system, the trichromat’s analogue of the S/M system found in most mammals. In mammals, including wallabies, S cones represent about 10% of the total number of cones (DeMonasterio, Schein, & McCrane, 1981; Long & Fisher, 1983; Müller & Peichl, 1989; Curcio et al., 1991; Szél & Röhlich, 1992; Juliusson, Bergstrom, Rohlich, Ehinger, & van Veen, 1994; Hemmi & Grünert, 1999). A specialized bipolar cell that makes exclusive contacts with short wavelength-sensitive S cones has been observed in primate retina (Kouyama & Marshak, 1992), and although direct evidence is scant, a similar bipolar cell type may exist in other mammals (e.g., Linberg, Suemune, & Fisher, 1996). In primate, this on-type S-cone bipolar cell makes connections with the inner dendritic tier of a small field bistratified ganglion cell (Dacey & Lee, 1994; Ghosh, Martin, & Grünert, 1997). The outer dendritic tier of this ganglion cell receives input from off-center bipolar cells with mixed L/M-cone inputs. It has been suggested that S/M+L color opponency in primates is generated by antagonistic bipolar cells (Dacey & Lee, 1994), although more recent evidence indicates that the bipolar cells may already be color opponent (Dacey, 2000). At the very least, color vision in wallabies requires a similar population of S-cone selective bipolar cells, although the sign of the responses of these cells need not be preserved.

How are we to account for the response polarity of the M-on/S-off ganglion cells in the wallaby? There are two obvious alternatives: (1) these retinas express an S-cone selective off-bipolar cell or (2) they have an S-cone selective on-bipolar cell similar to primates and squirrels, but the S signal is inverted after passing through an intervening inhibitory amacrine cell. We cannot yet rule out either alternative. It is interesting to note that the S-off input was delayed by about 15 ms with respect to the M-on signal, whereas the time course of the latter was similar to that seen in spectrally nonopponent cells. Advancing the S-off response so that it superimposed the M-on response showed that the time-courses were very similar and that the 15-ms delay stems from a pure delay in the response onset of the S-off channel. This result is reminiscent of results from monkey, with the difference that in the monkey the onset delays were similar but the time to peak of the on signals was 10 ms to 20 ms faster than for the off signals (Chichilnisky & Baylor, 1999). One would predict that, due to the mismatch in the time courses of the color opponent signals, “white” light stimuli, which drive the S and M pathways to a similar degree, will transiently excite the cells.

Rabbit, squirrel, and monkey are the only animals where M-on/S-off color opponent cells have been reported in the retina. This result documents another mammal where such cells are found. It is tempting to suppose that both classes of color opponent neuron exist in all mammals, M-on/S-off and S-on/M-off, and that this dichotomy is generated by two S-cone selective cells, an on-bipolar cell, and an on-amacrine cell. The latter could be generated simply by making selective contacts with the former. The M channel need not be cone selective due to the relatively small number of S cones. This simple model is attractive in requiring only a single class of S-cone selective bipolar cell. Further, it predicts that the S signal in both M-on/S-off cells and S-on/M-off cells should be blocked by application of APB (2-amino-4-phosphonobutyric acid), a pharmacological agent, which selectively blocks on-bipolar cell responses.

The authors thank Drs. Paul Martin and Ted Maddess for helpful comments on the manuscript, and K. Williams and M. Maier for taking care of the animals. We thank an anonymous referee for suggesting the use of the inverse phosphor-cone matrix, which we have extended to the transformation of second-order kernels. Commercial Relationships: None.