The nervous system faces a shifting problem. It has to shift its mode of operation from one state to another as it faces new demands (i.e., it has to shift its attention, its contrast sensitivity, its temporal integration time, etc.). How it achieves this isn't clear. Here we examined a case where it was possible to obtain an answer, and the answer was intriguingly simple: the system produced the shift by changing the gap junction coupling of one of its cell classes. The coupling acted as a way to inactivate the cell class, and, by doing so, change the system's behavior.
The findings are both surprising and exciting: surprising, because a seemingly complicated problem was solved with a simple mechanism, and exciting, because the mechanism is present not just in the retina, but throughout the brain, suggesting it might generalize to other network shifts. To be specific, gap junction coupled networks are present in visual cortex, motor cortex, frontal cortex, hippocampus, cerebellum, hypothalamus, and striatum, among many other places (Galarreta and Hestrin, 1999
; Bennett and Zukin, 2004
Furthermore, a regulator is also in place. In the retina, the regulator is a neuromodulator, dopamine: Light triggers the release of dopamine, which closes gap junctions via second messengers (McMahon et al., 1989
; Dong and McReynolds, 1991
; Weiler et al., 2000
). Dopamine, as well as noradrenaline and histamine, have been found to open and close gap junctions in several of these brain areas (Cepeda et al., 1989
; Yang and Hatton, 2002
; Onn et al., 2008
; Zsiros and Maccaferri, 2008
The possibility for generalization to other networks is substantial and straightforward to see:
(1) While the results in this paper show the mechanism in non-spiking neurons, it readily applies to spiking cells as well and thus to networks in the brain. This is because the mechanism involves only basic biophysics – a change in cells’ input resistance. Briefly, if a cell class is coupled by gap junctions, it has the potential to have its input resistance turned up and down. When the junctions are closed, the input resistance of the cells is high. This makes the cells more responsive to incoming signals and allows them to send strong signals out. When the junctions are opened, the input resistance drops. This makes the cells less responsive to incoming signals and allows them to send out only weak signals. In the case of spiking neurons, the signals can become so weak that the probability of firing can be reduced essentially to 0; i.e., the cells can be effectively turned off.
(2) The mechanism has the potential to affect many types of network operations. While the one presented in this paper was a negative feedback loop – the gap junction coupling provided a way to turn the feedback on or off (or up or down) – one can readily imagine many other operations that could be altered by turning the activity of a pivotal cell class in a network on or off, such as alterations in feedforward signaling, lateral signaling, recurrent signaling (e.g., the stabilization of attractors), to name a few.
(3) The timescale over which the mechanism operates, that is, the timescale over which the change in coupling occurs – a scale of seconds (McMahon et al., 1989
; McMahon and Mattson, 1996
) – is consistent with many state changes, such as changes in arousal, changes in attentional set, shifts in decision-making strategies, e.g., shifts in the weighting of priors, shifts to speed versus accuracy (Standage and Paré, 2009
), allowing it to mediate many behavioral processes.
(4) Since the cellular machinery for regulation of gap junction conductances is in place, the mechanism can evolve via a change in a single gene, a gene for a gap junction protein. This makes it an easy gain from an evolutionary standpoint. A powerful selective advantage – the ability to shift a network from one state to another – could be rapidly acquired, and, in addition, acquired independently in multiple networks. (For a review of gap junction proteins, see Bennett and Zukin, 2004
Figure emphasizes this latter point, that this gap junction coupling mechanism offers a single gene solution to a seemingly complicated set of problems, network state changes. To address this, we used, again, the horizontal cells, as an example. Specifically, we took the behavioral results from the wild-type and Cx57 knockout animals and imposed them on a predator-detection scenario. We filmed an approaching predator, restricting the movies to the temporal frequencies available to each genotype, as indicated in Figure D left. The results are shown in Figure . In day conditions the movies for the two genotypes are essentially the same; the predator can be seen when it is moving, i.e., when the movie is dominated by high temporal frequencies, and when it is still, i.e., when the movie is dominated by low temporal frequencies. In contrast, in night conditions, the movies diverge. In the movie filtered through the frequencies visible to the wild-type animal, the predator remains visible even when it is still; this is consistent with the wild-type's ability to shift to low temporal frequencies. In the movie filtered though the frequencies visible to the knockout, the predator disappears. Only a ghost is present (see Supplementary Material
for the complete movies). The wild-type's maintenance of visual contact with the predator gives it an obvious selective advantage.
Figure 4 The selective disadvantage of a Cx57 gene loss. (A) Movie of an approaching predator, filtered through the frequencies available to the wild-type animal, as provided by Figure D left. In day conditions, the predator can be seen both when (more ...)
Estimating the extent to which input resistance can be reduced by coupling
As discussed above, changes in coupling can act as a dial to turn the input resistance of a cell up or down. We can estimate the range of the dial as follows: The standard experimental measure of coupling is the length constant (Xin and Bloomfield, 1999
; Shelley et al., 2006
). Xin and Bloomfield measured the length constant of horizontal cells under several scotopic and photopic light levels and found the maximal difference to be a factor of ~3. The maximal difference occurred when the scotopic light level was 1–1.5 log units above rod threshold and the photopic light level was >3 log units above rod threshold, levels that we matched for this paper. Since, for 2-D coupling (Lamb, 1976
), input resistance is inversely proportional to the square of the length constant (detailed in Materials and Methods
and Appendix 2
), the input resistance of the horizontal cells at the scotopic light level is estimated to be about a factor of 9 less than that at the photopic light level.
In the general case, as with horizontal cells, the extent to which gap junction coupling can shunt a cell is the ratio of the total conductances of the gap junctions that can be modulated, to the cell's baseline (“leak”) conductances. Many factors – including the cell's geometry and the complement and distribution of channels and gap junctions – combine to determine this ratio. The example of horizontal cells shows that this can be as much as an order of magnitude.
Linking a behavior to a neural mechanism
Following a behavioral change down to the mechanism that underlies it is often not possible experimentally. It was possible here because of a confluence of factors: the relevant network could be identified and its component cell classes are known (as shown in Figures and ), and the protein around which the mechanism revolves, the particular gap junction protein, Cx57, is present only in one cell class (the horizontal cells) and not elsewhere in the brain (Hombach et al., 2004
), allowing the circuit to be selectively disrupted. The significance of the latter is that it allowed a direct connection to be made between the disruption in the circuit and the disruption in the behavior, since no other circuits were perturbed.
Potential alternative models for the shift toward low temporal frequencies
As an animal moves from a light-adapted to a dark-adapted state, several changes occur in the retina other than the change in horizontal cell coupling via the Cx57 gap junctions. How can we be sure that our result – the shift toward low temporal frequencies – is not produced by these other changes? Here we systematically go through them.
The most well known change is the shift from cone to rod photoreceptors. This can't account for our results, because the knockout undergoes the same cone-to-rod shift, and it doesn't undergo the shift to low frequencies (Figure ). In addition, it's well known that the cone-to-rod shift affects high frequencies, not low. We show this in Appendix 1
, Figure , specifically for our species, the mouse. As shown in the figure, the frequency response curves for the rod and cone are both flat below 0.5
Hz, meaning there is no frequency-dependent change in this region. In contrast, our results show a selective boost at frequencies below 0.5
Hz; that is, the system shifts to favor low frequencies. The shift from cones to rods can't account for this.
Figure 5 The frequency response difference between the rods and cones lies in the high frequencies, not the low. (A) Impulse responses of the two photoreceptors, reproduced from Nikonov et al. (2006) for cone and Luo and Yau (2005) for rod. (B) Frequency responses (more ...)
Another change that occurs during dark adaptation is rod–cone coupling (see Ribelayga et al., 2008
, for rod–cone coupling as a result of circadian rhythms; also Yang and Wu, 1989b
; Wang and Mangel, 1996
; Trumpler et al., 2008
). Rod–cone coupling, though, is mediated by gap junctions formed by Cx36, Cx35, and Cx34.7 (reviewed in Li et al. (2009
)), not Cx57 (Janssen-Bienhold et al., 2009
). Cx57 is not present in rods and cones (Hombach et al., 2004
; Janssen-Bienhold et al., 2009
) and thus the knockout is not perturbing these couplings.
Similarly, gap junction coupling in the inner retina likely plays a role in dark adaptation, since the AII amacrine cells of the rod pathway are coupled by gap junctions (Bloomfield et al., 1997
). However, Cx57 is not a gap junction in these cells (Janssen-Bienhold et al., 2009
), so changes in inner retinal coupling can not account for our results.
Recent reports have indicated that some gap junctions act as hemichannels (Kamermans et al., 2001
; Shields et al., 2007
). If Cx57 acted in this fashion, it could provide for ephaptic transmission of a feedback signal. However, the possibility that Cx57 is a hemichannel has been examined at the ultrastructural level, and ruled out (Janssen-Bienhold et al., 2009
). Furthermore, feedback to photoreceptors has been shown to be intact in the Cx57 knockout by two groups (Shelley et al., 2006
; Dedek et al., 2008
Finally, a standard concern with most or all knockout experiments is that knocking out a gene could lead to secondary developmental effects. While we can't completely rule this out, there is no evidence for altered development in the Cx57 knockout: retinal anatomy appears unperturbed (Hombach et al., 2004
; Shelley et al., 2006
), temporal tuning by day, as measured at the ganglion cell and behavioral level, remains intact, i.e., is the same as in wild-type (Figure D), and spatial processing, also measured at the ganglion cell and behavioral level, remains intact as well (Dedek et al., 2008
). While compensatory effects are possible, the likelihood that they would lead to such close matches along all these axes is very low.
Thus, while cone-to-rod shifts, photoreceptor coupling, and other factors contribute to dark adaptation, they can't account for the results presented here, and the probability that the results could be accounted for by developmental effects, as mentioned above, is very low.
One issue that we can't completely rule out, though, is the following: even though horizontal cell feedback to photoreceptors is known to be present and can account for our results, we can't completely rule out the possibility that the shunting of horizontal cell current causes the shift in tuning through some other action. For example, if horizontal cells were to act as a mediator between multiple circuits with different kinetics (e.g., different photoreceptor readout circuits), then the shunting of the horizontal cell current could shift tuning by causing a switch from one circuit to another. But note that any alternative model must be consistent with the known constraints: (a) the difference between wild-type and knockout is present under scotopic conditions (Figure ), where all responses are rod-driven, (b) the tuning shift involves low frequencies, (c) the mouse retina has only one kind of horizontal cell, and it serves both kinds of photoreceptors, and (d) connexin-57 is only involved in horizontal cell-to-horizontal cell coupling. We chose the horizontal cell feedback model shown in Figure because it is a parsimonious model that satisfies these constraints and is consistent with current known actions of horizontal cells.
We conclude by mentioning that in one species (the rabbit), when light levels are much lower, more than an order of magnitude below the scotopic level used in this study, gap junctions close (Xin and Bloomfield, 1999
) with no corresponding reversal of the shift in integration times (Nakatani et al., 1991
). This suggests that in this extreme range, other mechanisms must take over, mechanisms likely intrinsic to the photoreceptors, as described in Tamura et al. (1989
Relation of Cx57 to spatial processing in the dark- and light-adapted conditions
Horizontal cells provide negative feedback to photoreceptors (Werblin and Dowling, 1969
) and antagonistic feedforward to bipolar cells (Yang and Wu, 1991
), and it has long been thought that they contribute to the receptive field surround. One might expect, therefore, that eliminating coupling in these cells would alter spatial processing as well as temporal processing as the retina shifts from day to night vision. A previous study, though, shows that spatial tuning remains normal in the Cx57 knockout (Dedek et al., 2008
). The likely basis for this is the fact that the surround is generated by circuits in more than one layer – specifically, by amacrine cell circuits in the inner retina, as well as by horizontal cells in the outer retina (Cook and McReynolds, 1998
; Taylor, 1999
; Roska et al., 2000
; Flores-Herr et al., 2001
; McMahon et al., 2004
; Sinclair et al., 2004
). As mentioned in Dedek et al. (2008
), the lack of a change in spatial tuning in the knockout implies that inner retinal mechanisms dominate for the problem of adjusting spatial tuning to different light-adaptation levels.
Coupling as a mechanism to produce synchrony
We conclude by mentioning that gap junction coupling has also been proposed as a mechanism to create synchronous firing among neurons, e.g., for creating oscillations (for review, see Bennett and Zukin, 2004
). The idea presented in this paper – that changes in coupling serve as a way to inactivate a cell class or reduce its impact – is not mutually exclusive with this proposal. This is because the effect of coupling depends on the state of the cell. As mentioned above, when a cell becomes coupled to other cells, its input resistance drops. For spiking neurons, this means the probability of reaching threshold and firing is reduced. If, however, the cell receives strong enough input to allow it to cross threshold, its firing can produce synchronous spikes in coupled cells. Thus, gap junction coupling can potentially mediate more than one network operation.