Photoreceptors Across a Billion-Fold Range, and When They Fail

A mouse cone that has lost both of its calcium sensors, GCAPs and recoverin, still adapts to background light. If the textbook feedback loop is gone and the cell adapts anyway, what is doing the work?

The retina lesson gave you phototransduction and the calcium loop, and senses and cGMP gave you the cascade and its shutoff. Sensory coding introduced adaptation as a property of any sense. This lesson starts at gain control: how a photoreceptor resets its sensitivity across backgrounds, where the standard account is incomplete, and what happens when photoreceptors die.

Calcium as the gain knob, and what is left when you remove it

The loop runs like this. Light closes cyclic-nucleotide-gated (CNG) channels in the outer segment. Calcium entry through them stops, but a sodium-calcium-potassium exchanger (NCKX) keeps pumping calcium out, so outer-segment calcium falls. At low calcium, two calcium-sensing proteins change state. GCAP1 and GCAP2, the guanylyl cyclase activating proteins, speed cGMP synthesis, which reopens channels. Recoverin lets go of GRK1, the kinase that switches off activated visual pigment, so the pigment is shut off faster. Both effects push the cell back toward its dark state while the light is still on, and that push is gain control.

The test of a loop is to cut it. Vinberg and colleagues (2015) did this in mouse rods, recording ex vivo from isolated retina at 37 degrees Celsius with a transretinal electroretinogram (ERG), in which downstream synaptic responses are blocked so the voltage across the retina reports the photoreceptors. They compared wild-type rods with rods lacking the GCAPs, rods lacking recoverin, and rods lacking all three proteins. Calcium-dependent light adaptation persisted in the rods lacking both sensors.

Vinberg and Kefalov (2018) repeated the logic in cones. The mice also lacked rod transducin (a Gnat1 knockout), so rod phototransduction was gone and the recorded signal came from cones. All mice were 2 to 3 months old. They compared control cones with cones lacking GCAP1, GCAP2 and recoverin, and they repeated the measurement with extracellular calcium lowered from 1.2 mM to an estimated 30 nM. The knockout cones still adapted to background light. The authors interpreted this as a sign of a calcium-independent component, and also reported a further calcium-dependent mechanism in the same cones.

Hold that conclusion at its real strength. The low-calcium bath is a change in the solution around the retina, not a direct clamp of outer-segment calcium, and the inference rests partly on fitting a model. No molecule has been shown to be the residual mechanism. What would settle it is an experiment that holds intracellular calcium fixed while the background series is run in cells lacking each candidate in turn. Until then the honest statement is that GCAPs and recoverin contribute and do not account for all of the adaptation.

Reading an adaptation curve: half-sensitivity background and slope

Sensitivity here means response amplitude divided by the strength of a dim flash. As the background brightens, sensitivity falls. The cone study fitted this with a modified Weber-Fechner relation, where S is the flash sensitivity on a background of intensity I, and Sdark is the sensitivity in darkness:

S(I) / Sdark = 1 / (1 + (I / I0)^n)

I0 is the half-sensitivity background: when I equals I0, the bracket is 1 + 1 and sensitivity is 0.5 of its dark value. The exponent n sets the slope. On a log-log plot, at backgrounds well above I0, the curve falls with slope minus n. With n = 1, sensitivity is proportional to 1/I, so response size on a background stays about constant for a fixed contrast, which is Weber behaviour. With n above 1, sensitivity falls faster than Weber’s rule says.

Worked problem: two kinds of cone at one background. The fit values come from Vinberg and Kefalov (2018), isolated mouse retinas, 2 to 3 months old, ex vivo, 505-nm light. Control cones had I0 = 48,800 photons per square micrometre per second and n = 1.0. Cones lacking GCAPs and recoverin had I0 = 12,100 photons per square micrometre per second and n = 1.5. Take a background of I = 48,800 photons per square micrometre per second.

1. Control: I/I0 = 1, so S/Sdark = 1/(1 + 1) = 0.500.
2. Deficient: I/I0 = 48,800/12,100 = 4.03. Raise it to the power 1.5: 4.03^1.5 = 8.10. So S/Sdark = 1/(1 + 8.10) = 0.110.
3. For a response of fixed size, each preparation measured against its own dark sensitivity, the deficient cones need a flash 0.500/0.110 = 4.55 times stronger.

Check the units: both the ratio and the sensitivities are dimensionless. The limits matter. These are central fit values with no uncertainty carried through (the paper gives them as means with standard errors). Each sensitivity is normalized to its own genotype’s dark sensitivity, so 4.55 is not a ratio of absolute thresholds between genotypes. And it is a mouse ex vivo measurement. The same study reported dim-flash time to peak of 56 ms in control cones and 142 ms in the deficient cones (means, standard errors in the paper), so the deficient cones were also slower, consistent with weaker feedback.

Background I: 48,800 photons µm⁻² s⁻¹ (log scale; the 10³ to 10⁶ range is this sketch’s choice) Separate the two parameters (counterfactual: deficient cones with n = 1.0, keeping their I₀)

The slider snaps to 12,100 and 48,800, the two I₀ values.

Control cones, S/S_dark

0.500

Cones lacking GCAPs and recoverin (fit), S/S_dark

0.110

Ratio, control ÷ deficient (flash strength needed for a fixed-size response, each against its own dark sensitivity)

4.55

Slope on the log-log plot well above I₀ (control; deficient)

−1.0; −1.5

In this sketch: the curves are computed from the lesson’s fit, S/S_dark = 1/(1 + (I/I₀)ⁿ), using central fit values with no uncertainty carried through (control I₀ = 48,800 photons µm⁻² s⁻¹, n = 1.0; cones lacking GCAPs and recoverin I₀ = 12,100, n = 1.5). Isolated mouse retina 2 to 3 months old, 505-nm light, Gnat1 knockout background (Vinberg and Kefalov, 2018). The lesson does not give the range of backgrounds that was tested, so the whole curve is labelled fit extended.

Rule to keep: I₀ is the background that halves sensitivity and n sets the slope; each sensitivity is relative to its own dark value, so the ratio is not a ratio of absolute thresholds between genotypes, and the fit does not name the residual mechanism (the low-calcium bath is not a clamp of outer-segment calcium).

Rods in daylight: a saturation rule that needed revising

A standard story says rods saturate once background light is bright enough, and cones then carry vision. Earlier in this course, retinal cell types and intrinsically photosensitive ganglion cells described other routes for bright-light information; here the question is whether the rod route really closes.

Tikidji-Hamburyan, Reinhard, Storchi and colleagues (2017) tested it in mice. They recorded ganglion cells with a multi-electrode array in isolated retinas of cone-deficient mice (Cnga3 knockouts, 4.5 to 6 weeks old), measured ERGs (8 weeks), and recorded in the dorsal lateral geniculate nucleus of anaesthetized mice (about 6 to 8 weeks), with intact controls as well. Rod-driven responses to moderate-contrast stimuli were present at backgrounds up to 10^7 dark-adapted-equivalent R* per rod per second in isolated retina, and up to 10^6.35 in the thalamus in vivo. Contrast sensitivity fell when a bright background came on and then recovered with continued exposure, and recovery was faster at higher backgrounds. A phototransduction model suggested that changes in gain and bleaching adaptation, the loss of sensitivity as pigment is bleached until it is regenerated through the visual cycle, contribute at high backgrounds.

So the rule needs revising: in these mice, rod saturation depends on how long the retina has been exposed, not on irradiance alone. The limits are as important as the result. This is mouse, measured ex vivo and in thalamus. It sets no daylight cutoff for human rods, and photoisomerization rates cannot be turned into lux without assumptions this study does not supply. Whether human rods contribute to conscious vision in bright light would need a human test with rod-isolating conditions.

One photon: what 1942 showed and what 2016 added

Hecht, Shlaer and Pirenne (Journal of General Physiology 25:819-840, 1942) measured the least blue-green light that dark-adapted human observers could detect. Threshold flashes held 54 to 148 photons at the cornea. After correcting for losses in the eye, they estimated an upper limit of 5 to 14 photons absorbed by rods, and their frequency-of-seeing analysis gave 5 to 8 critical events at threshold. The arithmetic gives a feel for the losses: 5/54 and 14/148 are both near 0.09 to 0.10, so roughly one corneal photon in ten ended up absorbed by a rod, on their estimates. Hecht did not show that a person detects one delivered photon. The inference was statistical, about a handful of absorptions.

Tinsley, Molodtsov, Prevedel and colleagues (Nature Communications 7:12172, 2016) delivered photons one at a time. Alipasha Vaziri conceived and supervised the study. Three trained men in their twenties, dark-adapted, chose which of two time intervals contained a heralded light event. Among 30,767 trials, 2,420 were post-selected as single-photon events. The pooled probability of a correct choice was 0.516 plus or minus 0.010, against 0.5 for chance, with P = 0.0545. A high-confidence subset gave 0.60 plus or minus 0.03. Combining the single-photon and attenuated-light data with Fisher’s method gave P = 0.014. A rough check: if the pooled figure rests on those 2,420 events, the binomial standard error is the square root of 0.25/2,420, about 0.010, and 0.516 is about 1.6 standard errors above chance. The authors also interpreted a dependence of detection on recently seen photons as a possible priming process lasting seconds, and that is an interpretation.

So the evidence for single-photon perception is real but analysis-dependent: it comes from three men, it depends on how trials were selected, and it does not show reliable awareness of every photon that reaches the cornea. It also separates steps that are easy to merge: a photon delivered, a rod absorbing it, the observer’s confidence, and the forced choice.

When photoreceptors die: rod-cone dependence, and two trials

In typical retinitis pigmentosa (RP), a heterogeneous group of progressive dystrophies, rod degeneration precedes cone loss. Leber congenital amaurosis (LCA) typically causes profound visual impairment beginning in the first months of life, and it covers defects in phototransduction, the retinoid cycle (the pigment epithelium’s recycling of retinoids, in which RPE65 converts all-trans retinyl ester to 11-cis retinol) and ciliary transport.

Why do cones die when the faulty gene is a rod gene? Léveillard and colleagues (Nature Genetics 36:755-759, 2004) screened a mouse retinal cDNA library on cone-enriched cultures from embryonic chicken retina and found RdCVF, described as a truncated, thioredoxin-like protein. Aït-Ali and colleagues (Cell 161:817-832, 2015), using embryonic chicken cone cultures and the rd1 mouse, found that RdCVF binds basigin-1 on the photoreceptor surface, that basigin-1 associates with GLUT1, and that the pathway increases cone glucose entry and aerobic glycolysis. A missense mutant failed to bind basigin-1, stimulate glucose uptake or prevent secondary cone death in the model. RdCVF is enzymatically inactive and is not an antioxidant enzyme, and this is experimental work, not a human trial. Compare the metabolic coupling in glia and metabolic partners and the wider picture in neurodegeneration and cell death.

Two trials show what different designs can say. Russell and colleagues (Lancet, 2017) ran a randomized, controlled, open-label phase 3 trial of bilateral subretinal AAV2-RPE65 gene augmentation in people with biallelic RPE65 disease (age 3 years and over) who had enough viable retina; outcome grading was masked. The outcome was a multi-luminance mobility test (MLMT), a navigation course scored by light level. At one year the between-group difference in mean bilateral change was 1.6 light levels (95% CI 0.72 to 2.41), and 13 of 20 treated participants passed at 1 lux. The US FDA approved the product on 19 December 2017 according to its records. Pierce and colleagues (New England Journal of Medicine 390:1972-1984, 2024) delivered CRISPR editing machinery for CEP290 IVS26 to the worse eye of 14 participants (see CRISPR and functional genomics for the tool, and sequencing and genomes for finding such variants). Retinal implants and optogenetic vision restoration are for sensory prostheses.

Design problem

In the CEP290 editing study, 6 of 14 participants improved on cone-mediated sensitivity by full-field stimulus testing and 11 of 14 improved on at least one of four outcomes, in an open-label study with no sham control. The RPE65 augmentation trial was randomized and controlled, open-label, with masked outcome grading and a one-year between-group difference of 1.6 light levels (95% CI 0.72 to 2.41), with 13 of 20 treated participants passing at 1 lux. Interpret these data: which claims does each design support?

One way to do it

Start with what a single-arm study cannot control. Participants may improve with practice on the task, outcomes fluctuate naturally, and people who know they were treated may try harder or report more. Without a comparison group those effects cannot be separated from treatment. The untreated fellow eye is a useful within-participant comparison, because it shares the person and the testing days, but it is not a sham: it does not have the same procedure, and the two eyes may differ in disease stage. So the CEP290 numbers support a statement of safety (no dose-limiting toxic effects were recorded) and a preliminary signal of possible benefit, no more.

Next, separate the endpoints. Mobility, retinal sensitivity, acuity and patient-reported outcomes are different measurements. That 11 of 14 improved on at least one of four outcomes is a weaker statement than that 11 improved on all four, and a count of “at least one” rises with the number of outcomes. The 6 of 14 on cone-mediated sensitivity is a specific, more informative figure.

Randomization adds a control group that goes through the same period of practice and repeated testing, and masked grading limits grader bias. The trial was open-label, so participants knew their group and their own expectations were not equalized. That supports the RPE65 claim that treated participants did better on mobility than controls at one year, by 1.6 light levels. It does not support more: the finding is limited to participants with biallelic RPE65 disease and viable retina, to the MLMT as an endpoint, and to the follow-up of the trial. Neither study gives a prognosis, a claim of general benefit, or any advice about treatment for an individual.