Coincidence detection of single-photon responses in the inner retina at the sensitivity limit of vision

Abstract

Background: Vision in starlight relies on our ability to detect single absorbed photons. Indeed, the sensitivity of dark-adapted vision approaches limits set by the quantal nature of light. This sensitivity requires neural mechanisms that selectively transmit quantal responses and suppress noise. Such mechanisms face an inevitable tradeoff because signal and noise cannot be perfectly separated, and rejecting noise also means rejecting signal.

Results: We report measurements of single-photon responses in the output signals of the primate retina. We find that visual signals arising from a few absorbed photons are read out fundamentally differently by primate On and Off parasol ganglion cells, key retinal output neurons. Off parasol cells respond linearly to near-threshold flashes, retaining sensitivity to each absorbed photon but maintaining a high level of noise. On parasol cells respond nonlinearly due to thresholding of their excitatory synaptic inputs. This nonlinearity reduces neural noise but also limits information about single-photon absorptions.

Conclusions: The long-standing idea that information about each photon absorption is available for behavior at the sensitivity limit of vision is not universally true across retinal outputs. More generally, our work shows how a neural circuit balances the competing needs for sensitivity and noise rejection.

Fig. 1

On and Off parasol ganglion cells encode weak flashes differently. A, On parasol ganglion cell spike responses to dim flashes delivered at the time of the arrow. Each box shows 25 trials with flashes of constant nominal intensity as indicated in the lower left corner. B, Off parasol ganglion cell (neighboring cell in the same preparation as the cell in A) spike responses to the same flash strengths as in A. C, On parasol cell mean firing rates (PSTH, 10-ms time bin) for the same cell and flash strengths as in A. D, Off parasol cell mean firing rates for the same cell and flash strengths as in B. E, Two-alternative forced-choice task for the same On (black) and Off (grey) parasol cells as shown in panels A–D. Inset shows the average of all correct choices (continuous line) and incorrect choices (dashed line) for On (black, top panels) and Off (grey, bottom panels) parasols at the two lowest light intensities indicated by roman numerals (I & II) both in the main panel and inset.

Fig. 2

On but not Off parasol ganglion cells integrate single-photon responses nonlinearly. A, Response amplitude (increase in mean firing rate or mean excitatory current) as a function of flash strength for On parasol cell (black symbols) and for Off parasol cell (decrease in mean firing rate, grey symbols). The cells are the same example cells as in Fig. 1. The dashed line shows the expectation for linear integration. Inset: Collected data on the slope of the stimulus-response relation (mean ± SEM) in the nonlinear region (On parasols, black bar graphs) and in the same domain in Off parasols (grey bar graph). Slopes were determined over the range of flash intensities where flash sensitivity (response per photoisomerization, as shown in D) grew from 5 to 95% of maximum for On parasols. Symbols connected by lines are from the same cells. B, On parasol excitatory synaptic inputs to the same cell and flash strengths as in Fig. 1A. C, Mean excitatory currents (115 trials at each flash strength) for the same cell and flash strengths. D, Sensitivity of input currents - i.e. response divided by flash strength. E, A simple thresholding model can explain the shape of the stimulus-response relation. Data points show integrated excitatory synaptic inputs (mean ± SEM) for 7 On parasol cells that were probed using the weakest flashes producing < 1 R*/RGC. The nonlinear region of the stimulus-response relation occurs where changes in flash strength shift structure in the response distribution from below threshold to above threshold (e.g. the red data point). Peaks in the modeled response distributions are due to discrete photon counts arising from Poisson distribution of the absorbed photons. Spread in the peak the modeled distributions comes from additive Gaussian noise.

Fig. 3

Dim backgrounds eliminate the nonlinearity in On parasol responses. A, Stimulus-responses relations for spike responses and excitatory synaptic inputs in darkness (black lines and symbols) and in the presence of a dim background (green lines and symbols) (different cells for spike and input current measurements). A simple threshold model (inset) predicts relief of the nonlinearity by a background light that shifts the response distribution above the threshold. B, Summary of effect of background on the slope of the stimulus-response relation. C, The dependence of noise variance on background light in an example cell. The upper part shows example traces of current noise (left) and spiking noise (right) in the same cell at the same light levels (background intensities are indicated in units of R*/rod/sec). The current noise traces were matched-filtered by the cell’s impulse response prior variance estimation to focus on the most critical noise component (see the lower part of the plot). The dashed line plots the expectation if the variance increased linearly with increasing Poisson fluctuations due to the background. The x-axis intercept of this line provides an estimate of the dark noise (equivalent to 0.0027 R*/rod/sec). The steep decrease in the variance measured at the lowest backgrounds reflects suppression of noise due to the nonlinear threshold. D, Collected data on the ratio of the linearly predicted variance to the measured variance in darkness (grey symbols and black bars), and the variance measured under one dim background intensity, 0.008 R*/rod/sec, vs. that in darkness (green symbols and bars). In a subpopulation of the cells, we measured the variance only in darkness and at one background (0.008 R*/rod/sec). In such cases, the extrapolation was based on the variance measured at 0.008 R*/rod/sec and the x-axis intercept was fixed to −0.003 R*/rod/sec. Bar graphs show mean ± SEM: Var_DIM/Var_DARK = 62 ± 28 (spikes); 17 ± 5 (excitatory inputs); ${\mathit{Var}}_{\mathit{DARK}}^{\mathit{LIN}}/{\mathit{Var}}_{\mathit{DARK}}=19\pm 5$, (spikes); 5.3 ± 1.4 (excitatory inputs). E, Signal-to-noise ratio (mean response per photoisomerization/standard deviation of noise) as a function of flash strength. The measured signal-to-noise ratio has been scaled by that expected for a linear system, assuming the maximum flash sensitivity (cf. Fig. 2D) and noise as determined in panel C (dashed line) by extrapolation of the measured variance to darkness from backgrounds that relieve the nonlinearity. Open symbols show SNR (mean ± SEM) for excitatory inputs (n = 10 cells), and spikes output (n = 24) for On parasol cells where signal and noise were measured in the same cells. Closed symbols show SNR estimates for On parasol cells where noise (n = 17 cells, excitatory inputs; n = 29, spike output) and signals (n = 26, excitatory inputs; n = 58, spike output) were measured from different cells.

Fig. 4

The thresholding nonlinearity is located at the synapse of the On cone bipolar onto the On parasol cell. Panels A–D show that the nonlinearity is not part of the Off-parasol circuit, which shares with the On -parasol circuit the stages up to the AII amacrine cell, as indicated by blue shading in the diagram in panel A, Rod convergence at different levels of the circuit is shown in parentheses. B, Excitatory synaptic inputs to an Off parasol cell elicited by the same flashes intensities as used to probe On and Off parasol cells as in Fig. 1 and Fig. 2: (mean = black, SEM = grey, n = 21 traces). The traces plot the decrease elicited by a flash in tonic excitation in darkness (such that the strongest two flashes complete abolish tonic excitation similarly as they completely abolish spiking activity in the Off parasol shown in Fig. 1B and D). C, Stimulus-response relation for Off parasol inputs. D, Collected data comparing slopes of stimulus-response relations for excitatory inputs to On and Off parasol cells. Panels E–H show that rod- and cone-mediated signals share a common thresholding nonlinearity. E, Retinal circuits conveying rod and cone signals to an On parasol ganglion cell in the dark. The only shared element is the cone bipolar synaptic terminal onto On-parasol. F, Separation of rod and cone responses. Long-wavelength flashes elicited responses with a fast cone component and a slow rod component. Subtracting from this a pure rod response matching the rod component (elicited by an appropriately chosen short-wavelength flash) allowed isolation of the cone response (width of each trace = ± SEM). G, Stimulus-response relation for isolated cone responses (inset) in darkness (black) and in the presence of a dim background (green). H, Collected data on the slope of the cone stimulus-response relation in the dark and in the presence of the dim background.

Fig. 5

Predicted impact of the observed nonlinearity on sensitivity of the retinal output signals. A, Schematic of the model. Rod signals arising from single-photon absorptions and spontaneous photon-like noise events in a subset of rods (red) were pooled linearly. Additive synaptic noise (N) was added, and the pooled signals and noise were passed through a nonlinear threshold to represent the inner retinal subunit responses. Multiple subunits were pooled to generate modeled ganglion cell responses. B, Stimulus-response relations for models (see Experimental Procedures) with different thresholds ( ) and subunit sizes (n) compared to data from an example cell. C, Stimulus-response relations for different models in the presence of a background producing 0.008 R*/rod/sec. Symbols show the measured stimulus-response relation under background from the same cell as B. D, Flash/no flash discrimination performance for an observer of the modeled ganglion cell outputs, = 2. The modeled ganglion cell received (indirect) input from 4096 rods. E, False positive rate across flash strengths for an observer of the modeled ganglion cell outputs.

Fig. 6

Measurement of the time window for nonlinear interactions in On parasol cells. A, Excitatory synaptic currents elicited by a single flash (thick trace) or pairs of flashes delivered at different time offsets. B, Responses to the second flash of a pair with the response to the first flash subtracted. C, Interaction index, defined as the area of subtracted response in B divided by the area of the response to a single flash from A. Each point is the mean ± SEM across 5 cells. The smooth line is an exponential with a 50 ms time constant.