Saccadic modulation of stimulus processing in primary visual cortex

Abstract

Saccadic eye movements play a central role in primate vision. Yet, relatively little is known about their effects on the neural processing of visual inputs. Here we examine this question in primary visual cortex (V1) using receptive-field-based models, combined with an experimental design that leaves the retinal stimulus unaffected by saccades. This approach allows us to analyse V1 stimulus processing during saccades with unprecedented detail, revealing robust perisaccadic modulation. In particular, saccades produce biphasic firing rate changes that are composed of divisive gain suppression followed by an additive rate increase. Microsaccades produce similar, though smaller, modulations. We furthermore demonstrate that this modulation is likely inherited from the LGN, and is driven largely by extra-retinal signals. These results establish a foundation for integrating saccades into existing models of visual cortical stimulus processing, and highlight the importance of studying visual neuron function in the context of eye movements.

Figure 1. Saccades produce biphasic modulation of V1 neuron firing rates.

(a) Schematic of the behavioural task and stimulus. Top: the animal was required to maintain fixation on a target that made periodic jumps every 700 ms. Middle: random bar patterns (‘one-dimensional (1D) ternary noise' updated every 10 ms) were displayed, covering the recorded neurons' RFs, while the animal made ‘guided saccades' to maintain fixation on a visual target. Bottom: because the fixation target moved parallel to the random bar stimuli, the sequence of 1D noise patterns in the neurons' RFs (region highlighted by dashed red lines) were not affected by accurately executed saccades (timing of an example saccade indicated below). (b) Saccade-triggered average SU firing rates (normalized by each neuron's mean rate) showing biphasic modulation (n=84). Here, and in all subsequent figures, shaded regions show the interval mean±s.e.m. (c) Schematic showing the definitions of suppression and enhancement magnitudes (α_S and α_E, respectively), as well as their timing (τ_S and τ_E). (d) For neurons with significant modulation (circles; suppression: n=83 out of 84; enhancement: n=77 out of 84; see Methods), the strengths of perisaccadic suppression and enhancement were variable, but had similar magnitudes overall (suppression: 0.32, 0.24–0.44; enhancement: 0.29, 0.19–0.38; median, interquartile range). The timing of peak saccadic suppression was highly conserved across neurons (64, 56–72 ms), while the timing of peak enhancement was more variable (127, 112–154 ms). Small dots indicate non-significant peaks.

Figure 2. Saccades produce a monophasic suppression of gain and stimulus information.

(a) Schematic of the nonlinear LNLN cascade stimulus processing models. The summed input of a set of LN subunits gives the ‘generating signal' g(t), which is transformed into a firing rate by the spiking nonlinearity. (b) Example SU stimulus processing model with 4 excitatory (top) and 4 suppressive (bottom) filters. The ‘upstream nonlinearity' associated with each filter is shown below. Scale bar is 0.2°. (c). Firing rate (colour) as a joint function of the generating signal and time since saccade onset for the example neuron in b. Average firing rate at each time relative to saccade onset is shown below. (d) Vertical slices from the joint response function show that at the time of maximal saccadic suppression (dashed green line in c) the response gain is greatly suppressed (green trace), whereas at the time of maximal saccadic enhancement (dashed red line in c), the neuron's firing rate increases in a largely stimulus-independent manner (red trace). Shaded grey area indicates the distribution of g. (e) For the example neuron (top), as well as across the population (bottom; n=84), saccades produced biphasic firing rate modulation. Such firing rate modulation is decomposed into a multiplicative gain suppression (f) and an additive increase in firing rate ‘offset' (g). (h) As a result, the single-spike information (I_SS; magenta) showed a large monophasic suppression following saccades (top: example neuron; bottom: population avg.). Information rates (green), given by multiplying I_SS by saccade-triggered average firing rates, showed similar monophasic suppression.

Figure 3. Differential modulation of firing rate and stimulus selectivity.

(a) Perisaccadic I_SS, computed separately for those neurons showing strong I_SS suppression (blue; n=42) versus weak I_SS suppression (red; n=42), based on a median split. (b) Neurons showing strong I_SS suppression had much larger increases in firing rate offset, as expected. (c) Neurons with larger reductions in I_SS also showed stronger gain suppression. (d) Despite large differences in the magnitude of perisaccadic reductions in stimulus selectivity, saccades produced similar firing rate modulation for the two groups of neurons.

Figure 4. Microsaccades produce similar firing rate modulation, but weaker gain suppression.

(a) Saccades (blue) and microsaccades (red) produced similar biphasic firing rate modulation, with slightly weaker suppression (median relative suppression magnitude: 0.81-fold; n=84) and enhancement (0.77-fold) following microsaccades. (b) Suppression of response gains following microsaccades was qualitatively similar to, but weaker than, suppression following saccades. (c) The increase in firing rate ‘offset' following microsaccades was substantially weaker than that following saccades. (d) Stimulus information showed monophasic suppression following microsaccades, but the suppression was substantially weaker (0.54-fold) and shorter-lived compared with saccades.

Figure 5. Timing and laminar profile of perisaccadic firing rate modulation suggests upstream origin.

(a) Schematic showing three different sources of saccade-related modulatory signals: (red) signals targeting ‘upstream' LGN inputs to V1; (green) direct modulatory inputs to V1; (blue) signals targeting downstream cortical structures that provide feedback projections to V1. (b) Perisaccadic firing rate suppression occurred significantly later for neurons with more delayed stimulus responses (Spearman's rank correlation ρ=0.61; P=6.4 × 10⁻⁹; n=76). Red line shows linear fit and black line shows the diagonal. (c) Stimulus-onset-triggered current source density (CSD) profile from an example laminar probe recording. The location of a short-latency current sink is used to estimate the upper and lower boundaries of layer IV. (d) The time of maximal saccadic firing rate suppression was significantly delayed for multi-unit activity from supragranular electrodes (blue; n=58 across 9 recordings) compared with granular (red; n=55) and infragranular (black; n=85) electrodes. (e) Supragranular units also showed temporally delayed stimulus response profiles (see Methods).

Figure 6. Perisaccadic changes in V1 stimulus processing are driven by upstream gain suppression.

(a) Schematic illustrating the key differences between an ‘upstream' source of gain modulation (left, red) and modulation arising from direct inputs to V1 (right, green). In the former (‘pre-filter' modulation), perisaccadic suppression is already present in the stimulus-driven inputs to a V1 neuron, before application of its stimulus filters (illustrated here by a single temporal filter; black). Thus, the neuron's firing rate at a delay τ after a saccade will be less sensitive to some stimuli, producing an altered ‘effective temporal kernel' (shown for two example times; bottom left). Alternatively, direct perisaccadic suppression onto the V1 neuron, acting after temporal integration of its stimulus filters (‘post-filter' modulation), will produce a uniform scaling of the neuron's temporal kernel (bottom right). (b) The fundamental difference between these two sources of gain modulation is thus the pattern of perisaccaddic changes in the neuron's sensitivity to stimuli at different latencies (bottom). (c) Perisaccadic changes in the neurons' sensitivity to stimuli at different latencies (Methods); shown for an example neuron (left), and averaged across the population of neurons (right; n=84). The time of maximal perisaccadic suppression occurred systematically later for stimuli at larger latencies, consistent with the pre-filter suppression model. (d) The time of maximal saccadic suppression (red dots at left in c) systematically increased as a function of stimulus latency. Across the population, the slope of this relationship was much greater than 0 (median=0.82, blue arrow; P=3.3 × 10⁻⁹; n=67; see Methods), suggesting gain suppression occurs before the neurons' temporal filtering. Red and black lines show slopes of 0 and 1, respectively. (e) Schematic of the stimulus processing model with perisaccadic gain suppression incorporated either upstream of the stimulus filters (red) or after integration of subunit inputs (green). (f) The pre-filter gain model performed significantly better than the post-filter gain model (median relative LL improvement=22%, blue arrow; P=2.3 × 10⁻¹⁰; n=84; Methods). (g) Average pre-filter gain kernels (red), showing that the gain acting on stimuli presented over a short window following saccade onset was sharply suppressed. By comparison, post-filter gain kernels (green) were temporally delayed relative to the pre-filter model, reflecting the delay associated with processing by the stimulus filters.

Figure 7. Saccade modulation is due largely to extra-retinal signals.

(a) Schematic diagrams showing three different saccade conditions. Left: saccades made with a grey background. Middle left: saccades made with natural image backgrounds. Middle right: simulated saccades generated by translating the background image during static fixation. Right: saccades made in total darkness. (b) Comparison of modulation by saccades made with image backgrounds (red) versus grey backgrounds (blue). Clockwise from top-left, population averages (n=64) of perisaccadic firing rate, I_SS, gain and offset. Saccades with image backgrounds produced similar modulation of average rates, with slightly stronger suppression (median=1.22-fold; P=8.9 × 10⁻⁷), but equivalent enhancement (P=0.30). They also produced similar, but slightly weaker, reductions of I_SS (median=0.86-fold; P=2.6 × 10⁻³). (c) Similar to b, comparing the effects of real (red) and simulated (green) saccades (n=56). Simulated saccades produced biphasic firing rate modulation, though both suppression and enhancement were weaker (suppression: 0.52-fold, P=8.4 × 10⁻¹¹; enhancement: 0.57-fold, P=8.9 × 10⁻⁹), and temporally delayed (suppression: 1.31-fold, P=1.2 × 10⁻⁵; enhancement: 1.48-fold, P=3.0 × 10⁻⁸). Simulated saccades produced qualitatively different effects on gain and offset, such that they produced a slight increase in I_SS resulting from a reduction, rather than increase, in firing rate offset. (d) Top: average relative MU firing rates showed biphasic modulation following saccades made in complete darkness (n=96 MUs). Compared with saccade modulation during visual stimulation (magenta), saccades made in darkness (black) produced weaker modulation (particularly suppression), and a more prolonged period of post-saccadic enhancement (note wider time axis). Bottom: expanded view of the region highlighted by horizontal lines above.