Mechanisms and benefits of granule cell latency coding in the mouse olfactory bulb

Abstract

Inhibitory circuits are critical for shaping odor representations in the olfactory bulb. There, individual granule cells can respond to brief stimulation with extremely long (up to 1000 ms), input-specific latencies that are highly reliable. However, the mechanism and function of this long timescale activity remain unknown. We sought to elucidate the mechanism responsible for long-latency activity, and to understand the impact of widely distributed interneuron latencies on olfactory coding. We used a combination of electrophysiological, optical, and pharmacological techniques to show that long-latency inhibition is driven by late onset synaptic excitation to granule cells. This late excitation originates from tufted cells, which have intrinsic properties that favor longer latency spiking than mitral cells. Using computational modeling, we show that widely distributed interneuron latency increases the discriminability of similar stimuli. Thus, long-latency inhibition in the olfactory bulb requires a combination of circuit- and cellular-level mechanisms that function to improve stimulus representations.

Keywords: coding; inhibition; latency; olfactory bulb.

Figure 1

Long-latency granule cell activity. (A) Experimental setup: A brief, 300 μs stimulation pulse was applied via an extracellular stimulating electrode impaled into a glomerulus. Responses of inhibitory granule cells were recorded using patch clamp. (B) Average depolarization following glomerular stimulation. Arrowhead denotes stimulation time. (C) Example voltage traces from a granule cell following glomerular stimulation (arrowhead), eight trials shown. (D) Distribution of mean first spike latencies across granule cell population. (E) Relationship between granule cell first spike latency and spike probability. (F) Relationship between granule cell first spike latency and evoked rate.

Figure 2

Long-latency granule cell activity is driven by late onset excitatory inputs. (A) Membrane voltage and (B) membrane currents of a single granule cell following brief glomerular stimulation (arrowhead). Six trials shown. (C) Membrane potential (upper) and somatic currents (lower) averaged across trials. (D) Distribution of EPSC onset across cells. (E) Correlation between first spike latency and EPSC latency across cells.

Figure 3

Long-latency excitation comes from tufted, but not mitral cells. (A) Example of a mitral cell (black) and tufted cell (blue) response to glomerular stimulation (arrowhead). (B) Distribution of first spike latencies following glomerular stimulation in mitral (black) and tufted cells (blue). (C) Optical imaging setup. A stimulating electrode was impaled into a glomerulus and a nearby (but not connected) tufted cell was patched. The glomerulus was stimulated on every trial, while the tufted cell was made to fire on every other trial. Responses from three granule cells (D–F) to glomerular stimulation with (blue) and without (gray) additional tufted cell input were monitored optically. Black dash denotes timing of glomerular stimulation, blue dash denotes onset of tufted cell activity for tufted stimulation trials (blue).

Figure 4

Long-latency activation is an intrinsic property specific to tufted cells. (A) Example mitral (black), tufted (blue) and tufted cell + 4-AP responses to four amplitudes of injected current. (B) Firing rate/latency relationship in response to somatic current injection across the population of mitral cells (black triangles) and tufted cells (blue). Red circles denote tufted cells in the presence of 20 μM 4-AP; maroon triangles denote mitral cells in the presence of 20 μM 4-AP. Bars indicate s.e.m.

Figure 5

Inhibition with widely distributed latency improves stimulus discrimination. (A) Generating inhibition. For each stimulus, a population spiking template is generated (shown here in black). Model granule cell spike times are assigned such that the sum of all spikes is equal to the peri-stimulus time histogram template. While the total spiking is equal across model variants, they differ in the range of first spike latencies (shown with colored bars, below). First spike latencies ranged between 0 and (from bottom to top) 200, 400, 500, 600, 800, or 1000 ms. (B) Example of mitral cell spike trains receiving inhibition from each model variant shown in A. (C) First two principal components of mitral cell population firing for one pair of stimuli. Stimulus 1 = +; stimulus 2 = •. For each model variant, 100 trials are plotted per stimulus. (D) Classification accuracy of 100 stimuli as a function of population size. Colors correspond to model variants plotted in A–C. Bars denote s.e.m.

Figure 6

Latency-induced improvements are robust to changes in parameters and stimulus generation. (A) Heatmap of classification accuracy for varying ranges of latency spread at different behavioral time points. (B) Alternate stimulus generation. Each stimulus has a stimulus-specific latency pattern (as seen in Figure 5), as well as a stimulus-specific subset of active interneurons. Here, each stimulus recruits only 50% of model granule cells. Classification accuracy as a function of mitral cell population size. Colors correspond to latency ranges in Figure 5 (200, 400, 500, 600, 800, 1000 ms).