A primacy code for odor identity

Abstract

Humans can identify visual objects independently of view angle and lighting, words independently of volume and pitch, and smells independently of concentration. The computational principles underlying invariant object recognition remain mostly unknown. Here we propose that, in olfaction, a small and relatively stable set comprised of the earliest activated receptors forms a code for concentration-invariant odor identity. One prediction of this "primacy coding" scheme is that decisions based on odor identity can be made solely using early odor-evoked neural activity. Using an optogenetic masking paradigm, we define the sensory integration time necessary for odor identification and demonstrate that animals can use information occurring <100 ms after inhalation onset to identify odors. Using multi-electrode array recordings of odor responses in the olfactory bulb, we find that concentration-invariant units respond earliest and at latencies that are within this behaviorally-defined time window. We propose a computational model demonstrating how such a code can be read by neural circuits of the olfactory system.

Fig. 1

Primacy coding of odor identity. a Schematic presentation of the patterns of glomerulus (OSN) activation for three concentrations of the same odor. The total number of active glomeruli increases with an increase of odor concentration. b The temporal profiles of the odor concentration in the nose during inhalation for three concentrations of the presented stimuli. Dashed lines represent concentration thresholds for representative glomeruli (horizontal) and corresponding response latencies (vertical). c Temporal sequence of glomeruli activation for three different odor concentrations. d Optogenetic mask schematic demonstrating the effect of the optogenetic mask on temporal sequences when presented late and early relative to odor-evoked temporal pattern

Fig. 2

Neural response to optogenetic mask and effect on odor information. a Example raster and PSTH for two example MT cells. Inhalation onset corresponds to t = 0. Top: Example MT cell responses to optogenetic mask (25 mW, ON:2 ms—OFF:8 ms—ON:2 ms), light stimulation at 20 ms post-inhalation onset. Middle: Odor response (left: 2-hydroxyacetophenone, right: alpha-pinene). Bottom: Odor plus laser mask. b Baseline-subtracted mean of PSTHs for laser-responsive units (n = 113). c Cumulative distribution function (black line) and histogram of units’ response latencies to masking light stimulus. d Linear classifier cross-validation results for unmasked data for two odor presentations (limonene, pinene). Responses to unmasked odor presentations were time-binned (10 ms bins) and used to train support vector machine (SVM) classifiers. For each time point, SVMs were trained using the response vector inclusive of bins from t = 0 to that time. Each classifier’s performance is described by cross-validation on unmasked trials (blue) and testing on masked trials (green). Shaded areas indicate 95% confidence interval (Clopper–Pearson method)

Fig. 3

Optogenetic masking behavioral paradigm. a Behavioral task schematic: Mice were trained to respond to 2 odors "A" and "B" at 5 different concentrations to lick left or right water spout. Mask timed to the onset of the first inhalation after odor delivery was presented during subset of trials for two concentrations (asterisks). b Discrimination performance vs. mask latency for odors (2-hydroxyacetophenone and eugenol) at two concentrations. Mask stimulus presentation was initiated on the first inhalation of odorant and after the mask onset latency, t _mask, had elapsed. High concentration annotated in black markers, low concentration in grey. Error bars indicate 95% confidence interval estimates. Weibull fit to the data indicated with thick lines. Markers above performance curves indicate Weibull threshold latency values for each fit with 95% confidence interval estimates. c Mouse reaction time vs. performance in unmasked stimuli of two concentrations (as above). Dots indicate the data binned into bins of 125 trials. Lines indicate Weibull fit. Inset: difference between mask and reaction time threshold latencies. d Mouse reaction time vs. performance for unmasked trials (gray) compared with late masked trials with t _mask >  = 100 ms (blue) for the same data set. Points represent bins of 50 trials each. e Performance vs. mask latency for discrimination of pure (blue) carvone enantiomers vs. mixtures made with those odorants (green)

Fig. 4

Concentration invariant MT cell response latencies correspond with behavior. a Example raster and PSTH of two MT cells, which were responsive to odorant across 3 orders of magnitude concentration. t = 0 corresponds to inhalation onset. b Same plots for MT cells which respond to only a subset of concentrations. Dashed PSTH lines represent responses that did not cross the threshold for significance (see methods). c Cumulative distributions of MT cell response latencies. d Response latencies of MT cell-odor pairs across concentrations for concentration-stable and unstable units. Units’ responses across concentrations are connected. Area of each point in the plot is proportional to the Cohen’s D-score of the response over non-odor response (see “methods” section). Point positions are randomly jittered on the concentration axis for visibility. e Odor responses sorted by units’ response latencies at the highest concentration of odorant. Responses are represented as D-scores, where positive and negative scores are excitation and inhibition relative to blank inhalation responses

Fig. 5

Computational model of primacy decoder can explain behavioral effects of mask. a The schematic of the network included in the computational model. Pyramidal neurons in PC receive random excitatory connections from MT cells (black lines) and connected to a random subset of other pyramidal neurons via inhibitory connections (blue lines). The inhibitory connections are responsible for blocking of the inputs to PC after the initial (primary) representation is formed. b, c Responses of MCs (100 out of 300) and PC cells for late (b) and early (c) masks. MCs respond to Odor A, Odor B, and ChR2 (green, red, cyan). PC responses at the end of sniff cycle (gray squares) are shown for two odors in the same panel. For the early mask (c), PC responses are indistinguishable for two smells (yellow). d Discrimination success rate as a function of mask latency for three conditions: odors A vs. B—high concentrations (black line), odors A vs. B–low concentrations (gray line), odor mixtures 60% A + 40% B vs. 40% A + 60% B (blue line). Low/high concentration conditions were introduced by expanding/compressing the timing of MC odorant-related inputs within the sniff cycle. Mixtures were introduced by combining together high conc. sequences of inputs corresponding to odors A and B. The temporal dependencies reproduce the basic qualitative features of data (Fig. 2b, e), including concentration-dependence of the error rate at different ChR2 timings