Odor representations in the olfactory bulb evolve after the first breath and persist as an odor afterimage

Significance

Even when presented with steady stimuli, sensory systems respond dynamically, including adaptation or responses after the stimuli are gone (afterimages). To measure these dynamics in smell, we recorded electrical signals from the mouse brain. We found that the neuronal representation of odors changes between breaths, and not by simple adaptation, showing that the olfactory system also undergoes these dynamics. After the end of the odor, the brain continued responding, showing that there are olfactory afterimages. Finally, we tried to isolate where afterimages are generated and found that, while the nose has some contribution, afterimages are mainly maintained in the brain.

Abstract

Rodents can discriminate odors in one breath, and mammalian olfaction research has thus focused on the first breath. However, sensory representations dynamically change during and after stimuli. To investigate these dynamics, we recorded spike trains from the olfactory bulb of awake, head-fixed mice and found that some mitral cells’ odor representations changed following the first breath and others continued after odor cessation. Population analysis revealed that these postodor responses contained odor- and concentration-specific information—an odor afterimage. Using calcium imaging, we found that most olfactory glomerular activity was restricted to the odor presentation, implying that the afterimage is not primarily peripheral. The odor afterimage was not dependent on odorant physicochemical properties. To artificially induce aftereffects, we photostimulated mitral cells using channelrhodopsin and recorded centrally maintained persistent activity. The strength and persistence of the afterimage was dependent on the duration of both artificial and natural stimulation. In summary, we show that the odor representation evolves after the first breath and that there is a centrally maintained odor afterimage, similar to other sensory systems. These dynamics may help identify novel odorants in complex environments.

Sensory systems, even when presented with fixed stimuli, use dynamic neural representations that change over time. These changes range from simple potentiation or adaptation to more complex temporal patterns (1, 2). These dynamics can persist even after the cessation of stimulus, which can take the form of an off-response or prolonged aftereffect. Aftereffects have been intensely studied in vision (1, 3) and also observed in audition (4, 5), touch (6, 7), taste (here the afterimages may be due to persistent ligand binding) (8, 9), and insect olfaction (10, 11). In mammalian olfaction, the only reported aftereffect is a “persistent afterdischarge” following high concentration odors (12).

Olfaction is an active process that is coordinated by breathing (2, 13). Odorants bind to odorant receptor neurons (ORNs) in the epithelium, which synapse onto mitral/tufted (M/T) cells in the olfactory bulb (OB) at precisely defined glomeruli. Direct recordings of ORNs in rats show that ORNs are excited by odors and that activity can last for seconds after odor cessation (14). Others have measured the output of ORNs, by imaging glomeruli, and found that the response is gated by each breath and that the amplitude decreases with time (13, 15). The importance of breath segmentation for M/T cells has recently been shown in awake subjects, where neurons respond with precise, phasic firing patterns (16–18). However, these recordings have focused on how information is represented during the first breath, as rodents can identify odors in a single breath (19–21). Activity in piriform cortex is also segmented by breaths, and neurons there respond sparsely (22, 23), and with adaptation (24, 25). To date, no one has addressed, in the OB of awake rodents, how the odor representation evolves with each breath and how the odor representation changes after odor offset. Answering these questions will yield insight into how odor identity is maintained and how the olfactory network represents odors.

To investigate olfactory dynamics, we recorded from M/T cells in awake, head-fixed mice. We found that, during the odor, the odor representation shifts significantly after the first breath and not via simple attenuation. Furthermore, we observed a subset of cells that responded in an odor-specific manner after odor cessation—an olfactory afterimage. We performed calcium imaging on sensory neuron terminals, photostimulated channelrhodopsin-expressing M/T cells, and determined that the afterimage is primarily maintained centrally rather than peripherally. In conclusion, we have found that the odor response is dynamic and that one must consider all breaths of the response during an odor and afterward to characterize a cell's “odor receptive field.” The existence of an odor afterimage shows that afterimages may be a general, useful property of all sensory systems.

Results

Evolution of the Odor Representation in Single Cells.

We recorded from M/T cells of awake, head-fixed mice, using tetrode arrays, and measured respiration using a directional airflow sensor (average breath duration 327 ± 140 ms, SD; early trials with “fast-sniffing” were elided). We presented odorants (5% dilution in air) to the mice using a custom olfactometer for 2 or 2.5 s, with odor onset timed to follow the end of inspiration. Given the variability of breath duration in awake animals, we aligned each response by truncating spikes that arrive later than the mean breath duration and padded short breaths with empty time. This procedure ensures keeping the actual spike time unaltered for each breath (Materials and Methods).

Following realignment, many cells showed clear odor responses with precise firing within the sniff cycle (16, 17). Fig. 1A shows a cell with a low basal firing rate that is excited by odor presentation (Top and Middle), with most spikes coming 100–200 ms after inspiration onset (Bottom). This cell continued to fire at the same time for each subsequent breath. In addition to the cells that maintained their response for each breath, a subset of cells changed their odor response after the first breath. The cell in Fig. 1B fired phasically during all breaths but fired ∼40 ms later during the second breath than during the first (Bottom). Other cells' responses changed more dramatically: some were excited for the first breath before being inhibited (Fig. 1C) or were inhibited for just the first breath (Fig. 1D). Overall, we saw diverse changes in M/T cells' odor responses between the first and second breath.

Fig. 1.

The odor response of individual cells changes after the first breath. (A) This cell fires with the same phase for each breath in response to butanol (5% diluted in air). (Top) Histogram of odor response (bin = 20 ms). Inspiration onset marked by dashed blue lines, and odor above. (Middle) Average firing rate during each breath. (Bottom) Cumulative spike plot during the preodor and first–third odor breaths. (B) This cell shifts phase between the first and second odor breath, as seen in Bottom cumulative spike plot (KS, P < 0.02). (C) This cell is excited for one breath before being inhibited (ANOVA, P < 0.05). (D) This cell is inhibited for the first breath before firing phasically (ANOVA, P < 0.05). (E) Percentage of responsive cell-odor pairs for first and second breath (Left). Percentage of cells that respond exclusively during the first or second breath, or respond to both (Right). n = 516 cell-odor pairs from 10 mice. (F) Phase (circular mean) of responses during the first and second breath. The second breath response arrives later. n = 117 cell-odor pairs. (G) Magnitude of phasicness (circular variance) of response during first and second breath. There is no difference between the first and second breath. Same data as in F.

To quantify how cells responded, we calculated whether cells exhibit a rate or a temporal change during each breath. To identify rate responses (changes in firing rate during the odor), we averaged for each breath the firing rate over the entire breath and compared this rate with the average baseline firing rate (Fig. 1 A–D, Middle; ANOVA with Tukey’s post hoc test, P < 0.05). Of the 516 cell-odor pairs analyzed for Fig. 1, 21% (44 excited, 66 inhibited) responded with a rate change during the first breath, and 19% (36 excited, 61 inhibited) during the second breath (Fig. 1E). To detect changes in spike timing (i.e., a temporal change), we calculated the cumulative spike distribution over the breathing cycle (Fig. 1 A–D, Bottom) for baseline breaths (black line), the first odor breath (blue line), and subsequent breaths (red, green). We tested whether the spike distributions were different using a Kolmogorov–Smirnov (KS) test (Materials and Methods). Using this criterion, we found that 27% (137/516) of cell-odor pairs responded during the first breath, and 23% (118/516) responded during the second breath (Fig. 1E). Combining these two measurements, we found that 37% (193/516) of cell-odor pairs responded during the first breath with either a rate or temporal change, and 36% (187/516) responded during the second breath (Fig. 1E). These response percentages are similar to those observed previously (25% at 0.05% dilution in ref. 16; 60% at 2% dilution in ref. 17). Given that some cells do not respond during the first breath, but respond afterward, we calculated how many cell-odor pairs responded during the first breath alone (16%, 85/516), second breath alone (15%, 79/516), or both (20%, 108/516; Fig. 1E). Of the subset of cells that have a temporal response during the first and second breath, 23% (45/192; KS, P < 0.02) of cell-odor pairs changed phase between the first and second breath.

To more closely examine these temporal responses, we used circular statistics. For all cells that had both a temporal response and phasic firing during the first and second breath, we calculated the circular mean of firing (i.e., phase) during the breath. Responses during the second breath came, on average, 30 degrees (∼30 ms) later than the first breath (158 ± 6 vs. 190 ± 7 degrees; Watson–Williams test, P < 0.05; Fig. 1F). We also calculated the circular variance (magnitude of phasicness) for the first and second breath and found no systematic shift in the strength of the phasing (paired t test, P = 0.27; Fig. 1G). What happens to the odor representation beyond the second breath? We repeated the circular mean test for the second and third breaths and found that the changes in phase were smaller than those between the first and second breath (χ² test, P < 0.01). In conclusion, the odor representation becomes more stable after the first breath.

These dynamics could have been the result of an unstable stimulus. To characterize our pulse of odorants, we recorded the odor concentration, using a photoionization detector (PID), and performed multiunit recordings on the same day. The PID measurements showed that the concentration reached 80–90% of its final concentration within 150 ms (before the first inspiration), then gradually increased over a few hundred milliseconds (Fig. S1 A and B). The percentage of responsive cells and changes between breaths were similar for these new measurements compared to our first dataset (Fig. S1C). Seeing that the concentration could change 5–10% between the first and second breath, we were concerned that the representation change between these breaths was due to concentration and not neural dynamics, as these changes are near the intensity discrimination threshold for rats (26). Previous investigations into concentration coding in olfaction have looked at changes in concentration of orders of magnitude, not a few percent (27–29). Given that the concentration changes over the time range of the first breath, we used the variability in breathing as a proxy for small changes in odor concentration (Fig. S1 D and E). We found similar responses for short and long odor durations before the first odor breath, evidence that the changes between breaths are not due to small concentration changes, but representation dynamics.

Evolution of the Population Representation.

We next turned to the population level to see how the ensemble representation changed with each breath. We constructed a population vector that includes each cell's response for the entire breath (Fig. 2A) (27) and calculated the distance between the odor representations for each breath (Materials and Methods). We found that all odor breaths are different from control breaths (ANOVA, P < 0.01; Fig. 2B, top row). Furthermore, we found that the first odor breath was different from each subsequent breath whereas all of the later breaths were similar to each other (ANOVA, P < 0.05; Fig. 2B, second row). On the population level, the odor representation changes the most between the first and second breaths. We found similar results in the experiments following PID measurement (Fig. S1F).

Fig. 2.

The population representation changes after the first breath. (A) Diagram of population vector construction. Each cell's firing rate was calculated during eight bins (Top, ∼40 ms) for each breath before and during the odor. A population vector was then made containing every cell's whole-breath representation for each breath and odor. (B) We calculated the distance between preodor (Ctl) and odor breaths (1–5) for each odor and normalized to ctl–ctl distances. The preodor breaths were highly distant from all odor breaths (top row). The first odor breath was also distant from the second–fourth odor breaths (second row). However, the distance between odor breaths 2–5 was similar to the ctrl-ctrl distance. n = 345 cell-odor pairs (115 cells from six mice). Color bar: interbreath distance normalized to control breaths. (C) Odor prediction using a template from the same breath (black), first odor breath (blue), or fifth odor breath (red) (mean ± SEM of 50 permutations; n = 124 cells from five mice). Using the same breath as the template, the prediction was high for all breaths. Using the first or fifth breath as the template yielded accurate predictions only for nearby breaths.

As an alternative look at the population, we used a classifier to see whether population activity of individual trials could predict the odor. To do so, we calculated the distance between the population vector for a single trial (sample) and the average population vector of each odor (template); the smallest distance is the “predicted” odor (Materials and Methods). Using the same breath for both sample and template, the predictor is at chance during the preodor period and is able to accurately classify the odor for all odor breaths (Fig. 2C). As a second measure of how the odor representation evolves, we asked the predictor to classify odor responses using the first or fifth odor breath as a fixed template. Using the first breath as a template, the prediction is most accurate for the first breath but quickly diminishes (Fig. 2C); in contrast, using the fifth breath as a template yields a poor prediction for the first breath that improves with each breath (Fig. 2C). The fact that these cross-breath predictions are above chance probably reflects the subset of cells that have stable representations for each breath.

Olfactory Representations Following Odor Cessation.

In all other sensory systems, aftereffects have been observed both neurally and behaviorally (1, 3–9). To investigate odor coding after odor cessation, we analyzed cells' postodor firing. Some cells that responded to the odor continued to do so after the odor ceased, as if the odor was still present (Fig. 3A). Other cells fired phasically, but at a different phase from either preodor or odor firing (Fig. 3B). Many cells responded either with prolonged inhibition (Fig. 3C) or excitation (Fig. 3D). To see whether odor remained after the presentation, we measured the concentration using a PID. Trace odorant remained for a few hundred milliseconds after the stimulus (Fig. S2A); thus, for all statistics, we used the fourth postodor breath, which occurs ∼1 s after the end of the odor pulse, after the odor dissipated. Of all cell-odor pairs, we found that 30% of cells had a postodor response that differed from preodor firing (rate, 55/465; temporal, 73/465; both rate and temporal, 12/465; Fig. 3E, Left). The rate change during postodor responses was toward decreases in firing rate (55/85 decreased; 30/85 increased), similar to the odor responses. Thirty-four percent (158/465) of cell-odor pairs responded only during the odor, 20% (94/465) responded both during the odor and after, and 10% (46/465) responded only during the postodor (Fig. 3E, Right). To look at the phase of the postodor responses, we compared postodor responses to those during the odor (using odor breath 4 for comparison). The postodor breaths had an earlier phase than odor breaths (Watson–Williams test, P < 0.01). The circular variance of postodor responses was higher than the odor responses (paired t test, P < 0.01; Fig. S2B).

Fig. 3.

Individual cells contain odor-specific information after odor offset. (A) This cell fires phasically during the odor and postodor. (Left) PSTH (bin = 20 ms) during odor and afterward. The odor response was cut at 2 s to allow alignment of the first postodor breath. (Right) Cumulative spike plot of the preodor breath (black), fourth odor breath (blue), and the fourth and fifth postodor breaths (red, green). This cell maintains its phase during the postodor. (B) This cell fires phasically during the postodor, but with an earlier phase than the preodor or odor firing. (C) Histogram of a cell that fires phasically during the odor, but it is inhibited during the postodor. (D) Histogram of a cell that is inhibited during the odor and is excited postodor. (E, Left) Approximately 30% of cell-odor pairs respond during the postodor period (compared with preodor). All statistics were calculated using the fourth breath to avoid lingering odor contamination. (Right) Thirty-four percent of cell-odor pairs respond during the odor, 20% respond during both the odor and postodor, and 10% respond only during the postodor (n = 465 cell-odor pairs from 10 mice). (F) Phase of responses during the fourth odor breath and fourth postodor breath. The phase of responses is earlier during the postodor (n = 22 cell-odor pairs). (G) The ensemble contains odor-specific information after odor cessation. The odor was predicted using the same sample breath as the template (mean ± SEM of 50 permutations; Materials and Methods). The odor prediction was higher for awake mice than anesthetized, and, in anesthetized mice, for familiar odors over naïve odors (ANOVA, P < 0.05; n = 124 cells from six mice for awake; 168 cells from four mice for familiar; 194 cells from seven mice for naïve).

The postodor responses could contain information regarding the presented odor or could be a general reset response. To differentiate between these possibilities, we used our odor predictor to investigate the postodor breaths. Here, the predictor was able to correctly identify the presented odor above chance for >10 breaths (or >3 s), after the odor has dissipated, and gradually returned to baseline (Fig. 3G). To see whether the postodor representation is similar to the odor representation, we used the fifth odor breath as a template and found that the predictor was only above chance for a few breaths (ANOVA, P < 0.05; Fig. S2C). To look at the network evolution of the odor afterimage, we performed the breath distance analysis and found that the representation changes gradually with each breath (Fig. S2D). Furthermore, if we used the odor predictor with a fixed postodor breath template, the prediction was accurate only around the template breath (Fig. S2C). Thus, the postodor representation contains odor-specific information, is not a recapitulation of the odor representation, and gradually dissipates. Given these three properties, we propose that the postodor response is an olfactory afterimage.

So far, we have performed all of the experiments in awake mice, but temporal dynamics in other systems have been shown to be dependent on animal state, such as wakefulness or attention (30, 31). To see whether these top-down processes can influence the odor dynamics, we repeated the experiments in anesthetized mice, under two conditions: mice that were presented the odor while awake, and then anesthetized (familiar); and mice that were anesthetized before odor presentation (naïve). In anesthetized mice, the odor representation had interbreath dynamics similar to awake mice (Fig. S3A), and individual cells had postodor responses (Fig. S3B). After odor cessation, both familiar and naïve mice had postodor predictions above chance, but less than the awake prediction (ANOVA, P < 0.01; Fig. 3G). The familiar postodor prediction was also higher than the naïve prediction, although only for the first few postodor breaths (ANOVA, P < 0.05).

Olfactory Afterimages at Low Concentration.

All experiments performed so far were at a high odor concentration (5%), which evokes relatively dense activity patterns in glomeruli (28, 32). To check whether olfactory afterimages are present at lower concentrations, and to see whether the afterimages contain concentration-specific information, we repeated our experiments using three lower odor concentrations (0.1%, 0.4%, and 2%). Individual cells had postodor responses that were concentration-specific (Fig. 4A). Postodor responses were present at all concentrations tested, with more postodor responses at the highest concentration (Fig. 4B). To investigate concentration coding on the population level, we used the predictor to classify both the odor and concentration presented. During the odor, the predictor was able to identify the odor and concentration with >60% accuracy (Fig. 4C). After odor cessation, the predictor was still able to identify the odor and concentration combination above chance for >10 breaths. To ensure that the predictor was working even at the lowest concentration, we used the predictor to classify the odor within each concentration (Fig. 4D) and found that it was above chance at all concentrations. We next used the predictor to classify the concentration for each odor. Here, the classification success rate was lower than for odor prediction [because odor representations are similar for different concentrations (27, 33)] but was above chance both during and after the odor (Fig. 4E). The experiments at different concentrations allowed us to perform some controls. To see whether the interbreath dynamics are due to the high concentration used, we performed the breath-distance analysis on responses to 0.1% odor and found that the representation changes after the first breath (Fig. S4A). To see whether the postodor predictions were simply due to remaining odorant, we predicted postodor breaths using low-concentration odor templates (Fig. S4B). These templates were only above chance for a few breaths, like the same-concentration odor template, showing that the afterimage is not simply a low-concentration representation.

Fig. 4.

Postodor responses contain both odor and concentration information. (A) PSTHs (bin = 20 ms) from a cell that responds to amyl acetate (Aa) differently at different concentrations (odors diluted first to 10% in mineral oil, then to the final concentration in air). Breaths shown as dashed blue lines. (B) Percentage of responsive cell-odor pairs for the fourth postodor breath at each concentration (n = 107 cells from two mice). Same criterion as Fig. 3E. (C) The population can predict the odor and concentration during both the odor and the postodor period (mean ± SEM of 50 permutations, chance level = 11%). (D) The odor and postodor prediction is present at all concentrations (mean ± SEM of 50 permutations, chance level = 33%). (E) The postodor prediction contains concentration information for all odors (mean ± SEM of 50 permutations, chance level = 33%).

Site of Generation of the Olfactory Afterimage.

Some possible sources of the postodor responses include lingering input from ORNs, resonant activity within the OB, or feedback from higher order areas. Recordings of ORN activity in anesthetized rats have shown that activity can continue for a few seconds after odor cessation (14). To investigate postodor ORN activity, we performed calcium imaging of sensory terminals in awake, head-fixed OMP-GCaMP3 mice. During odor application, we were able to identify individual glomeruli that responded to the stimulus (Fig. 5A). Like M/T-cell activity, calcium responses in glomeruli are gated by breaths, with sharp increases in fluorescence following inspiration (Fig. 5B). Following odor cessation, the fluorescence in most glomeruli decayed with a single monotonic exponential. However, in a small minority of glomeruli, breath-by-breath activity continued for a few seconds after odor cessation (Fig. 5B, Center, top trace).

Fig. 5.

Sensory activity is nonpersistent after odor cessation. (A) Average olfactory maps obtained for three odorants (Ac, acetophenone; EB, ethylbutyrate; He, 3-Hexanone) in the same awake mouse. The fluorescence from GCaMP3, expressed under the control of the olfactory marker protein (OMP) promoter, is averaged over the 3 s of odor presentation and is represented as ∆F/F. (B) Calcium dynamics in sensory neuron terminals from the ROI drawn in A, for individual trials. The breathing of the animal is presented in the lower trace. The Center black trace (ROI 1, acetophenone) shows a rare example of a cycle-by-cycle modulation of the fluorescence signal after the end of the odor presentation. All of the odorants are diluted 20× in air and are presented for 3 s. (C) Normalized responses from glomeruli–odor pairs that respond to odor (n = 181). The amplitude of the fluorescence is measured for each breath and normalized to the amplitude of the first breath during the odor presentation (nyanza green box). A single exponential decay was fitted to and removed from the postodor data. The area around the curve represent ± SEM. (D) Average fraction of glomeruli that respond to a stimulus. For all six odorants, we calculated the percent of glomeruli that responded, including all glomeruli that responded to at least one odorant (n = 79 glomeruli). (E) Prediction of odor identity based on sensory inputs. Using a population vector consisting of all glomeruli that respond to an odor (n = 79 glomeruli), we predicted the odor during and after presentation. The predictor is able to identify the odor during presentation but returns to chance quickly afterward. Mean ± SEM of 50 bootstrapped samples using 50 glomeruli.

To quantify the responses, we calculated the normalized change in fluorescence (ΔF/F) during each breath for responsive glomeruli–odor pairs (Fig. 5C). The responses were largest during the first breath, as the signal had increased from baseline. Following odor offset, the ∆F/F returned to baseline in a few breaths. For glomeruli that respond to at least one odorant, we then calculated for each breath the percentage of glomeruli that showed a response above noise level (Fig. 5D). During the odor, ∼65% of the glomeruli responded during the first breath, and this percentage diminished with each breath. Following odor offset, by the fourth breath, ∼95% of the glomeruli stopped responding. We should note that this is not a precise estimate of glomeruli responsiveness: we are not considering glomeruli that did not respond to any odor; have selected odorants that activate the dorsal surface; and are missing glomeruli that are below the sensitivity of the calcium indicator. To enable more direct comparison of the imaging results to the electrophysiology, we used the odor predictor, this time building populations out of glomeruli activity. The predictor was able to identify the odor with >80% accuracy during the first breath, and >50% accuracy for all breaths (Fig. 5E). However, during the postodor period, the predictor was accurate only for a single breath. Although a small fraction of sensory neurons do respond after odor cessation, they do not appear to contain as much odor-specific information as M/T cells.

Odorant properties such as polarity, volatility, and sorptiveness can affect how quickly odorants are absorbed into the epithelium and have been shown to affect olfactory bulb activity and behavior (34–36). To see whether these properties could influence the odor afterimage, we repeated the experiments using odorants with a wide range of physicochemical properties. For both imaging and electrophysiology, we quantified the percentage of regions of interest (ROIs) or cells that responded during the fourth postodor breath for individual odors and plotted this percentage against odorant polarity, volatility, and sorptiveness (Fig. 6). There was no significant effect of any of these properties.

Fig. 6.

Odor physicochemical properties do not affect the odor afterimage. (A) Percentage of imaged ROIs with a postodor response during the fourth postodor breath, plotted as a function of polarity [water-octanol partition, log(K_ow)], vapor pressure (VP; log of VP in mmHg at 25 °C), and sorbptiveness (log of Henry's law constant). No property was significantly correlated with postodor responsiveness (linear fit; K_ow, P = 0.54; VP, P = 0.2; Henry, P = 0.92). For odorants and physicochemical properties see Table S1. (B) Percentage of cell-odor pairs with postodor response during the fourth postodor breath (including both rate and temporal responses). No property was significantly correlated with postodor responsiveness (linear fit; K_ow, P = 0.55; VP, P = 0.08; Henry, P = 0.86).

If the odor afterimage is not maintained peripherally, it must be maintained centrally. To try to observe the central representation, we recorded from the OB in the absence of peripheral input by blocking the nostril ipsilaterally to the recording site. To block the nostril, we applied a liquid silicone elastomer around the nostril and let it harden. Mice were anesthetized to reduce stress and to prevent them from removing the block. When we applied the plastic, the air pressure measured from the contralateral nostril doubled (Fig. 7 A and B).

Fig. 7.

Odor evokes activity in OB when the ipsilateral nostril is blocked. (A and B) Example air pressure (Upper) and LFPs (Lower) from open (A) and closed (B) nostril. Inspiration is up. Blocking the nostril causes inspiration-induced LFP deflections to change polarity. For both the open and blocked nostril, odor presentation increases LFP activity. (C) Power spectrum for the LFP shown in A and B. The power was calculated over 3- to 4-s epochs before, during, and after the odor. (D) Average power of all locations for open and blocked nostril (n = 9 recording sites from three mice). The power was higher during the odor and postodor than during the preodor (two-way ANOVA with factors of frequency and odor condition, F_{odor (2,12)} = 5.19, P < 0.05 for odor factor; odor vs. preodor, P < 0.05 with least significant difference post hoc test; odor vs. postodor, P < 0.05). The frequency was taken as the mean frequency of each recording at 5, 10, ..., 35 Hz).

With the nostril blocked, we recorded spikes from M/T cells (layer verified by odor responses after elastomer removal). Blocking the nostril nearly silenced neurons (37). When we applied odorants, only 4/423 cell-odor pairs responded during the first breath, showing that feedback alone is insufficient to drive OB activity for almost all neurons (38).

We also measured the local field potential (LFP) from the mitral cell layer, which allowed us to record subthreshold activity from both mitral and granule cells (39). At rest, with the nostril open, breathing induced theta-frequency deflections, and odors elicited gamma oscillations (Fig. 7 A and C) (40). In contrast, at rest, nostril blockade reversed the polarity of inspiration deflections and decreased the LFP power at all frequencies (Fig. 7 B and D). With the nostril blocked, odor presentation still caused theta and gamma activity in the LFP, despite lacking direct ORN input (Fig. 7B); the large inspiration-induced deflection was delayed ∼50ms. To quantify these changes, we calculated the LFP power for 3–4 s before, during, and after the odor. At some recording locations, with the nostril blocked, the LFP power at frequencies between 5 and 30 Hz increased both during the odor and during the postodor period (Fig. 7C). These frequencies are consistent with those for cortical feedback (40, 41). We averaged the LFP power from nine recording sites and found that, with the nostril blocked, the LFP power during the odor and postodor period was higher than the preodor period (Fig. 7D). One possible site of the odor afterimage is in the connectivity between the OB and olfactory cortices.

If the odor afterimage is maintained centrally, we wondered whether it would be possible to artificially stimulate the network and generate persistent activity without ORN input. We recorded from M/T cells in the medial OB of awake Thy1-ChR2 mice (42) and placed a light fiber coupled to blue laser (473 nm, 10–15 mW, 2-ms pulses) just above the dorsal surface of the OB, >1 mm away. However, when we flashed the light, the mice salivated and became agitated, making stable recording impossible. In the one awake mouse that we tried to record, light stimulation of more than one pulse induced fast sniffing (resting rate, 2.85 ± 0.7 Hz prelight; one pulse, 2.83 ± 0.57 Hz; >5 pulses, 6.3 ± 1.6 Hz; mean ± SD). Therefore, we shifted to anesthetized mice. We stimulated the OB at 10 or 20 Hz for 0.1–2.5 s. Light stimulation did not change the breathing rate in anesthetized mice (prelight, 2.58 ± 0.12 Hz; postlight, 2.77 ± 0.16 Hz). Following a single light pulse, the M/T cells' firing was basically unaffected. Each light pulse caused large deflections in the LFP (Fig. 8A). Following stimulation for more than 0.5 s, many M/T cells exhibited persistent firing that could last seconds (Fig. 8B); other cells underwent prolonged inhibition (Fig. 8C); and some cells were unaffected (Fig. S5A). We were often unable to record spikes during the light trains, either due to the large LFP changes making spike extraction difficult, or due to the M/T cells being silenced by inhibitory activity in the network. Photostimulation of wild-type mice yielded neither LFP fluctuations nor persistent firing (Fig. 8D).

Fig. 8.

Increasing stimulus strength drives more persistent network activity. (A, Upper) LFP recorded in mitral cell layer during Thy1-ChR2 stimulation at 20 Hz. (Lower) Light stimulation pulses. (B) PSTH from an M/T cell in a Thy1-ChR2 mouse during photostimulation at 20 Hz (bin = 20 ms). After stimulation, this cell was excited for a few seconds. (C) PSTH from an M/T cell that was inhibited after Thy1-ChR2 photostimulation. (D) Plot of integrated firing difference versus ChR2 stimulus duration (mean ± SEM) for Thy1-ChR2 mice (filled symbols) and WT mice (open symbols). Increasing the length or frequency of photostimulation increased the persistent activity for Thy1-ChR2 mice. n = 118 cells from five mice for Thy1-ChR2; 61 cells from two mice for WT. (E) PSTHs from one M/T cell from an awake mouse responding to short (0.5 s, Upper) and long (2 s, Lower) odor presentation (bin = 20 ms). The short presentation did not elicit a postodor response whereas the long presentation did (KS, P < 0.02). (F) Odor prediction following different odor duration (mean ± SEM of 50 bootstraps, chance level = 33%). Longer stimulus presentations elicited more accurate odor predictions (n = 81 cells from three mice).

To quantify the change in firing rate, we calculated the difference between the firing profile for prelight control breaths and postlight breaths (Fig. 8D; Materials and Methods). At 10 Hz, increasing the stimulus duration caused longer and stronger firing changes. Furthermore, for stimuli longer than 1 s, higher stimulation frequency caused larger firing differences. To better look at the duration of the effect, we calculated the firing difference breath-by-breath for each firing condition (Fig. S5B). At 10 Hz, the effect lasted only a few breaths whereas, at 20 Hz, the effect lasted >10 breaths.

Given that longer artificial stimulation yielded longer network activity, we asked whether natural stimuli could do so as well and varied olfactory drive by presenting odors to awake mice for 0.5–5 s. We found that some M/T cells' postodor responses differed between short and long odor application (Fig. 8E). On the population level, we used the odor predictor to classify postodor responses and found that the predictor was more accurate following longer odor application (ANOVA, P < 0.05; Fig. 8F). We measured the postodor odorant concentration using a PID and found that the concentrations were basically identical following different length odor presentation; thus, the improved postodor accuracy is not due to increased odorant concentration (Fig. S6A). In insects, following odor application of different lengths, the network returns to baseline via a common trajectory (10). To compare postodor trajectories, we calculated the distance between the postodor representation following 2 s and 5 s odor application (Fig. S6B). These representations were highly different, perhaps reflecting that the network responsible for the postodor effect had not settled into a fixed state. The long odor presentation also allowed us to compare the decrease in information during the odor to the decreases during the postodor (Fig. S6C). The breath-by-breath decrease in information was more variable between breaths during the afterimage, perhaps due to the smaller number of informative cells, but the magnitudes were similar. In summary, driving the OB network more strongly, either via natural or artificial stimulation, enhances persistent network activity.

Discussion

Although it has been shown that the sniff cycle is integral to the odor code, no one has focused on how the odor code changes with each breath in awake mammals. We recorded from M/T cells in awake, head-fixed mice and found that the odor code is dynamic during and after the odor. Between the first and second breath, individual cells could switch from inhibition to excitation, vice versa, or could shift the phase of their response (Fig. 1). On the population level, these differences mean that the odor code for the first breath is different from subsequent breaths (Fig. 2). After odor cessation, we identified an olfactory afterimage that contains odor- and concentration-specific information (Fig. 4) that also depends on stimulus duration (Fig. 8). Using calcium imaging and channelrhodopsin stimulation, we identified the primary site of the odor afterimage as central rather than peripheral (Figs. 6–8). The existence of olfactory afterimages shows that sensory afterimages are common to all sensory systems and may reflect that afterimages serve a functional purpose.

Dynamics of the Odor Response.

Some forms of olfactory dynamics have been observed before. In invertebrates, odors elicit complex temporal patterns (43, 44) that stabilize over time (10, 11). In contrast to the gradual changes observed in invertebrates, we observed an abrupt change in the representation between the first and second breath. This disparity may be due to the differences in the way invertebrates and mammals segment olfactory information: by motion and breaths, respectively. Turning to mammals, in anesthetized mice, our laboratory has previously reported that the population representation, using average firing rates, is most in flux at odor onset and offset but did not consider whether these dynamics contained odor information, or the source of these changes (33). We have extended those findings here to awake mice, single cells, and consideration of complex responses not reflected by the average firing rate, and have investigated the source of these dynamics. In the OB of rabbits, and the piriform cortex of anesthetized rats, attenuation has been observed (24, 25, 45). In contrast, we observe many forms of changes between the first and second breath. We should note that we have imaged only the dorsal OB and recorded from medial and dorsal OB, leaving open the possibility that activity is different in the ventral or lateral OB.

Multiple mechanisms could underlie the change in odor code between breaths. Recordings from drosophila and rat ORNs exhibit complex activity patterns, including adaptation, and delayed excitation (14, 46, 47). Imaging of glomeruli shows adaptation (Fig. 5) (15). One simple hypothesis would be that adaptation of ORNs would delay firing downstream. We found that some M/T-cell responses did shift to a later phase during the second breath (Fig. 1F), which could reflect attenuation of ORN input. Simple sensory adaptation, however, cannot explain the more dramatic changes in firing after the first breath (Fig. 1D).

Feedback within the OB or from higher areas could alter the response after the first breath. M/T cells form dendrodendritic synapses with GABAergic cells, including periglomerular and granule cells, which could mediate feedback within the bulb (48, 49). Besides synapses in the OB, granule cells also receive feedback from higher areas like piriform cortex and the anterior olfactory nucleus (41, 50–52). Given that the largest changes we see are between the first and second breath, it is possible that, before the first breath, the network is at rest; the first breath then activates the olfactory network, including feedback within the OB, and from cortex; then, during the second breath, ORN input is arriving into a different, active network. The delayed LFP activity in the blocked nostril condition (Fig. 7) may reflect the feedback delay. It has been shown that tufted cells are active earlier in the breathing cycle than mitral cells (53, 54); thus, the delayed activity in the blocked nostril condition could reflect that there are two olfactory processing streams, with tufted cells feeding direct olfactory information whereas mitral cells project to recurrent networks. Of note, this feedback appears to be automatic, and independent of attention, as we observed interbreath dynamics in both awake and anesthetized animals. Further experiments using transgenic mice to target granule cells or higher feedback areas will be necessary to isolate the source of the interbreath dynamics.

What cognitive use could this coding shift be useful for? Odor detection and discrimination can occur within a single breath (16, 19–21), and some have hypothesized that subsequent breaths are used for finer discrimination, as a speed–accuracy tradeoff (19, 20). From this view, one might hypothesize that the changes in the odor representation reflect better discriminability between odors; however, in our data, the best odor discrimination was during the first breath. Animals perceive odors clearly only for a few breaths (55), which may be due to the adaptation (not attenuation) that we observe after the first breath. This adaptation may then allow the detection of new odors in the background of the known odor. However, the significant change in representation between the first and second breath raises the issue of how the odor identity can be maintained between the breaths (56) although maintenance of odor identity is not a problem for differences in concentration (27).

The Olfactory Afterimage.

Following odor cessation, we recorded odor- and concentration-specific responses that we call an olfactory afterimage. The postodor responses during the first couple of breaths may be due to lingering odorants at low concentrations, but the odors quickly dissipate (Fig. S2). All of our single-cell analyses were performed using the fourth postodor breath, after the odor had gone; furthermore, the postodor prediction was above chance for >10 breaths, long after the odor was gone.

An excitatory persistent afterdischarge that lasts for tens of seconds has previously been reported in the M/T cells of anesthetized mice following odor presentation at 25–50% dilution (12), and short off-responses have been observed in frogs and hamsters (29, 57). In comparison, we observe varied postodor responses including phase changes and inhibition, and our olfactory afterimage lasts only a few seconds. We have furthermore extended the findings to awake mice and odor concentrations two orders of magnitude smaller, showing that the olfactory afterimage contains both odor- and concentration-specific information and is generated centrally.

Where is the olfactory afterimage maintained? The simplest explanation would be continued activity coming from ORNs. Some ORNs in flies have “ultra-prolonged” responses and continue firing seconds after the odor is gone (47, 58). However, when we performed calcium imaging of glomeruli in awake mice, only a small minority of glomeruli were active following the odor, which did not provide enough odor-specific information for classification (Fig. 5). Two-photon calcium imaging in dendritic tufts of mitral cells in anesthetized mice revealed that the calcium activity continues at a low level after odor cessation (59). These findings do not preclude ORN input from contributing to the afterimage but mean that whatever ORN input remains is at a low level.

Multiple lines of evidence point to a central representation. First, we observed that the olfactory afterimage is state-dependent for it is both stronger and longer in awake mice than anesthetized mice. Whether this difference is due to top-down processes like attention or anesthesia-induced changes in the network (28, 32) is impossible to disambiguate. We also observed that the olfactory afterimage is stronger in anesthetized mice for familiar odors over naïve odors although this effect was weak. The second argument for a central representation is that we observed activity in the olfactory bulb in the absence of direct olfactory input, even during the postodor period, which shows that there is feedback from higher areas (41, 52).

The final argument in favor of a central representation is that we were able to induce long-lasting activity in the olfactory network by photostimulating M/T cells with channelrhodopsin. Thy1 is expressed by mitral cells in the bulb (42), and stimulating them simultaneously should induce coactivation of granule cells, as well as activity in the contralateral olfactory bulb, AON, and piriform cortex. Our recording site in the medial bulb was >1 mm away from the fiber on the dorsal surface, showing that the network is not local. However, not all cells were affected by light stimulation (Fig. S5A), which could reflect input specificity to the dorsal bulb stimulation, or that only a subpopulation of cells are influenced by feedback. The observation that increasing light stimulation or odor presentation both yield longer aftereffects may reflect that the network is being driven into a dynamic steady state, which the network has to return from (10). The ChR2-induced activity lasted for >10 breaths and was stronger following prolonged light stimulation; similarly, the olfactory afterimage lasted for >10 breaths and was stronger following longer odor stimulation. The olfactory afterimage could be maintained by the same networks that shape interbreath dynamics. Further experiments targeting specific subpopulations will be necessary to dissect this circuit.

Sensory aftereffects have been observed in all sensory systems and have been most investigated in the visual system, including negative afterimages, the tilt aftereffect, and the motion aftereffect (1, 3, 60). Afterimages are distinguished from aftereffects by the fact that afterimages do not require a second stimulation to reveal their presence; given this definition, we refer to the postodor representations as afterimages. Intricate models have been proposed for the centrally maintained aftereffects, which rely on retinotopy and orientation maps in visual cortex, but olfactory networks lack such organization. The simplest model for visual aftereffects is that central neurons adapt to stimulation of their preferred orientation or direction (61). Although some cells did adapt during the odor (Figs. 1C and 3D), the postodor responses were too diverse for simple adaptation to explain the olfactory afterimage. Thus, central olfactory afterimages are likely maintained by olfaction-specific network properties.

The long duration and odor specificity of the olfactory afterimages lead us to hypothesize that they are a form of short-term memory (62). What is intriguing about this sensory memory is that there has been no report of a perceptual analog: that is, odor perception in the absence of odor. (It is difficult to design a behavioral paradigm to test postodor discrimination without the confound of odor discrimination.) In contrast, stimulus aftereffects in vision and audition are typically associated with perceptual analogs. This olfactory memory could be useful in contexts where an animal encounters multiple different odorants in a time-varying fashion and needs to distinguish between familiar and novel odorants.

Materials and Methods

For details, see SI Materials and Methods.

Animals.

For electrophysiology, we recorded from seven Slc6a4-Cre heterozygous mice in a C57BL/6J background, sixteen C57BL/6J mice, and five Thy1-ChR2 mice (B6.Cg-Tg(Thy1-ChR2/EYFP)^18Gfng/J, JAX 007612; The Jackson Laboratory). All mice were male, with the exception of two Thy1-ChR2 mice. Slc6a4-Cre (GENSAT) mice were infected with floxed-ChR2-YFP injected in the Raphe, but recordings were performed before photostimulation. All animals recorded were >8 wk old and maintained in 12-h day/night cycle. For calcium imaging, we used floxed-GCaMP3/OMP-Cre mice [Gt(ROSA)26Sor^{tm38(CAG-GCaMP3)Hze}/J (63): JAX 014538, crossed with OMP^tm4(cre)Mom (64); JAX 006668] in a C57BL/6J background, hemizygous for both alleles, and imaged at 6–7 wk old. All animal protocols were in accordance with the Swiss Federal Act on Animal Protection and the Swiss Animal Protection Ordinance and were approved by the University of Geneva and Geneva state ethics committees (authorization 1007/3387/2).

Surgery and Recording.

Steel head-posts were cemented to the mice's skulls 2–3 d before recording. Mice were then habituated to the recording apparatus for 2–3 training sessions before recording. We used silicon-based recording electrodes (A-4 × 2-Tet-5 mm-150–200-312; NeuroNexusTech) for extracellular recordings acquired at 32 kHz, and transferred to computers via DigitalLynx (Neuralynx). For calcium imaging, we thinned the skull over the olfactory bulb and placed a 5-mm coverglass over it. Images were acquired at 50 Hz for 10 s using a Micam Ultima system (Brainvision) mounted on a custom-built microscope (65). We presented odors using a custom-made olfactometer, coordinated via computer, as described previously (33, 65). For details of surgery, electrophysiology, calcium imaging, and detailed analysis, see SI Materials and Methods.