|Home | About | Journals | Submit | Contact Us | Français|
Neural activity underlying odor representations in the mammalian olfactory system is strongly patterned by respiratory behavior; these dynamics are central to many models of olfactory information processing. We have previously found that sensory inputs to the olfactory bulb change both their magnitude and temporal structure as a function of sniff frequency. Here, we asked how sniff frequency affects responses of mitral/tufted (MT) cells – the principal olfactory bulb output neuron. We recorded from MT cells in anesthetized rats while reproducing sniffs recorded previously from awake animals and varying sniff frequency. The dynamics of a sniff-evoked response were consistent from sniff to sniff but varied across cells. Compared to the dynamics of receptor neuron activation by the same sniffs, the MT response was shorter and faster, reflecting a temporal sharpening of sensory inputs. Increasing sniff frequency led to moderate attenuation of MT response magnitude and significant changes in the temporal structure of the sniff-evoked MT cell response. Most MT cells responded with a shorter duration and shorter rise-time spike burst as sniff frequency increased, reflecting increased temporal sharpening of inputs by the olfactory bulb. These temporal changes were necessary and sufficient to maintain respiratory modulation in the MT cell population across the range of sniff frequencies expressed during behavior. These results suggest that the input-output relationship in the olfactory bulb varies dynamically as a function of sniff frequency, and that one function of the postsynaptic network is to maintain robust temporal encoding of odor information across different odor sampling strategies.
Active sensing allows an animal to control the interaction between a stimulus and the sensory neurons detecting it. Whisking, licking, and saccadic eye movements are classical examples of active sensing behaviors; in olfaction in terrestrial vertebrates, odor sampling is controlled by the inhalation of air into the nose, which can occur in the course of baseline respiration or during active sensing (i.e., ‘sniffing’). The dependence of sensory neuron activation on respiration has profound consequences for olfactory system function.
First, respiration imposes a temporal structure on the activation patterns of olfactory receptor neurons (ORNs), with ORNs typically firing in 100 – 200 ms bursts of activity following inhalation (Spors et al., 2006; Carey et al., 2009). ORN input dynamics in turn appear to shape the temporal response patterns of mitral/tufted (MT) cells, the main output neurons of the olfactory bulb (Sobel and Tank, 1993). Temporal patterns of inhalation-driven ORN and MT cell spike bursts may by themselves robustly encode information about odorant identity and intensity (Chaput, 1986; Hopfield, 1995; Margrie and Schaefer, 2003; Schaefer et al., 2006; Cury and Uchida, 2010; Junek et al.).
Second, sniffing alters the strength and temporal organization of ORN activation patterns. Sniffing involves precise but complex control of the frequency and magnitude of inhalation during behavior (Welker, 1964; Youngentob et al., 1987; Kepecs et al., 2007; Wesson et al., 2008b). The most prominent form of active sniffing involves repeated inhalation at high frequencies (6 – 10 Hz); in rodents this behavior is expressed during active exploration, investigation of novel stimuli and odor learning (Welker, 1964; Macrides et al., 1982; Wesson et al., 2008a; Wesson et al., 2008b). Sustained high-frequency sniffing alters patterns of ORN activation by attenuating the magnitude of sniff-evoked responses and reducing the temporal coherence between sniffing and ORN spiking (Verhagen et al., 2007; Carey et al., 2009).
Changes in sniffing behavior during active sensing may also alter patterns of postsynaptic activity in the olfactory bulb (OB) and beyond. Indeed, changes in the frequency and temporal structure of ORN input are predicted to lead to qualitative changes in the way that the OB processes odor information during active sampling (Schoppa and Westbrook, 2001; Balu et al., 2004; Hayar et al., 2004a; Wachowiak and Shipley, 2006). However only a few studies have recorded MT cell responses during sniffing in the high-frequency (6 – 10 Hz) range (Bhalla and Bower, 1997; Kay and Laurent, 1999; Bathellier et al., 2008; Cury and Uchida, 2010), and none have provided a detailed characterization of how - or whether - the transition from passive respiration to sustained high-frequency sniffing systematically alters the response properties of OB neurons
Here, we recorded from MT cells in anesthetized rats while using a “sniff playback” device to reproduce sniffing behavior recorded previously from awake animals (Cheung et al., 2009). This approach allowed us to investigate how sniff frequency affects MT cell responses under naturalistic sampling conditions. We found that increasing sniff frequency altered the shape of the inhalation-evoked MT cell response transient and did so in a way that enabled a population of MT cells to maintain strong respiratory patterning across the range of sniff frequencies expressed during behavior.
Female Long Evans rats (9–15 wks old; mean body mass, 243 g) were anesthetized with urethane (1.7 g/kg) or isoflurane (1 – 3%) and kept on a heating pad for the duration of the experiment. The sniff playback technique used here – including sources and part numbers – is described in detail in Cheung et al (2008). Briefly, a tracheotomy was performed and a breathing tube was inserted toward the lungs. A second tube was inserted via the trachea into the nasopharynx and secured with surgical thread and VetBond. The sniff tube was connected through a short length of polyethylene tubing to a glass syringe, the piston of which was coupled to a computer-controlled linear solenoid actuator.
Playback command signals for the solenoid were derived from intranasal pressure signals previously recorded from awake head-fixed rats (Verhagen et al., 2007). We used two different playback traces; the first replicated a 10-sec recording of sniffing in an awake rat that consisted of both low- and high-frequency sniffing (Figure 1A), while the second trace consisted of bouts of sniffs repeated at each of five frequencies from 1 – 5 Hz with a pause of 2 sec in between each bout (Figure 1B). Because the waveform of individual sniffs varies with frequency (high-frequency sniffs tend to be shorter), for each frequency the sniff waveform in a bout was taken from the average of many (>100) sniffs recorded at that frequency. The average sniff for a given frequency range was generated by recording all sniffs from long (30 – 45 min) sessions of sniffing behavior (Carey et al., 2009) and binning each sniff by its preceding inter-sniff interval (ISI). Sniffs in each bin (100 ms wide; bin centers were 1000 ms, 500 ms, 400 ms, 300 ms, and 200 ms, corresponding to sniffing at 1, 2, 2.5, 3.3, and 5 Hz) were averaged and the resulting average sniff for each frequency bin used to synthesize the playback traces. For each playback trace, the sniff waveform used for any given sniff was the average sniff calculated for that sniff’s ISI. This approach allowed us to examine responses to a particular sequence of sniffs and to control for potential effects of changes in sniff waveform. The gain of each of the command traces was adjusted during playback such that the working range of the syringe was ~1.5 mL, an appropriate tidal volume for a ~250-g rat (Walker et al., 1997). The pressure transients delivered to the nose were recorded with a pressure sensor (24PCAFA6G, Honeywell, Morristown, NJ) to monitor the timing and reliable reproduction of the sniff playback signals. The animal’s intrinsic respiration was monitored by placing a photodiode under the chest; each breath modulated light access to the photodiode.
Odorants were presented using a custom flow-dilution olfactometer controlled by a computer (Verhagen et al., 2007). Pure liquid odorants (Sigma-Aldrich) were diluted in mineral oil such that the saturated vapor contained approximately 200 ppm of odorant; this vapor (in pure nitrogen) was further diluted by clean air in a mixing manifold to yield an odor stream containing ~2–10 ppm final odorant concentration. Odorant was presented for the entire duration of the sniff playback trace. Stability of odorant concentration for the duration of the presentation was confirmed using a portable photoionization detector (miniRAE2000, RAE Systems).
ORNs were loaded with calcium-sensitive dye and optical signals imaged from the dorsal OB as described previously (Wachowiak and Cohen, 2001; Verhagen et al., 2007)(Carey et al., 2009). Briefly, rats were anesthetized with ketamine/xylazine and ORNs in both nares loaded with calcium-sensitive dye (Oregon Green BAPTA-1 dextran, 10 kd m.w., Invitrogen, Carlsbad, CA). Imaging was performed through an optical window (~3 mm ant-post × 1.5 mm med-lat) made by thinning the bone over one OB. Optical signals were collected using a 4× 0.28 NA objective, an Olympus epifluorescence illumination turret (BX51), a 150W Xenon arc lamp (Opti-Quip) and appropriate filter sets (Verhagen et al., 2007). Images were acquired using a 256 × 256 pixel CCD and digitized at 25 Hz along with the sniff playback signals using an integrated hardware/software package (NeuroCCD SM-256 and NeuroPlex, RedShirtImaging LLC).
Analysis of imaging data was performed largely as described previously (Wachowiak and Cohen, 2001; Verhagen et al., 2007; Carey et al., 2009). Regions of interest representing distinct glomeruli were selected manually based on the discrete appearance of signal foci of appropriate size (75 – 150 µm). For each glomerulus, we automatically detected responses to each sniff using a custom curve-fitting algorithm implemented in Matlab (Carey et al., 2009). All glomerulus-odor pairs used for analysis were tested at subsaturating odorant concentrations (i.e., we confirmed that a higher concentration elicited stronger response amplitudes for each glomerulus). Raw calcium signals are shown for display in Figure 1.
For analysis of ORN input dynamics, we estimated temporal patterns of action potential firing across the ORN inputs to a glomerulus using temporal deconvolution as described by Yaksi and Friedrich (2006) and Verhagen et al. (2007). The average ORN response generated by a sniff was calculated from the deconvolved calcium signal by averaging across all sniffs repeated at a given frequency, aligned to the time of inhalation onset (Figure 1G); each ‘sniff-triggered’ average response was then fit with a double sigmoid curve to measure temporal dynamics parameters (Carey et al., 2009); see also below. Curve fitting was only performed for responses that were strongly modulated by sniffing (e.g., those with a ‘sniff modulation index’, defined below, of at least 0.5).
Coherence between ORN inputs and sniffing (Figure 1F) was calculated from the deconvolved ORN signal as in (Carey et al., 2009): first, to ensure the measurement was limited to the effect of sniff frequency (rather than sniff shape or amplitude), we replaced each sniff with an averaged sniff waveform. The coherence between this modified sniff waveform and the deconvolved ORN input waveform was then calculated for each glomerulus for the duration of each bout between the second and the last sniffs.
Extracellular recordings from OB units were performed in separate animals from the imaging experiments. Following the double-tracheotomy surgery (described above), a small (~1 × 1 mm) craniotomy was performed over one OB and the dura was removed. Unit recordings were obtained using a single-channel tungsten microelectrode (2.0 MΩ at 1 kHz; MicroProbes WE3PT12.0F3), amplified with a BMA-931 amplifier and Super-Z headstage (CWE), bandpass filtered from 1 Hz - 10 kHz and digitized at 20 kHz using a Micro1401 MkII DAQ board and Spike2 software (CED, Cambridge, UK). Additional highpass filtering was done in Spike2 at a cutoff frequency of 200 Hz. Custom scripts in Spike2 were also used to control odorant presentation.
Recordings from presumptive mitral/tufted (MT) cells were obtained by slowly lowering the recording electrode to the appropriate depth for the mitral cell layer of the dorsal or ventral OB; electrode depth was monitored with a piezoelectric micromanipulator with digital readout (Sutter MP-225). The location of the mitral cell layer was determined using a stereotaxic atlas and, in some experiments, confirmed by locating the inversion point of LFPs evoked by LOT stimulation. During this phase, sniff playback was used to generate ongoing ‘resting’ sniffs at 1 Hz, in the absence of odorant. We selected only units that appeared well-isolated, in the vicinity of the mitral cell layer and which showed clear spiking activity in the absence of odorant. Once a target unit was chosen, odorants effective at eliciting excitatory responses at low to moderate concentrations were identified using a rapid screening procedure (Davison and Katz, 2007). Premade mixtures of 6–7 odorants each were presented successively, after which the components from the most effective mixture were tested. The lowest effective concentration of the most effective odorant was then used for the duration of the recording for that target cell (typical concentration was 3–5 ppm; concentrations used ranged from 1–10 ppm). Odorants (and number of cells) used in the final dataset were: 2-butanone (1), 2-heptanone (3), 2-hexanone (2), 2-methyl-2-butenal (5), 2-methyl butyraldehyde (1), 2-octanone (3), acetophenone (1), benzaldehyde (5), butanal (1), heptaldehyde (4), heptanal (2), hexanal (2), isoamyl acetate (1), isovaleraldehyde (2), octanal (2), and valeraldehyde (2). The typical recording sequence for a target cell consisted of at least 15 trials, during each of which the odorant was presented continuously during the sniff playback command trace. In some cells, we varied this basic experiment by interleaving trials using different sniff playback command traces or with different odorant concentrations. In all cases, the inter-trial interval was at least 60 s.
Though we limited targeted cells to those that were well-isolated during baseline sniffing, odorant presentation tended to elicit excitatory responses in additional units, necessitating offline spike sorting. Spike sorting for each recording was performed in Spike2 and was based on spike shape and amplitude and confirmed by the distribution of inter-spike intervals. Only cells with spikes that were clearly separable by this criteria were included in the dataset. After spike sorting, all data other than spike times were discarded; subsequent analyses were performed using MATLAB (Mathworks). For each unit, we constructed sniff-triggered peri-sniff time histograms (PSTHs) for sniff-evoked responses at each frequency of sniffing by compiling spike counts in 10-ms time-bins for all sniffs repeated at a given frequency (see blue trace in Figure 2B). Thus, all PSTHs reflect average activity evoked by many (typically > 50) sniffs. For most analyses (except where stated) we used only the third and subsequent sniffs in each bout, as the first sniff had no definable frequency and the second was often difficult to distinguish from the first response, especially at the higher sniff frequencies. Spike counts were divided by the number of sniffs and the bin width and expressed as spikes/sec in all figures and subsequent analyses.
The dynamics of sniff-evoked responses were measured by fitting double-sigmoid curves to the PSTHs (for this case generated with 5-ms bins) for each unit at each sniff frequency (Fig. 2C; see also (Carey et al., 2009)). The curve-fitting was used to measure peak firing rate, rise-time (time from 10% to 90% of peak firing rate), and duration (response width at 50% of peak firing rate). For summary data involving durations and rise-times, units with multiple peaks (N = 6) were omitted. For Fig. 5, peak firing rate and total spike count were normalized for each MT cell by dividing by the value of that parameter for the 1-Hz response. We defined response onset time as the time from each sniff to the first spike for which the instantaneous firing rate (measured from each spike to the next) exceeded the spontaneous firing rate by one standard deviation. This spontaneous firing rate was calculated by averaging the firing rate in the first 200 ms after each sniff without odorant presentation. Response latencies reported are the mean onset time across all sniffs. The s.d. of these onset times reflected the precision of spike timing. Response dynamics and response variance were also measured from single trials by calculating the instantaneous firing rate at each spike time and calculating the mean and standard deviation of the resultant waveform across sniffs. As expected, the mean firing rate (shown in black in Fig. 2B) approximated a smoothed version of the PSTH.
To test whether sniff frequency alters MT response patterns, we compared the PSTHs for sniffs at different frequencies within the same MT cell using a 2-sample Kolmogorov-Smirnov test to compare the distributions of sniff-triggered spike times between 1- and 5-Hz sniff responses. For comparing response dynamics across frequency (Figures 8B, C), response latencies were normalized for each cell by subtracting the latency of the 1-Hz response and durations were normalized by dividing by the duration of the 1-Hz response. Latencies with respect to sniff phase were calculated by dividing the latencies by the inter-sniff interval and then subtracting the 1-Hz response latency. Durations with respect to sniff phase were calculated by dividing each duration by the inter-sniff interval.
A ‘sniff modulation index’ (SMI) was used to measure the degree to which MT cell activity was modulated by the sniff cycle, relative to peak firing rate. To calculate this index, first the PSTH (bin size, 10 ms) was slightly smoothed by convolving it with a Gaussian (s.d., 1 bin). The SMI was then defined as (PSTHmax - PSTHmin) / PSTHmax, where PSTHmax and PSTHmin equal the peak and subsequent trough, respectively, of the smoothed PSTH during a bout of repeated sniffs. Thus, SMI = 1 indicates that, for a given MT cell and sniff frequency, the PSTH minimum was 0 (i.e., there was no MT cell activity in at least one 10-ms bin per sniff cycle), and SMI = 0 indicates that the PSTH was completely flat over the entire cycle from one sniff to the next (i.e., there was no modulation by sniffing). Because the PSTH is the average response across many (> 50) sniffs, chance variability among bins is expected to be smoothed by averaging, and the contribution of this variability to the SMI low. However, to estimate the chance SMI level, a shuffled PSTH was constructed from spike rasters that were randomly shuffled in time and the SMI was calculated for this shuffled PSTH. As expected, the shuffled SMI was low: 0.24 ± 0.10 for 5 Hz sniffing; chance SMI for individual MT cells is indicted in the Figures.
Population spike histograms for the entire recorded MT cell population (Fig. 8A,B) were generated by normalizing (to the peak) the PSTHs for each MT cell and then averaging these waveforms. Synthetic response histograms were constructed for each cell using as a source the PSTH evoked during either 1 Hz sniffing or 5 Hz sniffing. Source PSTHs were normalized to their own peaks. The histogram (bin size, 5 ms) was initialized to zero, then synthesized by processing each sniff time in sequence: for each sniff, the source PSTH was attenuated based on the current value of the histogram bin in which that sniff occurred. The attenuation factor was 0 (no change) if the current value was 0.2 or less; the attenuation factor scaled linearly from 0 to 1 for current values from 0.2 to 1.2. Thus, synthesized responses saturated at 120% of peak firing rate of the source PSTH. This nonlinear synthesis algorithm was preferable to simple convolution of the source PSTH with sniff times as it avoided unrestricted summation of responses at higher sniff frequencies. SMI values for synthetic response histograms were calculated for individual cells in the same way as for experimental data.
To systematically evaluate how variations in sampling behavior – in particular, sniff frequency – shapes MT responses, we generated naturalistic inhalations (i.e., sniffs) resembling those expressed by awake animals. To achieve this we used a ‘sniff playback’ device consisting of a solenoid-driven syringe under the control of an analog command waveform. The device has been described previously (Cheung et al., 2009); however, for the purposes of this study the key features of sniff playback are a) the ability to reproduce the pressure transients characteristic of respiration or sniffing in the awake animal; b) the ability to generate resultant airflow patterns mimicking those seen in the awake animal; and c) the ability to reproduce these waveforms with high precision both within and across different animals to generate naturalistic sniff waveforms in the anesthetized animal. These features are illustrated in Figures 1A and 1B. Sniff playback allowed us to systematically explore sampling parameters and to compile data across trials and across animals without the confound of differences in odor sampling or changes in behavioral state.
Sniff playback command traces were derived from intranasal pressure measurements taken during low- or high-frequency sniffing in an awake, head-fixed rat (see Methods and (Verhagen et al., 2007)).We used two different sniff playback traces. The first was derived from a 10-sec epoch recorded from an awake, head-fixed rat which included resting respiration, a bout of high-frequency sniffing, followed by a return to resting respiration (Figure 1A). The second trace consisted of a series of sniff bouts at different frequencies.in which the same sniff was repeated 6 – 9 times at each of five frequencies from 1 – 5 Hz, with a pause of 2 sec in between each bout (Figure 1B). This first trace allowed us to record responses to a behaviorally-expressed, 'natural' sequence of sniffs generated in the behaving animal, while the second trace allowed us to more systematically assess how responses vary across frequency. The interval between each sniff bout in the second trace was chosen to allow evoked responses to return to baseline; the first sniff in each bout thus had no frequency and so could serve as a control for whether sniff waveform affected responses.
We first recorded responses of ORNs to the sniff playback traces using presynaptic calcium imaging, an approach we have used previously in awake rats (Verhagen et al., 2007; Carey et al., 2009). The purpose of these experiments was to confirm that our earlier observations in awake rats could be repeated using sniff playback in anesthetized rats. As expected, each sniff elicited a brief calcium signal transient reflecting a short (100 – 200 ms) burst of action potentials in the population of ORNs converging onto a glomerulus (Figures 1C, D). We compared key temporal parameters of these sniff playback-driven calcium responses (latency from sniff, rise-time, and duration) with those elicited in awake rats (Carey et al., 2009), and found that these parameters closely matched. Values for the sniff playback experiments were: latency, 142 ± 58 ms; rise-time, 111 ± 21 ms; duration, 433 ± 76 ms duration (N = 33 glomeruli), compared to the following from awake rats: latency, 154 ± 59 ms; rise-time, 91 ± 40 ms; duration, 392 ± 149 ms (Carey et al., 2009). At low sniff frequencies (1 – 2 Hz), responses repeated with little decrement in amplitude for each sniff in the bout. At higher frequencies (4 – 5 Hz), the initial response (e.g., to the first sniff) was the same, but responses to subsequent sniffs were attenuated in amplitude (Figures 1C – E). We also observed a loss of coherence between phasic ORN inputs and sniff timing as sniff frequency increased, although the degree of coherence loss differed for different odor-glomerulus pairs (Figures 1D, F). Aside from the decrease in response amplitude and coherence, the temporal structure of the sniff-evoked ORN response was relatively invariant to changes in sniff frequency (Figure 1G). These results recapitulate those seen in awake, behaving rats (Verhagen et al., 2007; Carey et al., 2009). Thus, the sniff playback approach is capable of reproducing the key features of ORN input responses across a range of sniff frequencies.
Next we recorded from individual OB units during sniff playback. We focused on well-isolated units that we presumed to be MT cells based on their location and additional criteria described in the Methods. We also selectively focused on units that showed excitatory responses to relatively low (< 10 ppm) concentrations of single odorants. MT cells showing purely inhibitory responses were observed but not studied further here. In total we recorded responses from 37 cells in 20 animals, under either urethane (N = 17 cells) or isoflurane (N = 20 cells) anesthesia. Most basic response properties did not differ between the two anesthetic regimens, except where reported below. Spontaneous spiking activity was lower and more stable using isoflurane (mean; 16 ± 24 Hz spontaneous firing rate for N = 17 cells recorded under urethane anesthesia; 14 ± 15 Hz for N = 20 cells recorded under isoflurane; see Methods).
In most MT cells, sniff playback for an excitatory odorant elicited a brief burst of action potentials (Figure 2A). At low (1 Hz) sniff frequencies, this burst was repeated reliably with each successive sniff and was independent of the intrinsic respiratory rhythm. Some (but not all) MT cells showed temporal patterning of spiking in the absence of odorant; in these neurons patterning was also linked to sniff timing and not the respiratory rhythm (not shown). We constructed peri-sniff time histograms (PSTHs) of the response of each MT cell to a single sniff (Figure 2B; see also Methods). In general, all cells showed a transient excitation following inhalation (we explicitly did not include cells that failed to show any excitation); however there was considerable diversity in the specific temporal pattern of this response across cells (Figure 2B). Sniff-triggered response patterns differed in latency to excitation, rise-time of the excitatory response, and duration of excitation; in addition, some cells showed a bimodal excitation pattern with a second peak in spike frequency after the initial response transient (e.g., Figure 2B cells 3 – 5). Despite variance from cell to cell, for a given cell the basic response pattern was repeated with little variance for multiple sniffs of a given odorant.
We quantified the following key parameters of the MT cell response: peak firing rate, rise-time and duration of the average sniff-driven response (Figure 2C), which were measured from the PSTHs, as well as mean response latency after inhalation, which was measured from individual sniffs. Details for how each parameter was quantified are given in the Methods. Peak firing rates across all recorded cells varied from 38.8 to 200 Hz (mean, 115 ± 35 Hz for N = 17 cells recorded under urethane anesthesia; 103 ± 43 for N = 20 cells under isoflurane). In a subset of cells (n = 7) we presented the same odorant at multiple concentrations (2 – 3 concentrations per cell, ranging from 2-fold to 50-fold change in concentration) (Figure 2D). In these, peak firing rates increased with concentration (Figure 2D; p < 0.05, paired t-test comparing lowest and highest concentrations, N = 7 cells). Response latencies decreased slightly in most cells (p < 0.05, paired t-test), while response rise-times showed no significant change (p = 0.16, paired t-test). These effects of odorant intensity are similar to those observed for ORN inputs in awake rats (Carey et al., 2009).
The distributions of the temporal parameters of all MT cell responses are shown in Figure 3. Because we used the same double-sigmoid fitting method as used to characterize the temporal patterning of sniff-driven ORN inputs (Carey et al., 2009), we could compare MT cell response dynamics with those of ORN inputs imaged using playback of the same sniff waveforms. In general, MT response latencies (Figure 3A) were similar to those of imaged ORN inputs (median MT versus ORN latency, 75 ms v. 88 ms), while MT response rise-times (Figure 3B) and durations (Figure 3C) were shorter (rise-times: 28 ms v. 60 ms; durations: 80 ms v. 154 ms). Notably, the distribution of PSTH durations was bimodal, with 19 cells showing durations of < 100 ms (mean, 66 ± 15 ms) and 12 cells showing durations of > 100 ms (mean, 175 ± 44 ms). Short- and longer-duration MT cells also had distinct rise-times (30 ± 18 ms for short- and 61 ± 48 ms for longer-duration; p < 0.05, t-test), and there was a high correlation between rise-time and duration (r = 0.63, p < 0.001; Figure 3D). There was no significant correlation between latency and duration (r = 0.05, p = 0.78), consistent with the idea that ORN input dynamics primarily determine MT response latency but not later aspects of the response.
Recent behavioral and electrophysiological studies have pointed to the importance of spike timing – in particular first spikes – to odor representations and odor perception (Margrie and Schaefer, 2003; Schaefer et al., 2006; Wesson et al., 2008a; Cury and Uchida, 2010; Junek et al., 2010) . We thus measured the temporal precision with which a sniff could repeatedly elicit spiking responses by measuring response onset time and its variance (s.d.) across repetitions of the same sniff waveform (Figure 4A; see Methods). Lower variance values reflect increased sniff-coupled precision of onset timing. Median precision was 15.6 ms, and ranged from 2 – 44 ms (Fig 4B; N = 35 cells). It is possible that a MT cell’s baseline firing rate or variance could affect onset precision. However, we did not find a significant correlation between onset precision and either the mean or s.d. of the baseline firing rate (r = 0.31 to baseline mean, Figure 4C, and r = 0.28 to baseline s.d.). In addition, precision was not correlated with response amplitude (r = −0.08) or latency (r = 0.17). Onset timing did, however, become more precise with increases in odorant concentration (p < 0.05, paired t-test comparing lowest and highest concentrations, N = 7 cells). Thus, MT cells show different degrees of precision of spike timing across sniffs, although in general this precision is relatively high.
We next asked whether and how changes in sniff frequency alter MT responses . We analyzed MT response dynamics in the same way as described above, but now for sniffs generated at frequencies ranging from 1 – 5 Hz. Figure 5 shows examples from two MT cells (Figure 5A); these cells show different apparent effects of increasing sniff frequency. One cell responds with a transient burst following each inhalation at all frequencies (Figure 5A, bottom). In contrast, a different cell responds reliably at sniffs from 1 – 3 Hz but begins to show attenuation in peak firing rate during sustained sniffing at 4 and 5 Hz (Figure 5A, top). This cell also fails to return to baseline firing rate after each response, resulting in less modulation of firing by the sniff cycle. For both cells, the response to the first sniff of each bout is the same, indicating that these effects are frequency-dependent and not due to the slightly different sniff waveforms at each frequency. Figure 5B shows an example of the response of the lower cell from Figure 5A to the natural sniff sequence; this cell responds reliably to each inhalation, even during the high-frequency sniff bout, which reaches sniff frequencies of ~6 Hz.
To test whether sniff frequency alters MT cell responses to the same odor, we constructed PSTHs for the average sniff-evoked response at each frequency (excluding the first sniff). PSTHs for two additional example cells are shown in Figures 5C and D. We compared PSTH shape for different frequency responses (see Methods). Every MT cell except one showed a significant change in spike distributions between and 1 and 5 Hz sniffing (36 of 37 cells significant at p < 0.05, Kolmogorov-Smirnov test). Thus, sniff frequency clearly alters MT cell response patterns.
Because ORN inputs show strong attenuation of response magnitude as sniff frequency increases (Verhagen et al., 2007; Carey et al., 2009), we predicted that MT cell responses would, on average, show a similar effect. Consistent with this prediction, we found that the peak firing rate (measured from the PSTH) decreased significantly (p < 0.001, paired t-test, N = 37 cells) at sniff frequencies of 3 – 5 Hz (Figure 5E); however the degree of the attenuation was lower than that of ORN inputs (e.g., ~25% vs. ~60% for 5 Hz sniffing). Total spike count within the 200 ms following inhalation also decreased (Figure 5F; p < 0.001, paired t-test, N = 37 cells). Responses to the first sniff in each bout showed no change across frequencies (p = 0.77 for effect of sniff frequency on peak firing rate, p = 0.75 for effect of sniff frequency on spike count; ANOVA, N = 37 cells). Increasing odorant concentration reliably increased peak firing rates, but there was no systematic effect on the degree of attenuation.
We also predicted that MT cells – like ORN inputs (Carey et al., 2009) – would show less modulation in firing rate by the sniff cycle at higher sniff frequencies. Across all cells, our index of firing rate modulation by sniffing (sniff modulation index, SMI; see Methods for definition) decreased significantly (Figure 5G; paired t-test between 1 and 5 Hz, p < 0.001, N = 37 cells) but only slightly, and less so than for ORN input signals. However, as is evident from the examples in Figure 5A, the ability of MT cells to maintain temporal patterning at high frequencies varied across the population. Most MT cells maintained high modulation of firing rate with the sniff cycle, with 26 of 37 cells showing almost no loss of modulation at 5 Hz (SMI > 0.90). However, the remaining cells did show a reduction in modulation, albeit a modest one (Figure 5H). There was no relationship between recording depth and SMI (r = −0.11), suggesting no systematic difference in sniff modulation between dorsal, lateral or ventral MT cells, nor was there a relationship between SMI and peak sniff-evoked firing rate (r = −0.0011).
In summary we found that, while the reduction in input magnitude during sustained high-frequency sniffing appears – on average – to translate into reduced MT response magnitudes, MT responses tend to maintain temporal patterning even in the face of sharply reduced temporal patterning of ORN inputs.
What parameters of MT responses change with sniff frequency? We first tested the effect of sniff frequency on the timing of MT response onsets. Previous studies have proposed that odor information may be robustly represented by the relative phase of MT cell activity within the sniff cycle (Chaput, 1986; Hopfield, 1995; Roux et al., 2006; Fantana et al., 2008); other studies have supported an alternative hypothesis of a temporal code based on absolute response latencies (Margrie and Schaefer, 2003); (Spors et al., 2006; Junek et al.). To distinguish between these two possibilities, we analyzed MT response latency and duration across frequencies, both with respect to phase of the sniff cycle and to linear time relative to inhalation onset. Figure 6A shows PSTHs for an example MT cell at each frequency, plotted in linear time, with only slight changes in overall shape. Across all cells (n = 31), response latencies increased slightly (by ~ 20 ms) with frequency (p < 0.001, ANOVA; Figure 6B), while response durations decreased slightly (p < 0.05; Figure 6C). In contrast, when plotted with respect to sniff phase, PSTHs changed dramatically with frequency (Figure 6D), as did response latencies and durations (Figures 6E, F). Thus, latency-based odor representations are relatively invariant (and thus more robust) to sniff frequency, while phase-based representations are highly dependent on frequency.
The duration and rise-time of the inhalation-driven MT response are likely critical for determining the ability of MT cells to follow sniffs at high frequency, as responses must reach a peak and return to baseline in the time between successive sniffs. We had earlier observed a bimodal distribution of response durations for MT responses at 1 Hz sniffing(e.g., Figure 3); we thus asked whether MT cells showing short- (<100 ms) and long- (>100 ms) duration responses showed differential effects of sniff frequency on response patterns. The examples in Figure 7A suggest that this is the case: the MT cell with a short-duration burst at 1 Hz shows relatively little change in PSTH shape as sniff frequency increases while the cell with a longer-duration burst shows a marked change, with an overall shortening of response duration (Figure 7A). Consistent with these examples, we found that MT cells with longer-duration responses at 1 Hz sniffing showed a significant decrease in response duration (Figures 7B; p < 0.001, one-way ANOVA across frequency) and in response rise-time (Figure 7C; p < 0.05, ANOVA) at higher sniff frequencies. Mean response duration changed from 152 ms (1 Hz) to 83 ms (5 Hz), while mean rise-time changed from 61 ms (1 Hz) to 36 ms (5 Hz). In contrast, MT cells with shorter-duration responses at 1 Hz showed no change in duration or rise-time with frequency (p > 0.2 for both, ANOVA). Consistent with this analysis, we also found a high concordance between those cells maintaining the highest modulation at 5 Hz and those showing short-duration responses at 1 Hz (17 of 19 short-duration cells showed low (SMI > 0.90) loss of modulation).
We also tested whether MT response latencies or onset precision changed with sniff frequency, separately analyzing MT cells with short- and longer-duration responses at 1 Hz (Figure 7D, E). There was a slight but significant increase in response latency with sniff frequency for both groups of MT cells (80 ms for 1 Hz to 101 ms for 5 Hz, Figure 7D; p < 0.05, ANOVA). We found no effect of sniff frequency on the precision of response onsets, however (Figure 7E).
To confirm that the effects of sniff frequency generalize to natural sniffing patterns, in a subset of cells we presented odorant while performing sniff playback using the naturalistic sniff trace shown in Figure 1A. For analysis of frequency effects using this trace, we grouped the sniffs in the trace into one group containing all low-frequency sniffs (5 sniffs per trace, mean sniff frequency, 1.6 Hz) and the second containing high-frequency sniffs (14 sniffs per trace, mean, 5.8 Hz). Between these two groups, peak firing rates decreased from a mean of 103 ± 39 Hz during low-frequency sniffing to 72 ± 28 Hz during high-frequency sniffing (N = 15 cells). In addition, there was a clear separation of frequency effects as a function of MT response duration: response durations did not differ for the shorter-duration MT cells (mean, 62 ± 13 ms for low-frequency sniffing versus 67 ± 18 ms for high-frequency sniffing; N = 11 cells), but decreased for longer-duration MT cells (mean, 137 ± 17 ms during low-frequency sniffing versus 80 ± 31 ms for high-frequency sniffing; N = 4 cells).
In summary, we found using both the synthetic and naturalistic sniff playback traces that higher sniff frequencies lead to a reduction in the duration of the inhalation-evoked spike burst as well as a reduction in the rise-time of the burst. Together, these effects lead to a shortening of the inhalation-driven burst, and they occur only for those MT cells that show longer-duration responses at low sniff frequency.
Given the diversity in response dynamics and the change in these dynamics with sniff frequency, we asked how these responses would appear across a population of MT cells. To visualize this we generated population spike histograms from all units, independent of odorant, using either the sniff bouts trace or the ‘natural’ sniffing trace (see Methods), with the rationale that this histogram would approximate the coherent temporal pattern of activity across a MT cell population during odorant sampling (Figure 8A). Despite the fact that different MT cells have a range of excitation onset latencies, rise-times and durations, the population activity pattern followed sniffs with high reliability (Figure 8A), showing sharp peaks in activity following each inhalation even during 5 Hz sniffing and during the high-frequency sniff bouts expressed in the awake rat (Figure 8C).
Do the changes in PSTH shape contribute to the temporal patterning of MT cells as sniff frequency increases? To test this, we constructed synthetic response histograms using, for each unit, the mean PSTH evoked during 1 Hz sniffing and convolving this response (with a small activity-dependent attenuation factor) with the sniff times in the higher-frequency bouts (see Methods). At high sniff frequencies the synthetic responses showed significantly less modulation by each sniff than the actual responses (mean SMI at 5 Hz sniffing, 0.80 ± 0.18 simulated responses vs 0.93 ± 0.10 actual responses, p < 0.001, paired t-test; N = 37 cells). The synthetic population histogram also showed less modulation than the actual population histogram (Figures 8B, C). Thus, the frequency-dependent change in shape of the inhalation-evoked PSTH has the effect of increasing the ability of a population of MT cells to maintain coherent temporal patterning in spike timing as sniff frequency increases.
Because awake, behaving rats sample odorants at frequencies up to 10 Hz (Kepecs et al., 2007; Wesson et al., 2008b; Wesson et al., 2009), we used the same approach to extrapolate population-level response dynamics at frequencies higher than the range that we could test experimentally using sniff playback. We extended the synthetic response histograms to 10 Hz sniffing (convolving each PSTH with sniff times repeated at 10 Hz) using either 1 Hz or 5 Hz PSTHs from the same cells (Figure 8D). For each, we measured the SMI at frequencies up to 10 Hz (Figure 8E). Population-level response patterning was almost entirely lost at 10 Hz if the 1-Hz PSTH responses were used, but patterning remained strong using the 5-Hz PSTH responses. These effects were clear for individual MT cells as well: nearly all cells showed greater modulation for 10-Hz sniffs when the synthetic responses were based on the 5-Hz PSTH than when based on the 1 Hz PSTH (Figure 8F). These analyses indicate that MT response patterns evoked by sniffing at 5 Hz are better able to follow sniffs at these and higher frequencies than are response patterns evoked by sniffing at 1 Hz.
In earlier studies characterizing the temporal structure of sensory input to the OB, we found that ORN inputs are tightly coupled to inhalation, have intrinsic differences in their dynamics and are qualitatively altered by changes in sniff frequency (Spors et al., 2006; Verhagen et al., 2007; Carey et al., 2009). These findings led to predictions about the relationship between sensory input dynamics and postsynaptic activity, and about how sniff frequency might affect MT responses. Here we tested those predictions by recording MT responses while manipulating sniffing in anesthetized rats, using sniff waveforms taken from awake rats. Several of our observations support earlier predictions: we found that inhalation-driven MT responses – like ORN inputs –show intrinsic differences in their response latency, rise-time, and duration that vary over a similar range, and that high-frequency sniffing attenuates the magnitude of the MT response. Other predictions – in particular, that temporal patterning of MT responses would diminish at high frequencies – were not supported. Instead most MT cells maintained strong temporal patterning of responses at all frequencies tested; this was accompanied by changes in the shape of sniff-evoked responses as sniff frequency increased.
High-frequency sniffing is a hallmark feature of active odor sampling, yet only a few studies have attempted to characterize how changes in sniff frequency affect olfactory processing. Recordings from awake, freely-moving rats have yielded mixed results on the effect of sniff frequency, with some reports that temporal patterning of OB neurons is diminished at high frequencies (Bhalla and Bower, 1997; Kay and Laurent, 1999) and one other reporting that patterning persists (Cury and Uchida, 2010). Resolving these differences is difficult because frequency effects were not analyzed systematically and because of the potential confound of behavioral state-dependent modulation of responses (Bhalla and Bower, 1997; Kay and Laurent, 1999). A previous study using artificial sniffing in anesthetized rats found that most MT cells maintain respiratory patterning at moderate (3 Hz) and high (6 Hz) sniff frequencies, with a reduction in response magnitude (Bathellier et al., 2008).
Our results extend these earlier reports by establishing that inhalation-driven responses change significantly with sniff frequency and that these changes are driven by a ‘bottom-up’ pathway related to active sampling rather than ‘top-down’, neuromodulatory pathways which can also shape MT spike timing in awake animals (Doucette et al., 2011). For many cells – in particular those that maintained the strongest temporal patterning at high sniff frequencies – the duration and the rise-time of the inhalation-driven response shortened, leading to a temporal sharpening of activity. Thus, odorant-evoked response patterns of MT cells change with sampling behavior. This conclusion contrasts with a recent characterization of MT responses in behaving rats, which reported that firing patterns were frequency-invariant (Cury and Uchida, 2010). That study, however, focused only on the first in a bout of high-frequency sniffs. Here, we found that the first sniff in any bout elicits equivalent responses regardless of the frequency of subsequent sniffs; we have previously reported the same result for ORN inputs (Verhagen et al., 2007). Responses to later sniffs, however, do depend on recent sniff history.
One parameter of MT response dynamics that remained invariant to sniff frequency was onset latency relative to inhalation. This invariance is consistent with MT response timing depending on ORN inputs, which also show frequency-invariant latencies (Carey et al., 2009). The fixed timing of MT responses relative to inhalation constrains models of temporal coding that rely on the phase of MT activity relative to the respiratory cycle (Chaput, 1986; Hopfield, 1995). These models have demonstrated robust phase-based coding of odor information by MT response onsets but have typically only considered cases of regular respiration rates. In behaving rodents, sniffs are rarely repeated at regular intervals and sniff frequency can vary dramatically from one sniff to the next (Welker, 1964; Kepecs et al., 2007; Wesson et al., 2008b). One proposal for adapting phase-based coding schemes to the diverse and irregular sniffing patterns has been to translate absolute timing into phase-based timing using the period between two successive inhalations (Roux et al., 2006; Fantana et al., 2008). We found that translating MT responses onto a phase timebase eliminated the frequency-invariance in response latencies as well as response durations. These results – along with others arising from awake, behaving animals – suggest that information coding by MT responses appears to be organized in terms of absolute timing relative to inhalation rather than relative to a respiratory ‘cycle’ (Margrie and Schaefer, 2003; Cury and Uchida, 2010; Junek et al. 2010).
At the population level, the net effect of frequency-dependent changes in MT responses was to enhance respiratory patterning. Responses evoked by 1-Hz sniffing were significantly less modulated when repeated at higher frequency than responses during 5-Hz sniffing. Furthermore, 5-Hz sniff-evoked responses were able to maintain significant patterning even when repeated at 10 Hz (e.g., Figure 8). Thus, the frequency-dependent temporal sharpening of MT spike bursts enables coherent population-level spiking dynamics to remain linked to each inhalation across the range of sniff frequencies expressed during behavior.
What changes in the inhalation-evoked MT response underlie their ability to maintain respiratory patterning across frequencies? The two parameters that changed most consistently were duration and rise-time: MT responses reached a peak more quickly and shut down more quickly at high sniff frequencies. MT response rise-times and durations were strongly correlated, with faster rise-times correlating with shorter durations. Several pieces of evidence suggest that these frequency-dependent changes are mediated by OB postsynaptic circuitry rather than frequency-dependent changes in sniff waveform or presynaptic activity patterns. First, MT responses to the first sniff in a bout did not change shape with frequency, despite the fact that sniff waveform did change slightly. Second, ORN inputs did not show substantial changes in shape as frequency increased (e.g., Figure 1G), and presynaptic inhibition of ORN inputs does not change with sniff frequency (Pírez and Wachowiak, 2008). Third, frequency-dependent effects on duration and rise-time were apparent only for a subset of MT cells – in particular, cells with relatively long (> 100 ms) response durations at low sniff frequencies.
One candidate circuit mediating these effects involves connections between external tufted (ET) cells, inhibitory periglomerular interneurons and MT cells. ET cells are excitatory interneurons receiving strong ORN input and directly exciting periglomerular interneurons within the same glomerulus (Hayar et al., 2004b). ET cells show entrainment of spike bursts to rhythmic ORN inputs (Hayar et al., 2004a). Thus, an increase in ET cell burst frequency as sniff frequency increases is predicted to cause stronger and more synchronous activation of periglomerular interneurons, leading in turn to a shortened MT cell response as sniff frequency increases and enhanced respiratory modulation at higher sniff frequencies (Wachowiak and Shipley, 2006). The results observed here support these predictions. Alternatively, temporal sharpening might be mediated by recurrent inhibition between MT and granule cells. Recurrent inhibition can enhance the temporal precision of MT cell firing in the face of rhythmic ORN inputs (Balu et al., 2004; Schoppa, 2006); (David et al., 2007), and there is evidence that recurrent inhibition is stronger at higher sniff frequencies (Young and Wilson, 1999). In either case, ORN inputs that are more synchronous would lead to faster activation of MT cells and also to stronger recruitment of the feedforward or recurrent inhibitory circuits that shorten the MT cell response, consistent with the correlation we measured between response rise-time, duration, and the degree of frequency-dependent temporal patterning (e.g., Figures 3 and and5).5). One set of predictions that follow from this model is that the rise-time of ORN inputs to a glomerulus should correlate with that of MT cells innervating the same glomerulus, and that glomeruli receiving faster ORN inputs should show stronger modulation of MT cell responses at high sniff frequencies. These predictions need to be tested by comparing ORN inputs to and MT cell responses for the same glomerulus.
Odorant sampling in behaving animals is a dynamic but precisely controlled behavior that presumably optimizes the detection and processing of sensory information. How might frequency-dependent attenuation and temporal sharpening of MT spike bursts optimize odor information processing? First, maintaining (and even enhancing) the coherence of MT spiking with sniffing has been shown to enhance the precision of MT spike timing, thus increasing the potential discriminatory power of MT spike trains (Hopfield, 1995; Schaefer et al., 2006). Second, temporally-patterned MT spiking appears to be important for information processing downstream of MT cells - for example, the temporal integration of inputs from multiple MT cells by pyramidal neurons in piriform cortex (Luna and Schoppa, 2008; Stokes and Isaacson, 2010). Third, frequency-dependent changes in MT response patterns may leave odor identity coding unchanged but enhance the ability of the population to process other aspects of odor information relevant to active sensing. For example, the reduction in MT response magnitude may reduce the salience of odorants that are constant across a high-frequency sniff bout while maintaining sensitivity to newly-encountered odorants (Verhagen et al., 2007). Finally, high-frequency, temporally-patterned MT bursts may be important in driving synaptic plasticity within the OB or its targets during odor learning. Further electrophysiological experiments combined with behavioral analyses will be important to test these possibilities.
The authors thank Ian Davison, Tristan Cenier, and Yusuke Tsuno for assistance with the extracellular recording protocol, and Don Katz and Markus Rothermel for discussion and comments on the manuscript. This work was supported by funding from NIH (R01 DC06441to MW and F31 DC010312 to RMC).
Conflicts of interest