Spatial and Temporal Organization of Ensemble Representations for Different Odor Classes in the Moth Antennal Lobe

Introduction

The spatial and temporal patterns of neural activity triggered by different scents at the initial stages of olfactory processing in the brain have been the subject of intense investigation in diverse animal models for well over half a century (Adrian, 1942). In an ongoing effort to decipher the neural coding strategies underlying the central processing of olfactory information, an increasing number of studies in the vertebrate olfactory bulb (OB) and insect antennal lobe (AL) rely on imaging methods that reveal odor-dependent patterns of activity across olfactory glomeruli. Odorants possessing different functional groups or hydrocarbon-chain features evoke unique but overlapping activity patterns, whereas the patterns for odorants with similar chemical structure are less distinguishable (Rubin and Katz, 1999; Uchida et al., 2000; Linster et al., 2001; Wachowiak and Cohen, 2001; Johnson et al., 2002; Sachse and Galizia, 2002; Meijerink et al., 2003). How, then, does the brain discriminate structurally similar odor molecules that evoke comparable spatial activation patterns? In an extension of previous models (Gelperin and Tank, 1990; Freeman, 1992), a proposed temporal coding mechanism could explain how molecularly similar odorants are discriminated in the insect AL (Stopfer et al., 1997; Laurent et al., 2001), but the well defined spatial aspects of the olfactory code are minimally represented in this model (Wehr and Laurent, 1996; Theunissen, 2003). Other studies in both vertebrates and invertebrates support an alternative model in which both the spatial and temporal aspects of the olfactory signals are inseparable components of the code (Spors and Grinvald, 2002; Christensen et al., 2003). To evaluate these hypotheses in greater detail, it is necessary to use a method that allows high-resolution monitoring of odor-evoked activity in both space and time.

Some studies have shown that mitral/tufted (M/T) cells or projection neurons (PNs) innervating the same glomerulus are more likely to show correlated firing in response to a stimulus (Buonviso and Chaput, 1990; Carlson et al., 2000; Lei et al., 2002; Christensen et al., 2003; Schoppa and Urban, 2003). Other studies have found, however, that the M/T cells innervating different glomeruli in the same region of the OB can also show strong correlations (Kashiwadani et al., 1999). Intraglomerular coactivity could play a role in amplifying the spatial representation of the specific chemosensory signal encoded in that glomerulus, whereas interglomerular synchrony could underlie a synthetic process that binds the different molecular features of the stimulus encoded in different information streams. Definitive coding roles for intraglomerular and interglomerular synchronization therefore have yet to be defined for any olfactory system.

To examine directly and simultaneously the roles of local and distributed network correlations in the neural representation of different olfactory stimuli at early stages of processing, we used a spatially defined array of 16 recording electrodes (Christensen et al., 2000; Daly et al., 2004b) to monitor both the spatial and temporal patterns of ensemble activity triggered by chemically different classes of odorants. Ensemble recordings revealed that each odor class evoked a different spatial focus of activity, but the correlated firing of action potentials within and between recording zones further distinguished the population codes for different odorants across the glomeruli of the main AL.

Materials and Methods

Experimental preparation. Adult male moths (Manduca sexta; Lepidoptera:Sphingidae) were reared in the laboratory on an artificial diet under a long-day (17/7 hr light/dark cycle) photoperiod. Animals were prepared for experiments 2-3 d after emergence, as described previously (Christensen et al., 2000). In preparation for recording, the moth was secured inside a plastic tube with dental wax, leaving the head and antennae exposed. The head was opened to expose the brain, and the tube was fixed to a recording platform attached to the vibration-isolation table. The preparation was oriented so that both ALs faced upward, and the tracheae and sheath overlying one AL were carefully removed with a pair of fine forceps. The brain was superfused slowly with physiological saline solution (in mm: 150 NaCl, 3 CaCl₂, 3 KCl, 10 N-tris[hydroxymethyl] methyl-2-aminoethanesulfonic acid buffer, and 25 sucrose, pH 6.9) for the duration of the experiment.

Olfactory stimulation. Olfactory stimuli were delivered to the preparation as reported previously (Christensen et al., 1993). Pulses of air from a constant air stream were diverted through a glass syringe containing a piece of filter paper, on which was deposited an odor compound (10 μg for plant-derived volatiles and 10 ng for pheromone components). The odor stimulus was pulsed by means of a solenoid-activated valve controlled by an electronic stimulator (W-P Instruments, Sarasota, FL). In each experiment, the outlet of the stimulus syringe was positioned ∼2 cm from and orthogonal to the center of the antennal flagellum ipsilateral to the AL. Stimulus duration was 200 msec, and five pulses were separated by an interval of 2 sec (supplemental material showing responses to longer stimulus durations is available at www.jneurosci.org).

Four classes of olfactory stimuli were used (see Fig. 1 A): (1) aliphatic aldehydes: nonanal (nal) and E-2-hexenal (t2h); (2) aromatics: phenylacetaldehyde (paa) and methyl salicylate (met); (3) monoterpenoids: (±) linalool (lin), nerol (ner), ocimene (oci), myrcene (myr), and geraniol (ger); and (4) sex pheromone components: E, Z-10,12-hexadecadienal (bombykal or bal; the primary component of the female sex pheromone), E, Z-11,13-pentadecadienal [c15; a mimic of a second component of the pheromone (Kaissling et al., 1989)], and the blend of the two components (bld). Some ensembles were not tested with c15 and bal individually. The control solvent for plant volatiles was mineral oil (ctr) and for pheromone components was cyclohexane (can). All of the compounds tested are biologically important volatiles that Manduca encounters in its natural environment (Shields and Hildebrand, 2001; Fraser et al., 2003). In addition, an olfactory stimulus that the moth would not encounter naturally, cyclohexanone (cyc), was included because olfactory conditioning studies have shown that Manduca learns to associate this odorant with sucrose reinforcement (Daly et al., 2001).

Ensemble recording and data analysis. The odor-evoked responses of 118 units were obtained in 12 male moths. Recordings were made with 16-channel silicon multielectrode recording arrays (MRAs) supplied by the Center for Neural Communication Technology at the University of Michigan (http://www.engin.umich.edu/facility/cnct/backind.html). To study the spatial distribution of odor-evoked activity, we selected an MRA design (a 4 × 4 array) that suits the dimensions of the AL in Manduca (see Fig. 2 A). These probes have four shanks spaced 125 μm apart, each with four recording sites 50 μm apart. The MRA was positioned under visual control using a stereo microscope. The four shanks were oriented in a line parallel to the antennal nerve, with the first shank inserted into the macroglomerular complex (MGC) (see Fig. 2 B). The MRA was advanced slowly through the AL using a micromanipulator (Leica Microsystems, Bannockburn, IL) until the uppermost recording sites were just below the surface of the AL. In this manner, the four shanks of the MRA defined four zones of glomerular neuropil across the AL.

Ensemble activity was recorded simultaneously from the 16 channels of the MRA using two Lynx-8 amplifiers (Neuralynx, Tucson, AZ). Spike data were extracted from the recorded signals and digitized at 20 kHz per channel using Discovery acquisition software (Data Wave Technologies, Longmont, CO) and a 2821-G 16SE analog-to-digital board (Data Translation, Marlboro, MA) on a personal computer platform (Data Wave Technologies). Filter settings (typically 0.6-3 kHz) and system gains (typically 5,000-20,000) were software adjustable on each channel. Spikes were sorted using a clustering algorithm based on the method of principal components (PCs) (Off-line Sorter; Plexon, Dallas, TX). Only those clusters that were separated after statistical verification in three-dimensional (PC1-PC3) space (multivariate ANOVA; p < 0.01) were used for additional analysis (6-15 units were isolated per ensemble; n = 12 ensembles in as many animals). Spikes arising from the same unit were visible on adjacent recording sites, thus providing geometric information about the spatial origin of the signals. Each spike in each cluster was time-stamped, and these data were used to create raster plots (see Fig. 3) and to calculate peristimulus time histograms (PSTHs), interspike interval histograms, cross-correlograms, and rate histograms. All analyses were performed with Neuroexplorer (Nex Technologies, Winston-Salem, NC) using a bin width of 5 msec, unless noted otherwise.

Conditional response probability in single units. To assess the validity of the arbitrary grouping of chemical classes and to examine the response preference of each unit for certain olfactory stimuli, we calculated the conditional probability of response to all members of the stimulus set (see Fig. 1 A) (Fletcher and Wilson, 2003). First, we grouped together all units that responded to a particular stimulus (e.g., reference stimulus = linalool), then calculated the probability that this subset of units also responded to the remaining members of the stimulus set (comparison stimuli). For example, the probability that linalool-responsive units also responded to nerol was calculated as follows: 1

where N_ner/lin was the number of nerol-responsive units within the pool of linalool-responsive units, and N_lin was the total number of linalool-responsive units. If two odorants activated the same pool of units, the conditional probability for the two odorants would be 1.0. To summarize results for the entire data set, the conditional probability scores for all olfactory stimuli were color-coded and displayed in a response-probability matrix, as shown in Figure 4.

Spatial distribution of odor-evoked activity. Sorted units were arranged according to which of the four AL zones (I-IV) defined by the MRA yielded each recording (see Figs. 2 A, 3). For each unit sorted, PSTHs were generated for all responses to each odor stimulus and the control (solvent only) stimulus. The response window was defined as the 500 msec period beginning at the onset of the 200 msec stimulus pulse. A unit was considered to be responsive if its control-subtracted PSTH was above (excitatory) or below (inhibitory) the 95% confidence limits derived from a cumulative sum (CUMSUM) test based on the Monte Carlo method (Ushiba et al., 2002). Inspection of the entire 118 unit data set revealed that a 500 msec window was sufficient to capture all of the 1000+ stimulus-evoked responses recorded. We quantified the control-corrected response for every unit by calculating a response index (RI) similar to that used to quantify the uptake of 2-deoxyglucose in the rat OB (Johnson et al., 2002). RI values reflect the deviation from the mean response of all units across all odors in one ensemble, as follows: 2

where R_odor is the number of spikes evoked by the test odor minus the number evoked by the control stimulus, R_m is the mean response, and SD is the SD across the data matrix. The RI values for the nonresponsive units fell between -2 and +2, based on the CUMSUM test. Given that >50% (on average) of recorded units in each ensemble were unresponsive, R_m approximated the background activity level, and thus negative values of the RI indicated response suppression. The RI values for all units were color-coded and arranged as an activity matrix with each row representing the ensemble response to a different odor stimulus (see Fig. 5). The RI had a range from -3.0 SDs (strongly inhibited units; cool colors) to +3.0 SDs (strongly excited units; warm colors).

Calculating response profiles and profile similarities (profile similarity index) between units. Variation across the responses to the different test stimuli was summarized as a color-coded correlation matrix for each unit (see Fig. 6C,D). This matrix was generated by calculating the 55 pairwise correlations between the PSTHs for each of the 11 stimuli and then translating the correlation coefficients into a color scale. To increase statistical efficiency (Kass et al., 2003), we applied a Gaussian filter (band-width of three bins) to the control-corrected PSTHs, each covering a time window starting at 200 msec before and ending at 1.5 sec after stimulus onset. Using a 5 msec bin width, each PSTH generated a vector containing 340 elements. Pairwise correlations of these vectors were then performed to obtain the 55 values for each matrix. We define the matrix as the overall “signature” of the neuron or a response profile that summarizes the similarity between the responses of the neuron to each pair of odor stimuli. The correlation coefficients represented on the color scale in Figure 6 are therefore a measure of response similarity, not response intensity as in Figure 5A. Next, to compare the response profiles between any two units, the two signature matrices were transformed into separate vectors, and the correlation coefficient between the two vectors was calculated (this coefficient was calculated for all pairs of units). We define this coefficient as the profile similarity index (PSI), which ranged from 0.0 (no similarity) to 1.0 (indistinguishable signatures) for each pair of units (see Figs. 6 F, 7). A PSI near 1.0 indicated that the two neurons were likely receiving common input.

Measuring stimulus-evoked coactivity between units. To calculate the temporal relationship between each pair of units, we used a cross-correlation analysis using the following formula: 3

where [CE]_RAW is the number of coincident events in the 5 msec cross-correlogram peak centered around t = 0, and [CE]_SHUFFLED is the number of coincident events after trial shuffling (shift predictor method) to correct for coincidences attributable to chance and an increased firing rate (Aertsen et al., 1989). The corrected correlograms were calculated by averaging over four trial shifts and subtracting the result from the raw correlogram. T is the total response time over which spikes were counted (500 msec), and N₁ and N₂ are the total number of spikes recorded from units 1 and 2 during time T. The coactivity index (C_%) therefore reflects the percentage of coincident spikes relative to the total number of spikes recorded from the two neurons (see Fig. 6 F). All calculations were implemented in Matlab 6.5 (The Mathworks, Natick, MA) or Neuroexplorer (Nex Technologies).

To visualize the odor-dependent coactivity pattern within an ensemble, we arranged individual units in a circular array similar to the “connectivity matrix” used to describe ensemble patterns in the hippocampus (Wilson and McNaughton, 1994). Each pair of units was connected with a line that depicted the magnitude of the correlation (C_%): values from 10-19%, dashed line; values ≥20%, solid line (see Fig. 8 B, C). In addition, each unit was color-coded to reflect its response magnitude, and these values were used to make a direct comparison between the spatial and temporal activity patterns evoked by different odorants (see Fig. 8C). To measure the ensemble-wide similarity between coactivity patterns evoked by different olfactory stimuli, we transformed the coactivity patterns composed of positive C_% values for all unit pairs in each ensemble into one-dimensional vectors (one for each odorant) and then calculated the correlation coefficient between the two vectors. This transformation is shown for one ensemble in Figure 8 D and repeated for the entire data set in Figure 8 E. Population data were calculated only within, not between, ensembles.

Histological identification of recording probes. To examine the precise location of the recording probes, the brain was excised and immersed in 1-2% glutaraldehyde in 0.1 m phosphate buffer to increase tissue contrast and facilitate locating probe tracks. Brains were fixed for at least 1 hr, then dehydrated with a graded ethanol series, cleared in methyl salicylate, and finally imaged as whole mounts with a laser-scanning confocal microscope (Nikon PCM 2000 equipped with a 457 nm argon laser). Optical sections were 5 μm, and this method reliably revealed the tracks of the four MRA shanks in the AL without the need for tissue staining (see Fig. 2B). The identification of individual glomeruli is an important prerequisite for assigning functional significance to a given glomerulus, but this is beyond the scope of this study.

Results

Discussion

Odor representations across olfactory glomeruli constitute a complex mosaic of excitatory and inhibitory activity, and neural-ensemble recordings provide a uniquely powerful tool to examine the dynamic and distributed character of these odor-dependent network representations in the brain (Christensen et al., 2000; Lehmkuhle et al., 2003; Stopfer et al., 2003; Daly et al., 2004a,b). Our ensemble data show that each of the plant-derived odorants that we tested evoked a stereotyped and combinatorial pattern of spiking activity across the glomeruli in the moth AL. These patterns did, however, show considerable spatial overlap, and some members of the same chemical class produced highly overlapping activity patterns (Fig. 5). Consistent with molecular evidence that olfactory neurons expressing homologous odorant receptors project to neighboring glomeruli (Tsuboi et al., 1999), support for spatially organized odor representations comes from numerous studies in widely differing species (for review, see Christensen and Hildebrand, 2002; Korsching, 2002; Keller and Vosshall, 2003). Our population data now provide direct evidence that the representations of nonpheromonal olfactory stimuli involve a combinatorial pattern of both excitatory and inhibitory responses distributed across each coding ensemble. Moreover, we found that approximately one-third of the neurons we sampled showed high selectivity (Fig. 1B) and were localized to specific zones of the AL neuropil (Fig. 5). The spatial activity patterns derived from our MRA data also match closely with previous functional mapping studies in the same species. Optical imaging using calcium-sensitive dyes showed that the spatial patterns triggered by aromatic volatiles were localized to the most medial portion of the AL (Hansson et al., 2003), and the same region in our study (zone IV) represents the excitatory activity focus for the aromatics met and paa (Fig. 5). A similar correspondence between imaging and ensemble-recording data also exists for excitatory responses to monoterpenoids (e.g., lin) (Fig. 5). These two independent methods therefore confirm for Manduca that different classes of odorants are represented as overlapping but distinct spatial patterns of activity in the AL glomeruli. Accumulating evidence from the vertebrate OB (Rubin and Katz, 1999; Fuss and Korsching, 2001; Nikonov and Caprio, 2001) and insect AL (Gao et al., 2000; Vosshall et al., 2000; Galizia and Menzel, 2001) also supports this combinatorial spatial model.

From an information-processing perspective, chemotopic mapping of olfactory glomeruli offers the organism at least three key advantages: (1) the convergence of a large population of functionally equivalent olfactory receptor cells onto a single glomerulus reduces the detection threshold for weak signals above a noisy background; (2) a combinatorial glomerular map increases exponentially the coding capacity of an olfactory network, especially one like the moth's AL that consists of relatively few glomeruli; and (3) the hardwired connection between different levels of processing in the olfactory pathway increases the speed of information transmission. Experimental evidence that structurally similar stimulus compounds can be discriminated through spatial coding is growing steadily (Rubin and Katz, 1999; Johnson and Leon, 2000a,b; Uchida et al., 2000; Inaki et al., 2002; Meijerink et al., 2003; Xu et al., 2003). Moreover, in Drosophila, the chemotopic arrangement of glomeruli is also preserved at the next stage of processing in the protocerebrum (Marin et al., 2002; Wong et al., 2002).

The PN activity map also undergoes considerable transformation as a consequence of inhibition mediated by GABAergic LNs. Inhibitory, spiking LNs interconnect many, if not all, glomeruli in the moth's AL (Christensen et al., 1993), and both ultrastructural and electrophysiological data indicate that LNs receive direct synaptic input from receptor cells (Tolbert and Hildebrand, 1981; Christensen et al., 1993). Inputs to the same LN from functionally different sub-populations of receptors therefore converge on different dendritic targets on the LN as defined by their specific branching patterns in the glomeruli. The spread of inhibition from wide-field LNs to PNs contributes substantially to the spatially distributed response of the PN ensemble before this information is relayed to the protocerebrum. Morphometric modeling studies indicate that the lateral interactions mediated by LNs will be stronger between adjacent glomerular zones (Christensen et al., 2001), but longer-distance interactions are also possible. In the present study, we provide experimental evidence to support the computer models: lateral inhibition, triggered by sex pheromonal input to zone I, can spread across the entire AL to suppress the firing of units in zones II, III, and even IV (Fig. 5B). In the same manner, stimulation with the aromatic compound paa produced a focus of excitatory activity in zone IV, but the same odorant evoked inhibitory responses in the remaining three zones, including zone I (Fig. 5B). Monoterpenoid responses revealed a “center-surround”-type organization, with an excitatory center in zones II and III and inhibitory surrounds in zones I and IV (Fig. 5B). Other inhibited units, however, were found in the same zones as the activated units. These data therefore provide direct evidence for both interzonal and intrazonal inhibition in that inhibition can spread long distances across the AL to influence the activity of distant glomeruli, but it can also function locally, providing inhibitory regulation to PNs within the activated glomerulus or glomeruli. Our data are furthermore consistent with recent imaging studies in other insects (Ng et al., 2002; Sachse and Galizia, 2002). Both local and global inhibition has been reported in the honeybee's AL, where these two inhibitory systems may use different neurotransmitters that operate in concert during the response to odor (Sachse and Galizia, 2002). Global inhibition may serve to enhance contrast in the odor representation by suppressing the activity of glomeruli in the regions surrounding the central focus of excitatory activity (Figs. 3, 5) (Yokoi et al., 1995; Aungst et al., 2003). In contrast, local inhibition likely serves to enhance the timing precision of spiking between PNs innervating the same glomerulus (Lei et al., 2002) (see below).

Long-term ensemble recording allows testing of multiple and diverse olfactory stimuli, and our results demonstrate that each individual AL unit displays a unique signature response across a wide range of different volatiles, thus confirming that individual glomeruli participate in the coding of multiple odorants (Fig. 6C,D). We extended this observation by demonstrating that the strength of synchronous firing between units also increased as a function of similarity in their response profiles (Figs. 6F, 7C). Likewise, none of the pairs with low response similarity displayed significant coactivity. This result agrees with an increasing number of studies showing that specific temporal relationships among functionally defined neurons may be used by the brain to recognize and discriminate different features of an olfactory stimulus, although the mechanisms underlying these temporal relationships may vary in different olfactory systems (Christensen et al., 2000, 2003; Laurent et al., 2001; Lei et al., 2002; Sachse and Galizia, 2002). Manduca PNs, unlike those in locusts (Laurent et al., 2001), exhibit transient synchronization in response to olfactory stimulation, not involving fast oscillations (Christensen et al., 2003; Daly et al., 2004a,b).

Importantly, the patterns of coactivity need not involve many spikes from the units in the coding ensemble. In fact, our sample of 535 within-ensemble pairs of recorded units revealed that, on average, <12% of the odor-evoked spikes in each pair were temporally correlated within 5 msec (Table 1). This is consistent with the transient nature of the coordinated ensemble response to brief olfactory stimuli (Fig. 3). It is also important to note that this stimulus condition is closely analogous to the situation in breathing animals. Numerous studies in mammals have shown, for example, that odors evoke respiration-coupled activity in M/T cells (for review, see Sobel and Tank, 1993; Cang and Isaacson, 2003; Margrie and Schaefer, 2003). Moreover, the brief responses of these vertebrate M/T cells bear a striking resemblance to the PN responses we see in Manduca. Although first observed in PNs responding to sex pheromone, stimulus-coupled responses have recently been recorded in moth PNs encoding both plant odor (Reisenman et al., 2004) and CO₂ (Guerenstein et al., 2004) signals. Collectively, these data demonstrate that both M/T and PN response patterns are, to a great extent, delimited by the temporal pattern of peripheral input to the glomerulus.

In one recent study examining pairs of M/T cells separated by 300-500 μm (the width of two to three glomeruli), 27% of the pairs displayed odor-evoked synchronization, supporting the presence of local processing among glomeruli occupying the same region of the OB (Kashiwadani et al., 1999). That study did not, however, involve systematic measurement of interelectrode distances. In the present study, we used a spatially distributed and precise recording array that allowed us to compare the response profiles and examine the temporal correlations between adjacent and nonadjacent cells simultaneously across the entire AL. Using this technique, we found direct evidence for local processing in Manduca, showing precise timing between nearby cells with a similar response profile, analogous to what has been reported in the mammalian OB (Mori et al., 1992). Of the 125 unit pairs with a C_% value ≥20 (Fig. 7A), 95% were localized to the same or neighboring zones of glomerular neuropil. These local interactions have been predicted from both empirical and modeling studies in Manduca (Christensen et al., 2001; Reisenman et al., 2004; Guerenstein et al., 2004), and they are furthermore consistent with paired recording results from other vertebrate and invertebrate studies (Schoppa and Westbrook, 2001; Lei et al., 2002). Additional analysis of the temporal relationships between units revealed for each stimulus a distinct pattern of unit coactivity across the ensemble, and each pattern could be clearly distinguished by which units were correlated as well as by the strength of these correlations (Fig. 8E) (supplemental Fig. 1, available at www.jneurosci.org as supplemental material). The strongest synchrony occurred between neurons that showed statistically significant responses based on their PSTHs, but weaker correlations also occurred between nonresponsive neurons, suggesting that even neurons that are not statistically responsive may nevertheless participate in the coding ensemble through temporal coordination of their firing (Fig. 8C). Multichannel ensemble recording thus revealed odor-dependent “subthreshold” temporal interactions that may be below the resolution of popular activity-labeling methods. Collectively, our findings using neural-ensemble recording support the hypothesis that the spatial extent of stimulus-evoked activity in the olfactory system is defined primarily by the particular subset of glomeruli activated by a given stimulus. Moreover, they provide new evidence that local processing within and between neighboring glomeruli may further enhance discrimination, especially among chemically similar odorants.