Rapid Odor Processing in the Honeybee Antennal Lobe Network

In their natural environment, many insects need to identify and evaluate behaviorally relevant odorants on a rich and dynamic olfactory background. Behavioral studies have demonstrated that bees recognize learned odors within <200 ms, indicating a rapid processing of olfactory input in the sensory pathway. We studied the role of the honeybee antennal lobe network in constructing a fast and reliable code of odor identity using in vivo intracellular recordings of individual projection neurons (PNs) and local interneurons (LNs). We found a complementary ensemble code where odor identity is encoded in the spatio-temporal pattern of response latencies as well as in the pattern of activated and inactivated PN firing. This coding scheme rapidly reaches a stable representation within 50–150 ms after stimulus onset. Testing an odor mixture versus its individual compounds revealed different representations in the two morphologically distinct types of lateral- and median PNs (l- and m-PNs). Individual m-PNs mixture responses were dominated by the most effective compound (elemental representation) whereas l-PNs showed suppressed responses to the mixture but not to its individual compounds (synthetic representation). The onset of inhibition in the membrane potential of l-PNs coincided with the responses of putative inhibitory interneurons that responded significantly faster than PNs. Taken together, our results suggest that processing within the LN network of the AL is an essential component of constructing the antennal lobe population code.


INTRODUCTION
Honeybees (Apis mellifera) are social insects that use odors as communication signals. These odors are of simple structure, typically composed of a single chemical compound. On the contrary, naturally occurring odorants that become relevant for the bee during foraging are generally complex blends of a large number of volatile compounds (Pichersky and Gershenzon, 2002) with time-varying composition and concentrations. Psychophysical studies in bees (Chandra and Smith, 1998;Deisig et al., 2003), catfi sh (Valenticic et al., 2000), and humans (Livermore and Laing, 1998) have shown that the neural representation of an odor mixture can be synthetic or elemental whereby the mixture either acquires a novel perceptual quality or it resembles its elemental compounds. In vertebrates, odor mixture interactions were observed at different levels of the olfactory system, in the periphery (Duchamp-Viret et al., 2003), in the olfactory bulb (Davison and Katz, 2007;Giraudet et al., 2002;Lin et al., 2006;Tabor et al., 2004) and in the olfactory cortex (Lei et al., 2006;Zou and Buck, 2006). Studies of mixture interaction in the honeybee antennal lobe using Ca 2+ -imaging of glomerular activity revealed suppressive interactions in which the response to an odor mixture can be much lower than the response to the single compounds (Deisig et al., 2006;Joerges et al., 1997;Silbering and Galizia, 2007). Such a suppression effect was not observed in the moth Spodoptera littoralis where glomerular Ca responses to mixtures always demonstrated hypoadditive interactions resembling the response to the strongest compounds within the mixture (Carlsson et al., 2007).
Rapid odor processing in the honeybee antennal lobe network Sabine Krofczik 1,2 , Randolf Menzel 1,2 and Martin P. Nawrot 1,2 * 1 Institut für Biologie -Neurobiologie, Freie Universität Berlin, Germany 2 Bernstein Center for Computational Neuroscience, Berlin, Germany In their natural environment, many insects need to identify and evaluate behaviorally relevant odorants on a rich and dynamic olfactory background. Behavioral studies have demonstrated that bees recognize learned odors within <200 ms, indicating a rapid processing of olfactory input in the sensory pathway. We studied the role of the honeybee antennal lobe network in constructing a fast and reliable code of odor identity using in vivo intracellular recordings of individual projection neurons (PNs) and local interneurons (LNs). We found a complementary ensemble code where odor identity is encoded in the spatio-temporal pattern of response latencies as well as in the pattern of activated and inactivated PN fi ring. This coding scheme rapidly reaches a stable representation within 50-150 ms after stimulus onset. Testing an odor mixture versus its individual compounds revealed different representations in the two morphologically distinct types of lateral-and median PNs (l-and m-PNs). Individual m-PNs mixture responses were dominated by the most effective compound (elemental representation) whereas l-PNs showed suppressed responses to the mixture but not to its individual compounds (synthetic representation). The onset of inhibition in the membrane potential of l-PNs coincided with the responses of putative inhibitory interneurons that responded signifi cantly faster than PNs. Taken together, our results suggest that processing within the LN network of the AL is an essential component of constructing the antennal lobe population code.

Keywords: antennal lobe, Apis mellifera, latency code, local interneurons, olfaction, odor mixture, projection neurons, temporal coding
Animals have evolved sophisticated mechanisms to distinguish naturally occurring odors, and to extract meaningful information from a rich olfactory environment. Recent behavioral studies could show that a perceptual discrimination of odors occurs rapidly after the onset of an odor stimulation. Mice are able to discriminate dissimilar odors within <250 ms (Abraham et al., 2004) and they can distinguish a novel odor from learned odors within only 140 ms (Wesson et al., 2008). A recent study in honeybees has demonstrated their ability to distinguish between different olfactory stimuli that are as short as 200 ms (Wright et al., 2009). Consistent with this result, we found that mushroom body (MB) extrinsic neurons in the honeybee reliably signal the presence of a conditioned odor within <200 ms after stimulus onset (Pamir et al., 2008;Strube-Bloss et al., 2008). Thus, odor processing in the olfactory pathway must be fast and lead to a reliable code of odor identity at the input level to the central brain structures.
The olfactory system of the honeybee comprises 60,000 olfactory sensory neurons (OSNs; Esslen and Kaissling, 1976), which are located predominantly in pore plate sensillae on the antennae. Each OSN conveys the olfactory information through one of the four antennal tracts (T1-4, Figure 1A) and innervates neurons in one of ∼160 glomeruli (Arnold et al., 1985;Flanagan and Mercer, 1989;Kelber et al., 2006). Glomeruli receiving input from T1 OSN are located in the ventral-rostral hemisphere of the AL whereas glomeruli which receive input from T2-4 glomeruli in the dorsal-caudal hemisphere ( Figure 1A; Kirschner et al., 2006). Glomeruli represent synaptic interaction sites between the OSN, a total of ∼4000 AL-intrinsic local interneurons (LNs, Figure 1B, depicted in red; Fonta et al., 1993) and ∼950 projection neurons (PNs, Figures 1C,D; Kirschner et al., 2006;Rybak and Eichmueller, 1993). Anatomically we distinguish two types of PNs. The lateral PNs (l-PNs) receive input exclusively from T1 glomeruli ( Figure  1A, green circle) and send their axons along the lateral antennocerebralis tract to the higher order brain centers, the lateral horn (LH) and the MB. The median PNs (m-PNs) exclusively originate in T2-4 glomeruli ( Figure 1A, purple circle) and project along the median antennocerebralis tract, fi rst to the MB and then to the LH (Abel et al., 2001;Bicker et al., 1993;Kirschner et al., 2006;Mobbs, 1982;Müller et al., 2002). Each uniglomerular PN is therefore identifi ed by the respective glomerulus and numbered according to the 3-D Atlas of the honeybee AL (Figures 1C,D; Galizia et al., 1999).
In the present study, we performed intracellular recordings of single l-PNs, m-PNs and LNs and analyzed their spatio-temporal responses to odors and odor mixtures. We specifi cally addressed the following questions. How fast does an ensemble code evolve Rapid odor processing in antennal lobe (T1 = ventral-rostral, T2-4 = dorsal-caudal). LNs could easily be identifi ed either by their specifi c intracellular action potential waveform and their fi ring pattern which clearly differ from those of the PNs (Galizia and Kimmerle, 2004). Previous imaging studies of the AL could only resolve signals from the T1 glomeruli but not form T2-4. We therefore preferentially targeted m-PNs that innervate T2-4. In total we identifi ed and analyzed PNs: N = 7 (l-type), N = 23 (m-type); and LNs: N = 7.

ODOR STIMULATION
The animal's antennae were exposed to a constant active charcoalfi ltered stream of room air (airfl ow rate 10 ml/s) guided through a glass tube with an inner diameter of 14 cm placed 1 cm from the antenna. For each odor, 4 µl were placed on a fi lter paper in a glass syringe, which was introduced into the continuous air stream. The control stimulus was a glass syringe plus fi lter paper with mineral oil (Sigma-Aldrich, Germany) delivering a constant airfl ow rate of 10 ml/s. To avoid mechanical stimulation a computer-controlled valve switched between control stimulus and odor-loaden syringe such that the antenna was always exposed to an air stream with a fl ow rate of 20 ml/s. The stimulus timing was computer-controlled and each odor stimulation lasted for 2 s. The delay from the onset trigger for valve opening to the arrival of the odorant at the antenna was estimated as ∼200-220 ms. For response latency analyses and display purposes we therefore subtracted a fi x delay of 200 ms from the measured latencies. All odors were diluted 1:10 in mineral oil and checked for purity and mixture composition by gas chromatography (Trace GC Ultra, Thermo, Electron Corporation, Waltham, MA, USA). The animals' antennae were stimulated with a tertiary mixture and its elemental compounds as well as other single compound odors and complex odor mixtures (Sigma-Aldrich, Germany). The tertiary mixture was composed of the blossom compounds nonanol and hexanol and 2-heptanon, which is known to be a compound of the alarm pheromone in bees (Balderrama et al., 2002). Additionally natural fl oral blends as well as single compounds exhibiting different functional groups were tested. In some animals multiple-trial presentations of the same odor were used to test response reliability, whereas in other animals one-trial presentation (in some cases with few repetitions) of different odors (between 2 and 30) were performed in a randomized fashion. We analyzed only stable recordings that lasted for at least 10 min.

DATA ANALYSIS
Matlab and the Matlab Statistics Toolbox (The MathWorks, Inc.) were used for the analysis of the intracellular voltage traces and spike trains. We implemented all custom-written algorithms used for analysis in the present manuscript in the FIND open source Matlab toolbox for neural data analysis (Meier et al., 2008; http:// fi nd.bccn.uni-freiburg.de/).

Firing rate analysis
Intracellular spikes times were detected with a threshold criterion. Firing rates were measured by convoluting single trial spike trains with a kernel of asymmetric shape k(t) = t × exp(−t/τ) and time resolution τ, defi ned on a fi nite support (Nawrot et al., 1999;Parzen, 1962). We aligned the resulting rate function at time w that indicates the center of mass of the kernel (w = 42 ms for τ = 25 ms). For in a population of PNs? Is there a functional distinction between the two anatomically distinct types of PNs? With regard to mixture encoding we confi rmed the hypothesis that l-PNs can show mixture suppression as previously reported in Ca-imaging studies at the glomerular level, and we tested whether this effect does carry over to the m-PNs. Finally, we asked whether local inhibition with in the AL network underlies response suppression in individual PNs.

HONEYBEES
Worker bees (A. mellifera carnica) were caught at the hive entrance or in an indoor fl ight room, immobilized by cooling, and mounted in plastic tubes. The bees were fed with sucrose solution and kept in the dark at 20°C and high humidity. On the following day, the head was fi xed with wax and opened between the median ocellus and the base of the antennae. Glands and tracheal sheaths were removed. A second hole was cut to expose the esophagus and to stretch it aiming to reduce movement of the brain. The brain was kept wet at all times by dribbling drops of bee physiological saline solution onto (in mmol -1 : 130 NaCl, 6 KCl, 2 MgCl 2 , 10 HEPES, 17 glucose, 6 fructose, 160 sucrose, pH 6.7).

ELECTROPHYSIOLOGY
Glass electrodes were pulled with a horizontal laser puller (P-2000 Sutter Instruments, Novato, CA, USA) and their tips were fi lled with 4% tetramethylrhodamine-biotin dextran (Micro-Ruby; Invitrogen, Germany) in 0.2 M potassium acetate. The electrodes (resistance 120-200 MΩ) were positioned at the top of the AL by using a micromanipulator (Leitz) and lowered posteriorly into the neuropil until a neuron could be penetrated. A reference electrode, a chlorized silver wire, was inserted into the eye. The recordings were done in bridge mode using an intracellular BRAMP 1 amplifi er (NPI Electronics, Tamm, Germany). Data were, digitized using a 1401 interface and stored on a PC using spike2.5 software (Cambridge Electronic Design, UK).

NEURON IDENTIFICATION
Single l-and m-PNs were intracellular stained by injecting Micro-Ruby iontophoretically using depolarizing current pulses with a duration of 0.2 s applied at 1-2 Hz. Successful staining allowed the identifi cation of the PN type (l or m) by location of their uniglomerular projections (Figures 1C,D) (l-PNs receive input from T1 glomeruli whereas m-PNs from T2-T4 glomeruli, see Figure 1A). Intracellular recording and fi lling (approximate recording time >10 min, approximate staining time ∼20 min) was performed intra-axonally. Due to the small axon diameter (∼3-5 µm), recording conditions were extremely diffi cult, strongly limiting the overall experiment time. We kept stimulus protocols short to ensure suffi cient time for the iontophoresis. In cases of highly stable recording conditions, we increased the stimulus protocol to either comprise a larger stimulus set or to increase number of repetitions for a subset of the stimuli. Successful staining could be achieved in about 50% of all neurons. Since the antennal lobe of the honeybee is well described (Abel et al., 2001;Galizia et al., 1999;Kirschner et al., 2006) identifi cation of l-and m-PNs was alternatively performed by the spatial position of the electrode with respect to the hemispherical segregation of T1 and T2-4 glomeruli color-coded representation of excitatory and inhibitory responses, we subtracted from the response the baseline rate measured during 1 s preceding the stimulus. We then represent the rate response of a neuronal ensemble with n neurons to odor stimulus a at any given point in time as the n-dimensional rate vector v a . We then defi ne the distance between two rate vectors

Response latency analysis
To estimate absolute and relative trial-by-trial latencies we used a method that we described in detail elsewhere (Meier et al., 2008). A demo script for this method together with a demo data set is available online at the offi cial website of the FIND toolbox (http://fi nd. bccn.uni-freiburg.de/?n=Docu.Tutorial). In a fi rst step, we estimated from a given single trial spike train the time derivative of the signal by convolution with an asymmetric Savitzky-Golay fi lter (Savitzky and Golay, 1964) (polynomial order 2, width 300 ms, Welch windowed). In a second step, we estimated the N-1 relative latencies of the individual spike trains. For this we used an algorithm that allows to optimally align all N single trial derivatives (taking into account the fi rst 700 ms) such that their pair-wise cross-correlation is maximized (Nawrot et al., 2003). The resulting optimal N-1 time-shifts correspond to the relative latencies and we measured their standard deviation σ to quantify the cross-trial latency variability. In a third step, we aligned in time all individual single trials according to the optimal time shifts (with a zero mean shift to preserve the center of mass) and then merged all single trial spike trains into 1.
For each odor, we thus obtained one trial-aligned and merged spike train. To obtain the relative latencies across different odors we repeated the same procedure but now using the merged spike trains. We again used the standard deviation to quantify latency variability across odors. To test for the signifi cant difference of two distributions of latency differences we used the Kolmogorov-Smirnov test. To obtain the absolute response latency for each neuron as averaged across all odors we optimally aligned the odor-specifi c spike trains as described above and merged all spike trains. We then computed the time derivative of the fi ring activity as explained above and measured the peak time of the maximum. We defi ned this point in time at which the rise of the response was steepest as the absolute single neuron latency. Absolute latency estimates have a small but fi x bias toward slightly higher values as we subtracted the lower limit of the estimated delay for odor delivery (200 ms), and because of the asymmetry of the Savitzky-Golay fi lter kernel.
Before estimating the latency of intracellular inhibition the membrane potential of PNs was low-pass fi ltered with a cut-off frequency of 100 Hz using an exponential impulse response kernel (decay time constant 1.6 ms). Response onset was then determined as the time of crossing a threshold defi ned as the average baseline response (during 5 s preceding stimulus onset) minus either one or two times the standard deviation during baseline period (excluding APs).

Specifi city index
We defi ned the specifi city index s of a neuron as the number of odors that evoked either an excitatory (s ex ) or inhibitory (s in ) response (within the fi rst second after stimulus onset) divided by the total number of tested odors. For each type of PNs we computed the specifi city index separately for single compound and odor mixture presentations.

Mixture index
To quantify the response difference between mixture and most effective compound we calculated the mixture index κ = (y − x)/ (x + y) where x denotes the absolute number of response spikes for the individual odor compound that yielded the strongest response and y denotes the absolute number of spikes in the response to the mixture, counted in a fi x interval from 200 to 700 ms after stimulus onset at t = 0. κ Will be close to 0 in the case of a hypoadditive mixture response, and κ assumes values close to −1 (+1) in the case of a suppressive (additive) mixture effect. The distribution of κ values may be strongly skewed, in particular for values close to −1/1. Therefore, before averaging values of κ we applied Fisher's z-transform. We tested for a deviation from the null hypothesis of two samples of κ having the same median using the Wilcoxon rank sum test.

l-AND m-PN ODOR RESPONSE PROFILES
We performed intracellular recordings of single PNs and tested their dynamic responses elicited by single compound odors. l-and m-PNs were identifi ed either by their uniglomerular arborizations, (Figures 1C,D) or by the location of the particular recording site (see Section "Materials and Methods"). First, we characterized the response kinetics and response reliability of PNs. Both types of PNs typically exhibited a phasic-tonic spike response pattern that outlasted the stimulus presentation. Figures 2A-D shows the response characteristics for a l-PN receiving input from the T1-22 glomerulus (Figures 2A,B) and a m-PN receiving input from the T3-45 glomerulus (Figures 2C,D). The responses were highly reliable across trials in all neurons tested. If a response was observed during one stimulus presentation, then it could reliably be detected in all of the trials. This is demonstrated in the spike raster diagrams of Figures 2B,D. Based on this result we reduced the odor stimulus protocol to single or few trials of one and the same odor to test a larger set of odors within the limited experimental time. On average, the dynamic response profi le of l-and m-PNs was very similar ( Figure 2E). The spontaneous fi ring rate was typically low with a mean of 6.8 spikes/s (median 3.4/s) and the average maximum responses peaked around 70 spikes/s. Several neurons showed no spontaneous spiking activity. Average response latencies of individual neurons were not signifi cantly different (P = 0.7, Wilcoxon rank sum test) in l-PNs (148.3 ± 33 ms) and m-PNs (123.5 ± 14.6 ms) ( Figure 2F).

ODOR-SPECIFIC RESPONSE LATENCIES OF l-AND m-PNs
From behavioral experiments, it becomes evident that a code of odor identity must be rapidly established at the level of the AL output (see Section "Introduction"). We here consider one possible fast coding strategy, namely the encoding of odor identity in an odor-specifi c pattern of response latencies across an activated PN ensemble. To this end, we compared the differences in response latency as measured across different odors with the differences in latency across trials of repeated stimulus presentation. The experimental limitations of recording time made it necessary that we typically tested a diverse set of odors with very few trials or only very few odors during many trials. For a statistical evaluation, we therefore pooled all relative across-trial latencies (see Section "Materials and Methods") separately for l-and m-PNs (Figures 3A,B). The trial-to-trial variability of response onset was quantifi ed by its standard deviation σ which was similar for both, l-PNs ( Figure 3A) and m-PNs ( Figure 3B) with σ l trial = 22.5 ms and σ m trial = 28.3 ms, respectively. Next, we quantifi ed the relative latencies across odors for each neuron and then pooled these again separately for the two types of PNs. The standard deviation of latencies across odors was considerably larger with σ l odor = 41.6 ms ( Figure 3A) and σ m odor = 49.2 ms ( Figure 3B). These latency differences across odors are signifi cantly larger than across trials in both, l-and m-PNs (see Section "Materials and Methods"). Thus, on a statistical basis, response onset is odor-specifi c in both types of PNs. For the population of PNs this result immediately implies that there exists an odor-specifi c spatial latency pattern in the activated ensemble of uniglomerular PNs (cf. Figure 4A) establishing a fast code of odor identity.

l-AND m-PN RESPONSES IN SPACE AND TIME
Results from Ca 2+ imaging studies in the honeybee AL have shown odor-specifi c activation pattern of T1-glomeruli which are innervated by l-PNs. This suggests that odor identity is encoded by the activation of an odor-specifi c ensemble of PNs (Galizia and Menzel, 2000;Joerges et al., 1997;Sachse et al., 1999). To test this hypothesis on the level of individual l-and m-PNs we performed two analyses. First, we concentrated on individual PNs and quantifi ed their odor selectivity on the basis of mean responses during the fi rst 800 ms after stimulus onset. For l-and m-PNs separately, we classifi ed each response as inhibitory, excitatory or no response and defi ned the specifi city index s, which represents the relative number of odors that elicited an excitatory (s ex ) or inhibitory response (s in ) (see Section "Materials and Methods"). Our results show that on average l-PNs show an excitatory response to about half of the odors tested while m-PNs are less odor-specifi c with excitatory responses to about 2/3 rd of the tested odors (cf. Table 1; for examples see Figure 1 in Supplementary Material). Inhibitory responses (i.e., a suppressed spike rate) were observed in l-PNs for either 12% or 13% of the odors but they were very rarely observed in m-PNs and only for single compound odors but never in the case of an odor mixture. The odor-specifi c mean response classifi cation of individual neurons necessarily translates into an odor-specifi c binary response pattern across the PN population.
Next, we examined the spatio-temporal activation in a pseudo population of non-simultaneously recorded l-and m-PNs and their responses to three single compound odors, the ketone 2-heptanon (2,7on) and the alcohols hexanol (6ol) and nonanol (9ol) (Figure 4A). The color-coded response rate profi les in Figure 4A indeed indicate that there exists a unique spatio-temporal pattern associated with each odor. A mean response classifi cation (excitatory, inhibitory, no response) resulted in an odor-specifi c fi ngerprint (matrix in Figure 4A). Since responses were highly reliable across repeated stimulations, this odor-specifi c spatial activation pattern will be identical in each trial. To investigate how the separations of odor-specifi c ensemble fi ring patterns evolve over time we computed the vector distance d in time-resolved fashion. In each point in time d ab measures the distance between the vector of neuronal fi ring rates for two odors a and b. In each panel of Figure 4B we show the average distance d′ = (d ab + d ac )/2 between the respective odor a and the two other compounds. The units of d represent the difference in fi ring rate per neuron, i.e., the number of spikes per second per neuron (see Section "Materials and Methods"). The resulting profi les in Figure 4B display a highly dynamic behavior. The vector distance increases rapidly after stimulus onset and peaks at about 150∼250 ms. The highest divergence of odor response patterns naturally occurs only during the initial phasic responses. After about 300-500 ms the patterns again converge to reach an almost constant distance of ∼20 spikes/s associated with the tonic response that outlasted the stimulus.
The spatial identifi cation of l-and m-PNs by their uniglomerular arborizations further allowed us to establish a schematic map of the glomerular activation patterns within the AL. In Figure 4C the spatio-temporal fi ring rate of identifi ed PNs during the stimulation with 2-heptanon (2,7on) is shown as a glomerular activation pattern in eight successive time windows. Each time window represents in a color code the odor response for each individual PN at the location of the corresponding innervated glomerulus, resulting in a spatio-temporal activation pattern across all identifi ed glomeruli. This interpretation assumes that the average PN activity of several uniglomerular neurons is highly similar to the response profi le of the individual recorded neuron. Conversely, such a mapping can make a direct prediction for the glomerular response pattern in Ca-imaging studies that measure the mean activity of identifi ed glomeruli.

DIFFERENTIAL ODOR MIXTURE INTERACTIONS IN l-AND m-PNs
Imaging studies in the honeybee AL revealed that T1-glomeruli often exhibit suppressed responses to a mixture but not to the elemental mixture components when presented alone (see Section "Introduction"). To test the prediction that this suppression effect is also evident at the level of single l-PNs, and to test whether it is also observable in individual m-PNs, which are not accessible by Ca-imaging in the AL, we compared individual l-and m-PN responses to an odor mixture and to its three elemental components when presented alone. Figures 5A,B show the dynamic odor response profi les of one l-and one m-PN elicited by the tertiary mixture (black) and its compounds, the ketone 2-heptanon (green), and the alcohols hexanol (orange) and nonanol (blue). The l-PN showed excitatory responses to all three elemental compounds but was suppressed by their tertiary mixture ( Figure 5A). Note that the suppression is rapid such that not a single response action potential is elicited. The m-PN, however, showed excitatory responses to the odors hexanol, 2-heptanon and to the mixture, where the response to the mixtures resembled the stronger response to 2-heptanon ( Figure 5B).
The color-coded response profi les in Figure 5C show the odor mixture responses across all PNs tested. The l-PNs were mostly inhibited or showed no response. In comparison, the majority of m-PNs showed a pronounced excitatory response.  in (A,B). The differences in latencies across trials were not signifi cantly different for the two types of PNs (P > 0.05). See Section "Materials and Methods" for details on latency estimation.  In Figure 5D, we show the time resolved vector distance d. In each point in time, it measures the difference of the neuronal ensemble response to the mixture and its response to each of the individual compounds (cf. Figure 4) as indicated by line color. The spatial response pattern for the mixture rapidly and strongly diverged from the response patterns of hexanol (orange) and nonanol (blue), and to a lesser degree from the response pattern of 2-heptanon (green).
We classifi ed odor mixture responses according to three categories of mixture interactions: higher (synergy), equal (hypoadditivity) or lower (suppression) than the response to the most effective compound, following the conventions in (Duchamp-Viret et al., 2003). For each l-and m-PN we determined the most effective compound, i.e., with the strongest response, and compared it with the response to the odor mixture. The bar diagram in Figure 5E shows the response magnitude elicited by

FIGURE 4 | Population rate code in projection neuron ensembles. (A)
Rate responses of PN populations to three different odors. Color code indicates rate deviation from the baseline as measured during the 500 ms preceding stimulus onset. Each row represents the fi ring rate of one individual neuron. The identifi cation of each neuron is indicated on the left of each panel. A subpopulation of N = 17 neurons (from bottom to horizontal black line) was tested with all three odors. The horizontal bars to the right of each panel indicate the response spike count within the fi rst 500 ms after stimulus onset. The matrix on the right classifi es for each neuron (vertical axis) and for each odor (horizontal) the type of rate response: excitatory (red), inhibited (blue) or neutral (white). The vertical columns thus represent a spatial fi ngerprint for the respective odor. (B) The time resolved distance d′ of the population rate vectors indicates the difference in fi ring rate per neuron (see text). In each panel, the dotted gray line and associated time stamp indicate the crossing of a threshold defi ned as mean plus fi ve times the standard deviation as measured before stimulus onset. The dashed gray line indicates the time at which 90% of the maximum value was reached. Inset in left panel shows asymmetric shape of kernel function for rate estimation (τ = 25 ms) with correct temporal scale. (C) We translated the ensemble fi ring rates into a spatio-temporal rate code in the antennal lobe space by assigning to each neuron the corresponding 2-D section of the identifi ed glomerulus (see Section "Materials and Methods"). The gray patch indicates a plane through the antennal lobe in the standard atlas of the honeybee brain. After each step of time from left to right we updated the color code according to the fi ring rate at the time indicated. A spatial map of the identifi ed glomeruli (see Figure 2 in Supplementary Material) and additional video clips are available online in the "Supplementary Material. " the odor mixture (white bars) and its strongest compound (black bars) indicating clear suppressive odor mixture interactions for all three l-PNs that we had tested for all compounds. In contrast, hypoadditive odor mixture interaction was observed in all m-PNs. For a quantitative analysis we determined the mixture interaction index κ as the relative difference of response magnitude between the mixture and the strongest compound (see Section "Materials and Methods"). The κ index reveals either synergistic (1), hypoadditive (0) or suppressive (−1) odor mixture interaction in each l-and m-PN tested. The l-PNs showed a clear suppressive interaction (κ = −0.98) in line with the prediction from imaging studies. Surprisingly, the m-PNs exhibited a clear hypoadditive interaction with average κ = −0.09 (Figure 5F). This difference between l-and m-PN mixture interactions was signifi cant (P < 0.01, Wilcoxon rank sum test). Synergistic odor mixture interactions were found in neither of the tested PNs. In Table 1 we listed the specifi city index separately for single chemical compounds and mixtures that combined at least three such elementary compounds. It shows that only l-PNs exhibited inhibited responses to mixtures in 12% of all cases. Thus, our results for the l-PNs meet the prediction of previous imaging studies. The fact that we did not observe any suppressed response in all our m-PN recordings is a surprising and unexpected result, marking a functional difference of the two types of PNs.

FAST INHIBITION IN l-PNs COINCIDES WITH LN RESPONSES
What could be the cause for the observed response suppression in l-PNs following odor mixture stimulation? Does suppression become effective at the level of OSNs or could local inhibition within the AL network explain the effect? To answer this question we fi rst tested the intracellular membrane potential of PNs for signs of inhibition. The experimental setting for intracellular recordings from PNs makes it necessary to record from their individual axons. Thus, we recorded the membrane potential distal from the integrating segments and individual synaptic inputs could not be resolved. APs, however, were always detectable and in some cases we could observe clear hyperpolarizing deviations of the membrane potential that can be attributed to inhibitory inputs. In Figure 6A we show six example traces from l-PNs with marked inhibition following stimulus onset. We can now see that the l-PN that innervates the T1-22 glomerulus is actually inhibited in response to the previously examined tertiary mixture. This was Krofczik et al.
Rapid odor processing in antennal lobe not visible in the rate response in Figure 5C because of the low spontaneous activity level. We estimated the onset of the inhibitory compound by thresholding the membrane potential with respect to the mean and standard deviation prior to stimulus onset (see Section "Materials and Methods"). Surprisingly, we found that the response latency of the observed inhibition is on the order of only ∼60-90 ms and thus considerably shorter than the observed excitatory spike responses in PNs (Figure 2). This explains why during mixture response suppression we did not observe a single response spike even in the initial response phase. Could this fast inhibition of PNs be mediated by LNs? To answer this question we analyzed the onset latencies of excitatory spike responses in LNs. Figure 6C shows the average dynamic rate profi le of LNs (green) and PNs (gray). Indeed, the average latencies (vertical dotted lines) of LNs are signifi cantly shorter than those of PNs with an average delay of Δt≈60 ms (Figure 6D). The fast Rapid odor processing in antennal lobe excitatory response of LNs to odor onset is also evident in the individual single trial traces shown in Figure 6B. We may conclude that inhibitory interneurons in the AL can exhibit fast odor responses and are thus capable of rapidly and effectively suppressing PN responses.

DISCUSSION
We investigated the nature and dynamics of the encoding of single compound odors and odor mixtures in two morphological distinct types of PNs of the honeybee using in vivo intracellular recordings.
Our results show that each odor activated a subset of PNs with each PN responding rather non-specifi cally to a number of odors. The spatial representation of odor identity in the PN ensemble evolves rapidly and peaks at about 150 ms after stimulus onset. PN activation involves a distinct odor-specifi c response latency, which implies an odor-specifi c pattern of relative latencies across the activated PN ensemble that is conveyed to higher order olfactory neuropils, the MB and the LH. Odor mixture representation depends on the type of PN, signifying a functional specialization of the two anatomically distinct neuron types. Our results indicate that l-PNs encode odor mixtures synthetically through a suppressed response to the mixture -an observation that was previously made on the level of average glomerular activity using Ca-imaging techniques -whereas m-PNs were never suppressed but encoded odor mixtures elementally, representing the strongest compound within the mixture (hypoadditive mixture interaction). Moreover, our results showed that synthetic odor mixture representation is associated with a prolonged hyperpolarization of the membrane potential, the onset of which in turn matched the stimulus-response latencies of LNs, which are signifi cantly and considerably shorter than the ones observed in PNs. Therefore, we postulate that suppressive odor mixture interactions as observed in l-PNs arise from fast lateral inhibition mediated by the LN network.

COMPARISON WITH IMAGING STUDIES
Part our results show that excitatory single neuron responses to individual odors organize into spatial patterns that involve about half of the PN population. Mixture suppression and inhibitory responses to complex odorants were confi ned to l-PNs. This part of our results is consistent with earlier imaging studies that could demonstrate the spatial representation of odors across T1 (l-PN) glomeruli in the honeybee AL (Deisig et al., 2006;Galizia et al., 1999;Joerges et al., 1997;Sachse et al., 1999) and suppressive odor mixture interaction in T1 glomeruli (Deisig et al., 2006;Guerrieri et al., 2005;Joerges et al., 1997). However, the Ca-imaging methods used to monitor AL activity imposed several technical limitations. First, Ca-imaging yields only a limited temporal resolution that makes it diffi cult to assess the detailed dynamics with which a reliable representation emerges in the glomerular space. Second, during bath application of the Ca dye one generally cannot distinguish the types of neurons that were stained, PNs, LNs (excitatory or inhibitory), or glia (Heil et al., 2007).
Retrograde staining methods are likely to stain several PNs per glomerulus and thus average responses rather than single neuron responses are observed. Also, deliberate staining of LNs alone is diffi cult and thus no distinct conclusions on the dynamic involvement of LNs in the AL computations could be drawn. Finally, the imaging techniques used to measure glomerular activity could only resolve signals from the superfi cially located T1 glomeruli but not from T2-4 glomeruli. Thus, the odor response activity of local populations of m-PNs could not be tested in any of the mentioned studies. To overcome these limitations we carried out intracellular recordings. This allowed to monitor the activity of single identifi ed neuron types (l-PNs, m-PNs, LNs) with single AP resolution and permitted an analysis of the intracellular membrane potential. We preferentially targeted axons in the median antennocerebralis tract to specifi cally study m-PN function. Taken together, the combined results from Ca-imaging and single neuron electrophysiology contribute to a refi ned picture of odorinduced processing in the honeybee AL.

RAPID DUAL ENCODING OF ODOR IDENTITY IN THE PN ENSEMBLE
Our analyses indicate that the PN population in the honeybee encodes odor identity in the fi ring pattern across activated and inactivated PNs as well as in the spatial pattern of relative response latencies in the activated PN ensemble. The spatial pattern of PN fi ring rates evolved rapidly after stimulus onset. Different odors result in a signifi cantly different pattern within few tens of milliseconds as demonstrated by our time-resolved measure of vector distance d. Ninety percent of the maximum distance was reached after about 150 ms. This result is in line with previous studies in the locust where odor classifi cation in extracellularly recorded PN ensembles reached the near-maximum after 100-150 ms (Mazor and Laurent, 2005), in Drosophila where odor tuning in individual PNs was strongest in the period of 130-230 ms after stimulus onset , and in Bombyx mori where a maximal separation of the PN population activity was reached 100-150 ms after response onset (Namiki and Kanzaki, 2008). Note, that we used a non-causal kernel function for fi ring rate estimation (see Section "Materials and Methods"). Strict causality for d in Figures 4 and 5 (τ = 25 ms) adds 42 ms to the estimated times. Both aspects of this coding scheme, rate and latency, also satisfy the behavioral observation of fast odor discrimination within <200 ms (Pamir et al., 2008;Wright et al., 2009). Stimulus encoding by the temporal pattern of stimulus-response latencies has been postulated for different systems (e.g., Chase and Young, 2007;Hopfi eld, 1995). In the olfactory system of rodents, it was recently shown that the response latencies of mitral cells in the olfactory bulb are odor-specifi c (for review see Schaefer and Margrie, 2007). The benefi ts of a latency code are obvious. By definition, the response onset latencies represent the fi rst datum of an ensemble response. A subset of early responding neurons will yield an initially incomplete representation of odor identity that will be available much sooner than the average onset latency, which we estimated to be ∼130 ms (cf. Figures 2 and 6; note that our method has a tendency of slightly overestimating absolute latencies, see Section "Materials and Methods"). Moreover, the pattern of response latencies might readily yield a concentration-invariant representation of odor identity. At the same time, it seems diffi cult to implement a reliable latency code for both, stimulus identity and concentration. We may thus hypothesize that the fi ring rate pattern across the complete PN population acts complementary to the latency code and, with a short delay, provides a refi ned code of odor identity and concentration.
From a functional point of view the important question is whether the information present in the ensemble latency pattern is actually used in the next stage of the insect olfactory system, the MB and the LH, and how a useful read out could be accomplished. For rodents it was hypothesized that downstream neurons could read out a latency code through specifi c delays that involve multi-synaptic transmission and subsequent coincidence detection. However, there are two reasons why the mechanisms of transmitting and reading a latency code are likely to be fundamentally different in insects. In rodents, dynamics of olfactory coding and perception are naturally determined by the sniff cycle, which itself can be controlled by the animal during active sensing (Schaefer and Margrie, 2007;Wesson et al., 2008). In insects, such a dominant rhythm does not exist. Moreover, the anatomical connection scheme is much more complex in rodents than in insects.
In the honeybee, PNs provide input to yet unidentifi ed neurons in the LH and to the Kenyon cells (KCs), the principal cells of the MB. Here the relatively small number of PNs (∼950 in the honeybee) diverge onto a much larger subpopulation of the total ∼160,000 KCs that are assumed to integrate olfactory information. Conversely, each of these KCs receives convergent input from a number of PNs. For the so-called clawed KCs this number was estimated in the order of 10 (Szyszka et al., 2005). Recently (Turner et al., 2008) provided a similar estimate of about 10∼15 PN inputs per KC for the Drosophila. In locust the statistics of convergence might be different (Turner et al., 2008). There, each KC receives input from ∼50% of the 800 PNs with small individual synaptic weights (Jortner et al., 2007). Based on results in the locust we may assume that only a small proportion of KCs respond to a given odor (spatial sparseness, Perez-Orive et al., 2002). Individual KCs responses to single odor stimuli are brief with only a few spikes (temporal sparseness, Perez-Orive et al., 2002;Szyszka et al., 2005). In the locust a mechanism of feedforward inhibition was suggested to explain truncation of KC responses (Assisi et al., 2007;Perez-Orive et al., 2002). In the bee this may be a result of the inhibitory feedback onto the presynaptic boutons of KCs as revealed by EM studies (Ganeshina and Menzel, 2001) and Ca 2+ imaging of the PN boutons (N. Yamagada, personal communication). If we assume a truncated integration time window for KCs, then a KC may detect the fi rst near-coincident inputs from synchronously activated PNs, while desynchronized or delayed inputs may fail to elicit a response. Thus, KC responses might strongly emphasize dynamic changes of the olfactory input, i.e., changes in odorant composition and concentration. In our experimental setting where we probed the responses to a single static stimulus, this property would readily translate the response latency pattern at the level of PNs into a spatio-temporal pattern of briefl y activated KCs in response to the stimulus onset, while the static presence of that stimulus would be largely ignored. In a natural environment, olfactory input changes dynamically (Budick and Dickinson, 2006;Vickers, 2000;Vickers et al., 2001). This would lead to a dynamic pattern of PN activation and inactivation and thus to a dynamically changing composition of the activated KC population that follows signifi cant transients in the input. This scenario predicts that under dynamic input conditions, the number of activated KC may remain low at any given point in time, i.e., preserve a sparse spatial representation in the high-dimensional space of KCs. The spiking activity of individual KCs, however, would increase and exhibit a more complex temporal pattern as was observed for temporally overlapping odor presentations in the locust (Broome et al., 2006).

LOCAL COMPUTATION IN THE AL NETWORK
As to date, our anatomical and functional picture of the AL network with its estimated 4000 LNs is still incomplete. Other than in Drosophila (Wilson and Laurent, 2005) but similar to lobsters and moths (Christensen et al., 1993;Schmidt and Ache, 1996) the circuitry of the honeybee AL involves two morphological distinct types of spiking inhibitory LNs that are distinct by their heterogeneous and homogenous branching patterns (Fonta et al., 1993). Moreover, there exists a division into a GABAergic neuron population (Bicker et al., 1993;Sachse and Galizia, 2002;Sachse et al., 2006) and neurons that may involve histaminergic or/and glutaminergic transmission (Barbara et al., 2005;Bicker et al., 1988;Sachse et al., 2006). The existence of excitatory neurons has recently been reported in Drosophila Root et al., 2007;Shang et al., 2007). Their existence has also been predicted for the honeybee AL for various reasons (Malaka et al., 1995), but clear experimental evidence is still lacking. Several of our observations support this hypothesis. First, we report that the PN response onset, on average, is delayed by about 60 ms with respect to the faster interneuron response onset (cf. Figure 6D). This fi ts well to the assumption that PN activation is at least to some extent mediated via excitatory interneurons. Second, excitatory input from LNs can explain how fast responding inhibitory interneurons could effectively suppress responses in PNs before excitatory input drives PN fi ring (cf. Figure 6A). Third, we observed broad response profi les of individual uniglomerular l-PN and m-PNs across chemically diverse compounds, an observation that has been made repeatedly in insects. This broad tuning could readily be explained if excitation is mediated by LNs . Note that the existence of excitatory interneurons would imply that the individual identifi ed LNs analyzed in our study could be of either type, inhibitory or excitatory.
The most apparent effect of local computation in the AL is the mixture suppression effect where individual uniglomerular l-PNs are inhibited in response to a mixture but not to its individual components. This is supported by a recent Ca 2+ imaging study of PN boutons in the honeybee that suggests that odor complexity is decoded by a subset of KCs preferentially postsynaptic to l-PN boutons (N. Yamagada, personal communication). Our results showed that the onset of intracellular inhibition in l-PNs matched the excitatory response onset of LNs, which in turn was signifi cantly shorter than in PNs (Figure 6). From this we conclude that lateral inhibition is responsible for the observed mixture suppression effect. From Ca 2+ -imaging studies of T1 glomeruli we know that increasing the number of components in the mixture enhanced suppressive interglomerular interaction (Joerges et al., 1997). What could be the behavioral relevance of a mixture-specifi c code? For social insects like the honeybee a unique cue on the complexity of an odor might indeed be advantageous. In particular, this could allow to distinguish simple odors that are relevant in inter-individual communication from natural plant odorants of high complexity.
Taken together, our recordings indicate that interglomerular connectivity within the AL is not merely a means to regulate and normalize the overall activation level of PNs. Rather, LNs are