Sensory Processing in the Drosophila Antennal Lobe Increases the Reliability and Separability of Ensemble Odor Representations

Odor responses are more reliable in PNs than in ORNs

The variability of a neuronal response can be quantified as the variability in the number of spikes evoked by a sensory stimulus. In most sensory systems, the spike-count variability of stimulus-evoked responses increases at each successive processing level in a sensory system^17–19. However, in the Drosophila antennal lobe, ~40 ORNs with the same receptive field converge onto ~4 PNs in each glomerulus^20–22. Thus, by pooling across these inputs, PNs might be able to reduce their response variability. We therefore compared the reliability of odor-evoked spike counts in ORN and PNs.

We presented an odor stimulus in multiple consecutive trials to the same cell (a “block” of trials, Fig. 1a). To quantify spike count reliability across trials, we divided each set of repeated responses into 50-ms windows that overlapped by 25-ms. In each time window, we computed the mean and the standard deviation (SD) of the spike count across repeated responses by the same cell to the same odor. Odors typically evoked more vigorous responses in PNs than in ORNs (Fig. 1a,b; see also Supplementary Fig. 2). So, although the typical SD of PN responses is slightly larger than that of ORN responses (Fig. 1c), PN responses are less variable in proportion to the magnitude of the response (Fig. 1d; p<10⁻¹⁵ whether comparing over the entire stimulus period [gray bar] or the 100 ms epoch at the response peak [black bar], Mann-Whitney U-test, n=779 ORN responses and 843 PN responses). Thus, individual PNs are more reliable than individual ORNs, which should tend to make their responses more informative.

We also compared the SD of ORN and PN spike counts as a function of the mean spike count for each time window. For all mean spike counts, PNs have a lower SD than ORNs (Fig. 1e). Furthermore, the SD is not strongly dependent on the mean, and so stronger responses have a lower CV. Because PN responses are on average stronger than ORN responses (Fig. 1b), this also tends to make PNs more reliable than ORNs.

PNs preferentially transmit rising phase of ORN signals

ORN responses typically do not peak until 100–300 ms after odor onset²¹. This is probably because spiking is coupled to odorant receptor activation by the G-protein-mediated generation of second messengers. However, odors can trigger rapid behavioral responses in flies, with a total latency from stimulus to motor reaction of less than 300 ms²³. This suggests that neurons in the brain are preferentially tuned to detect the rising phase of ORN signals, rather than the response peak. This motivated us to compare the onset kinetics of odor responses in synaptically connected ORNs and PNs.

Comparing peri-stimulus time histograms averaged across all odor responses in all cells, we noted that PN responses rise more rapidly than ORN responses (Fig. 2a). Furthermore, PN responses begin to decay while ORN responses are still growing. This is also clear in most direct comparisons between synaptically connected ORNs and PNs responding to the same odor (Fig. 2b,c and Fig. 1a). Overall, PN responses peak significantly faster than ORN responses (Fig. 2d, p<10⁻⁷, paired t-test, n=69 odor/glomerulus combinations), and the time to half-decay of the response is shorter for PNs than for their presynaptic ORNs (Fig. 2e, p<10⁻⁵, paired t-test). Taken together, faster rise and faster decay mean that more excitatory drive to third-order neurons occurs within an early epoch of the odor response (Fig. 2f, p<10⁻¹¹, paired t-test). Thus PNs act as high-pass filters which preferentially signal the rising phase of the ORN response.

Because PN responses accommodate rapidly, we have chosen to quantify response magnitudes in PNs by measuring average firing rates during an early epoch of the response (a 100-ms time window beginning 100 ms after odor onset, see black bar in Fig. 1b). Because a fruit fly can respond rapidly after encountering an odor, this early epoch should be particularly informative to downstream neurons. Throughout this study, we have also quantified PN responses in a different way: following other investigators^{21, 24}, we have measured average spike rates over the entire 500-ms stimulus period. The major conclusions of this study are the same for both of these response metrics.

ORNs and PNs differ in selectivity and odor preferences

Now setting aside the issues of trial-to-trial reliability and response kinetics, we examined the average response magnitudes for each cell type to our odor stimuli. How does the response profile of each PN type compare to the response profile of its corresponding ORNs? We began by asking simply if these responses are linearly correlated. For each glomerulus we found a statistically significant correlation between the ORN and PN response profile (p<0.05 for all 7 glomerular comparisons, Pearson’s correlation, Supplementary Table 1), but r² values are only in the range of 0.26–0.81. This means that a linear scaling of ORN responses only explains 26–81% of the odor-dependent variance in PN responses.

Two features of PN odor responses diminish this linear correlation. First, for each glomerulus, PNs are less selective than their presynaptic ORNs (p<0.05, Wilcoxon signed rank test, Fig. 3b, similar results in Supplementary Fig. 3). To test the generality of this result, we also compared ORN and PN selectivity for glomerulus DM4 at three different odor concentrations. Like our standard concentration (1:1,000 dilution), weaker stimuli (1:10,000 and 1:100,000 dilutions) produce PN response profiles that are less selective than the corresponding ORN response profiles (Fig. 4). Other investigators using identical stimulus conditions have shown that ORN responses are very sparse at the 1:100,000 dilution, indicating that this concentration is near the bottom of this system’s dynamic range^{21, 22, 24, 25}. These results show that broad PN tuning is not a phenomenon limited to high odor concentrations.

Another factor that diminishes this linear correlation is that the rank order of odor preferences differs for ORNs and PNs. For example, whereas ethyl butyrate is the 3^rd-ranked odor of DL1 PNs, it is only ranked 16^th among the odor responses of DL1 ORNs (Fig. 3a). Some of this difference is merely due to errors in estimating each average response profile on the basis of a limited sample of individual experiments. However, sampling error cannot completely account for this result. This can be shown by pairing an individual ORN with a corresponding individual PN and computing the correlation between their odor ranks, and then comparing the distribution of these correlations with the correlations obtained from ORN/ORN or PN/PN pairings. Because we were not able to test every odor in every experiment, we assembled many simulated response profiles by drawing randomly from a normal distribution defined by the mean and standard deviation of each average response profile (illustrated in Fig. 5a, see also Supplementary Methods). The median correlation between ORN and PN ranks was only 0.47, which is substantially lower than the correlation between ORNs of the same type or between PNs of the same type (0.65 and 0.61 respectively, Fig. 5). The simplest explanation for this result is that a PN’s odor preferences are significantly influenced by lateral connections between glomeruli^{26, 27}.

A nonlinear transformation function for each glomerulus

Clearly, the output of a glomerulus is not a simple a linear scaling of its inputs. Furthermore, because ORNs and PNs differ in their ranked odor preferences, no monotonic function can describe the relationship between the ORN and PN response profile for a glomerulus. We therefore asked if there is any systematic relationship between ORN and PN responses. For each glomerulus, we plotted PN responses to each odor as a function of ORN responses to the same odor (Fig. 6a). This revealed a consistent transformation function for each glomerulus, albeit with some scatter. These functions have a similar shape for most glomeruli: they initially slope steeply, meaning that the gain of the transformation function is high for weak inputs. As ORN input levels increase these curves flatten, meaning that the gain of the transformation function decreases. (This is true for all glomeruli we tested except DL1.) In sum, these plots show that PNs inherit much of their tuning from their presynaptic ORNs, but the transformation is highly nonlinear.

This type of transformation function may be useful because it makes better use of a PN’s available response range. This is illustrated by projecting the points in each plot onto both the x- and y-axes (Fig. 6a, green and magenta symbols). ORNs are not using all parts of their dynamic range with equal frequency, since most of our stimuli elicit responses confined to a narrow range of firing rates. However, two odors that elicit similarly weak levels of activity in an ORN (green) can elicit quite different levels of activity in a postsynaptic PN (magenta), because each glomerular transformation function shows high gain at low ORN input levels. This tends to distribute the responses to these stimuli more uniformly throughout the response range of each PN.

This type of sensory transformation has been termed “histogram equalization”²⁸ because it produces a flatter histogram of response intensities. To examine whether this is occurring across the entire population of cells in our data set, we plotted the distribution of response intensities for ORNs and PNs, accumulated across all odors and all glomeruli (Fig. 6b). The ORN response histogram shows a large peak at low response intensities, and a long flat tail covering the rest of the ORN dynamic range. The PN response histogram, by contrast, is much flatter, meaning that all available response intensities are used with more uniform frequency. In this sense, PNs encode our odor stimuli more efficiently than ORNs do.

Although there is clearly an overall relationship between ORN and PN responses for each glomerulus, it is also important to note that these functions do not predict PN odor responses completely. Because ORNs and PNs differ in their ranked odor preferences, no monotonic function will account for all this data.

Odor representations in multi-glomerular coding space

Third-order neurons receive convergent input from multiple PN types^29–32. It is therefore important to examine odor representations in multi-glomerular coding space. In the simplest case, histogram equalization in one dimension should also produce a more uniform distribution of odors in multiple dimensions. In order to visualize odor representations in 7-dimensional space, we reduced the dimensionality of this space by performing principal components analysis (PCA). The first two principal components define the two-dimensional projection that maximizes the variance of the data. In this projection, most odors are still clustered near the origin of ORN space, with only a few odors located far from this cluster (Fig. 7a). In the equivalent PN space, odors fill the available coding space more uniformly (Fig. 7b).

Thus, as a result of the ORN-to-PN transformation, odor representations are distributed more efficiently in multi-glomerular coding space. In concrete terms, the ensemble patterns of spiking activity elicited by any two odors become more different. We quantified this by measuring Euclidean distances between odors in 7-dimensional space for all possible pair-wise combinations of odors ([18 choose 2]=153 pairs). During the early epoch of odor responses (black bar), distances are significantly larger in PN space as compared to ORN space (p<0.0001, paired t-test, n=153). Moreover, the distribution of distances is narrower for PNs than for ORNs over the entire stimulus period (note inter-quartile range in Fig. 7c versus 7d). This means that odors are distributed more uniformly in PN coding space. Some odor distances decrease, but others increase.

Is the separation of odors in multi-dimensional space larger or smaller than we would predict simply on the basis of independent odor separation in each one-dimensional coding channel? The answer depends on the degree of correlation between different glomeruli. Lateral connections clearly shape PN odor responses (Fig. 5 and refs. ^{26, 27}); if these connections increased correlations between different PN types, this would decrease inter-odor distances in multi-dimensional space. To address this issue, we constructed a simulated data set that preserves the distribution of response magnitudes for each glomerular cell type, but breaks any dependencies between odor responses in different glomeruli. We achieved this by independently shuffling the odor labels on each glomerular response profile. We then measured inter-odor distances in 7-dimensional space for all possible pair-wise combinations of odors. When we repeated this simulation many times, the range of distances we obtained was indistinguishable from the distances we measured in our real data set (Fig. 7c,d). This means the separation between ensemble odor representations is roughly what we would predict based simply on histogram equalization in each glomerulus individually.

Correlations between cell types and odors

We have seen that PNs use all portions of their dynamic range with approximately equal frequency, and in this sense encode odors more efficiently than ORNs do²⁸. However, the term “efficient coding” has also been applied to the idea that the responses of different neurons should be maximally independent from each other³³. We measured the independence of different glomerular coding channels by computing the percentage of the variance in the ensemble odor responses that is captured by each of the 7 principal components (PCs) of the 7-dimensional ORN or PN coding space. If all 7 cell types were completely correlated, then the first PC would account for 100% of the variance in the data. In other words, all the data would lie along a single line in multi-dimensional space. Conversely, if all cell types were perfectly decorrelated, and if the data was distributed according to a multi-dimensional Gaussian, then each PC would account for an equal amount of the total variance (100% ÷ 7 = ~14%). (Even in this case, we would need a very large odor set to discern this perfect decorrelation.)

The PCs of our data set fall between these hypothetical extremes (Fig. 8a). In part, this simply reflects the limited size of our odor set and the non-Gaussian distribution of the ORN response histograms (see Fig. 6b). We demonstrated this by independently and randomly shuffling the odor labels on each of the 7 ORN response profiles and recomputing the PCs of this simulated data set. These simulations always produced a first PC that accounted for a disproportionately large share of the variance (usually 30–40%, Fig. 8a). However, the PCs of the real (non-shuffled) data set are even more skewed, with the first PC accounting for 54% of the variance. This means that ORNs in different glomeruli have odor preferences that are more highly correlated than we would expect based simply on the distribution of response magnitudes in each glomerular channel.

Similarly, about half the variance in the PN data is captured by the first PC (51%, Fig. 8b). This is mainly due to the limited size of our odor set and the non-Gaussian distribution of the PN response histograms, since shuffling odor labels on each PN response profile always produced a skewed distribution of PC contributions (Fig. 8b). Because real PN data produced a distribution that was even more skewed than the simulated data, PNs (like ORNs) are more correlated than we would expect based simply on the distribution of response magnitudes for each PN type.

In sum, sensory processing in the Drosophila antennal lobe does not change the degree of independence between different glomerular coding channels. The conclusions of this analysis are similar regardless of whether we measure spike rates around the response peak (Fig. 8), or over the entire stimulus period (Supplementary Fig. 6).

PN odor responses are more separable than ORN responses

Increased PN reliability and more uniform odor distances in PN coding space should mean that odors are more discriminable on the basis of PN spike trains as compared to an equivalent number of ORN spike trains. We tested this prediction by measuring the ability of an algorithm to identify the odor stimulus based on the ensemble neural response elicited by that odor. Because our data come from single (not multi-unit) ORN and PN recordings, we simulated “multi-unit” responses by assembling data from different glomerular classes. Each simulated data set comprised 90 multi-unit responses (18 odors, 5 spike trains per odor per cell). We performed linear discriminant analysis to identify the linear combinations of input variables that best separated all 18 odor response clusters from each other. To evaluate the quality of these discriminations, we held out one multi-unit odor response from the data set, trained the algorithm with the remaining 89, and predicted the odor corresponding to the one held-out response. The predicted odor was then compared to the actual odor. We repeated this analysis with many independently assembled multi-unit responses at each time point in the odor response.

Prior to odor onset, the prediction success rate hovers near chance (Fig. 9a). (The success rate is slightly above chance because different cells had somewhat different spontaneous firing rates, and spontaneous firing rates sometimes drifted during an experiment; thus, spontaneous firing rates were slightly “predictive” of the odor because successive trials with an odor were presented consecutively rather than interleaved). After odor stimulus onset, success rates rise rapidly. As expected, including more glomerular classes in the data set produces higher success rates (Fig. 9b). Because success rates using PN data hit a ceiling at 100% for some classifications using data from >3 glomeruli, this procedure underestimates the difference in discriminability between ORN and PN responses. Nevertheless, success rates are significantly higher for PN data than for ORN data for all conditions in Fig. 9b (p<0.005, Mann-Whitney U-tests, n=20 runs of the classification procedure for ORNs and PNs for each condition). This demonstrates that a linear discriminator can classify odors more accurately with responses from several PNs than with responses from the same number of ORNs.

An improved signal-to-noise ratio

Studies in other systems have implied that the variability of stimulus-evoked spike counts almost always increases at successively higher levels of sensory processing³⁴. For example, visual responses of higher cortical neurons are often very noisy¹⁷, in contrast to the reliability of retinal ganglion cells³⁵. A direct comparison of responses evoked by identical stimuli in the retina, thalamus, and visual cortex has confirmed that spike-count reliability decreases at each successively higher level of the visual stream¹⁸. This is despite the fact that a simple cell in primary visual cortex pools signals from ~30 thalamic neurons³⁶, which ought to improve its reliability. Similarly, a direct comparison of spike trains at successive levels of an insect auditory circuit has found that noise increases at successively higher levels¹⁹. Our results show a different trend: spike counts in individual PNs are significantly more consistent than spike counts in individual ORNs. This is partly because PNs tend to fire more vigorously than their presynaptic ORNs in response to the same stimulus, and stronger responses are more reliable for both ORNs and PNs. This may imply an increasingly deterministic control of spike timing at high firing rates due to intrinsic refractoriness³⁷. However, even at the same firing rates, PN responses are more reliable than ORN responses. This may reflect the benefits of pooling: each PN is postsynaptic to many ORNs, and all these ORNs respond similarly to odors^{21, 22, 38}. If noise is uncorrelated across ORNs, then pooling these inputs should tend to improve the reliability of PN responses.

On balance, the improvement in reliability is smaller than one might predict. Each glomerulus corresponds to ~40 ORNs and ~4 PNs; this means the average PN pools inputs from 10 to 40 ORNs (depending on whether each ORN contacts all PNs in a glomerulus). Since pooling N ORN inputs should decrease the variability of the pooled average by √N, we would expect the coefficient of variation to improve by √10 to √40. The effect we describe is on the low end of this range, suggesting that each ORN contacts only a single PN, or that PNs receive additional noise from other neuronal sources.

High-pass filtering of olfactory signals

Our results show that PNs can be extremely sensitive to small differences between weak levels of ORN input. Even a small increase in ORN spike rate above baseline can produce a robust response in postsynaptic PNs. As a result, PN responses rise rapidly even when ORN responses build slowly. This is particularly useful because the onset kinetics of ORNs are intrinsically limited by the speed of signal transduction cascades linking odorant receptor activation to spike initiation. PN responses then rapidly decline while ORN spike rates are still rising. This means that PNs act as high-pass filters, transmitting the rising phase of ORN responses preferentially over the tonic component of ORN responses. This rapid accommodation might be due to any of several mechanisms, including for example short-term synaptic depression at the ORN-to-PN synapse.

Taken together, a faster rise and a faster decay should sharpen the estimate of odor arrival time by downstream neurons. For a fly in flight, this should translate into an improved estimate of odor plume location. Notably, Drosophila can turn in flight less than 300 ms after encountering an odor plume²³. A similar phenomenon operates in the visual system, where sluggish photoreceptor responses trigger speedy depolarizations in downstream neurons³⁹ and ultimately rapid behavioral responses to visual stimuli.

We note that Drosophila PN responses differ from the responses of locust PNs, which typically show more complex temporal patterning^{29, 40, 41}. Locust PNs also show a higher average level of maintained activity throughout the odor response (relative to response peak) and also often display excitatory responses to odor offset⁴¹. By contrast, Drosophila PNs accommodate rapidly and typically do not burst after stimulus offset (but for some exceptions see Supplementary Fig. 2f).

A nonlinear transformation increases coding efficiency

A major finding of this study is that although PNs inherit a substantial portion of their odor tuning from their presynaptic ORNs, this relationship is highly nonlinear. This nonlinearity disproportionately amplifies small differences between weak levels of ORN input. By contrast, small differences between strong ORN inputs are not amplified to the same degree. Most ORN odor responses cluster in the weak end of the ORN’s dynamic range. As a result of this nonlinear transformation, PNs use their dynamic range more uniformly than ORNs do. If all portions of a neuron’s dynamic range are used with equal frequency, the carrying capacity of that information channel is maximized because the entropy of the neuron’s response is maximized. This tends to protect signals from contamination by noise added at later stages in the processing channel⁴². This has long been recognized to be a useful computation in sensory processing²⁸. If broader tuning curves might be useful, why has evolution not simply produced broadly tuned ORNs? ORN responses are directly linked to the way odorant receptor proteins interact with odor molecules; therefore broadening ORN tuning might require changing the biophysics of odorant receptors in ways that are unfavorable for other reasons.

Broad PN tuning may seem counter-intuitive: we tend to think of higher-order neurons as being more selective than their presynaptic inputs. These expectations are founded in part on the paradigm of visual processing, where successive layers of higher-order neurons are increasingly specialized to represent complex conjunctions of visual features. However, there is also a huge expansion in the number of higher-order neurons devoted to representing complex visual features as compared to the number of retinal ganglion cells. Thus the dimensionality of higher visual representations is increasingly large, so these brain regions can afford to code information increasingly sparsely as ensembles of narrowly-tuned neurons. By contrast, in early olfactory processing, the dimensionality of the second-order representation is the same as the dimensionality of the first-order representation. Thus, total coding space cannot increase (unless time is used as an additional coding dimension⁴³).

It is important to note that, in a truly efficient coding scheme, neurons should efficiently encode natural stimulus distributions, not arbitrary stimulus distributions²⁸. Although our odor set is chemically diverse and relatively large, it may not be representative of the odors a wild fly would encounter. In the future, it would be interesting to learn whether the principles of olfactory processing we describe here also apply to a more naturalistic distribution of odor stimuli.

Another caveat is that we have not sampled all antennal lobe glomeruli. However, because most of the glomeruli in our data set showed a similar nonlinear transformation function, our conclusions likely generalize to many glomeruli outside our data set. An interesting special case is glomerulus DA1, which was not included in our data set. ORNs projecting to this glomerulus respond weakly but highly selectively to a Drosophila pheromone. Their postsynaptic PNs are also highly selective for this odor, but respond much more robustly⁴⁴. Thus, this processing channel exhibits amplification without a change in response selectivity.

Ensemble odor responses

Extending our analysis from one glomerular channel to multi-glomerular ensembles, we found that odors are more uniformly separated in ensemble PN coding space as compared to ensemble ORN coding space. This separation is approximately what we would expect based on the increased separation between odors within each glomerular coding channel. We demonstrated this using a simulation that made the odor preferences of each glomerulus independent from each other, while preserving the characteristic response magnitudes of each cell type.

We also found that sensory processing in the antennal lobe does not substantially alter the independence of different glomerular coding channels. The degree of correlation between different PN types is similar to the degree of correlation between different ORN types. Interestingly, ORNs have more correlated odor preferences than we would expect simply on the basis of their tuning breadth and the size of our odor set. We also obtained the same result by re-analyzing a large published data set²⁴ comprising 24 Drosophila ORN types and 110 odors (result not shown). This result may be due to the common evolutionary origin of different Drosophila odorant receptors in gene duplication events⁴⁵. In addition, some odors are intrinsically more volatile than others, which will tend to produce similarities in the odor preferences of different glomeruli. Like ORNs, PNs are also somewhat more correlated than we would expect based on their tuning breadth and the size of our odor set. This may simply reflect correlations inherited from ORNs.

The role of lateral connections between glomeruli

Glomeruli in the Drosophila antennal lobe are connected by both GABAergic interneurons^{20, 46} and cholinergic interneurons^{26, 27}. What is the role of these connections in the transformations we have described?

We have noted that the rank order of PN odor preferences is different from the order of ORN odor preferences. We show that this difference is too large to be explained by the uncertainty in our estimates of each average odor response profile (Fig. 5). Therefore, some of this difference is probably caused by lateral inter-glomerular connections, because lateral inputs would have an odor tuning that reflects the odor preferences of ORNs presynaptic to other glomeruli. In principle, either inhibitory or excitatory lateral connections could cause this phenomenon. The computational significance of this phenomenon is not clear, since it does not seem to decorrelate the responses of PNs in different glomeruli (Fig. 8).

It is easy to see how lateral connections could cause scatter around each glomerular transformation function. However, lateral connections may also play an important role in determining the underlying shape of these transformation functions. For example, all of these functions have a y-intercept >0 (Fig. 6). This reflects the tendency of PNs to respond weakly to an odor even when their presynaptic ORNs are not responding at all. Lateral excitatory connections are strong enough to trigger these responses^{26, 27}. Moreover, lateral inhibition could act on ORN axon terminals to govern the probability of neurotransmitter release, and thereby contribute to a nonlinear relationship between pre- and postsynaptic activity. At the analogous synapse in the olfactory bulb, ORN-to-mitral cell synapses show strong frequency-dependent short-term plasticity that is modulated by presynaptic inhibition via GABAergic local neurons⁴⁷. Thus, we should not assume that the systematic relationships in Fig. 6 are purely intrinsic to each glomerulus. More mechanistic experiments will be required to disentangle the role of intra- versus inter-glomerular mechanisms in shaping these transformation functions.

Odor discriminability

We have shown that a linear discriminator can identify odors more accurately on the basis of PN spike trains than on the basis of an equivalent number of ORN spike trains. This is likely due to both the increased distances between odors in PN space, and the improved signal-to-noise ratio among PNs. It is important to point out that linear discriminant analysis is not meant to emulate a biologically plausible downstream neuron, and that real third-order neurons will be subject to more constraints than our algorithm is. Also, this is not an optimal decoder, and so its performance may not reflect the total amount of information in the responses it decodes. Finally, the total amount of information in the entire PN ensemble cannot, of course, exceed the total amount of information in the entire ORN ensemble. What we have shown here is that the information in a limited subset of the PN ensemble is more useful for a linear decoder than the information in an equivalent number of ORNs. This is an important demonstration of the potential functional consequence of increased PN reliability, combined with increased inter-odor distances in PN coding space.

In conclusion, we have described two fundamental tasks accomplished by the first stage of the olfactory processing stream. On a neuron-to-neuron basis, our comparisons show that signal reproducibility is increased, and distinctions between the responses to different stimuli are enhanced. The details of odor processing in the vertebrate olfactory bulb could be very different, especially because the number of glomeruli in vertebrates is much larger. Nevertheless, most organisms share a common olfactory processing architecture, suggesting that some of the basic principles we have demonstrated in flies may also apply to vertebrates.

Fly stocks

Flies were reared at room temperature on conventional cornmeal agar. All experiments were performed on adult female flies 2–7 days post-eclosion. Supplementary Table 4 lists genotypes for all experiments. See Supplementary Methods for stock origins.

ORN recordings

Flies were immobilized in the trimmed end of a plastic pipette tip under a 50× air objective mounted on a Olympus BX51WI microscope. A reference electrode filled with saline was inserted into the eye, and a sharp saline-filled glass capillary (tip diameter < 1μm) was inserted into a sensillum. Recordings were obtained with an A-M Systems Model 2400 amplifier, low-pass filtered at 2 kHz, and digitized at 10 kHz. ORN spikes were detected using routines in IgorPro (Wavemetrics). See Supplementary Methods for details.

PN recordings

Whole-cell recordings from PN somata were performed in vivo as previously described⁴⁶. One neuron was recorded per brain and the morphology of each cell was visualized post hoc with biocytin-streptavidin and nc82 histochemistry as described previously⁴⁶, except that in the secondary incubation we used 1:250 goat anti-mouse:AlexaFluor633 and 1:1000 streptavidin:AlexaFluor568 (Molecular Probes). See Supplementary Methods for details.

Olfactory stimulation

We chose a panel of 18 odors to maximize the chemical diversity of our stimuli, and to maximize overlap with odors used in other studies of the same ORNs^{24, 22}. For all experiments (except in Fig. 4), odors were diluted 1:100 v/v in paraffin oil (J.T. Baker, VWR #JTS894), except 3-methylthio-1-propanol, which was diluted 1:100 v/v in water, and 4-methyl phenol, which was diluted 1:100 w/v in water. In Fig. 4, odors were diluted 1:100, 1:1000, or 1:10,000 in paraffin oil. See Supplementary Methods for details.

Data analysis

See Supplementary Methods for data analysis details.