A topographical organization in the primary olfactory cortex

Abstract

A key principle of most sensory cortices is their topographic organization, in which neurons in proximity exhibit similar tuning properties, typically varying along a specific stimulus feature. However, such an organization has not yet been identified in the olfactory system. Here, we developed a method that reveals the set of glomerular inputs to multiple Piriform Cortex neurons. We found that, on average, each Piriform neuron receives input from ~60 glomeruli. Putative interneurons tend to receive input from more glomeruli. Most neurons were activated when multiple distinct subsets of 2-4 glomeruli were active. Notably, the input glomeruli of nearby Piriform neurons were positioned close to each other, sharing a few glomeruli. The similarity of glomerular input maps decreased with increased distance between Piriform neurons. Furthermore, this similarity principle aligned with an increase in odor-tuning similarity of nearby neurons in both awake and anesthetized mice. These results indicate that the primary olfactory cortex adheres to an organizational principle based on how it processes glomerular inputs.

Here, the authors develop a method which reveals that, contrary to the prevailing assumptions of spatial randomness in piriform cortex, there is a weak but significant topographical organization in which nearby neurons receive input from nearby glomerular patterns and exhibit more similar odor responses.

Introduction

Topographical mapping is a fundamental organizational principle of the mammalian sensory cortices. For example, in vision, adjacent locations in space are mapped to adjacent locations in the primary visual cortex in a retinotopic fashion. In contrast, olfaction may involve a weak chemotopic map in the olfactory bulb^1–3, but there is no evidence for retaining such mapping in downstream olfactory regions. In fact, a preponderance of anatomical and physiological evidence argues against a spatial organization in the olfactory cortex^4–13. However, organization in the olfactory cortex may be based on a principle that cannot be captured with current anatomical tools or odor stimulations. Several recent studies reported organization in the projections to the Piriform cortex (PC)^14–17, strongly suggesting that location within PC matters and that there is some functional organization within PC. This functional organizational principle has not yet been found.

Here, we revisited the question of structural organization in the olfactory cortex circuit. We hypothesized that the organizational principle is not based on physicochemical properties but on how PC neurons integrate inputs from the olfactory bulb (OB). Using optogenetics, we examined the computations of PC neurons in response to their OB inputs. We developed a technique and methodology to uncover the receptive fields (RFs) of multiple recorded olfactory cortical neurons by light-stimulating multiple glomeruli in mice expressing Channelrhodopsin in their Mitral/Tufted (MT) cells¹⁸. The RF of a sensory neuron is the region of the sensory space (e.g., the body surface or the retina) in which a stimulus alters the neuron’s activity. We extended the concept of RF to functional connectivity between neurons: If many neurons form synapses with a single downstream cell, they collectively form that cell’s RF. Once the RF of a neuron is revealed, it can be used to estimate the strength and sign of the neuron’s functional connections with each of its inputs and to predict the neuron’s response to novel stimuli¹⁹. Applying this method to anterior PC (aPC) neurons, here we found that aPC neurons receive input from several dozen glomeruli and respond when several distinct subsets of these glomeruli are active. Narrow-fast-spiking neurons receive input from a larger number of glomeruli than non-narrow-fast-spiking neurons. Strikingly, we found that aPC is topographically organized: Spatially close aPC neurons tend to have more similar RFs and to share more glomeruli than expected by chance. We further show that this increased similarity in RF maps is consistent with the increased similarity of odor tuning curves of nearby aPC neurons in both awake and anesthetized mice.

Results

A method for revealing the receptive field of olfactory cortex neurons

Under the assumptions of the linear-nonlinear Poisson model, the RF of a neuron can be estimated using the Spike-Triggered-Average (STA) technique²⁰. In this method, a large number of stimuli that span the stimulus space is applied. Computing the weighted sum of stimuli provides a characterization of the RF of all the simultaneously recorded neurons (Fig. 1a and Methods). The outcome of this process is a 2-dimensional map in which the value of each pixel is proportional to the synaptic strength between the neurons below that pixel and the recorded target neuron. Since odor stimuli cannot be delivered systematically to cover the odor space or in their thousands, we used optogenetics. Optogenetics converts the OB into a retina-like system, thereby circumventing this obstacle. We delivered thousands of random light patterns while recording single and multi-neuron spiking activity in anesthetized mice (Methods). This technique differs from previously used sequential or laser-scanning techniques^18,21–23 because it reveals all possible patterns that activate the neuron; it provides a model that can predict the neuron’s response to novel stimuli and allows comparison of the RFs of all simultaneously recorded neurons.

Fig. 1. Piriform Cortex Neurons Receptive Fields.

a Schematic diagram illustrating the experimental setup. Tbet-Cre mice expressing Channelrhodopsin2 in MT cells. Middle: Illustration of the stimulation sequence of random sets of glomeruli on the OB dorsal surface while recording neural activity in the aPC. Light patterns containing multiple square light patches are projected on the exposed OB surface. The spikes elicited by each light pattern and the light signal are illustrated in the lower traces. The pixel resolution is 20 µm. Right panel: Illustration of a hypothetical glomerulus. The center of a light patch can hit this glomerulus at 25 different locations. b Three examples of aPC neuron RF maps from three different mice. The first row shows the raw RF maps, the second row shows the z-scored maps, and the lower row shows the glomeruli-fitted maps. The color bars here and in subsequent figures represent the following: The color bar of the raw RF map denotes the mean firing rate of each pixel over all light patterns; this includes patterns that did not elicit a response, and therefore, the mean firing rate is lower than the actual value. The color bar of the z-scored map denotes the z-values computed from the RF map (Methods). The color bar of the glomeruli-fitted map displays the values of the 2-D Gaussians fitted to each peak in the z-scored map (arbitrary units - AU). c Upper: Distribution of the excitatory and inhibitory glomeruli number in the glomeruli-fitted maps of all converging aPC neurons (N = 50). The 25th–75th interquartile range is indicated by a thin vertical line. The median is marked by a dot, and its confidence interval is shown as a thicker bar; individual data points are overlaid. Lower. Number of excitatory and inhibitory glomeruli in each of the 50 neurons. Overlapping points were slightly jittered for visibility. The number of inhibitory and excitatory glomeruli is positively correlated (Pearson correlation). d Z-score values of all excitatory and inhibitory synaptic weights of the reconstructed glomeruli, sorted by the strength of all recorded neurons. The thick black line marks the mean. Only one neuron had more than four inhibitory glomeruli.

To tune the technique parameters, we first applied it to MT neurons in anesthetized mice, where we expected to find one excitatory glomerulus in the vicinity of the electrode and several inhibitory glomeruli²⁴. We used square light patches of 100, 120, 140, or 160 µm². The light patches could occur anywhere on the OB dorsal surface, but did not overlap. Previous studies used a sequential scan of the dorsal OB surface, dividing it into n by m non-overlapping patches, typically of the size of a glomerulus^{18,22,25–27}. Our projector system has a resolution of 20 µm/pixel, and since a glomerulus is estimated to be ~100 µm in diameter, our stimulation protocol could hit a glomerulus at ~25 locations (Fig. 1a). This high-resolution map nicely revealed the glomerular ellipsoid structure (Supplementary Fig. 1a, left panels).

To estimate the location and number of glomeruli in this relatively high-resolution RF map, we converted the RF map values to z-values. We then filtered out isolated pixels, assuming that a suspected glomerulus must contain at least several pixels, and set values below a certain threshold to zero (Methods). We termed this RF map the ‘z-scored’ map (Supplementary Fig. 1a, middle panels). To estimate glomerular locations from the z-scored map, we developed an iterative algorithm that fits 2D ellipsoid Gaussians to each local maximum and minimum (Methods). The Gaussian fit was performed on 80% of the data. We evaluated fit quality by assessing how well the reconstructed map predicted the firing rates elicited by the remaining 20% of light patterns. We termed this map the ‘glomeruli-fitted’ map.

As expected, most recorded MT cells had an RF-map composed of one excitatory glomerulus (mean ± std = 1.1 ± 0.3, median = 1, N = 15 MT cells from five mice; Supplementary Fig. 1b). The number of inhibitory glomeruli was low (0.5 ± 0.8, median = 0, Supplementary Fig. 1b), which is consistent with recent papers showing activity-dependent weak lateral inhibition between MT cells^{23,24,26–28}. The inhibitory effect could alter the temporal dynamics of the MT cell response—delaying the MT response or shortening the response time without affecting the peak response (Supplementary Fig. 1c, d). Interestingly, two MTs had two adjacent excitatory glomeruli (Supplementary Fig. 1e). Finding more than one glomerulus in a small number of OB neurons could be the result of an artifact caused by the stimulation method (light hitting the superficial dendrites of the recorded MT in its vicinity or reflected from the electrode placed in the OB), clustering two MT cells from two different glomeruli as one, or the sensitivity of our fitting algorithm that might consider an elongated glomerulus as two glomeruli. The median evoked firing rate elicited by light stimulations hitting the estimated location of the glomerulus above the recorded MT was 15.3 spikes/s (mean 17.8 spikes/s), which is within the range of previously reported odor-evoked responses in anesthetized mice²³ (Supplementary Fig. 1f).

PC neurons receive excitatory and inhibitory inputs

Next, we applied the STA method to aPC neurons. We recorded 34 aPC neurons using a tungsten electrode in anesthetized mice and 38 using a custom-made tetrode bundle. Figure 1b shows three examples of PC neurons’ RF-, z-scored-, and glomeruli-fitted- maps from three different mice. Each neuron receives input from a different number of glomeruli dispersed over the OB surface. Both excitatory and suppressive glomeruli were detected. Unlike MT neurons, PC neurons require more light-patterning stimuli to reach a stable RF map. We considered a neuron stable-convergent if its RF and z-scored map constructed from the first half of the stimuli resembled the map built from the second half (see Methods). Of the 72 recorded aPC neurons, 50 (69%) met this criterion (Methods). The convergence scores varied considerably (Supplementary Fig. 2). The main reasons for the lack of convergence are the use of too few light patterns or insufficient neuron stability (Supplementary Fig. 2; Methods). Analyzing all converging neurons, we found that each aPC neuron received, on average, excitatory input from 4.9 ± 2.4 glomeruli (mean ± std, median = 4) and inhibitory input from 2.2 ± 1.2 glomeruli (median = 2, Fig. 1c upper panel). The number of inhibitory glomeruli scaled with the number of excitatory glomeruli (r = 0.37, P = 0.008, N = 50, Pearson correlation, Fig. 1c lower panel). There was no correlation between the number of glomeruli found and the light patch sizes (P > 0.3). To examine whether the spatial distribution of glomeruli is randomly scattered on the OB surface or has some organization, we compared the pairwise distances between all pairs of glomeruli within an RF map to the distances expected by assuming a random location. We found that these two distributions did not differ significantly (median P across 30 simulations ~0.2; KS test, Supplementary Fig. 1g), suggesting that aPC neurons integrate from random sets of glomeruli. Examining how the response to light stimuli depends on the respiration phase revealed that some neurons responded to light stimuli at all phases, others elicited a stronger response at the neuron’s preferred phase, and some at the neuron’s non-preferred phase (Supplementary Fig. 3a). Overall, the response to light was moderately correlated with the respiratory dynamic (Supplementary Fig. 3b). The convergence scores were similar when we computed them based on light stimulations that occurred in the first and second half of the respiration phase (r = 0.95, P = 4.4e-38, N = 72, Pearson correlation, Supplementary Fig. 3c) suggesting that the exact respiration phase of the light stimulation does not strongly affect the RF map. We conclude that aPC neurons are connected to a relatively small number of glomeruli that are randomly dispersed on the dorsal OB surface, and that these glomeruli can exert either excitatory or inhibitory effects.

PC neurons respond to multiple distinct patterns composed of small subsets of glomeruli

The STA method can reveal how stimulating a single glomerulus or a combination of glomeruli affects an aPC neuronal response. To examine this, we plotted the neuron’s PSTH while restricting our analysis to light patterns containing light patches that hit a particular glomerulus or glomeruli combination. Figure 2a shows the responses of an aPC neuron for which we have identified presumably seven excitatory contributing glomeruli. Plotting the response of this neuron to all light patterns that contained a light patch that hit one of the six selected glomeruli that did not hit the other five, and the response to light patterns that hit all six glomeruli simultaneously shows that this PC neuron did not respond when one glomerulus was stimulated but responded strongly when all six glomeruli were stimulated (Fig. 2a left panel). Considering all possible combinations of the six glomeruli shows that this aPC neuron started to respond when at least three glomeruli were stimulated, and its response increased as more glomeruli were stimulated (Fig. 2a, middle and right panels). Not all triples of glomeruli activated the neuron. The example in Fig. 2b shows the response of an aPC neuron that receives input from presumably six excitatory glomeruli. This neuron responded weakly when one of these glomeruli was light-stimulated, and the response increased moderately as two or more glomeruli were stimulated. The response peaked when all glomeruli were photo-stimulated. The example in Fig. 2c shows a neuron receiving input from presumably nine excitatory glomeruli. This neuron responded weakly when three or four glomeruli were light-activated. Examining neurons containing inhibitory glomeruli in their RF map suggests that light activation of inhibitory glomeruli can reduce the neuron’s firing rate or change its temporal dynamics (Supplementary Figs. 4a, b). Plotting the normalized response of all aPC neurons as a function of the number of excitatory glomeruli out of six indicates that aPC neuron responses increase moderately with the number of stimulated glomeruli (Fig. 2d). The rate of increase may appear to saturate as more glomeruli are activated. However, since the number of trials in which light patterns activate more than six glomeruli decreases exponentially, this could be due to an analysis based on a small number of points.

Fig. 2. Individual aPC neurons respond to the activation of distinct glomeruli subsets.

a–c Analysis of the response of three aPC neurons to light stimulation that hit different glomeruli patterns. The left panel shows the PSTHs (mean ± 2 S.E.M) of the responses to light patterns that hit each glomerulus without hitting the other glomeruli, and when the light pattern stimulated all marked glomeruli together. For clarity, only the SEM of the response to all six glomeruli is shown (gray shading around the black line). The inset shows the glomeruli-fitted map, with the selected glomeruli highlighted. The numbers in the legends indicate the number of light patterns that contain light patches hitting these specific glomeruli. The middle panel shows the response to all glomeruli combinations sorted by ascending order of the number of stimulated glomeruli. The number of stimulated glomeruli is marked on each row. The right panel shows the evoked firing rate as a function of the number of glomeruli that were light-stimulated. Only combinations with at least five repetitions are shown. Red circles mark the mean. d aPC response increased as a function of the number of excited glomeruli. Shown are the normalized responses (mean ± 2 SD) of all aPC neurons to light-pattern combinations targeting the excitatory glomeruli (N = 50). The maximum number of glomeruli that could be investigated was six because the probability of hitting seven or more glomeruli with light patterns containing ~10 light patches is low, given the number of light stimulations we used (Methods). e Firing rate as a function of the sum of the synaptic strengths of the light-stimulated glomeruli in the neuron shown in (a). Pearson correlation value is shown. Colors mark the number of glomeruli stimulated. f aPC response increases as the sum of the glomeruli’s synaptic strengths increases. Comparison of linear regression models explaining the elicited firing rate quantified by the adjusted R². The regression model based on the sum of synaptic strengths as a predictor is better than the model based on the number of light-activated glomeruli (P = 0.0004, two-sided paired t-test, df = 48, t = −3.6, N = 49 neurons with at least two excitatory glomeruli).

The response of an aPC neuron is likely to depend more on the total synaptic input it receives than on the number of activated glomeruli. To examine this, we plotted the aPC firing rate of the example shown in Fig. 2a as a function of the sum of the estimated synaptic weights of the light-stimulated glomeruli. This analysis demonstrated that this neuron response is strongly correlated with the sum of the synaptic weights (r = 0.88, P = 3.0e-22, Pearson correlation, Fig. 2e). For example, this neuron response was higher when two strongly connected glomeruli were stimulated and did not respond to the stimulation of three weakly connected glomeruli (compare red and green circles in Fig. 2e). To examine this across all neurons, we compared how the number of glomeruli or the synaptic weights sum explains the neuron firing rate using a linear regression model. We found that a model based on the sum of synaptic weights better explains the firing rate than a model based on the number of glomeruli (Fig. 2f, P = 0.0004, df = 48, 2-sample paired t-test of the adjusted R² values). Fitting a quadratic model did not improve the prediction power compared to the linear model. We summarize that an aPC neuron is activated even when a subset of its connected glomeruli is active, if the synaptic strengths of these connections are sufficiently high. Its response scales moderately with the number of stimulated excitatory glomeruli and synaptic strengths. Several distinct sets of glomeruli can activate the same aPC neuron, indicating that aPC neurons are not tuned to detect only one unique glomeruli pattern.

Neighboring aPC neurons have overlapping receptive fields

We next examined the relationship between RFs of adjacent and distant aPC neurons. Figure 3a shows the z-scored and glomeruli-fitted maps of two pairs of units recorded on the same electrode. As can be seen, there is considerable similarity between these maps. The input regions in the two maps are in similar OB regions (e.g., the left upper corner in the right example and the right part of the OB in the left example). Computing the similarity between the two z-scored maps by computing the correlation between the two matrices reveals that these maps share a substantial part of their activation regions (Fig. 2; r = 0.24, P « 0.001, N = 2501 pixels and r = 0.42, P « 0.001, N = 2911 pixels, Pearson correlation—All correlation values above 0.1 or below −0.1 were highly significant, due to the large number of pixels in each map). The spike sorting quality parameters and the mean waveforms of these units are shown in Supplementary Fig. 5a, indicating that these maps are highly likely to represent two different neuron pairs and not the same neuron that has been mistakenly clustered into two different neurons (see also the next section and the Methods section for additional analysis and discussion on this issue). Additional examples of units recorded on the same electrode with varying levels of overlapping regions and adjacent regions are shown in Supplementary Fig. 5b. Interestingly, some neuron pairs had significant negative correlations (Supplementary Fig. 5c), suggesting that close-by neurons can have anti-correlated maps: One neuron receives excitatory input from one OB region, whereas the second neuron receives inhibitory input from the same region. Plotting the cumulative distribution function of the correlation values revealed higher rates of both negative and positive correlation values for neurons recorded on the same electrode (Supplementary Fig. 5d-e). To take this into account, we computed the correlation between the absolute values of the z-scored maps (values shown in parenthesizes in Supplementary Fig. 5c). Comparing the correlations between all pairwise absolute z-scored map values recorded on the same electrode to those recorded on different electrodes or experiments suggests that spatially close neurons are more likely to have overlapping input regions (mean correlation = 0.16 in neurons recorded on the same electrodes/tetrode bundles compared to 0.06 for neurons recorded on different electrodes/sessions, N = 50 neurons, Wilcoxon rank-sum test, Fig. 3b, c). When we compared all maps across all sizes (by clipping the larger map), the P value was 5e-16 (Z = 8.1, r = 0.22, 2-sided Wilcoxon rank-sum test).

Fig. 3. Receptive fields of spatially close neurons are more similar than distant ones.

a Two examples from (two mice) of neuron pairs recorded on the same electrode. The maps of the two neurons have overlapping or nearby input regions. Pearson correlation values between the z-scored maps are marked. The glomeruli-fitted maps have one overlapping glomerulus in one example and nearby glomeruli in the second. b Pairwise correlations between all neurons’ absolute z-scored maps in the single electrode (left) and tetrode bundle (right) experiments. Black thin rectangles on the diagonals mark neurons recorded on the same electrode or tetrode bundle. White rectangles along the diagonals indicate neurons recorded from the same animal, typically near one another. Map similarities are significantly higher in single- and tetrode recordings for neurons recorded on the same electrode/tetrode (Z values = 2.81 and 2.93). c Left: Distribution of the correlation values between the absolute z-scored maps of all neurons shown in b recorded on the same electrode/tetrodes (blue) and neurons recorded on different electrodes/sessions (red, N = 50 neurons). Right: Same data plotted in a box plot. A comparison was conducted between neurons whose maps had identical dimensions to avoid potential bias (P = 2.2e-7, Z = 5.1, r = 0.29). d Same as the right panel in c when we computed the similarity between the glomeruli-fitted maps (N = 50 neurons, Z = 4.44, r = 0.25). e Percentage of glomeruli overlap in neurons recorded on the same or different electrodes (N = 50 neurons, Z = 4.8, r = 0.29). f Distribution of the distances between glomeruli patterns based on computing the distance between the centroid of each pattern (N = 50 neurons, Z = 2.2, r = 0.06). g Distribution of the distances between glomeruli patterns based on computing, for each glomerulus and pattern, the mean shortest distance to one of the glomeruli in the other pattern (N = 50, Z = 5.03, r = 0.14). *all statistical tests in b-g are based on 2-sided Wilcoxon rank-sum test; r is the rank-based effect size (r = Z/√N).

A significantly increased correlation was detected even when we considered the single and tetrode recordings separately (Fig. 3b). Occasional high correlation values were observed between neurons recorded on different sessions in the same mouse in consecutive experiments in which the recording site was near the previous one (Fig. 3b, values outside of the black rectangles and inside the white rectangles). Comparing the correlation between all glomeruli-fitted maps revealed that glomeruli maps of neurons recorded on the same electrode/tetrode bundle are more similar than glomeruli maps recorded on different electrodes/tetrodes (N = 50, Fig. 3d, Wilcoxon rank-sum test). This increased similarity strongly suggests that close-by aPC neurons tend to share glomeruli or have more overlapping glomeruli than distant ones. Computing the percentage of overlap between glomeruli, we found that, on average, 18% of the glomerular area was shared when the units were recorded on the same electrode, compared to 12% when recorded on different electrodes/sessions (N = 50; Fig. 3e; Wilcoxon rank-sum test).

The analyses above quantify similarity between maps based on overlapping regions. However, it cannot capture map similarity when input regions are close but do not overlap. To quantify this, we estimated the distances between glomeruli patterns of all neuron pairs using two different metrics. The first metric was based on computing the Euclidean distance between the two patterns’ centroids. We found that glomeruli pattern centroids tend to be closer in neurons recorded on the same electrode or tetrode bundle than on different electrodes or recording sessions (N = 50, Wilcoxon rank-sum test, Fig. 3f). The centroid-based metric is good if the two patterns are located in similar OB regions, but it can give inaccurate results when the two patterns are spread over different parts of the OB and have similar centers. We therefore used a second metric, in which we computed for each glomerulus and each pattern the shortest Euclidean distance to one of the glomeruli in the other pattern. We defined the distance between two patterns as the average distance across all glomeruli-pair distances for the two patterns. This analysis revealed a significant difference in the glomerular pattern distances between close-by and distant units (Wilcoxon rank-sum test, Fig. 3g). Taken together, these analyses suggest that close-by neurons have more similar glomerular patterns because they share a few glomeruli or because the glomeruli are located in nearby regions. These glomeruli can exhibit the same sign (i.e., excitatory or inhibitory) in both neurons and occasionally opposing signs.

The more distant aPC neurons are, the less similar their RF maps

Recording with a single electrode or a tetrode bundle cannot provide insight into how the distance between neurons correlates with their RFs. To examine this, we recorded an additional set of aPC neurons using NeuroNexus probes with 32 contacts, for which the contact spacing is known. We used two different electrode configurations (Supplementary Fig. 6a) and conducted 15 recording sessions in nine anesthetized mice. We obtained 411 neurons, of which 244 met our convergence criterion. Recomputing the number of excitatory and inhibitory glomeruli in this experiment yielded results similar to those obtained in the single electrode/tetrode experiments (Supplementary Fig. 6b and compare with Fig. 1c). Plotting the similarity between the absolute z-scored maps against the neurons’ estimated pairwise distances of one recording session revealed that, consistent with our previous observation, nearby neurons can have similar and less similar maps (Fig. 4b). However, there was a trend towards a negative correlation between neurons estimated pairwise distances and their map similarities (r = −0.33, P = 0.07, N = 8 converging neurons out of 20 recorded, Pearson correlation; Fig. 4a) suggesting that neurons in closer proximity are more likely to have similar RF maps. Plotting the similarity between the absolute z-scored maps against the neurons’ estimated pairwise distances of all recorded neurons revealed a significant negative correlation (r = −0.09, P = 5.3e-7, Pearson correlation, Fig. 4c). Applying a statistical test with a mixed-effects model to verify that the significant correlation did not result from animal-to-animal variabilities and within session dependencies confirmed this correlation is significant (P = 6.2e-5, linear mixed-effects model with slope and intersect as random effects, Fig. 4c). This negative correlation was robust as it did not depend on the criterion used to define a map as converging (P ≤ 0.003 for all threshold values ranging from zero to 0.3, linear mixed-effects model, Supplementary Fig. 6d). The negative correlation was not biased by conducting multiple recording sessions in the same mice, as restricting the analysis to recordings that were conducted in different mice gave similar results (48 possible different nine recording sessions out of the 15 sessions; correlation value ranged from −0.07 to −0.23 and P value ranged from 0.0001 to 0.02). Excluding neurons that have a small number of excitatory glomeruli, which can spuriously affect the correlation value, increased the strength and significance of the correlation between distance and RF similarity (r = −0.11 and r = −0.12, when we removed neurons with only one or two excitatory glomeruli, respectively, P < 1e-6 for both, linear mixed-effects model). Examining each of the recording sessions revealed that in 12 out of 15 there was a negative correlation between neuron distances and their map similarity which is unlikely to occur by chance (P = 0.017, binomial test, Fig. 4d). A negative correlation between map similarity and unit pairwise distances was also observed when we split the recording sessions into recordings in which the electrode spanned the aPC in the, medial-lateral (ML) and the dorsal-ventral (DV) axes but not in the anterior-posterior (AP) axis, (r = −0.06, −0.07, −017; P = 0.0004, 0.27, 0.0009, for ML, AP, and DV respectively; Supplementary Fig. 6c).

Fig. 4. Receptive field similarity is inversely correlated with the neurons’ distance.

a Details of one of the experiments using the silicon probe. Maps’ similarities are low when neurons’ distances are ~400–600 µm. We applied 10,000 light patterns, each containing ten light spots of size 120 × 120 µm². The Pearson correlation values for all map pairs are shown in the right panel (N = 8 converging neurons out of 20 recorded). b The z-scored maps of the neurons in a, in correspondence to their location on the probe. Cyan circles mark the locations of the eight converging neurons on the probe. The z-scored maps are slightly jittered for better visibility. c Analysis of the similarity between the absolute z-scored maps and the estimated distances between all 244 neurons that crossed the 0.1 threshold out of 411 recorded. The black line is the least-squares linear fit. Pearson correlation value and linear mixed-effects model P values are indicated. The three colored arrows are related to the colored frames in e. 15 recording sessions in nine anesthetized mice. d Least squares lines for each of the 15 recording sessions. The number of converging neurons and the total number of neurons recorded in each session are shown in the legend. e Three examples of neuron pairs that have similar z-scored maps that are located relatively far from each other. The rectangle color corresponds to the arrow’s colors in b. Correlation and distances are marked on the top (r and d, respectively). AP anterior-posterior, ML medial-lateral. The z-scored and glomeruli-fitted maps for each unit are shown in the first and second panels. The kilosort4 template waveforms captured across all 32 channels are shown in the lower panel.

One possible caveat of this analysis is that neurons located close to each other might be more contaminated than distant ones, leading to two neurons that are considered different but are the same, thereby biasing the similarity of nearby neuron RF maps toward high levels. To rule this out, we examined how the correlation between neuron distances and map similarity is affected when we restrict our analysis to neurons at or above a given distance. We found that the correlation was significantly negative, even when we considered neurons that are 100, 150, 200, and even 250 μm apart (Supplementary Fig. 6e). Neurons that are more than ~100 μm apart are very unlikely to represent the same neuron, as non-overlapping channels capture them. These analyses strongly indicate that the farther the neurons are, the less similar their RFs are.

Although map similarity reduced with distance, there were notable exceptions. First, many of the nearby neurons could have dissimilar maps (dots around zero in Fig. 4c). Second, several distant neurons exhibit high similarity. For example, Fig. 4e shows the z-scored and glomeruli-fitted maps of three neurons estimated to be located 363, 371, and 405 μm apart. These maps are highly similar (r = 0.92, 0.8, and 0.76, respectively, Pearson correlation of absolute RF maps, color arrowheads in Fig. 4c corresponding to color panels in Fig. 4e). The glomeruli-fitted maps suggest that both units share at least one glomerulus and differ in the location of one or more glomeruli. These z-score values are considerably high, ruling out the possibility of falsely detected glomeruli (Fig. 4e, upper panels). These units were detected by different groups of channels that are far apart (Fig. 4e, lower panels) and are thus very unlikely to be contaminated or the same neuron split into two units.

Subsets of aPC neurons receive input from different numbers of glomeruli

Anatomical studies suggested that inhibitory neurons in aPC receive input from a larger set of MT cells compared with excitatory neurons⁷. Recordings in PC are likely to capture pyramidal and semilunar cells, and to a lesser extent, inhibitory neurons^29–32. Discriminating between pyramidal and semilunar neurons based on their spike waveforms is not yet considered accurate²⁹. However, several studies have attempted to classify aPC neurons as putative excitatory or inhibitory based on waveform width and firing rate^29,31,33,34. The exact threshold values for spike width and firing rate used in these studies ranged from 0.2 to 0.5 ms and from 0 to 10 spikes/s. This is probably because estimating spike widths depends on the recording setup, the recording methods, and the type and range of the applied filters (Supplementary Fig. 7; Methods). To avoid setting arbitrary threshold values and capture other possible groupings, we compared the number of excitatory glomeruli in the RF maps across a range of baseline firing rates and spike-width thresholds (N = 267 neurons recorded with tetrodes or silicon probes; Methods). This analysis revealed two clusters in which the number of excitatory glomeruli in the RF differs significantly (P < 0.05, Wilcoxon rank-sum test; Fig. 4f; Methods). The first cluster (marked with *) separates aPC neurons into narrow-spiking, relatively high baseline firing-rate neurons (firing rate ≥9 spikes/s and waveform width ≤9/32 = 0.281 ms) versus wide-spiking, low-firing-rate neurons. High-firing and narrow-spiking neurons are commonly identified as interneurons. Only 18/267 (6.7%) of the recorded neurons were classified as putative interneurons, consistent with recent reports suggesting that recordings in aPC yield only a small number of inhibitory neurons^29–31,33. Our analysis thus suggests that narrow-width high-firing neurons receive input from ~27% more glomeruli on average (4.9 versus 3.85; purple shading; Fig. 5b, left two box plots). Interestingly, this analysis revealed a second cluster (marked with ** in Fig. 5a) that separates neurons into those with very low firing rates (<1 spikes/s) and those with firing rates of at least 1 spike/s (“fr ≥ 1”, N = 175 neurons; “fr <1”, N = 92 neurons). The number of excitatory glomeruli was significantly higher in the low-firing-rate group than in the high-firing-rate group (P = 0.0001, two-sided Wilcoxon rank-sum test, Fig. 5b). One possible explanation for this result is that low-firing-rate neurons converge better and, as a result, reveal more glomeruli. However, convergence level and firing rate were not correlated (P = 0.4, Spearman correlation, N = 267). Furthermore, the number of excitatory glomeruli did not correlate with the neuron firing rates (P = 0.26, Spearman correlation). Together, these findings suggest that low-firing-rate neurons and putative interneurons (high-firing but with narrow spike widths) are two groups of neurons that receive input from a greater number of excitatory glomeruli, despite their very different firing rates. The group of neurons with a baseline firing rate of at least one spike/s and that are not interneurons has the lowest average number of excitatory glomeruli. It is worth noting that a recent study showed that some high-firing, narrow-spiking neurons could be semilunar cells²⁹. We did not find significant differences in the number of inhibitory glomeruli across baseline firing-rate or waveform-width groups.

Fig. 5. Subsets of aPC neurons receive input from different numbers of glomeruli.

a Left panel: A heatmap showing log10 P-values comparing the number of excitatory glomeruli in neurons with a firing rate of at least a given value and a spike width narrower than a given threshold (two-sided Wilcoxon rank-sum test, only groups that contained ≥ four neurons were considered). For visibility only, comparisons with a P value < 0.05 are shown. Two clusters of significant values are visible. The * marks neurons that have a firing rate of at least ~9 spikes/s and waveform width narrower than at least ~9/32 = 0.281 ms (putative interneurons, purple shadings). The cluster marked with ** represents neurons with a firing rate of at least one spike/s and not too sensitive to the waveform width. The probability of obtaining this effect size or a larger effect is less than 0.0012 (corrected for multiple comparisons, Methods). Right panel: waveform width distribution as a function of baseline firing rate. The purple rectangle shows the group of putative interneurons (high firing rate and narrow spiking). b Statistical comparison of the number of excitatory glomeruli in the RFs of neurons in the groups defined by the thresholds marked by * and ** in fr. The red line inside the box plot shows the median, the black dot denotes the mean, and the numbers within indicate the size of each group. The putative interneuron group has the highest average number of excitatory glomeruli, and the group of neurons with at least one spike/s baseline firing rate and that are not interneurons has the lowest.

Receptive fields in the awake state

The animal state shapes neural dynamics. In the PC of an awake mouse, there is an increase in baseline activity, inhibition load, and recurrent activity, which likely contribute to an improved odor representation³⁵. To examine the RF of PC neurons in the awake state, we recorded from 181 neurons using silicon probes (three mice, five sessions, 11,526 light stimulations on average). We found that the rate of RF convergence was much lower in the awake state (66/181) than in the anesthetized state (244/411; P < 1e-8, proportion test). This lower rate could result from greater recording instability, including frequent changes in firing rate and animal activity, observed in awake mice passively stimulated with light for long periods. Examining how repeated light stimulation affected the neurons’ firing rate suggested that PC neurons in the awake state are less responsive and more unstable to light stimulations (Supplementary Figs. 8a, b; see “Methods”). However, the neurons were less susceptible to habituation, defined as a reduction in firing rate due to repeated light stimulation (quantified by the negativity of the firing rate slope, Supplementary Figs. 8c, d), likely because they responded less to our light stimulation protocol. The mean number of excitatory glomeruli was significantly lower in the awake state (P = 2e-8, df = 253, Z = 5.8, 2-sample t-test, Supplementary Figs. 9b, c). The mean number of inhibitory glomeruli did not differ significantly between the two animal states (P = 0.099). As in the anesthetized state, the number of inhibitory glomeruli tended to scale with the number of excitatory glomeruli (r = 0.23, P = 0.06, N = 66, Pearson correlation, Supplementary Fig. 9d), and there was no correlation between spike waveform width and the number of connected glomeruli (Supplementary Fig. 7c, P = 0.47, N = 66, Pearson correlation). There was no correlation between neuron-pairwise distances and RF map similarity (r = 0.11, P = 0.9, linear mixed-effects model; Supplementary Fig. 9e). This lack of correlation in the wake state could be due to RF map instability or the use of a small dataset, or the result of comparing maps with a small number of glomeruli, which increases spurious map similarities. Alternatively, it may suggest that, in the awake state, nearby neurons have more dissimilar RF maps than in the anesthetized state due to stronger recurrent activities and a higher inhibitory load.

Receptive field similarity aligns with weak odor response similarity

Our results suggest that, in an anesthetized state, nearby aPC neurons tend to receive input from more nearby regions and share more glomeruli than more distant neurons. In a few cases, nearby aPC neurons with overlapping input regions can have opposing synaptic valences (e.g., Supplementary Fig. 4c), and distant neurons can have similar RF, possibly through long-range connections (Fig. 4c, e). This similarity structure suggests that the odor-tuning curves (OTCs) of nearby aPC neurons will be dissimilar, as has been reported in several previous studies of aPC^6,8,9,11 and posterior PC¹³. However, an exact quantification of OTC similarity as a function of the neuron distances in aPC is still missing. A recent study reported the responses of two sets of aPC neurons to 15 and 13 odors⁶. Neurons were recorded in awake mice using four-shank silicon probes in which the distance between shanks is 200 µm. Each shank contained eight contacts spread in 140–150 µm range. Their analysis suggested that the OTC similarity of neurons recorded on the same shank is similar to that recorded on different shanks (Fig. 7a in Iurilli and Datta⁶). However, comparison of OTCs as a function of neuron distances was not reported. We re-analyzed these datasets, focusing on how the distance between neurons relates to their OTCs. Analyzing the larger dataset reporting the responses of 199 aPC neurons to 15 odors revealed that OTCs of neurons recorded on the same shank are significantly more similar than those recorded on different shanks (mean correlation = 0.049, median = 0.061, P = 8.6e-14, two-sided Wilcoxon rank-sum test, N = 199 neurons, Fig. 6a). Although this mean correlation value is small, it cannot be explained by neural fluctuations or other external factors, as applying the exact computation on the same time duration before odor onset resulted in a correlation value that is not significant (mean = 0.005, P = 0.44, Z = 7.4, two-sided Wilcoxon rank-sum test). Furthermore, we found that the similarity between neuron OTCs is negatively correlated with the number of shanks separating them, suggesting that OTC similarity decreases with increasing distance between neurons (Fig. 6b; r = −0.99, P = 0.009 when considering the mean correlation values within each shank distance; Pearson correlation). Moreover, OTC mean similarity values were significantly above zero when comparing neurons located on the same shank or one shank from each other (Fig. 6b, P < 0.002 for shank distances 0 and 1, one-sided t-test) and below zero when the distance was four shanks (P = 0.02, two-sided t-test). The results were similar when we restricted the computation to neurons that responded to at least one odor (Supplementary Fig. 10a, b; P = 0.0008 when comparing the correlation on the same and different electrodes; P = 0.78 when applying the exact computation across the same time duration before odor onset). These results suggest that neurons located up to 200 µm apart exhibit odor responses that are more similar than expected by chance, and that this similarity decreases with distance. It is important to note that neurons separated by more than ~100 µm are very unlikely to be recorded on the same contact. Analyzing the dataset based on 13 natural odors provided some additional support for this conclusion, as the OTCs of neurons recorded on the same shank were significantly higher than those recorded on the different shanks (P = 9.7e-9, Z = 5.7, two-sided Wilcoxon rank-sum test, Supplementary Fig. 10c-f). However, the mean OTC similarity did not correlate with the distance due to an unclear significant increase in OTC similarity between neurons located four shanks apart from each other (Supplementary Fig. 10c, d), and there was a significant difference when applying the exact computation on the same time duration before odor onset (P = 0.03, two-sided Wilcoxon rank-sum test). Similar results were obtained when considering only neurons that responded to at least one odor (P = 8.6e-14 compared to P = 0.59 when applying the exact computation on the same time duration before odor onset; Supplementary Fig. 10e, f).

Fig. 6. Odor Responses of nearby aPC neurons are more similar than distant ones.

a Comparison between the similarity of neuron OTCs recorded in the same and different shanks. Left panel: Neurons recorded on the same shank exhibited highly significant, more similar OTCs than OTCs of neurons recorded on different shanks (P = 8.6e-14, Z = 7.4, two-sided Wilcoxon rank-sum test; number of pairs in the same and different conditions: 1971 and 18799). Right panel: The correlations when the same analysis is applied to neural activity before odor onset, i.e., baseline activity (P = 0.4, Z = 0.8). The dataset is based on 199 neurons responding to 15 monomolecular odors, averaged across 10 trials from ref. ⁶. b The similarity of odor tuning curves degrades with increasing distance between neurons. The distance between the shanks is measured in units of shank difference. The OTC similarity is negatively correlated with the distance between neurons (red line: r = −0.99, P = 0.009, Pearson correlation of mean values). The mean OTC correlation value is significantly above zero in the first two shank distance comparisons (P = 3.1e-7, 0.001; t = 5.1, 3.1; df = 901, 876; one-sided t-test), and below zero for neurons that are four shank away (P = 0.02, t = 2.2, df = 215, one-sided t-test). Error bars represent mean ± S.E.M. c Comparison between the similarity of neuron OTCs recorded with silicon probes in an anesthetized mouse. Each dot is a pairwise comparison of two neurons’ OTCs in one experiment. The panel on the right displays the least-squares linear fit lines for each recording session, along with the number of neurons in each session. Number of neurons, Pearson correlation, and the linear mixed-model P values are indicated. Eight sessions were recorded from four mice. d Same as in c for awake mice recordings. Seven sessions were recorded from three mice. e A model description of RF similarity of aPC neurons. Glomeruli patterns (circles shown inside the rectangles colored according to the aPC neuron color) of nearby aPC neurons tend to be in similar OB regions, with some overlapping glomeruli. f Comparison between OTC similarity and z-scored map similarity. The right panel shows the least-squares fit lines for each of the three experiments from two mice. The number of converging neurons out of the total recorded neurons in each session is indicated in the legend of the right panel. Pearson correlation and linear mixed-model P values are shown (N = 84 converging neurons out of 140).

The two datasets above did not report the exact location of each contact on each shank, limiting the analysis to distances between shanks. We therefore recorded the response to 15–20 odors of an additional 286 aPC neurons in eight experiments in four anesthetized mice and 185 aPC neurons in seven experiments in three awake mice. We found that OTC similarity was significantly correlated with the pairwise distances of the neurons in both the awake and anesthetized datasets (Fig. 6c,d, P < 9e-5 in both, linear mixed-effects model). This result cannot be explained by mistakenly clustering a single neuron into two nearby neurons, as the correlation remained highly significant even when considering only neuron pairs separated by at least 200 μm in both anesthetized and awake mice (Supplementary Fig. 10f). The results were similar when we the analysis was restricted to neurons that responded to at least one odor and when Spearman’s correlation was computed instead of Pearson’s.

Taken together, these results suggest that, on average, close-by neurons tend to have more similar OTCs than distant ones. This is consistent with the finding that adjacent aPC neurons receive odor information from glomerular sets with small overlapping regions, sharing only a small number of glomeruli and sometimes opposing synaptic valence. Figure 5e summarizes our suggested model of how aPC neurons integrate glomeruli inputs.

Neurons that have similar RF maps have similar odor tuning curves

The finding that nearby aPC neurons have more similar OTCs and RF maps suggests that neurons that have similar RF maps should also tend to have more similar OTCs. To test this, we analyzed a subset of three recorded sessions from two anesthetized mice, in which we recorded the neuron’s RF and responses to a set of odors (20 odors in one mouse and 18 in the second mouse). We then compared the neurons’ OTC pairwise similarity with the neurons’ pairwise z-scored map similarity. Strikingly, we found a significant positive correlation between the z-scored maps and OTC similarities (r = 0.07, P = 0.007, linear mixed-effects model, Fig. 6f). We conclude that nearby neurons exhibit greater similarity in their OTCs and RF maps, indicating a topographic organizational principle among aPC neurons.

Discussion

The approach of finding the RF of target neurons revolutionized vision research, paving the way to a better understanding of this sensory system on the anatomical, system, and cognitive levels, revealing how simple features are processed to form more complicated ones. The RF approach has been used in vision, auditory, and motor systems, and, recently, in the barrel cortex^36–38. However, applying this approach to olfaction research is impossible, as odors are discrete, complex, multidimensional stimuli, making it unfeasible to systematically vary them along a single dimension^13,39. In this study, we developed a new method that reveals the dorsal bulb RF of aPC neurons. The RF is defined here as the set of glomeruli that affect the target aPC neurons and thus characterize the functional connectivity between the PC and the MT cells. We used this method to determine the number, locations, and synaptic strengths of glomeruli. In addition, we examined how glomeruli interact to drive aPC neurons (Fig. 2), and how the RFs correlate with spatial position and relate to odor responses (Figs. 4–6).

We found that aPC neurons receive excitatory inputs from 0-11 glomeruli (median = 4) and inhibitory inputs from 0 to 6 glomeruli (median = 2) located on the dorsal OB (Fig. 1c and Supplementary Fig. 6e). The surgical exposure of the dorsal OB exposes ~150–200 glomeruli⁴⁰, which is about ~10% of the estimated ~1800–2000 OB glomeruli⁴¹. Since anatomical studies suggest that aPC neurons sample glomeruli from all over the bulb without any preference^4,7,10, it is reasonable to estimate that each aPC neuron receives inputs from ~40 excitatory and ~20 inhibitory glomeruli. These numbers are within the range suggested by previous studies^7,23 and provide a comprehensive estimation of the number of functional glomeruli from which each aPC receives input in an anesthetized state. The relatively low number of inhibitory glomeruli we observed may reflect the fact that inhibition exerts a weaker influence on neuronal firing rates than excitation, since the neuron’s baseline firing rate constrains the maximum reduction. Another contributing factor could be the use of anesthesia, which is known to alter OB and PC odor responses^35,42,43. The higher prevalence of odor-evoked inhibition reported in the aPC may arise from reduced input from MT cells. In contrast, our light stimulation predominantly excites MTs, suggesting that the inhibition in the aPC in our experiments could result solely from local circuitry. Moreover, odor stimulation activates glomeruli across the entire bulb surface, while our light stimulation is restricted to dorsal glomeruli, thereby engaging fewer aPC neurons and likely producing less recurrent inhibition. Together, these considerations suggest that our method underestimates the number of inhibitory glomeruli. The number of excitatory glomeruli may also be underestimated, as our method does not detect low-synaptic-efficacy glomeruli and may merge overlapping glomeruli into a single one. Our light stimuli were not aligned with a specific respiratory phase. This might obscure some interesting phase-dependent responses that are likely to exist (Supplementary Fig. 3). Furthermore, it has been shown that the sequence in which glomeruli are activated can affect both the aPC neuron responses and mouse behavior^18,44,45. Our method does not account for these dynamics, and therefore, the actual RF map may be more complex. The number of glomeruli observed in the awake state was considerably lower than in the anesthetized state. This suggests that only a subset of the connected glomeruli is affecting the PC neuron during wakefulness, probably reflecting higher gating of aPC activity, allowing only the few strongest active glomeruli to drive the aPC neurons, as suggested by primacy^46,47 or saliency coding, as the glomeruli with the most significant responses form a stable and predominant component of the odor representation^48,49.

It has been suggested that an aPC neuron acts as a coincidence detector, requiring the activation of multiple glomeruli to respond^23,50. However, we found that this number is not large, as it could be driven even when a relatively small subset of glomeruli is activated (~2-4 glomeruli, Fig. 2). A larger number of glomeruli is required to drive an aPC neuron when the synaptic weights between the glomeruli and the aPC neuron are weak (Fig. 2e, f). It is important to note that the synaptic weights revealed by the STA method are a function of the connection strength between the glomeruli and the aPC neurons and the strength of the MT cell’s response to light activation and therefore do not necessarily reflect the actual connectivity strength distribution. The relatively low number of active glomeruli required to activate an aPC neuron is consistent with previous studies showing that MTs form strong synaptic connections with aPC neurons and that stimulation of a few MT cells can drive aPC neurons^18,23,51. This implies that an aPC neuron connected to N glomeruli responds to many of these sub-patterns. As such, aPC neurons do not detect a specific pattern but rather a family of patterns. The implications of such a coding system on how odors are encoded in aPC require further investigation.

The number of connected glomeruli and their synaptic strength found here do not necessarily reflect direct connections, as it is possible that the observed spikes in an aPC neuron were elicited indirectly by one or more additional neurons from within the PC or from other regions. In the PC, pyramidal cells emit recurrent axon collaterals that form excitatory synapses on other pyramidal cells, and these lateral connections can span a considerable distance^52,53. This circuitry can increase the similarity of responses between distant neurons and the dissimilarity of responses between nearby neurons. This finding can explain why, in some of our recordings, we found similar RF maps in relatively distant neurons and dissimilar ones in nearby neurons. It is worth noting that assessing connectivity by examining the cross-correlation function requires several hours of baseline recordings³¹ and is biased toward inferring connections between unconnected but highly correlated neurons^54–58. Thus, the exact interactions between aPC neurons and their RF require further investigation.

Previous studies reported that inhibitory neurons receive input from a larger set of MT cells and have broader odor tuning curves compared with excitatory neurons^7,59. Analyzing our data across all possible groupings of neurons by firing rate and spike-width range revealed two groups with significantly different RF maps. A small group (18/267 = 6.7%) of high-firing-rate and narrow-spiking neurons received input from 27% more excitatory glomeruli than the neurons that did not have a high firing rate and narrow spike width. Previous studies reported that the proportion of inhibitory aPC neurons detected in extracellular recordings is low (~3–7%,^29,31,33,60). Thus, the characteristics of this group of neurons match those of interneurons. A second group of neurons could be characterized by having a very low firing rate (below one spike/s). Low firing-rate neurons received input from a larger number of glomeruli than those firing above this threshold.

We grouped aPC neurons into two clusters based on spike width and baseline firing rate. This approach has two limitations. First, exploring multiple parameterizations can inflate the risk of Type I error. Second, it does not exhaustively evaluate alternative groupings, including potentially non-linear boundaries, as suggested by ref. ³¹. While the second cluster remained significant after correction for multiple comparisons (Methods), the first did not, raising the possibility of a false-positive result. Nevertheless, the first cluster corresponds to high-firing, narrow-spiking neurons, a class commonly identified as putative interneurons. Importantly, the associated effect is not confined to a single parameter choice but is observed across several adjacent parameter values (Fig. 5a). Taken together, this consistency and biological plausibility suggest that the finding is likely robust, although we interpret it cautiously given the multiple-comparisons context.

A central finding in neuroscience is that neurons are topographically organized such that spatially adjacent stimuli on sensory receptor surfaces are represented in adjacent positions in the cortex. The organizational principles in the olfactory cortex have long been sought for⁶¹. Alternative organizational principles suggest that olfactory neurons are organized by valence, and that posterior PC encodes odor quality, whereas anterior PC encodes odor identity remove^{14–21,62,63}. Here, we show that although neurons in aPC form random connections with glomeruli in the OB, they nevertheless adhere to an organization principle: spatially proximate aPC neurons tend to exhibit more similar RF maps than distant ones. This similarity decreases as the distance between neurons increases (Fig. 4). We found that maps of nearby aPC neurons are similar because they share some glomeruli in their RFs, and the glomerular patterns are located close to each other. Thus, the topographical map we found in aPC is not a simple one-to-one mapping from the OB to the PC but rather a many-to-one mapping, with similar glomerular patterns tending to project to nearby PC neurons (Fig. 6e).

Glomeruli’s sensitivity to a specific molecular feature tends to form clusters on the OB surface^1–3. Therefore, the overlap and adjacency of glomeruli activation patterns could lead to more similar odor responses in target aPC neurons, thereby enhancing response similarity to chemically related odorants, as has been reported recently¹³. Alternatively, the associative network in aPC may decorrelate odor responses even if they share many glomeruli^{30,35,64–66}. Re-analyzing previously published datasets of odor responses of neurons recorded on four shanks in awake state, we found a highly significant correlation between neuron distances and their OTCs (Fig. 6). We further verified this result with new data recorded in anesthetized and awake states using a larger set of odors and more accurate distance estimates (Fig. 6c, d). Although the correlation was highly significant, its value was relatively small. This is consistent with the hypothesis that nearby aPC neurons share only a small number of glomeruli with sometimes opposing synaptic valence.

What is the advantage of this topographical organization for odor processing? A possible explanation for the increased glomerular pattern similarity and odor responses in nearby aPC neurons is based on the wiring optimization principle^61,67. According to this principle, evolution minimizes the length of the brain’s axons and dendrites. This principle explains why many sensory systems are organized in maps that reflect the spatial arrangement of the sensory space. Applying this principle in olfaction, where PC neurons receive input from a random group of glomeruli, suggests that nearby PC neurons should receive input from nearby and overlapping groups of glomeruli.

Finally, the STA method developed here can be extended to other direct targets of OB projection neurons, including the anterior olfactory nucleus, cortical amygdala, and entorhinal cortex. As each region is thought to participate in distinct odor-guided tasks, neurons in these areas are likely to follow different rules of glomerular integration and organizational principles. This approach can also be applied to examine how neuronal receptive fields change during learning. More broadly, it provides a general framework for investigating how downstream neurons process activity patterns from upstream circuits.

Methods

Surgical procedure

All surgical and experimental procedures were conducted in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and Bar Ilan University guidelines for the use and care of laboratory animals in research, and were approved and supervised by the Institutional Animal Care and Use Committee (IACUC) of Bar-Ilan University. Mice were housed under controlled environmental conditions with a reversed light/dark cycle, at an ambient temperature of 22 °C and relative humidity of approximately 52%. Mice were housed in a group cage and received no experimental treatment other than genotyping. We used 30 Tbet-Cre mice (Strain #:024507, Jackson Labs) crossed with the Ai32 strain (Strain #:024109, Jackson Labs), which expresses Channelrhodopsin-2 in mitral/tufted (MT) cells. Animals were anesthetized with ketamine/medetomidine (60/0.5 mg/kg, intraperitoneal) and placed in a stereotaxic frame. The bone overlying the dorsal olfactory bulbs (OBs) was removed to expose the largest possible dorsal surface. To allow electrode access, the bone overlying the anterior piriform cortex (aPC) was also removed. Supplemental anesthesia was administered as needed (30–50% of the initial dose of ketamine/medetomidine). Body temperature was maintained at 36–37 °C using a homeothermic blanket system (Harvard Apparatus). We used seven WT (C57BL/6JOlaHsd, Harlan Labs, Israel) mice for the odor experiments. We did not limit our experiment to males or females.

In vivo extracellular recordings

Overall, this study is based on 1130 aPC and 15 MT neurons recorded in our lab, and analyzes of an additional 477 aPC neurons taken from previously published datasets. The detailed breakdown of the number of neurons recorded in each experimental condition, the number of mice used, and the number of recording sessions per mouse is provided in Supplementary Table 1. Receptive fields were computed for 483 neurons in anesthetized mice and 181 in awake mice. Odor response analysis was applied to 466 neurons recorded in our lab and 477 neurons from Iurilli and Datta⁶ study. The majority of the RF experiments were conducted under anesthesia to increase recording stability and minimize variability in neural activity caused by the animal’s state (e.g., motivation, stress, motor actions, changes in respiration frequency). However, we also recorded 181 neurons in the awake state (five sessions from three mice).

Electrodes were typically inserted into the anterior piriform cortex (aPC) at ~1.8–2.4 mm anterior to bregma and 1.2–2.1 mm lateral from the midline (depending on the anterior coordinates). Recordings began ~20 min after probe insertion to allow stabilization.

In the OB and some aPC experiments, tungsten electrodes (5–10 MΩ) were used. Neural signals were amplified and filtered between 300 and 5000 Hz (AM-Systems 1800), sampled at 40 kHz (National Instruments, Austin, TX), and stored on a computer. Spikes were sorted offline using MClust (MATLAB, A.D. Redish). In some single-electrode recordings, the same neuron was recorded across multiple sessions and then combined based on waveform similarity. This was done to assess responsiveness to light and guide electrode relocation if necessary. This procedure biases the estimated rate of converging neurons in single-electrode recordings. Using tungsten electrodes, we recorded 15 mitral/tufted (MT) neurons from five mice and 34 aPC neurons from 17 mice.

In addition, we employed recordings in aPC using custom-made bundles of one to eight tetrodes. Neural signals were amplified and filtered between 300 and 6000 Hz (Neuralynx), sampled at 32 kHz, and stored. Spikes were sorted offline using SpikeSort3D by visually identifying dissociable clusters for each electrode or tetrode, and validating them with auto- and cross-correlograms across all neurons recorded in the same session. Using custom tetrodes, we recorded 38 aPC neurons from four mice.

In additional RF experiments, we recorded aPC neurons using NeuroNexus 32-channel silicon probes (A1X32-Poly3-5mm-25s-177-CM32 and Buzsaki32). We recorded 411 aPC neurons across 15 sessions from nine mice. In three recordings, we positioned the four shanks along the anterior–posterior axis (1.22–1.82 mm and 1.6–2.2 mm). In seven experiments, we recorded from the same cortical region after moving the probe, typically ~200–300 µm posterior to the original location. In one recording, we sampled the same region but detected different neurons. The results of the analysis remained consistent when this dataset was excluded and when we analyzed only experiments recorded on different mice (48 possible combinations; see text). For the NeuroNexus array, we spike sorted the data using KiloSort4 and curated it with Phy. Kilosort4 considers the contact arrangement and ensures the integrity of the recorded data by preventing neuron redundancy across multiple channels.

Spike waveform analysis

We computed spike waveform width as the interval from the peak to the trough. The peak was defined as the largest extremum, and the trough as the subsequent extremum. This approach performs well for regular and inverted spikes, as well as for spikes with a slight rising edge preceding the central peak (Supplementary Fig. 7a). When a spike was detected on multiple channels, we computed the width using the channel with the most prominent peak. The mean ± std waveform width in the Tungsten electrode, tetrodes, and silicon probes was 0.49 ± 0.14, 0.31 ± 0.06, and 0.38 ± 0.16 ms, respectively (mean ± std, Supplementary Fig. 7b). The mean spike widths in the tetrode and Silicon probe recordings were not significantly different, and we therefore combined them when we analyzed spike widths. Several recent studies reported a median spike-wave width ranging from ~0.3 to ~0.75 ms^{29,31,34,68,69}. The spike waveform width depends on several parameters, including the recording method, recording device, electrode type, sampling rate, and filter types. Indeed, the mean waveform width in the tetrode and silicon probe recordings was narrower than in the Tungsten electrode (P = 6e-5, 2-sample t-test, df = 483), suggesting that different filtering methods (analog versus digital) can affect spike widths. Our multi-channel recording system is based on Digital Lynx Sx (Neuralynx), in which we applied an online 256-tap FIR 300-6000 Hz filter, which can result in a narrower waveform width due to the application of DC filter combined with a high-pass filter. To verify this in an independent recording, we computed the waveform width of 28 neurons recorded in Dan Rokni’s lab from mice’s aPC using tetrodes and applying a 256-tap FIR filter between 300 and 5000 Hz. We found that the mean waveform width in this dataset was shorter than the range we obtained in the silicon probe recordings. A narrower average waveform width was also reported in a recording done with the Neuralynx system in rats aPC while applying a 600–6000 Hz filter⁶⁹.

To find possible differences in the number of connected glomeruli between different groups of neurons, we examined all possible combinations of firing rate (≥0 to ≥20 spikes/s) and waveform width (≤5/32 to ≤32/32 ms) values. This procedure revealed two clusters of parameter combinations that crossed the P < 0.05 (shown in Fig. 4f). The lowest P value in the group marked with ** was 6.61e-6, which is significant even after Bonferroni correction for multiple comparisons (P = 0.0012). It also survives a shuffling analysis (P = 1/300). The lowest P value in the group marked by * is 0.028, which does not exceed the multiple-comparison Bonferroni correction or a shuffling analysis. To ensure that baseline activity is not affected by the light stimulation, we computed it during the 10 s starting 10 s after the last light stimulation.

STA model

We assumed a linear-nonlinear-Poisson model²⁰, wherein on each time step, the components of the stimulus vector are linearly combined using a weight vector, k. The response of this linear filter is then passed through a nonlinear function f, whose output determines the instantaneous firing rate of the Poisson spike generator. Under these assumptions, the STA method enables us to estimate k and f.

Optical stimulation of the olfactory bulb

Optical stimulation of the OB was performed as previously described¹⁸. Briefly, spatially precise light patterns were projected onto the dorsal surface of the OB using a digital micro-mirror device (OPTOMA X600 DLP projector) coupled to an optical imaging system. A blue filter was placed above the collimating lens (f = 75 mm, achromatic doublet, Thorlabs), positioned 20 cm from the projector. Under this configuration, each projected pixel corresponded to an area of ~20 µm², with an estimated light intensity of ~1 mW/mm². Stimulation timing was controlled using the MATLAB Psychophysics toolbox. A photodiode (FDS1010, 400-ns rise time, Thorlabs) was used to record timestamps for each light stimulus.

The craniotomy size determined the total scanned area for each experiment. Light stimulation lasted 80–100 ms, followed by an inter-stimulus interval of ≥200 ms, which is sufficient for MT cell recovery¹⁸. In most single- and tetrode-based recordings, we applied 8–12 light patches (140 × 140 or 160 × 160 µm²), with the number and size of patches fixed within each experiment. In two recordings, we used four larger patches (200 × 200 or 240 × 240 µm²), and in two others, 25 smaller patches (120 × 120 µm²). The number of light stimuli ranged from 3290 to 18,481 (median = 9983; mean = 10,155; mode = 15,000), with variability reflecting adjustments made to achieve convergence. Some experiments were terminated prematurely due to animal conditions or technical issues.

In silicon probe experiments, we applied 10–12 patches (120 × 120, 140 × 140, or 160 × 160 µm²) and typically delivered 10,000–15,000 trials (median = 10,000; mean = 10,770; maximum = 15,000). All neurons crossing the convergence score threshold were included in the analysis, irrespective of patch size, number of patches, or total stimuli (see definition of convergence below). Because stimulation patterns were randomized, the same MT cell was unlikely to be repeatedly activated.

Following the presentation of N random light patterns, we computed neuronal RFs using the STA method, which calculates the weighted average of all stimuli^20,70:

$$S=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{S}}_{\mathrm{i}}{\mathrm{r}}_{\mathrm{i}}}{{\sum }_{\mathrm{i}}^{\mathrm{N}}{\mathrm{r}}_{\mathrm{i}}}$$

Where Sᵢ denotes the two-dimensional light pattern i and rᵢ is the firing rate of the recorded neuron in response to stimulus Sᵢ. Thus, S is the weighted average stimulus. Notably, the STA technique can detect both supra- and sub-threshold connections. This arises because some light stimulations simultaneously activate supra- and sub-threshold inputs, producing firing rates distinct from those elicited by stimulations that drive only supra-threshold connections. For the same reason, the STA method can also reveal inhibitory connections, even in neurons with low spontaneous activity, provided that a substantial number of excitatory glomeruli are present.

STA significance

The significance of the STA vector was assessed using a bootstrap procedure, as described by ref. ⁷⁰. Specifically, we randomly shuffled the light patterns and recomputed the STA. This process was repeated three times to generate a bootstrap distribution. For each pixel, we then estimated the probability of observing an STA magnitude greater than the 98th percentile or less than the 2nd percentile of the bootstrap distribution (corresponding to a z-value of ~2.1). Pixels that did not meet this threshold were set to zero. To further reduce noise, the resulting map was convolved with a two-dimensional linear filter, which reduced isolated or small patches of significant pixels. We refer to this final map as the z-scored RF map.

STA map convergence

To evaluate how well the RF maps represent the “true” receptive field, we defined a convergence score (also referred to as a stability score) as the correlation between maps constructed from the first and second halves of the stimuli. We computed this value for both the raw and z-scored RF maps. Because this measure is based on only half the trials, it underestimates the convergence value obtained using all trials. To establish a reasonable threshold for convergence, we computed convergence scores for all neurons after shuffling the light events. We then defined our threshold as values that exceed the 95th percentile (or fall below the 5th percentile) of the shuffled distribution. This procedure yielded a threshold of 0.07. For simplicity and to further exclude noisy maps, we set the threshold to 0.1. This cutoff effectively removed the left tail of the convergence score distributions (Supplementary Fig. 2b). The remaining neurons exhibited relatively high prediction values on the test set (Supplementary Fig. 2c). Excluding neurons below threshold did not alter the number of mice reported in each dataset. Importantly, our results were robust to the choice of threshold, as similar outcomes were obtained across thresholds of 0, 0.05, 0.1, 0.15, 0.2, 0.25, and 0.3 (Supplementary Fig. 4b).

The critical factor determining whether a neuron reached convergence was whether enough light patterns elicited substantial changes in its firing rate. When this number was low, the change in firing rate was small, or the neuron altered its response properties during the recording due to changes in animal state, the neuron typically exhibited a low convergence value. The correlation between convergence values and the number of light patterns applied was 0.31 (P = 0.1, N = 72, Pearson correlation; Supplementary Fig. 2d).

Several neurons responded to many light patterns with inhibition. Neurons displaying inhibitory responses required more light stimuli to converge, since inhibitory effects are usually smaller than excitatory responses and more challenging to detect when baseline activity is low (Supplementary Fig. 2e). Conversely, increasing the number of light patterns also raised the likelihood of losing neurons due to changes in their response properties over time.

To reduce the potential habituation of aPC neurons from multiple consecutive light stimulations, we paused the light stimulations for 2–3 s after every ~20–30 trials. Supplementary Fig. 8 shows the effect of multiple consecutive light stimulations on the neurons’ firing rate in two anesthetized and two awake experiments for both convergent and non-convergent RF maps. As shown, some neurons exhibit a substantial reduction in firing rate upon repeated stimulation. However, even neurons that undergo a significant decrease in firing rate reached a relatively stable RF map (convergence score >0.1). The pauses in light stimulation were also used to safeguard against missing or phantom light events. When a light event was missing or extra, we removed the block (or blocks) of light events whose number of events did not match the expected value, rather than excluding the experiment.

Non-convergent neurons may be connected to glomeruli that are not located in the dorsal OB or have synaptic weights too low to drive substantial responses with the current protocol. Some neurons may not have converged because their response properties do not conform to the assumptions of the STA framework (e.g., the linear–nonlinear Poisson model).

Estimating glomeruli number and their locations

We developed an iterative algorithm to fit two-dimensional Gaussian ellipsoids to local maxima and minima in the z-scored response map. Fitting was performed using 80% of the data, while the remaining 20% was used to evaluate prediction quality. Only glomeruli that improved the predictive power (i.e., >0.0001 from the current prediction) of the reconstructed map on the test data were retained. This procedure was repeated 30 times, and a glomerulus was considered valid if it was detected in at least 50% of the iterations and its center was located within half a glomerular diameter (~60 µm). This tolerance allowed for the detection of partially overlapping glomeruli. The final RF map prediction value was defined as the mean prediction accuracy across the 30 iterations.

The algorithm for a given neuron RF proceeded as follows:

1. Randomly split the dataset into training (80%) and test (20%) trials.

2. Construct the z-scored map from the training group.

3. Identify all local maxima and minima in the map.

4. Fit a Gaussian to each local maximum and update the glomerular map.

5. Evaluate the prediction of test responses using the updated map.

6. If prediction accuracy improved, keep the Gaussian; otherwise, discard it and proceed to the next local maximum.

7. Repeat the procedure for local minima.

8. Iterate this process 30 times to generate 30 candidate glomerular maps.

9. Cluster all candidate glomeruli across iterations; define a glomerulus as valid if detected in ≥50% of iterations, with centers allowed to vary within three pixels.

Because single-spot stimulation can elicit responses that broaden the estimated receptive field, the STA procedure for OB neurons revealed larger glomerular sizes. To account for this, we used a slightly larger Gaussian sigma (6 vs. 5) and applied two convolution steps to the raw maps.

Comparing RF maps

To estimate map similarity, we converted the two maps into vectors and computed their Pearson correlation. In a subset of experiments, we performed repeated recordings from the same region by moving the electrode or tetrode bundle by ~200 μm in the dorsal-ventral or the anterior-posterior axes. This procedure could artificially increase mean similarity across sessions. Nevertheless, despite this potential bias, the similarity between neurons recorded on the same electrodes remained significantly higher. We computed the correlation between map similarity and pairwise neuronal distance using Pearson’s correlation and obtained results comparable to those obtained with Spearman’s correlation. To account for dependencies in pairwise distance comparisons, we verified significance using the Mantel test (P < 1/1000 across 1000 iterations for the RF and OTC similarity analyses). Additionally, to account for dependencies among neurons recorded within the same session, we applied a linear mixed-effects model treating slope and intercept as random effects. All P values remained highly significant.

Spike clusters quality analysis

We assessed spike sorting quality using multiple criteria, including visual inspection of cluster separation in feature space, differences in spike waveforms, violations of interspike interval (ISI) constraints, cross- and autocorrelations, and temporal stability. In the silicon-probe experiments, the mean ISI violation rate (2 ms) was 0.19%, (median = 0.09%, maximum = 1.6%). In the single- and tetrode recordings, the mean ISI violation rate (2 ms) was 0.8% (median = 0.5%, maximum = 3.8%). In these recordings, the median signal-to-noise ratio (SNR), as defined by Suner et al. (2005), was 5.6. In the odor experiments, the mean ISI violation rate was 0.21% (median = 0.09%, maximum = 2.2%). Electrode positioning was validated both by detecting clear respiration-locked firing during recordings and by post hoc histological verification.

The increased similarity between RFs of neurons recorded on the same Tungsten electrode or on nearby tetrodes could occur when one neuron is mistakenly clustered into two or more units. While this is unlikely, given the apparent difference in neuron clusters and waveforms, we also conducted several additional analyses. First, in the single- and tetrode recordings, we compared the clustering quality (L-ratio and Isolation distance) of neurons recorded on the same electrode with that of neurons recorded on different electrodes. We observed no significant correlation between map similarity and clustering quality. Second, the L-ratio and Isolation distance values did not significantly differ in units recorded on the same electrode, and these recorded on different electrodes (Supplementary Fig. 5f). Third, given that anterior piriform cortex (aPC) neurons are phase-locked to specific respiration phases, especially in the anesthetized state^6,9,71, a neuron erroneously clustered as two distinct units should exhibit similar preferred phases. Consequently, we examined whether neurons recorded from the same electrode demonstrated more similar preferred phases than those recorded from different electrodes. We found no significant difference between the phase differences (Supplementary Fig. 5g). Thus, improperly clustering the same neuron into two units is unlikely to explain our results. We note that it is generally accepted that neurons recorded in the aPC with contacts >100 µm are likely different. This is clearly demonstrated in the silicon probe recordings.

Odors application

Odors were delivered using a custom-built olfactometer. Odorants were diluted 1:100 in mineral oil and stored in sealed plastic tubes; this concentration was selected to reliably elicit detectable responses. Fresh odor solutions (5–10 ml) were prepared before each experiment. Airflow was regulated with a mass flow controller (Agilent, Alimc-2LSPM) and set to 0.5 l/min. Between stimulations, air circulated freely to minimize residual odor contamination, and a vent positioned behind the mouse removed remaining odors. All odorants were obtained from Sigma-Aldrich at the highest available purity. Stimulation times and sequences were controlled by custom LabVIEW scripts. Each odor stimulus lasted 1.5 s, with an inter-trial interval of 10 s. The number of trials ranged from 10 to 15, and the number of odorants per experiment ranged from 15 to 20. A blank stimulus was included to verify that responses were not due to mechanical stimulation. Odor-evoked responses were quantified across the stimulation window beginning at odor onset.

Odorants used:

Odor Id	Odor name	CAS	Concentration (V/V)
3	2-phenylethanol	60-12-8	1:100
4	Eugenol	97-53-0	1:100
5	air
6	Hexyl octanoate	1117-55-1	1:100
7	Amyl butyrate	540-18-1	1:100
8	Ethyl valerate	539-82-2	1:100
9	2-butanone	78-93-3	1:100
10	Ethyl tiglate	5837-78-5	1:100
11	Citral	5392-40-5	1:100
12	3-hexanone	589-38-8	1:100
13	R-(+)-Citronellol	1117-61-9	1:100
14	Hexanal	66-25-1	1:100
15	air
16	L-Carvone	6485-40-1	1:100
17	2-heptanone	110-43-0	1:100
18	Octane	111-65-9	1:100
19	Pentyl acetate	628-63-7	1:100
20	Ethyl butyrate	105-54-4	1:100
21	Geraniol	106-24-1	1:100
22	3-Hexanol	623-37-0	1:100
23	Acetophenone	98-86-2	1:100
24	Cineole	470-82-6	1:100

Analysis of odor-tuning curves

We first analyzed the datasets published by Iurilli and Datta⁶. One dataset contained 199 aPC neurons responding to 15 odors, and the second contained 278 aPC neurons responding to 13 odors. Each odor was presented ten times. To quantify similarity between OTCs, we calculated the mean firing rate across the ten trials over a one-second window, subtracting baseline spike counts from the same time interval before odor onset. The pairwise signal correlation, reflecting the similarity of neuronal tuning profiles, was computed as Pearson’s correlation coefficient between the two baseline-subtracted response vectors, as in the original study. Excluding neurons that did not respond to any odors yielded similar results (P < 0.05, two-sided two-sample t-test).

We applied the same analysis to data we recorded in four anesthetized and three awake mice (eight and seven recording sessions, respectively). Because OTCs are typically not normally distributed, we used Spearman correlation as the primary measure; repeating the analysis with Pearson correlation produced similar results. In these odor experiments, the silicon probes spanned the ML axis (1.5–2.1 mm).

Statistical analysis

All statistical tests and analyses were performed using built-in functions or custom code in MATLAB R2024b. Details for each test, including the P value and the number of data points, are provided in the Results section, figure legends, or directly within the figures. In general, we used two-sided t-tests or Pearson correlations when the data could be assumed to follow a normal distribution, and Wilcoxon rank-sum tests or Spearman correlations otherwise. Additional statistical tests were applied as needed and are described in the text.

Box plots are based on MATLAB convention (box boundaries show the 25% and 75% percentiles (Q1 and Q3), red line denotes the median, lower and upper whiskers extend from the box edges to the most extreme data point within 1.5 × (Q3−Q1) of the box, “+” represent outliers).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

S.T. and R.H. conceptualized the experiments. S.T., O.B., T.D., and K.B. performed the experiments. S.T., O.B., and T.D. analyzed the data. S.T., T.D., and R.H. wrote the paper. S.T., O.B., T.D., and K.B. contributed equally.

Peer review

Peer review information

Nature Communications thanks Gabriel Lepousez and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data generated in this study have been deposited in the Zenodo database under accession codes: 10.5281/zenodo.18473332.

Code availability

The code generated in this study has been deposited in GitHub labrafihaddad/STA_2026_Paper_Code.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-70356-9.