Abstract
Recording and analysis of neural activity are often biased toward detecting sparse subsets of highly active neurons, masking important signals carried in low-magnitude and variable responses. To investigate the contribution of seemingly noisy activity to odor encoding, we used mesoscale calcium imaging from mice of both sexes to record odor responses from the dorsal surface of bilateral olfactory bulbs (OBs). The outer layer of the mouse OB is comprised of dendrites organized into discrete “glomeruli,” which are defined by odor receptor-specific sensory neuron input. We extracted activity from a large population of glomeruli and used logistic regression to classify odors from individual trials with high accuracy. We then used add-in and dropout analyses to determine subsets of glomeruli necessary and sufficient for odor classification. Classifiers successfully predicted odor identity even after excluding sparse, highly active glomeruli, indicating that odor information is redundantly represented across a large population of glomeruli. Additionally, we found that random forest (RF) feature selection informed by Gini inequality (RF Gini impurity, RFGI) reliably ranked glomeruli by their contribution to overall odor classification. RFGI provided a measure of “feature importance” for each glomerulus that correlated with intuitive features like response magnitude. Finally, in agreement with previous work, we found that odor information persists in glomerular activity after the odor offset. Together, our findings support a model of OB odor coding where sparse activity is sufficient for odor identification, but information is widely, redundantly available across a large population of glomeruli, with each glomerulus representing information about more than one odor.
- linear classifiers
- mesoscale calcium imaging
- odor information
- olfactory bulb
- random forest feature importance
- redundancy
Significance Statement
This study leverages mesoscale imaging and machine learning to investigate how odor information is first represented in the brain. Typically, recordings of neuronal activity focus on active individual cells, potentially overlooking broader variations in neuronal responses across populations. Our results demonstrate that a considerable amount of olfactory information is redundantly distributed across a large proportion of olfactory bulb glomeruli. Even after excluding a majority of glomeruli, odor identification remained possible. These findings indicate that, although a few glomeruli are sufficient for odor recognition, an abundance of additional information is represented across a broad population. Understanding how the brain manages redundant olfactory information will shed light on its adaptive mechanisms for navigating diverse real-world circumstances and responding to fluctuating internal states.
Introduction
A fundamental question in neuroscience is how features of the external world are converted to neuronal activity and represented in the brain. Odor encoding requires sampling a complex chemical sensory space and converting this information to neuronal activity. In mammals, odors are detected by sensory neurons expressing a single type of olfactory receptor, which transmit odor information to the olfactory bulb (OB) via synapses in olfactory receptor-specific glomeruli (Buck and Axel, 1991; Vassar et al., 1994). Odor response patterns in the glomerular layer (GL) of the OB are spatially stereotyped due to receptor-specific projection input from sensory neurons (Ressler et al., 1994; Mombaerts et al., 1996; Bozza et al., 2002; Soucy et al., 2009). This gives rise to a labeled line model of olfactory encoding where odors are spatially segregated among relatively sparse populations of finely tuned OB projection neurons, with some odors, especially at low concentrations, evoking activity at just a single glomerulus per OB (Rinberg et al., 2006a; Davison and Katz, 2007; Burton et al., 2022).
Activating sparse subsets of glomeruli in the OB is sufficient to drive odor-guided behavior (Chong et al., 2020), and glomeruli responding to the lowest latencies have the most impact driving downstream behaviors (Gill et al., 2020), suggesting that odor information is largely represented by sparse, low-latency glomerular activity (Wilson et al., 2017). Sparse codes offer computational advantages by utilizing a small subset of neurons to represent information, reducing required neuronal activity (Levy and Baxter, 1996). They enhance selectivity and specificity in responses, facilitating precise information processing (Zetsche, 1990; Field, 1994; Olshausen and Field, 2004). However, sparse codes are more sensitive to changes in participating neurons, making them less robust to alterations in input, gain, background activity, and circuitry compared with denser, more redundant codes (Spanne and Jörntell, 2015).
Nevertheless, odors remain discriminable when presented on variable and complex backgrounds (Vinograd et al., 2017; Y. Li et al., 2023), odor-evoked activity changes depending on behavioral context (Koldaeva et al., 2019; Kudryavitskaya et al., 2020), and significant changes in sensory input do not always impair odor discrimination (Slotnick and Bodyak, 2002; Knott et al., 2012), suggesting that odor coding is robust to these alterations. The OB displays both sparse, low-latency odor–evoked activity and dense, temporally complex, state-dependent activity during and after odor exposure (Spors and Grinvald, 2002; Fletcher et al., 2009; Kato et al., 2012; Vincis et al., 2012; Patterson et al., 2013; Cazakoff et al., 2014; Adefuin et al., 2022; Shani-Narkiss et al., 2023). However, the role and information content of dense activity in odor encoding remain unknown.
Here we used mesoscale two-photon calcium imaging to record odor-evoked responses from hundreds of glomeruli across bilateral OBs in mice. This approach allowed us to continuously image population odor responses with high spatial and temporal resolution. We then applied logistic regression to decode and predict odor identity from patterns of glomerular activity. We used classifier accuracy to quantify (1) the value added (VA) of individual glomerular responses to classifier predictions, (2) the extent to which odors were classifiable from different subsets of glomeruli at different time points, and (3) which aspects of glomerular population activity, beyond the first responding glomeruli, were necessary or sufficient for odor classification. VA analyses—adding or removing glomeruli from classification analysis one at a time—demonstrated that large populations of relatively low-value glomeruli contributed significantly to odor decoding and were sufficient for odor decoding even when top-ranked glomeruli were removed from the analysis. Additionally, odors were correctly classified from glomerular activity for an extended period after odor exposure. Together, these data suggest that odor information is redundantly represented across a dense population of glomeruli.
Materials and Methods
Mice
All experimental procedures were approved by the Baylor College of Medicine Institutional Animal Care and Use Committee. Thy1-GCamp6f 5.11 (Jackson Laboratory, 025393) mice of both sexes were used for two-photon imaging experiments. Mice were housed in a standard 12 h light/dark cycle and had ad libitum access to food and water.
Cranial window surgery
Chronic cranial windows were created in mice by removing a 4-mm-diameter section of the skull over the OB and inserting a glass coverslip. Before surgery, mice were treated with 5 mg/kg meloxicam. Anesthesia was induced and maintained with isoflurane during the surgical procedure. After induction of anesthesia, the scalp was injected subcutaneously with 0.05 ml bupivacaine and then cleaned and removed over the OB and dorsal skull. A 0.16″ thick stainless-steel shim (McMaster-Carr, A370-974) was centered over the OB and attached to the exposed skull with dental cement (C&B Metabond). A 4-mm-diameter piece of the skull was removed by carefully drilling though the skull. A 4 mm glass coverslip (Warner Instruments) was placed over the exposed brain and sealed in place with tissue adhesive (3 M Vetbond). The window was then fully sealed and fixed in place with cyanoacrylate superglue (Loctite). After the glue fully cured, the coverslip was protected with a cap of Kwik-Cast silicone elastomer (World Precision Instruments) that was removed immediately before imaging.
Odor delivery
All imaging was performed on awake mice head-fixed on a running wheel. Mice were head-fixed using a custom headplate (Liu et al., 2019) designed to attach to the stainless-steel shim implanted on the skull during the cranial window surgery. Mice were habituated to the head fixation and imaging setup for at least 30 min prior to testing each day. For odor delivery, a multichannel olfactometer (Burton et al., 2019) was placed 6 cm in front of the mouse. The olfactometer provided a constant stream of room air into which experimental odors were injected. Odors were mixed into the central airstream before delivery to the mouse by an eductor positioned at the output of the olfactometer. Stimulus panels for mouse imaging experiments consisted of 11 odors diluted to 10% v/v mineral oil and mineral oil alone. Odors were obtained from Sigma-Aldrich and included 1,4-cineole, 2-methylpyrazine, 2-ethylphenol, isoeugenol, trans-cinnamaldehyde (TCN), anisole, anisaldehyde, methylsalicylate, allyl sulfide, acetophenone, and eugenol. Diluted odors were loaded into reservoirs on the olfactometer, and air was passed through the reservoir and injected into the central stream to deliver odors. Injection in the central stream further diluted odors to ∼1% of their reservoir concentration for a final odor concentration of 0.1% delivered to the mouse. Odor delivery was controlled and synchronized with imaging via the custom LabVIEW software. Odors were delivered for 1 s followed by a 19 s delay where clean air continued to flow in the central air stream. Each session consisted of 480 trials where the 12 stimuli were presented in a pseudorandom order for 20–60 trials each. Odors were scavenged after delivery by a vacuum positioned 10 cm behind the mouse in line with the central air stream of the olfactometer.
Two-photon imaging
Two-photon imaging was performed on a Thorlabs/Janelia RAM mesoscope (Sofroniew et al., 2016) with a 16× water immersion objective and Tiberius tunable Ti:sapphire laser (Thorlabs). The laser wavelength was set to 920 nm to image GCaMP6f signals. Imaging parameters were controlled with the ScanImage software. To maximize frame rates in each imaging session, we defined fields of for acquisition that included only the dorsal surface of bilateral OBs (1,800 µm long × 1,800–2,500 µm wide). A single-plane 50–150 µm below the surface of the OB was imaged from each mouse. Images were acquired continuously throughout experiments (during odor presentations and intertrial intervals), with 5 µm/pixel resolution at the fastest possible frame rate allowed by the imaging parameters (15–18 Hz). After imaging, videos were raster corrected and motion corrected. Motion correction was done via two-iteration phase correlation to a template created by averaging the scan. Glomerular ROIs were segmented manually using the average image of all scan frames (Fig. 1B). Each ROI was drawn to encompass the cross section of a single glomerulus in the imaging plane, including all mitral and tufted cell apical dendrites within that area. The Thy1-GCaMP6f line also expresses GCaMP in piriform cortex excitatory neurons which project back to the OB, but any potential fluorescence from these projection terminals is likely diffuse, desynchronized, and low amplitude compared with GL fluorescence from MTC dendrites. Identification of glomeruli within the field of view was occasionally limited by the position of the window or curvature of the OB. Across datasets (N = 5 mice), 383 ± 47 glomeruli were segmented. Following segmentation of glomerular ROIs, fluorescence traces were extracted from ROIs and deconvolved with CaImAn (v1.0; Giovannucci et al., 2019). Proximity between glomeruli was compared with the correlation in their activity patterns to identify whether single glomeruli were systematically being erroneously segmented into multiple glomeruli. Proximity was measured as the Euclidian distance between the centroids of each glomerulus. Pearson's correlations were calculated using z-scored fluorescence traces from each glomerulus over the entire imaging session. Glomerular response vectors were constructed by averaging deconvolved traces over the duration of an odor presentation for each glomerulus–odor pair. For analyses over time, glomerulus response vectors at each time point were calculated by averaging deconvolved traces over a window of three imaging frames starting one prior to the reference frame (t − 1) and ending one after the reference frame (t + 1), creating a three-frame rolling window.
Population and lifetime sparseness
For both population and lifetime sparseness, deconvolved odor responses were summed over the duration of an odor presentation. To quantify lifetime sparseness, we used a threshold-independent sparseness index (SI) ranging from 0 (for a uniform distribution) to 1 for maximally sparse activity (Quiroga and Kreiman, 2010). To calculate the SI for an individual glomerulus, the full range of odor response magnitudes from that glomerulus was divided into 100 increments. Those increments were then used as thresholds for determining odor responses to individual odor trials. This created a cumulative distribution where each glomerulus responded to a single odor at the highest threshold and all odors at the lowest threshold. Using the area under the curve (AUC) of the cumulative distribution, the lifetime SI (LSI) for each glomerulus was calculated as
Machine learning methods
We trained multiclass linear classification models to identify 12 stimulus categories (11 odors and mineral oil) using the packages Scikit-learn (Pedregosa et al., 2011) and Keras (François, 2015), open-source machine learning frameworks in Python (v.3.7.14). We trained one-versus-rest (OVR) logistic regression with L2 regularization to avoid overfitting. We also trained a support vector classifier (SVC) with a linear kernel, and then we evaluated the models using stratified fivefold cross-validation and a confusion matrix (Wright, 1995; Menard, 2001; Kleinbaum et al., 2002). We also applied nonlinear classifiers including convolutional neural networks (CNNs; Z. Li et al., 2021), random forests (RFs; Breiman, 2001), and XGBoost (Chen and Guestrin, 2016). For the CNN, we built a fully connected neural network with one convolutional layer, followed by a ReLu activation function, a pooling layer, and a fully connected layer. We used a batch size of 32 and learning rate of 0.001 with an Adam optimizer (Kingma and Ba, 2014). The RF model used 1,000 estimators and a maximum depth of 200. XGBoost used 100 trees and a learning rate of 0.1.
RF feature importance calculation
One approach we took to rank glomeruli was to use an RF to quantify “feature importance” (FI). An RF is an ensemble method that combines multiple decision trees to improve predictive performance and control overfitting. Each decision tree is trained on a random subset of the data and consists of a set of internal nodes and leaves. At each internal node, the most significant feature is used to split the data into subsets based on feature values, aiming to increase the difference of the resulting subsets. We chose to use Gini impurity as a measure of the difference between subsets. Gini impurity measures the frequency at which any element of the dataset would be misclassified if it was randomly labeled according to the distribution of labels in the subset. In our case, Gini impurity ranges from 0 (pure) to 0.92 (maximum impurity for a 1 of the 12 classifications with balanced classes). Assume at node m of the binary tree t the best split is derived based on Gini impurity
The data are then split into sets based on the Gini impurity. Finally, for each tree t and for each feature i, the importance score at node mis calculated as follows:
Glomerular ranking for necessity and sufficiency analyses
For necessity and sufficiency analyses, we ranked glomeruli based on three criteria: response magnitude, classifier weight, or RF Gini impurity (RFGI) FI. Glomeruli (features) were ranked from the most significant to the least significant. Rank orders were compared across ranking methods using Kendall tau rank correlations between different methods applied to shuffled and ranked versions of the datasets (Figs. 4B, 5C). Kendall tau correlation coefficients and significance values were calculated in Python using the SciPy package (Virtanen et al., 2020). To define minimal subsets of glomeruli sufficient for odor classification, we trained and tested classifiers using only the top-ranked glomerulus. We then repeated this analysis adding in the next most significant glomerulus until all glomeruli were added (training and testing classifiers on the complete dataset). To define minimal subsets of necessary glomeruli, we used the same glomeruli rank orders and started by training and testing classifiers on the full dataset. We then repeated the analysis while excluding the top-ranked glomerulus, then excluding the next highest ranked glomerulus until all glomeruli were excluded.
Experimental design and statistical analysis
For distance versus correlation analyses, distance was calculated as the Euclidian distance between ROI centroids. Correlations between z-scores of GCaMP time series were calculated using Pearson's correlation coefficient in MATLAB across the entire imaging session. Comparisons were plotted, and best fit lines were calculated in PRISM 10 (GraphPad Software). Kendall tau rank correlations and significance values were calculated in Python as described above. Comparisons of Kendall tau values between ranked and shuffled datasets were made using paired t tests. For correlations between importance scores, response magnitudes, and classifier weights, Spearman R values were calculated using Prism 10 (GraphPad Software). For each comparison, Spearman R values were then compared with a null hypothesis of 0 using a one-way t test. Logistic regression analyses were carried out five times for each dataset (N = 5 imaging sessions across five mice, with one dataset analyzed per mouse). Design and implementation of logistic regression analysis are described in detail in Materials and Methods, Machine learning methods. Accuracy curves were averaged across the five replicates to obtain an average add-in and a dropout accuracy curve for each mouse. Average accuracy curves for each dataset were used to calculate the across-dataset mean accuracy curve and 95% confidence interval.
Results
Mesoscale two-photon imaging in the mouse OB reveals dense odor-evoked activity
To investigate the density and distribution of odor information across glomeruli requires (1) the ability to record odor-evoked activity across a large population of glomeruli with high signal-to-noise ratios to capture weakly active glomeruli and (2) methods to quantify odor information represented in large populations of glomeruli. To accomplish this, we used mesoscale two-photon calcium imaging to record odor-evoked activity from the GL of the mouse OB (Fig. 1A). This enabled us to record the activity of hundreds of glomeruli simultaneously, from bilateral OBs, continuously across hundreds of odor presentations and interstimulus intervals. Mice expressed GCaMP6f under the control of the Thy1 promoter (Thy1-GCaMP6f 5.11; JAX, 025393) which allowed us to visualize mitral and tufted cell apical dendrites in the GL and segment images into regions of interest (ROIs) covering individual glomeruli (Fig. 1B). Average fluorescence images from each imaging session were to segment of glomerular ROIs (383 ± 47 glomeruli per dataset, N = 5). During imaging, awake mice were passively presented with 1 of the 11 monomolecular odors or mineral oil diluted in a stream of clean air. Each odor was presented 20–60 times for a total of 480 odor presentation trials per session. We extracted fluorescence traces from each glomerular ROI, aligned them to odor presentations (Fig. 1C), and deconvolved them to estimate the timing of calcium signaling events more precisely (Fig. 1D). Notably, across pairs of neighboring glomeruli, activity patterns were poorly correlated (Fig. 1E). This confirms that single glomeruli are not systematically segmented into multiple ROIs, which would lead to highly correlated activity between ROIs with very small pairwise distances. The low pairwise correlations also indicate that widespread, low-amplitude activity is heterogeneous across glomeruli.
Mesoscale two-photon imaging of odor responses in the OB of awake mice. A, Schematic showing mesoscale two-photon imaging during odor presentations. Twelve different stimuli including 11 odors and mineral oil were presented in a random order for 1 s each separated by a delay of 19 s. Each imaging session included 480 odor presentation trials in a random order. B, Left, Average of motion-corrected imaging frames showing Thy1-GCaMP6f expression in the GL of the mouse OB. Right, The same average image with overlaid glomerular ROIs (blue). C, Activity from all glomeruli across four odor presentations (colored bars) thresholded so that time bins with z-scored activity >3 appear black. D, Deconvolved fluorescence traces from a subset of glomeruli (inset in G) showing response to TCN (colored bar). E, Comparison between Pearson’s correlations of raw GCaMP traces from pairs of glomeruli and their pairwise distance, showing only pairs separated by <300 µm. Thick lines represent exponential fits of all pairs separated by <300 µm. F, Cumulative probability distribution (CPD) of lifetime sparseness, calculated for each glomerulus using the threshold-independent SI described in. Each dataset is plotted separately. Thick lines represent the empirical cumulative distribution functions for each dataset. Shaded areas represent SEM. G, CPDs of population sparseness, calculated as the percentage of glomeruli responding to each odor presentation. Each dataset is plotted separately. Thick lines represent the empirical cumulative distribution functions for each dataset. Shaded areas represent SEM. H, Example heatmaps of dF/F averaged across the duration of a single presentation of each of 11 odors and mineral oil. Baseline fluorescence was calculated for each pixel as the average fluorescence over 5 s prior to odor presentations. Scale bars, 200 µm. Color bar, dF/F.
To assess the overall sparseness of glomerular odor responses, we calculated the sparseness of activity for each glomerulus over time (lifetime sparseness) as well as the sparseness of the population response to each odor presentation (population sparseness). To calculate lifetime sparseness, we used a threshold-independent SI ranging from 0 (for a uniform distribution) to 1 for maximally sparse activity (Quiroga and Kreiman, 2010). Across mice, the median lifetime sparseness of glomeruli was 0.67 ± 0.04 (Fig. 1F). This value is less sparse than what is seen for very low concentration odors (<0.1%) but in line with previous studies using higher concentration odors (Tan et al., 2010; Burton et al., 2022). We calculated population sparseness as the percentage of glomeruli responding above the threshold to each odor presentation. We set thresholds for each glomerulus at two times the standard deviation above their mean activity for all odor presentations. Across all mice, median population sparseness was 2.06 ± 0.66% (Fig. 1G). This method for calculating population sparseness is more appropriate than kurtosis for non-negative neuronal activity data (Willmore et al., 2011). However, a disadvantage of this method is that it relies on a strict threshold of activity defined across odor responses. Therefore, it necessarily ignores glomeruli with lower response amplitudes and lower signal-to-noise ratios which may cause the population response to appear artificially sparse. This method also ignores glomeruli suppressed by odor presentations which may also increase sparsity (Shani-Narkiss et al., 2023). Further illustrating the relative density of our observed odor responses, averaging dF/F across the duration of a single odor trial revealed a small population of highly active glomeruli as well as a larger population of lowly active glomeruli across odors (Fig. 1H).
Understanding how widespread, low-amplitude activity contributes to odor decoding requires a threshold-independent method for quantifying odor information in glomerular population activity. To accomplish this, we trained classifiers to decode odors from population-level glomerular response patterns on individual trials. In this approach, classifier accuracy provided a quantifiable measure of the extent to which odor information was present in glomerular activity. A particular advantage of our classifier-based analysis is that it provides a way to quantify the odor information available in different subsets of glomerular response patterns partitioned across space and time. To define glomerular odor response patterns, we summed the deconvolved fluorescence traces from individual glomeruli across the duration of individual odor presentations. Time-averaged glomerular response patterns, sorted by odor, showed robust odor-specific response patterns across glomeruli (Fig. 2A). To classify odors from glomerular response patterns, we employed supervised linear and nonlinear machine learning methods. We tested five architectures on deconvolved calcium traces including OVR logistic regression, SVC, CNN, RF, and XGC Boost (Chen and Guestrin, 2016). We used a train–test ratio of 80% training trials and 20% testing trials from within the same imaging session, repeating each experiment five times. Comparing accuracy data across imaging sessions from five mice, we found the linear method, OVR logistic regression, was the best performer. Notably, nonlinear models (CNN, 74.1 ± 4.75%; RF, 73.3 ± 4.25%; XGC Boost, 74.6 ± 4.75% accuracy) were generally outperformed by linear methods (SVC, 89.8 ± 3.93%; OVR, 90.5 ± 3.00% accuracy). This was unexpected but can likely be attributed to the nature of our datasets, which are constrained by a limited number of unique odors and trials, as well as the high data demands of nonlinear models. OVR logistic regression was highly accurate across the panel of 11 odors and the odorless mineral oil control when trained and tested on full glomerular odor response patterns (Fig. 2B). In subsequent classifier analyses, we applied the OVR logistic regression configuration described above.
Diverse classifiers discriminate odors from mouse OB glomerular responses. A, Heatmap showing deconvolved fluorescence activity summed across 1 s odor presentations for every glomerulus (columns) and odor presentation (rows). Rows show glomerulus response vectors of individual trials, sorted by odor. B, Confusion matrix summarizing OVR SoftMax logistic regression classifier predictions on the x axis versus true odor identities on the y axis. Color bar and text show counts of classifier predictions versus true odor identities normalized to the number of true odor presentations for each odor and averaged across five datasets.
Odor information is densely, redundantly represented across a large proportion of glomeruli
To determine how odor information is represented in subsets of glomeruli, we ranked glomeruli according to different response features (i.e., odor response magnitudes or weights initially assigned by OVR logistic regression). Then we either added glomeruli into the classifier analysis one at a time (starting with one) to test sufficiency (Fig. 3A) or removed them from the analysis one at a time (starting with all) to test necessity (Fig. 3B). First, we added and removed glomeruli from classifier analyses in a random order—without sorting or ranking by odor reposes. Adding glomeruli in a random order resulted in a rapid increase in classifier accuracy as the first 10% of glomeruli were added and a gradual increase in accuracy as the remaining 90% were added (Fig. 3C). When glomeruli were removed in a random order, accuracy failed to decrease until the last 20% of glomeruli were removed (Fig. 3D). Together, the fast rise in classifier accuracy with random addition and the slow decline in accuracy with random subtraction imply redundancy in the odor representation, where a large proportion of glomeruli are sufficient to represent odor identity, but few are individually necessary.
VA analyses reveal dense and redundant representations of odors in the mouse OB. A, A diagram showing addition of glomeruli to a OVR classifier analysis based on either maximum odor response magnitude or VA—defined as the weight assigned to each glomerulus by a classifier trained once on the full subset of glomeruli. OVR classifiers are first trained using the full subset of glomeruli. Then, in each iteration, one glomerulus is added into the analysis based on its response magnitude or VA ranking, and the classifier is retrained and retested. This repeats until all glomeruli have been added. B, A diagram showing the removal of glomeruli based on the criteria in A. Removal follows the same iterative approach where, in each iteration, one glomerulus is removed and classifier retrained and retested, repeating until no glomeruli remain. C, Add-in control showing how classifier accuracy changes when glomeruli are added into the classifier analysis one at a time in a random order, unrelated to their odor response properties. Thin lines are averages of 10 training and testing runs for one mouse. Thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. Mean and CI from this panel are duplicated in panels E and G for comparison. D, Dropout control showing how classifier accuracy changes when glomeruli are removed from the classifier analysis one at a time in a random order, unrelated to their odor response properties. Thin lines are averages of 10 training and testing runs for one mouse. Thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. Mean and CI from this panel are duplicated in panels F and H for comparison. E, Change in classifier accuracy when glomeruli are added to classifier training and testing one at time based on their maximum odor response magnitude. Thin lines are averages of 10 training and testing runs for one mouse. The thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. The light gray line and shaded area show mean and CI of the random-order add-in analysis in C. F, Change in classifier testing accuracy when glomeruli are removed from classifier training and testing one at time based on their maximum odor response magnitude. Thin lines are averages of 10 training and testing runs for one mouse. The thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. The light gray line and shaded area show mean and CI of the random-order dropout analysis in D. G, Classifier testing accuracy when glomeruli are added to classifier training and testing one at time based on their VA rank. Thin lines are averages of 10 training and testing runs for one mouse. Thick line is average across mice. Shaded area is 95% confidence interval across mice. Dotted line is chance. The light gray line and shaded area show mean and CI of the random-order add-in analysis in C. H, Classifier testing accuracy when glomeruli are removed from classifier training and testing one at time based on their VA rank. Thin lines are averages of 10 training and testing runs for one mouse. The thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. The light gray line and shaded area show mean and CI of the random-order dropout analysis in D.
To test the redundancy of odor representations in mouse OB glomerular activity, we next sought to determine minimal subsets of glomeruli necessary and/or sufficient for odor classification. To do this, we ranked glomeruli according to different features of their odor responses. As an intuitive and interpretable measure, we first ranked glomeruli based on the magnitude of their largest response to any odor. This targeted the highest responding glomeruli for addition or removal. We found that adding glomeruli based on response magnitude caused a sharp increase in classifier accuracy with a small subset of glomeruli added (Fig. 3E). Adding glomeruli based on random ranking resulted in a much slower, shallower increase in classifier accuracy (Fig. 3E, gray reproduces Fig. 3C mean). Including just the single highest ranked glomerulus led to classifier accuracy above chance. This agrees with previous studies demonstrating that sparse glomerular activation is sufficient for odor decoding and olfactory discrimination. Interestingly, it further implies that the summed activity of a single glomerulus over the course of an odor presentation is sufficient to classify more than one odor above chance. This supports earlier work showing that mice are able to discriminate distinct optogenetic manipulations of a single glomerulus (Smear et al., 2013).
Next, we tested whether odor information was restricted to this small subset of sufficient glomeruli. To do this, we performed a similar analysis removing glomeruli one at a time based on response magnitude (Fig. 3B). If odor information is concentrated in a few glomeruli, then we would expect that the subset of glomeruli sufficient for odor decoding would also be necessary and that removing them from the classifier analysis would cause testing accuracy to rapidly decrease to chance. However, when we removed glomeruli one at a time based on response magnitude, we found an unexpectedly gradual decrease in classifier test accuracy (Fig. 3F). Removing glomeruli based on random ranking resulted in an even slower, convex decrease in classifier accuracy (Fig. 3F, gray reproduces Fig. 3D mean) indicating that higher-ranked glomeruli are indeed more important for odor decoding. However, this analysis suggests that odor information is distributed across a broad population of glomeruli with a wide range of odor response magnitudes.
During odor presentations, a subset of glomeruli are strongly activated across all trials of the same odor, but a large population are activated at a lower magnitude with more variability across trials. To determine whether a more general metric of a glomerulus's contribution to odor decoding would provide a better estimate of the minimal necessary subset of glomeruli (i.e., a steeper decrease in accuracy with glomerulus removal), we performed a similar classifier-based necessity and sufficiency analysis ranking glomeruli based on their initial weights assigned by the OVR classifier. Similar to ranking glomeruli by response magnitude, we found that ranking glomeruli by their VA to the classifier led to a sharp increase in accuracy as glomeruli were added into the analysis one at a time (Fig. 3G). Removing glomeruli based on VA, we again found a slow, gradual decrease in accuracy (Fig. 3H). Together these data demonstrate that methods ranking glomeruli based on response magnitude or VA can identify high-information glomeruli that are sufficient for odor classification, but they fail to show that these same populations of glomeruli are necessary for odor classification. This means that either (1) odor information is broadly and redundantly distributed across a large population of glomeruli or (2) response magnitude and classifier weight are not effectively identifying the glomeruli that are most necessary for odor decoding even if they are effectively identifying sufficient glomeruli. In either case, the discrepancy between the analyses adding glomeruli and removing glomeruli suggest that odor information is denser and more redundant across glomeruli than has been previously appreciated.
Having found that response magnitude and classifier weight fail to define sparse populations of necessary glomeruli for odor decoding, we sought a more robust method to identify necessary subsets. To accomplish this, we applied a recursive approach using greedy optimization to rank the glomeruli. We added/removed one glomerulus at a time based on their contribution to classifier performance—repeatedly creating new models using the remaining glomeruli until all glomeruli were added/removed. Then, we ranked the glomeruli based on the order of their addition/elimination (Fig. 4A). The recursive VA approach, by definition, generated the fastest possible increase in classifier accuracy with glomerular addition and the fastest decrease in accuracy with glomerular subtraction because glomeruli are added or subtracted based on the degree to which they affect accuracy. Thus, the recursive VA approach reveals a ground truth ranking of which glomeruli are most informative to the OVR classifier. First, we compared glomerular ranks assigned by the recursive VA method with ranks assigned by the nonrecursive VA method (i.e., the initial classifier weights) and found heterogeneous similarity across datasets (Fig. 4B, p value range, 10−27–0.1). Comparison of true rank similarity (rank similarity, Kendall tau, 0.14 ± 0.07; p = 0.029 ± 0.02) with rank similarity when lists were shuffled (rank similarity, Kendall tau, −0.01 ± 0.01; p = 0.51 ± 0.10) revealed a statistically insignificant difference between the shuffled and unshuffled conditions (paired t test, t(4) = 2.172; p = 0.096) indicating that recursive and nonrecursive VA ranking methods identify different subsets of highly important glomeruli. We then performed similar classifier-based tests to determine necessary and sufficient subsets of glomeruli based on the recursive VA ranking. Similar to analyses ranking glomeruli by response magnitude and nonrecursive VA, we found that adding glomeruli based on their recursive VA rank sharply increased classifier accuracy (Fig. 4C). Removing glomeruli by recursive VA rank still revealed a gradual decline in classifier accuracy (Fig. 4D). However, examining the rate of decay in classifier accuracy, we found a sharper decline during removal of the top 20% of glomeruli, indicating that they contributed proportionally more to odor decoding. Accuracy reduced more gradually upon removal of the last 80%. Importantly, the improvement of the recursive over the nonrecursive approach indicates that as individual glomeruli are masked, the information represented by others become disproportionately more important to the OVR classifier. This suggests that odors are redundantly represented across populations of glomeruli.
Recursive ranking better identifies minimal subsets of glomeruli necessary for classification. A, A diagram illustrating recursive addition/removal of glomeruli based on ranking by recursively determining classifier weights assigned to individual glomeruli (VA). In each iteration, one glomerulus is added/removed, then the model performance is calculated, and then weights are recalculated for the remaining glomeruli. B, Kendall rank correlation coefficient (tau) showing similarity of rankings obtained with recursive and nonrecursive VA compared with shuffled rankings of the same datasets. Colors indicate the significance of rank similarities within datasets. Comparing Kendall tau values for ranked datasets to Kendall tau values for the same datasets shuffled with a paired t test shows no significant difference between true and shuffled rank correlations (p = 0.096). C, Classifier testing accuracy when glomeruli are added to OVR logistic regression training and testing one at time based on their recursive VA. Thin lines are averages of 10 training and testing runs for one mouse; the thick line is the average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. D, Classifier testing accuracy when glomeruli are removed from OVR logistic regression training and testing one at time based on their recursive VA. Thin lines are averages of 10 training and testing runs for one mouse; the thick line is the average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance.
RF feature selection using Gini impurity accurately identifies subsets of glomeruli necessary for odor decoding
An important caveat of the recursive VA approach is that it relies on defining glomerular importance based on contribution to classifier accuracy. Thus, rankings may be an artifact of the classifier used to perform the analysis. To circumvent this, we next sought to measure the contribution of individual glomeruli to odor decoding (1) without predefining important response features (i.e., response magnitude), (2) without using information from the classifier analysis itself, and (3) without recursively recalculating measures of importance. To accomplish this, we used information theoretic measures to determine the importance of individual glomeruli for odor decoding. We employed RF feature selection which is also recursive and has the advantage of being highly applicable to high-dimensional data (Menze et al., 2009). RF is an ensemble classifier that uses a variety of decision trees to subdivide data according to specific criteria (Breiman, 2001). At each split in a set of decision trees, a selected criterion is used to divide the data into two sets. The splits are then used to calculate the importance of individual features in the data for correct classification. In the current study, we used a measure of inequality—Gini impurity—as the criteria to determine splits in the RF decision trees (RFGI). Gini impurity is a measure of entropy or randomness, so reducing Gini impurity provides a way to determine which features in a dataset are less random and contribute more to correct classification (Archer and Kimes, 2008). In the RFGI method, at each node in the decision tree, the split in the data that provides the highest decrease in Gini impurity is selected, ultimately providing a measure of importance (FI score) for each individual glomerulus in the dataset (Fig. 5A). We hypothesized that the more comprehensive RFGI method would estimate the recursive VA approach in identifying minimal subsets of necessary and sufficient glomeruli.
RF feature selection using Gini impurity approximates recursive VA ranking. A, Distribution of RFGI FI scores across datasets. B, A diagram showing recursive removal of glomeruli based on RFGI importance score. In each iteration, one glomerulus is removed based on ranking by importance score, then the classifier's performance is calculated, and then importance scores are recalculated for remaining glomeruli. C, Kendall rank correlation coefficient (tau) showing similarity of rankings obtained with recursive VA versus FI score (left) compared with shuffled rankings of the same dataset. Colors indicate significance of rank similarities within datasets. ***p < 0.001, pairwise t test comparing Kendall tau values for ranked datasets to Kendall tau values for the same datasets shuffled. D, Classifier accuracy as glomeruli were added into the classifier analysis one at a time based on FI score. Thin lines are averages of 10 training and testing runs for one mouse. The thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. E, Classifier accuracy as glomeruli were removed from the classifier analysis one at a time based on FI score. Thin lines are averages of 10 training and testing runs for one mouse. The thick line is average across mice. The shaded area is 95% confidence interval across mice. The dotted line is chance. F, Comparison of FI score with mean odor response magnitude (average dF/F across all odor presentations). Individual points indicate glomeruli. Colors indicate datasets. G, Comparison of FI score with weight initially assigned by the classifier (OVR weight). Individual points indicate glomeruli. Colors indicate datasets. H, Spearman correlation between FI score and response features in B (mean response magnitude) and C (OVR weight), for each dataset. Circles represent datasets. Lines and error bars show mean and SEM across datasets. ***p < 0.001 one-sample t test; **p < 0.01 one-sample t test.
Using RFGI FI scores to rank glomeruli, we then performed a similar classifier-based necessity and sufficiency analysis. We added or removed glomeruli one at time based on their initial (nonrecursive) FI ranking (Fig. 5B). Notably, ranking based on FI was highly similar to ranking based on the recursive VA determination previously described (Fig. 5C). Glomerular ranks according to the recursive VA and nonrecursive RF approach were compared using the Kendall rank correlation coefficient. Comparison of rank similarity (rank similarity, Kendall tau, 0.44 ± 0.04; p = 4.4 × 10−24 ± 1 × 10−23) with rank similarity when lists were shuffled (rank similarity, Kendall tau, 0.009 ± 0.01; p = 0.58 ± 0.14) revealed significantly higher similarity in the unshuffled condition for all datasets (paired t test, t(4) = 11.11; p = 0.0004). The similarity in ranks determined by the recursive VA and RFGI approaches implies that RFGI provides a more accurate ranking of glomerular importance for odor decoding than either the intuitive measure of response magnitude or initial classifier weight (nonrecursive VA).
Similar to analyses ranking glomeruli by response magnitude and both recursive and nonrecursive VA, we found that adding glomeruli based on FI sharply increased classifier accuracy (Fig. 5D). This further supports that RFGI FI scores accurately identify a small subset of glomeruli sufficient for odor classification. Removing glomeruli by FI score again revealed a gradual decline in classifier accuracy (Fig. 5E). However, examining the rate of decay in classifier accuracy, we found the decline is sharper during removal of the top 20% of glomeruli, indicating that they contribute proportionally more to odor decoding. Accuracy reduced more gradually upon removal of the middle 60% of glomeruli (from 20 to 80%) though overall accuracy remained well above chance. Accuracy again declined steeply upon removing the last 20% of glomeruli. In contrast to ranking glomeruli based on response magnitude or nonrecursive VA, the sharper initial decline in accuracy when removing glomeruli based on FI suggests that this method better identifies necessary and important glomeruli. RFGI feature selection likely relies on more complex aspects of glomerular population activity to define which glomeruli are contributing to odor decoding. Accordingly, this method outperforms methods relying on response magnitude alone or nonrecursive VA and can approximate recursive VA analysis effectively. Further investigating the relationship between FI scores and other methods for ranking glomeruli, including response magnitude (Fig. 5F) and weight (Fig. 5G), we found that importance scores were moderately correlated with both response magnitude (Fig. 5H, one-sample t test, t(4) = 8.202; p = 0.0012) and classifier weight (Fig. 5H, one-sample t test, t(4) = 11.30; p = 0.0003). Together, these results suggest that (1) the overall importance of specific glomeruli to odor decoding is, unsurprisingly, related to the magnitude of their odor responses and (2) more surprisingly, RFGI identifies a different subset of important glomeruli compared with the OVR classifier trained on all glomeruli simultaneously, while it identified a similar subset of important glomeruli to the classifier recursively retrained with each glomerular removal. Thus, RFGI provides a powerful tool for identifying minimal subsets of necessary and sufficient glomeruli within large populations.
Redundancy of odor information decreases after odor offset, though odor-specific activity persists
Finally, even though glomerular activity evolves over the course of an odor presentation, it maintains odor specificity throughout odor delivery and after odor offset (Patterson et al., 2013). To test whether the extent of the redundancy and distribution of odor encoding evolved over the duration of an odor recording and after odor offset, we performed a similar classifier analysis using glomerular activity from sliding time window of three imaging frames across times during odor presentations and after odor offset. First, including the entire population of glomeruli in the analysis, we found that classifier accuracy peaked during the odor presentation but remained above chance for seconds after odor offset (Fig. 6A). This is consistent with previous studies showing that odor information persists in OB activity after odor offset, even though the lingering odor-specific activity is distinct from directly odor-evoked activity (Patterson et al., 2013; Ling et al., 2023). Notably, accuracy decays unevenly across odors after odor offset (Fig. 6B), suggesting that the persistence of odor information after odor offset is odor dependent. To examine how the density of odor information evolved over time, we carried out a similar VA analysis based on removing glomeruli based on RFGI FI. We found that the density of odor information sharply declines after odor offset. For example, using glomerular activity from odor offset to 1 s after odor offset (Fig. 6C, 1 s postodor), we found that classifier accuracy decreased to chance more rapidly as glomeruli were removed from the analysis compared with when the time window analyzed was during the odor presentation (Fig. 6C, during odor). Comparing RFGI analyses across time windows, we found lower overall accuracy using later time windows and faster declines to chance (Fig. 6D). Together, these data indicate that after odor offset, odor information persists in glomerular activity, though it is maintained by a smaller subset of glomeruli.
Odor information after offset is increasingly concentrated in a small population of glomeruli. A, Classifier accuracy when trained and tested on a three-frame sliding window relative to odor onset including preodor (1 s prior to odor onset), during odor (at odor onset), 1 s postodor, 2 s postodor, 3 s postodor, and 7 s postodor. Vertical lines show odor onset and offset. Thin lines are averages of 10 classifier training and testing runs for an individual mouse. The thick line is average across mice, and the shaded area shows the 95% confidence interval across mice. B, Classifier confusion by odor when using different time windows relative to odor onset to train and test the classifier. Time windows include preodor, during odor, 1 s postodor, 2 s postodor, and 3 s postodor. The x axis represents true odor identities, and the y axis represents odors predicted by the classifier analysis. Colors represent normalized counts of what each odor presentation was predicted as by the classifier. The diagonal line from the top left to bottom right represents accurate predictions. C, Importance of score-based VA analysis using different time windows to define glomerular activity patterns for classifier training and testing including preodor, during odor, 1 s postodor, 2 s postodor, and 3 s postodor. Individual mice are shown as thin lines. Averages across mice and SEM are shown as thick lines and shaded area, respectively. Insets show average and SEM across mice for 0–20% of glomeruli removed. D, Overlay of mouse averages and SEM of classifier accuracy removing 0–20% of glomeruli. Colors represent different time windows relative to odor presentations: preodor (purple), during odor presentation (magenta), 1 s postodor (orange), 2 s postodor (green), and 3 s postodor (blue).
Discussion
Overlap in odor-evoked activity patterns in the OB has been linked to perceptual similarity, suggesting that sparse odor responses enable odor discrimination (Gschwend et al., 2015). Paradoxically when odors are more complex, higher concentration, or naturalistic, odor-evoked activity is denser and more temporally complex. This raises questions of how odors are reliably encoded in relatively dense and complex patterns of OB activity at the earliest stages of olfactory transduction. To address these questions, we recorded population-level odor responses from the mouse OB with mesoscale two-photon calcium imaging, and we applied information theoretic measures to map odor information in glomerular activity patterns over space and time.
Previous studies using imaging or electrophysiological methods to record neuronal activity in the OB have largely relied on quantifying trial-averaged responses from a small subset of reliably time-locked glomeruli with high-magnitude responses and/or recording from small populations of active neurons (Ressler et al., 1994; Mombaerts et al., 1996; Bozza et al., 2002; Rinberg et al., 2006a; Davison and Katz, 2007; Soucy et al., 2009; Burton et al., 2022). These approaches can mask biologically relevant variability and lower-magnitude glomerular activity in population-level odor responses, emphasizing a sparse subset of active units (glomeruli or neurons). Additionally, stimulating sparse units in the OB to drive odor-guided behavior (Chong et al., 2020; Gill et al., 2020) highlights glomerular activity patterns sufficient for odor decoding but does not reveal which patterns are necessary or used under physiological circumstances. Supporting this, recent imaging of mitral cell odor responses suggests that suppressive responses are important carriers of odor information, increasing the distribution and discriminability of odors from mitral cell activity (Shani-Narkiss et al., 2023). This contrasts with earlier work showing sparser and less discriminable odor responses in mitral cells of awake mice when only excited mitral cells are considered (Kato et al., 2012). Calcium imaging of mitral cell bodies is limited to a spatially constrained subset of mitral cells, and focusing on “responsive cells” further biases responses to appear sparser. On the other hand, electrophysiological approaches, while offering high temporal and cellular resolution, either do not capture suppression or subthreshold activation or are limited in the number of cells that can be simultaneously recorded. These contrasts demonstrate how experimental and analytical approaches can bias measures of information distribution across a neuronal population, usually causing representations to appear overly sparse. Our glomerular imaging and analysis methods overcome some of these limitations by (1) grouping dendrites from functionally related sister cells and imaging large populations of glomeruli directly, (2) including subthreshold activation and suppression as measurable with calcium imaging, and (3) applying a threshold-independent measure of lifetime sparseness. Therefore, we provide a more comprehensive measure of total odor information available in OB glomerular activity over time.
In the current study, we directly examine information available in individual odor responses by using in silico classifiers to predict odor discrimination from population glomerular activity. Notably, similar decoding-based approaches also reveal denser odor representations in the piriform cortex—where the odor code was previously thought to be highly sparse (Bolding and Franks, 2017). While an in silico decoder-based approach does not directly reflect behavioral odor discrimination or olfactory perception, it does indicate whether there is sufficient information available from which odor discrimination would be possible. Our current study expands on the idea that odors are encoded by sparse subsets of glomeruli by providing evidence that relatively dense, low-magnitude, and variable-latency responses also contain extensive odor information, both during and after odor exposures.
It is possible that early olfactory circuits use both sparse and dense odor representations to achieve different computational goals (Bolding and Franks, 2017). In early olfactory circuits, sparse and low-latency glomerular activation allows rapid odor identification on the timescale of a single sniff (Uchida and Mainen, 2003; Abraham et al., 2004, 2010; Rinberg et al., 2006b). On the other hand, dense activity and redundant coding may enable robustness to changing contexts, environments, neuromodulation, behavioral states, and circuit plasticity. Additionally, the temporal persistence of odor information may provide a lingering representation of previously encountered odors (Patterson et al., 2013) and enable aggregation of odor information over sniff cycles. Supporting this, it has been found that mice take slightly longer to effectively discriminate very similar odors (Abraham et al., 2004) and discrimination accuracy increases when longer odor sampling times are enforced (Rinberg et al., 2006b). A limitation of our study is that it does not include behavior or circuit manipulations to examine these possibilities directly. An advantage, though, is that it generates theory-guided predictions. In the future, it will be important to address how redundant odor information is utilized in these contexts to impact olfactory perception and odor-guided behavior.
While sparse, low-latency glomerular activity is directly triggered by OSN input, the density and temporal characteristics of redundant glomerular activity suggest that it is unlikely to be directly downstream OSN input. Indeed, the persistence of odor information after odor offset has been previously shown to be independent of persistent OSN input, implying that odor-specific patterns of activity are centrally maintained instead of driven by ongoing activity from OSNs (Patterson et al., 2013). Odor-specific activity may be centrally maintained within the OB by cell intrinsic properties or persistent local circuit activity. In insect models, it has been shown that persistent odor-specific activity relies on a combination of cellular adaptation and local inhibitory feedback circuits (Saha et al., 2017). In the mouse OB, local circuits include both excitatory and inhibitory neurons (Nagayama et al., 2014), and recruitment of these circuits, particularly feedforward excitation via external tufted cells, is an intriguing candidate mechanism for driving secondary, dense, odor-specific activity in the OB (De Saint Jan, 2022). At the same time, the mouse OB receives extensive top–down input from olfactory areas including the piriform cortex and anterior olfactory nucleus (Carson, 1984; Shipley and Adamek, 1984; Boyd et al., 2012; Rothermel and Wachowiak, 2014), as well as from neuromodulatory centers like the locus ceruleus and basal forebrain (Shipley et al., 1985; Zaborszky et al., 1986; Gielow and Zaborszky, 2017). Long-range glutamatergic feedback from the piriform cortex is a less likely candidate for maintaining persistent odor representations as piriform feedback has been shown to affect gain but not timing of OB odor responses (Bolding and Franks, 2018). Nevertheless, feedback from olfactory cortical areas exhibits odor-specific activity, and it remains possible that it drives odor-specific activity in the OB without changing the timing of the population response. Neuromodulatory feedback, while not odor specific, may serve to amplify or alter the properties of odor-specific cortical feedback or local circuit activity—providing a mechanism for modulating secondary odor-specific activity separately from activity directly downstream of OSNs. Future work can address this directly by manipulating feedback to the OB and measuring its impact on the density of odor representations.
One important limitation of our approach is the disparity between the high dimensionality of the imaging data and the comparatively lower dimensionality of the stimulus panel. Although our large field of view imaging allows for the theoretical exploration of a much larger stimulus space, we are practically constrained by the number of odors and trials that can be presented in a single imaging session. This undersampling of the stimulus space likely leads to artificially high classifier accuracies. Importantly, our data suggest that some odor information is available in redundant activity, as evidenced by classifier accuracy above chance levels. However, we do not claim, nor do we believe, that redundant activity precisely subdivides the entire stimulus space—rather, our results suggest that it provides enough information to narrow down the stimulus space. This is reflected in our low-but-above-chance accuracies, even with our limited stimulus panel. We anticipate that this would remain true with larger odor panels and more comprehensive coverage of the stimulus space. Another related limitation is that our stimulus panel does not explore the dimension of odor concentration. Future work is necessary to determine whether and how concentration information is redundantly represented in the OB.
Together our data indicate that odor information is redundantly and persistently represented in the activity of a large population of glomeruli in the mouse OB. In contrast to previous work framing odor encoding as exclusively sparse, we suggest a model in which sparse coding is largely sufficient for odor identification, but redundant information may make odor coding more robust across different internal and environmental variables and over time. Future work will be needed to determine the extent to which dense, redundant, odor coding is the result of local circuit or feedback activity, how dense activity is transformed between glomerular input and MTC output, how this activity contributes to odor coding across different states and contexts, to what extent it applies to complex mixtures of odors, and to how it is influenced by neuromodulation. Ultimately, uncovering mechanisms that drive and modulate the redundancy, distribution, and persistence of odor information in early olfactory circuits will help reveal how these circuits optimize odor encoding for perception under widely variable real-world conditions and changing internal states.
Footnotes
This work was supported by the National Institutes of Health through National Institute of Neurological Disorders and Stroke and the BRAIN Initiative (U01 NS111692 to B.R.A. and P.P.) and National Institute on Deafness and Other Communication Disorders and the BRAIN Initiative (R00DC019505 to E.H.M.).
↵*D.P. and C.L.S. contributed equally to this work.
The authors declare no competing financial interests.
- Correspondence should be addressed to Elizabeth H. Moss at mosse{at}ohsu.edu or Ankit Patel at ankit.patel{at}rice.edu.
