We hypothesize that the ongoing activity of the mouse olfactory bulb (OB) encodes information about action and environment in order to contextualize odor-driven input from the nose (Freeman, 1978). This hypothesis predicts that the OB tracks variables such as behavioral state and place, even in the absence of an experimental task. To capture spontaneous behavior and neural dynamics, we implanted mice (n=4) with intranasal thermistors and silicon electrode arrays in the OB, and tracked their movements in a 40 by 15 cm behavioral arena from video under ambient light (see Methods; Fig 1a). We did not impose olfactory stimuli, task structure, or rewards, so that mice experienced only ambient stimuli and generated spontaneous behavior. Most of our recording sessions included a period of head fixation on a stationary platform for comparison with prior experiments (Shusterman et al., 2011), followed by a freely moving period, and then a second head-fixed period, which lasted between 60-90 minutes in total.

Even in this minimal experimental paradigm, mice exhibited consistently structured behaviors. Mouse breathing is coupled with orofacial and locomotor movements during natural behavior (Findley et al., 2021; Kurnikova et al., 2017; Weinreb, Pearl, et al., 2024). As expected from previous work, breathing rates were overall higher during free movement than during head fixation (Fig 1B, top), and breathing rates were correlated with movement speed. In addition to replicating these expected observations, we uncovered novel features of spontaneous behavioral structure. First, we found that instantaneous breathing rates in both conditions were multimodal. During free movement the distribution of sniff frequencies was well fit by a mixture of three log-normal distributions, while during head-fixed conditions, by two (Fig 1B, Fig 1- figure supplement 1). Further, these multiple modes of breathing frequency were associated with distinct movement speeds, such that the joint distribution of sniff frequency and speed formed discrete clusters that recur across sessions and animals (Fig 1B; Fig 1- figure supplement 1). Thus, the relation between sniffing and movement was more complicated than a simple linear correlation.

In addition to the patterning in instantaneous behavior, breathing rhythms and movement speed are structured at longer timescales. The time series of sniff frequency and movement shows stateful organization over timescales of minutes (Fig 1C; Fig 1 - figure supplement 1). In contrast, these persistent states of breathing rhythms are not apparent in head-fixed conditions. To quantify these observations, we computed the autocorrelation of instantaneous breathing frequency and found that the autocorrelation functions had significantly longer timescales in freely moving than head-fixed behavior (Fig 1C). Similar long timescale structure was present in the autocorrelation of movement speed as well as the cross-correlation between breathing and movement. Our analyses demonstrate that even in task-free, ambient-odor conditions, mice perform behaviors structured at multiple timescales.

Continuous, time-varying behaviors can be described as ethograms (Branson et al., 2009; Renner, 2022; Tinbergen, 1965), which divide the time series into discrete behavioral motifs, and provide a useful partition for subsequent analyses (Findley et al., 2021; Markowitz et al., 2023; Weinreb, Pearl, et al., 2024). Motivated by the clustered and long-timescale behavioral structure we observed, we fit the breathing rhythms and movement data from all mice with a Hidden Markov Model (HMM). The model was fit to the behavioral data preprocessed to extract the moving average of speed and the distribution of breathing frequencies, both evaluated in 5 second windows. Model selection was performed using Bayesian Information Criterion (BIC; fig 1 - figure supplement 2). We find that a three-state model well describes free-moving behavioral data and that these three states effectively separate the behavioral clusters (Fig 1D). Through observation of labeled behavioral data, we name these three states “rest”, “grooming”, and “exploration”. These discrete breathing states are seen across all animals but differ in their usage across individual sessions. The average persistence times of states are on the order of tens of seconds to minutes (fig 1 - figure supplement 2). Taken together, these behavioral analyses demonstrate that mice breathe and move in consistently structured ways, even when the experimental paradigm does not impose structure upon their behavior.

We next recorded spiking activity in the OB during our task-free, ambient-stimulus paradigm. We recorded from OB with extracellular electrodes, including in our analyses units that passed quality control criteria of fewer than 5 % refractory period violations, and fewer than 10% amplitude cutoff violations (see Methods and Fig 2 – figure supplement 1). Comparing mean firing rates of individual units between head-fixed and freely moving conditions, we found that across the population the mean firing rates were only slightly although significantly different (Fig 2A; head-fixed vs freely moving median = 4.51 vs 5.31 spikes per s; p=0.002, rank sum test ).

We predicted that ongoing neuronal activity of OB would reflect the structure of spontaneous behavior. Using the behavioral HMM to partition the sessions, we ask whether behavioral states can describe the similarity of co-occurring population activity. When viewing neural activity alongside behavior, it is apparent that the population vectors are similarly organized into time-varying states (Fig 2B, center). To quantify the similarity of the activity patterns across different time bins in the recording, we computed the cosine distance between population activity in time bins of 5 seconds width, to form a similarity matrix across time throughout a session (Fig 2B, bottom). The apparent block structure of the similarity matrix supports the impression of statefulness in the neural activity. We next compared this structure to the ethograms generated by our behavioral HMM. Importantly, the slow variation in both the behavioral and neural data raises the possibility of a nonsense correlation (Harris, 2021; Meijer, 2021). To quantify similarity relative to that expected from the slow variation in the data, we scored cosine distance as the number of standard deviations away from the mean of a null distribution formed by circularly shifting the time series (see Methods). Taking the grand mean across animals, we find that within a state, activity patterns overlap more than expected under the null distribution, while across states, activity patterns overlap less than predicted by this null hypothesis (Fig 2C). To quantify how well the behavioral HMM clusters the neural data, we calculated a silhouette score, a measure of consistency within clusters, for each session with respect to a circular shift null distribution (see Methods). Most sessions differed significantly from the null prediction (Fig 2D; mean score 6 sigma; 21/25 p<0.01). These analyses show that a model based only on behavioral variables – sniff frequency and movement speed – can effectively cluster neural data from OB, more so than expected from a nonsense correlation arising from the slow variation in behavior and neural activity. Thus, OB activity tracks behavioral structure, even when the behavior is not influenced by experimental stimuli or incentivized by rewards.

OB activity is already known to be strongly modulated by breathing at the level of individual sniffs (Fukunaga et al., 2012; Macrides & Chorover, 1972; Onoda & Mori, 1980). To visualize the relationship between sniff frequency and unit activity, we aligned spike rasters to inhalation times, and sorted inhalations vertically in descending order of instantaneous sniff frequency (Fig 3A). While most units respond at a consistent latency, some fire with uniform amplitude across sniff frequencies (e.g., Fig 3A, Unit 1), while others fire preferentially during specific frequency ranges, around high (>8 sniff per s; Unit 2), middle (4-8 sniffs per s; Unit 3) or low frequencies (<4 sniffs per s; Unit 4). To capture the joint relationship between inhalation timing and breathing frequency, we describe OB unit activity using “sniff fields” (SnFs), the averaged firing rate as a two-dimensional function of latency from inhalation and instantaneous sniff frequency (Fig 4B). Units displayed a diversity of tuning to frequency, demonstrating that this tuning is not merely a monotonic scaling with sniff frequency. Thus, we observe that variation in breathing rhythm modulates the firing rate of inhalation-synchronized responses in OB units.

Across the population, SnFs can be classified into a limited number of types. We extracted SnF latency and frequency response profiles from all units that were significantly predictable by a latency/frequency Generalized Linear Model (GLM) (913/1111 units from sessions with a head-fixed period; p<0.01, sign rank test; see Methods), quantified the similarities of these profiles by calculating earth mover’s distance matrices, and used k-means clustering on these matrices to identify subtypes of SnFs. Among units which significantly encoded latency from inhalation (853/913; p<0.01, Sign rank test), the variety of SnF latency profiles can be captured with two clusters (Fig 4C). One cluster has one peak at <100 ms after inhalation, and another cluster has two peaks, one at <100 ms, and one >200 ms. These two response profiles are consistent with those demonstrated in many studies in anesthetized and awake mammals, and have been shown to correspond with tufted and mitral cell morphology, respectively (Fukunaga et al., 2012; Onoda & Mori, 1980). Separately, among the units that significantly encoded sniff frequency (638/913; p<0.01, sign rank test), clustering the SnF frequency profiles revealed three clusters preferring low, medium, or high sniff frequencies (Fig 4B). The latency and frequency profile subtypes are fairly independent; examples of both latency profile types can be found in all three frequency profile types (Fig 4 – figure supplement 1). Further, instantaneous sniff frequency has a smaller, less consistent relationship with spiking during head fixation than during free-moving conditions (Fig 4B). Taken together, we show that OB neurons not only synchronize their spiking to inhalation, but also track variation in the frequency of the breathing rhythm by varying the amplitude of their inhalation-locked ongoing firing.

To test the extent to which sniff frequency modulation can be explained by inhalation latency modulation, we used a Generalized Linear Model (GLM) to predict individual unit spiking based on behavioral variables. By comparing variables in isolation and in combination, we can ask whether a given variable uniquely contributes to a predictive model of unit firing. We tested models on held out data compared against a null, mean firing rate model by constructing a log-likelihood increase (LLHi) metric (see Methods; (Hardcastle et al., 2017). We perform 10-fold cross validation and calculate statistics on the distribution of LLHi scores. We compared the LLHi scores of GLMs trained on sniff frequency and latency from inhalation, individually and in combination. Consistent with the strong tuning apparent in SnF visualizations, including frequency in the model significantly improves the prediction in 638/913 units, whereas including latency improves the prediction in 853/913 units (Fig 5A). Thus, the OB correlation with sniff frequency is not simply explained by the previously-established synchronization of unit activity to inhalation.

We next considered how OB units correlate with movement speed. The behavioral model described above (Fig 1) predicts population activity with both sniff frequency and head movement speed. The correlation between these parameters creates a confounding ambiguity: is the correlation between behavior and neural activity best explained by sniffing, movement, or both? To resolve this ambiguity, we used the same GLM approach. Models based on SnF parameters (a combined frequency/latency model; see Methods) predict firing significantly better than the null model in 913/1153 of units. Models based on movement speed predict unit activity in 249/1153 units (p<0.01, sign rank test), a smaller but still considerable fraction of the population. Thus, as with many other sensory areas of the brain, activity in OB correlates with movement (Parker et al., 2020). However, if we quantify the contribution of these two variables in a combined sniff field/speed model, we find that movement speed improves the predictions relative to that of a model based on SnF parameters in 102/1153 of units (p<0.01, sign rank test; Fig 5B). Further, 12/1153 of the units in our sample were more predictable by speed than by the SnF parameters (Fig 5B bottom). Thus, although movement speed is strongly correlated with OB activity, this correlation is largely redundant and reflects more the correlation between breathing and movement than movement itself.

Motivated by the ethological relevance of the relationship between olfactory signals and internal spatial representations (Baker et al., 2018; Gagliardo, 2013; Jackson et al., 2020; Jacobs, 2023; Matheson et al., 2022; Raithel & Gottfried, 2021), and given previous observations of conjunctive odor/place coding in hippocampus (Fischler-Ruiz et al., 2021; Komorowski et al., 2009) and piriform cortex (Kehl et al., 2024; Mena et al., 2023; Poo et al., 2022), we investigated the relation between OB activity and place. Strikingly, individual OB units show apparent place selectivity during free behavior (Fig 6A): spiking activity is spatially modulated for many units. We wondered how this spatial selectivity compared to that of hippocampal neurons, whose place field properties have been extensively studied (Best et al., 2001). However, direct comparison to hippocampal place fields described in the literature is difficult because most recording studies in the hippocampus use experimental paradigms that differ from ours in important ways. First, the behavioral arena we used is smaller than that of most hippocampal studies. Second, hippocampal experiments typically incentivize exploration by distributing food pellets or training on a maze task. Lastly, these experiments often exclude data in which the animals are not moving above a criterion speed. While these design choices successfully establish a focus on canonical place cells, they run counter to the goals of our study. We therefore recorded from neurons in the hippocampus (HPC) of mice in the same arena and task-free experimental paradigm as our OB recordings.

We compared place selectivity between OB and HPC. Place fields of units recorded from HPC appeared to be more specific than those of OB (Fig 6b, Methods). To quantify the spatial selectivity of individual units in OB and HPC, we modified a traditional metric of place selectivity, spatial information (Skaggs et al., 1992). This information theoretic measure effectively captures the selectivity of canonical place cells, which have very low ongoing firing rates. In contrast, this measure poorly captures the selectivity apparent in neurons with higher ongoing firing rates, such as hippocampal interneurons (Frank et al., 2001; Wilent & Nitz, 2007) or OB neurons (Fig 2A). To better generalize this metric to neurons with ongoing activity, for each unit we compared spatial information to a null distribution formed by circular shifting the position time series 1000 times, and expressed the selectivity as the Significance of Spatial Information (SSI; Stefanini et al., 2020; C. Yang et al., 2024), defined as the number of standard deviations of the real data from the null distribution. This metric, as illustrated by the example cells in Figure 6A, reveals that a substantial minority of OB neurons are spatially selective (196/1557 units, p<0.01), but a significantly smaller fraction than in HPC recorded under the same conditions (270/468 units; Fig 7A).

To evaluate spatial information at the level of OB and HPC populations, we trained a decoder model on simultaneous estimates of location extracted from video tracking and population activity, and tested the model’s performance on held-out data at predicting the mouse’s position based on the population activity (Fig 7B; Stefanini et al., 2020). We quantified model performance as the mean error between the decoded position and the actual position (see Methods). For 18/31 sessions from OB and 12/13 sessions from HPC, the model decoded the mouse’s position better than chance (Fig 7C). These analyses demonstrate that in task-free, ambient stimuli conditions, neuronal activity in HPC tracks an animal’s location in an environment. Importantly, we show here, for the first time, that OB neurons also track an animal’s location, consistent with the idea that the olfactory system plays an integral role in navigation (Baker et al., 2018; Dittman & Quinn, 1996; Gagliardo, 2013).

We have shown that OB neurons track breathing rhythms and place. Importantly, breathing rhythms and the states extracted from our behavioral model are not uniformly distributed in allocentric space (Fig 8, figure supplement 1). It is possible that the apparent place information we observe in OB could be explained by differential use of breathing rhythms in different regions in the arena. To test this hypothesis, we used GLMs to ask if inclusion of place significantly improves prediction of single unit spiking activity over a model based on latency from inhalation and sniff frequency (Figs 3 and 4). We found that adding the place covariate to a model with sniff field covariates significantly increased the log-likelihood of held-out data in 160/1153 units (p<0.01, sign rank test), and 81/1153 were better predicted by place than by sniff field (Fig 8A). The unique predictive contribution of place is inconsistent with the hypothesis that place selectivity is an epiphenomenon of the sniff field.

We next considered the possibility that apparent place selectivity could result from responses to ambient odorants. Although we did not apply odor stimuli in these experiments, ambient odors from the mouse and the environment are unavoidable and unevenly distributed in space. The most obvious candidate odor source would be the mouse’s own scent marks: mice, along with many organisms, scent mark their environment, and these marks can contribute to navigational behavior (Drickamer, 2001; Hurst et al., 2001; Khan et al., 2012; Means et al., 1992; Wallace et al., 2002). To test whether scent marks influence spatial selectivity, we performed floor rotation control experiments, in which we rotated a transparent floor mat 180 degrees halfway through a recording session. If place selectivity reflected the location of scent marks, then the “place fields” should rotate to reflect the new distribution of scent marks. Inconsistent with this hypothesis, we observed that place fields often maintained the same location selectivity before and after a floor rotation (Fig 9A). Across the population, the place selectivity of most units maintained a higher correlation across the floor rotation rather than correlating with the new position of the scent marks (Fig 9B,C). Additionally, we asked whether the place decoding models generalize across floor rotation conditions (see Methods). If responses to scent marks drove correlations with place, decoders trained on pre- and tested on post-floor rotation should perform significantly worse than those trained and tested on post-floor rotation data. However, we find that these decoders perform equally well (Fig 9D). Taken together, these findings suggest that place selectivity in the OB does not reflect the distribution of scent marks. However, it is important to recognize that these results do not exclude the possibility that other distal sources of ambient odor explain the place selectivity we observe. In either case, we show that OB contains decodable information about place. Inevitably, this activity will combine with odor-driven activity from the nose and be broadcast to the OB’s numerous postsynaptic targets, most of which send centrifugal feedback to the OB, and several of which are reciprocally connected with the hippocampus (Aqrabawi & Kim, 2018; Padmanabhan et al., 2019; Price, 1985; Reinert & Fukunaga, 2022; Shipley et al., 2008; Vanderwolf, 2001).