Odor identity decoding by mitral/tufted cells in the olfactory bulb from large-scale pooled datasets

Summary

The mechanisms by which the olfactory bulb encodes odor information are fundamental to our understanding of sensory processing in the brain. Here, we analyze large-scale pooled electrophysiological recordings and respiratory data from awake mice to explore how mitral and tufted cells (M/Ts) represent odor identity. Our results demonstrate that, while odor-evoked changes in firing rate are relatively sparse in the awake state, the temporal firing patterns of M/Ts, particularly those aligned with the respiratory cycle, carry significant information for accurate odor decoding. Importantly, the reliability of odor identity decoding improves as more neurons are sampled, and integrating information across multiple respiratory cycles further enhances decoding performance. These findings highlight the essential role of temporal coding and population dynamics in olfactory processing, offering new insights into the strategies used by the olfactory system to represent complex sensory stimuli.

Graphical abstract

Highlights

• •Respiration-locked firing patterns of M/T cells encode odor identity
• •Odor-decoding accuracy improves with more recorded M/T cells
• •Decoding performance peaks ∼80 ms after inhalation onset
• •Cumulative spike count scheme best captures odor information

Bioinformatics; Neuroscience

Introduction

One of the fundamental tasks of the brain is to accurately perceive stimuli from both the internal and external environment. This requires sensory centers at various levels to precisely encode specific stimuli and transmit this information to higher-order sensory or integrative centers, ultimately leading to perception. Accordingly, each sensory center contains specialized neurons responsible for encoding particular features of sensory stimuli. For example, in the primary visual cortex, dedicated neurons represent object features such as color and shape,¹ while in the auditory midbrain (inferior colliculus), distinct neurons encode sound frequency and temporal structure.² In the case of olfaction, the mitral and tufted cells (M/Ts)—the principal output neurons of the olfactory bulb (OB), the first central station in the olfactory system—represent odor identity and concentration through specialized neural circuits and coding strategies.^3,4,5

In recent years, studies using in vivo electrophysiological recordings from the OB of awake, behaving animals have revealed that M/Ts do not effectively encode odor identity through changes in firing rate alone.^6,7 Instead, specific temporal firing patterns within individual inhalation cycles allow for accurate decoding of odor type.^8,9,10 These findings suggest that the temporal sequence of M/Ts firing, in coordination with the animal’s respiratory cycle, carries critical information about odor identity. However, many details about how M/Ts employ this strategy to represent odor information remain unclear.

Research on M/Ts odor coding typically relies on standard single-unit in vivo electrophysiological recording techniques. Due to technical constraints, only a limited number of neurons—ranging from a few to a few dozen—can be recorded simultaneously from a single animal.^9,11 It remains an open question to what extent such a small sample of neurons can faithfully represent odor information. Within the OB, is it possible for a small subset of neurons to accurately encode odor identity, or is a larger population of neurons required to achieve reliable information through population coding? Determining the minimum number of neurons needed for precise odor decoding is an important topic for further investigation. Moreover, since odor identity decoding relies not only on the timing of neuronal firing but also on the phase of the animal’s respiratory cycle,^3,5,10,12 it is also unclear how the number of respiratory cycles affects the accuracy of odor decoding by M/Ts. Addressing these questions requires large-scale datasets and comparative analyses.

This study builds on in vivo electrophysiological and respiratory data collected in our laboratory over the past several years across multiple research projects.^{13,14,15,16,17} Compared to previous decoding studies that typically relied on fewer than 20 mice and approximately 100–200 M/Ts^6,15,18, the scale of our dataset—436 well-isolated M/Ts from 45 mice—enabled a deeper investigation into the above question. Our findings provide new evidence and insights into how the olfactory system utilizes temporal firing pattern strategies to represent odor identity.

Results

Single-unit responses of mitral/tufted cells to odorants in awake mice

Single-unit activity of M/Ts in the OB has been shown to carry important information on odor identity. To investigate how the activity of M/Ts represents odor identity, we simultaneously recorded extracellular spikes from M/Ts and monitored sniffing signals in awake, head-fixed mice during odor delivery (Figure 1A). Spikes were recorded using microelectrodes or tetrodes inserted into the mitral cell layer of the OB, and single units were isolated from the raw data (Figure 1B). In total, we collected 45 mice with complete respiratory recordings across seven independent projects, yielding a dataset of 436 well-isolated single units. Five of these projects have been previously published (Table 1), and a total of 21 mice were excluded due to incomplete respiratory recordings.

Figure 1.

Odor-induced firing activity of M/Ts in the OB during in vivo experiments

(A) A schematic diagram of the experimental setup. Electrodes were implanted into the mitral cell layer of the OB to capture firing patterns of M/Ts in awake mice, while a sniffing cannula was inserted into the nasal cavity to monitor sniffing signals.

(B) An example of spike detection and sorting on extracellular electrode recordings, resulting in the successful separation of two distinct single units (unit a, red; unit b, blue).

(C1–C3) Raster plots (up) and mean firing rates (down) of the odor-evoked excitatory (C1), inhibitory (C2) and no (C3) responses of M/Ts. Each panel depicts the responses of a single unit to multiple trials of an odor. The odor is present in 0–2 s.

(D1) Percentage of unit-odor pairs that showed excitatory (red), inhibitory (blue), or no response (gray). The data encompasses 1,744 unit-odor pairs (436 M/Ts, 4 odors). Stacked bar plots illustrate the odor-evoked response percentages across different odors (D2) and projects (D3). (D4) The percentage of units responding to a specific number of odors.

(E1 and E2) Mean firing rate of odor-evoked responses across different odors (E1, Kruskal-Wallis test, ${\chi }_{0.05,3}^2$ > H [0.17], p = 0.98, n = 436 units from 45 mice) and projects (E2, Kruskal-Wallis test, ${\chi }_{0.05,6}^2$ < H [324.70], p < 0.0001, the number of units in each project was listed in Table 1). ∗∗∗∗p < 0.0001, n.s. not significant.

Table 1.

Detailed information on the datasets for odor decoding analysis, including the number of mice and M/Ts recorded in each project

Data name	Number of mice	Number of units	Source
Published data 1	6	51	Chen et al., 202113
Published data 2	4	36	Wu et al.,202214
Published data 3	9	108	Wu et al.,202315
Published data 4	12	78	Chen et al., 202416
Published data 5	5	40	Geng et al., 202517
Unpublished data 1	5	50	–
Unpublished data 2	4	73	–

5 projects were derived from published data, and 2 were obtained from unpublished sources.

The mice were exposed to a fixed panel of four odorants: isoamyl acetate, 2-heptanone, benzaldehyde, and heptanoic acid. Compared to the blank control period (0–2 s before odor valve opening), some M/Ts exhibited clear increases or decreases in spike counts during odor stimulation (0–2 s after valve opening) (Figure 1C). Across the population, about one-third of the unit-odor pairs showed significant changes in firing rate, while approximately two-thirds showed no significant change (Figure 1D1). This distribution was consistent across all four odors (Figure 1D2) and all seven projects (Figure 1D3). Although nearly 70% of the units exhibited a significant change in firing rate to at least one odorant, less than 10% responded significantly to all four odorants (Figure 1D4), indicating that odors evoke sparse responses among M/Ts in the OB of awake mice.

Further analysis revealed no significant difference in mean firing rate (MFR) among different odorants (Figure 1E1). However, the MFR of M/Ts varied significantly across projects (Figure 1E2), which may be attributable to differences in recording electrodes. Taken together, these results demonstrate that in awake, head-fixed mice, the mean firing rate of M/Ts carries only limited information about odor identity. Other firing properties, such as spike patterns or temporal dynamics, may contribute to the neural representation of odor identity.

Inhalation-coupled mitral/tufted cell population responses convey substantial odor information

To further analyze how the temporal patterns of M/Ts spiking encode odor identity, we focused on the odor-evoked temporal information. Previous studies in rats have shown that the firing activity of M/Ts, synchronized with respiration, carries essential information for odor discrimination.¹⁸ Therefore, we specifically examined spikes occurring within the first three respiratory cycles. For each inhalation, we divided the spike counts within the subsequent 160 ms into eight non-overlapping 20 ms bins (Figure 2A, corresponding to 6.25 Hz).

Figure 2.

The odor decoding performance of an individual mouse is positively correlated with the number of recorded M/Ts when classified using inhalation-coupled spike counts

(A) Diagram of feature extraction for odor decoding. The upper and lower traces represent the raw spike (black) and respiration (green) signals, respectively. Classification features of the experimental group were the firing responses locked to the first three inhalation onset after odor stimulation.

(B1–D2) The performance of classification across three different coding schemes, as illustrated in the schematic on the left. (B1) “Instantaneous” bins, features were obtained by extracting the spike counts within a given 20 ms bin. (C1) “Concatenated” bins, features are composed of incrementally concatenated non-overlapping bins of 20 ms duration from 0 to t. (D1) “Cumulative spike count,” features consist of spike counts within a time window from 0 to t. The classification performance of three different coding schemes exhibits a positive correlation with the number of recorded M/Ts in each mouse (B2, linear regression, r = 0.0058, p = 0.0015, n = 45 mice; C2, linear regression, r = 0.016, p < 0.0001, n = 45 mice; D2, linear regression, r = 0.014, p < 0.0001, n = 45 mice). The black dashed lines represent 25% chance level.

Recognizing that the brain likely decodes sensory information by integrating population activity in real time, we assessed the ability of M/Ts population to discriminate odor identity on a single-trial basis. Instantaneous population features for classification were defined as the spike counts from all recorded M/Ts within each 20 ms bin on individual trials (Figure 2B1, instantaneous). To evaluate how well these features could predict odor identity, we applied a support vector machine (SVM) with a linear kernel, using leave-one-out cross-validation. For each iteration, one trial was held out for testing while the remaining trials were used for training, and this process was repeated until all trials had been tested.

Because neuronal activity at different time points may contribute to odor decoding, we also considered features derived from multiple adjacent, non-overlapping time bins (Figure 2C1, concatenated; Figure 2D1, cumulative spike count). Across all mice, classifier performance varied substantially: in some animals, decoding accuracy exceeded 60%, while in others it approached chance level (25%). Notably, the number of recorded units per mouse also varied, ranging from 3 to 25.

We hypothesized that the variability in decoding performance across mice was due to differences in the number of recorded units. Consistent with this idea, we found a significant positive correlation between classification accuracy and the number of recorded M/Ts (Figures 2B2, C2, D2). These results indicate that inhalation-coupled population responses of M/Ts encode information sufficient for odor discrimination, and that the discriminability depends on the size of the recorded M/Ts population in each mouse.

Odor decoding performance of olfactory bulb correlates with the number of recorded mitral/tufted cells

Because the number of M/Ts recorded in a single mouse was limited (typically fewer than 30), we further examined how decoding performance scales with the size of the recorded M/T population. To this end, we pooled all M/Ts recorded within each project, yielding larger ensembles ranging from 36 to 108 units per project. Using the three coding schemes described above, we found that odor decoding performance remained highly variable across projects (Figures 3A1–A3).

Figure 3.

The odor-decoding performance of projects containing different numbers of M/Ts

(A1–A3) The performance of classification across three different coding schemes (as shown in Figure 2). Each curve represents the result of a project. The time windows for the “concatenated” and “cumulative” schemes keep increasing over time, with a bin size of 0.02 s for each step.

(B1–C3) The optimal decoding accuracy of each project is related to the number of recorded M/Ts (B1, linear regression, r = 0.0027, p = 0.18, n = 7 projects; B2, linear regression, r = 0.0063, p = 0.053, n = 7 projects; B3, linear regression, r = 0.0064, p = 0.026, n = 7 projects; C1, unpaired t test, t₍₅₎ = 3.17, p = 0.025, n = 4 projects for small group, n = 3 projects for larger group; C2, unpaired t test, t₍₅₎ = 4.10, p = 0.0093, n = 4 projects for small group, n = 3 projects for larger group; C3, unpaired t test, t₍₅₎ = 4.41, p = 0.0070, n = 4 projects for small group, n = 3 projects for larger group). Data are presented as mean ± SEM. ∗p < 0.05, ∗∗p < 0.01. The black dashed lines represent 25% chance level.

To assess whether this variability was related to the number of recorded units, we analyzed the correlation between classification accuracy and unit count (Figures 3B1–B3). We observed a clear trend: projects with more recorded M/Ts tended to exhibit higher decoding accuracy. To further test this relationship, we divided the seven projects into two groups according to ensemble size: a “large” group (n = 3 projects, >60 units) and a “small” group (n = 4 projects, <60 units). Decoding accuracy was significantly higher in the large-unit group compared to the small-unit group (Figures 3C1–C3).

These findings demonstrate that although inhalation-coupled M/Ts spiking is crucial for encoding odor identity, the accuracy of odor decoding is strongly dependent on the size of the recorded M/Ts population. This highlights the importance of sampling larger ensembles to more accurately assess the coding capacity of the OB.

Decoding performance approaches saturation when the number of recorded mitral/tufted cells exceeds 150

To investigate how many M/Ts recordings are required to achieve optimal decoding performance, we pooled data from 436 M/Ts recorded across 45 mice for further analysis. We examined classification accuracy using three distinct coding schemes. Decoding performance based on instantaneous activity patterns progressively increased with each inhalation phase, but subsequently declined to baseline (chance level, 25%) during the remainder of the respiratory cycle (Figure 4A1). In contrast, decoding accuracy based on concatenated bins and cumulative spike count showed a steady increase from baseline, reflecting the temporal accumulation of odor information over time (Figures 4A2 and A3).

Figure 4.

The odor decoding performance approaches to saturation when the number of recorded M/Ts surpasses 150

The M/Ts recorded in all mice were aggregated together (n = 436 units from 45 mice).

(A1–A3) The performance of classification across three different coding schemes (as shown in Figure 2).

(B1–B3) The accuracy of the three coding schemes were compared in the first (B1), second (B2) and third sniffing (B3) after the odor stimulus (B1, one-way ANOVA, F_{(1.562, 10.93)} = 2.99, p = 0.10; Tukey’s multiple comparisons test; instantaneous (SVM) versus concatenated (SVM), p = 0.34; instantaneous (SVM) versus cumulative (SVM), p = 0.72; concatenated (SVM) versus cumulative (SVM), p = 0.10; instantaneous (SVM) versus instantaneous (MD), p = 0.46; concatenated (SVM) versus instantaneous (MD), p = 0.19; cumulative (SVM) verse instantaneous (MD), p = 0.80; n = 8 bins.

(B2) One-way ANOVA, F_{(1.371, 9.600)} = 34.50, p < 0.0001; Tukey’s multiple comparisons test; instantaneous (SVM) versus concatenated (SVM), p = 0.99; instantaneous (SVM) versus cumulative (SVM), p = 0.0010; concatenated (SVM) versus cumulative (SVM), p < 0.0001; instantaneous (SVM) versus instantaneous (MD), p = 0.99; concatenated (SVM) versus instantaneous (MD), p = 0.99; cumulative (SVM) verse instantaneous (MD), p = 0.00070; n = 8 bins.

(B3) One-way ANOVA, F_{(1.117, 7.818)} = 91.59, p < 0.0001; Tukey’s multiple comparisons test; instantaneous (SVM) versus concatenated (SVM), p = 0.074; instantaneous (SVM) versus cumulative (SVM), p < 0.0001; concatenated (SVM) versus cumulative (SVM), p < 0.0001; instantaneous (SVM) versus instantaneous (MD), p = 0.80; concatenated (SVM) versus instantaneous (MD), p = 0.11; cumulative (SVM) verse instantaneous (MD), p < 0.0001; n = 8 bins).

(C1–C3) The classification performance of three different coding schemes increases with the number of recorded M/Ts (the number of M/Ts ranges from 10 to 430, increasing in steps of 10), shown by red (C1), black (C2), and blue (C3) lines respectively. The specific decoding accuracy values for different M/Ts numbers are: (C1) 0.28, 0.28, 0.32, 0.37, 0.33, 0.36, 0.37, 0.39, 0.37, 0.43, 0.46, 0.44, 0.45, 0.44, 0.45, 0.47, 0.44, 0.48, 0.47, 0.48, 0.48, 0.49, 0.48, 0.49, 0.49, 0.52, 0.47, 0.49, 0.52, 0.52, 0.53, 0.53, 0.51, 0.52, 0.53, 0.53, 0.53, 0.53, 0.53, 0.53, 0.53, 0.53, 0.53; (C2) 0.32, 0.31, 0.35, 0.38, 0.36, 0.41, 0.39, 0.40, 0.39, 0.38, 0.38, 0.41, 0.44, 0.42, 0.41, 0.41, 0.44, 0.43, 0.43, 0.43, 0.43, 0.43, 0.43, 0.44, 0.45, 0.44, 0.43, 0.42, 0.44, 0.44, 0.43, 0.43, 0.45, 0.43, 0.44, 0.43, 0.44, 0.43, 0.44, 0.45, 0.44, 0.44, 0.44; (C3) 0.38, 0.43, 0.53, 0.50, 0.56, 0.57, 0.58, 0.59, 0.59, 0.61, 0.66, 0.64, 0.67, 0.66, 0.66, 0.67, 0.69, 0.70, 0.68, 0.72, 0.71, 0.71, 0.72, 0.73, 0.72, 0.72, 0.73, 0.73, 0.72, 0.74, 0.72, 0.73, 0.73, 0.75, 0.75, 0.75, 0.75, 0.75, 0.75, 0.75, 0.74, 0.75, 0.76. The binomial fitting result for each accuracy curve was indicated by dashed line in corresponding colors. The green circle on each fitting line signifies the point where classification performance reaches 90% to maximum accuracy (C1, n = 200 M/Ts; C2, n = 120 M/Ts; C3, n = 170 M/Ts). Data are presented as mean ± SEM. ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. The black dashed line illustrates 25% chance level.

Comparing the decoding performance of the three schemes across different respiratory phases, we found no significant difference during the first respiratory cycle (Figure 4B1). However, in the second and third respiratory cycles, the cumulative spike count scheme significantly outperformed the other two (Figures 4B2 and B3).

Building on our previous observation that decoding accuracy is strongly dependent on the number of recorded M/Ts, we further quantified this relationship (Figures 4C1–C3). As the number of M/Ts increased, decoding accuracy improved steadily from baseline. Notably, decoding based on cumulative spike count consistently outperformed the other schemes, reaching a maximum accuracy of approximately 75%. Fitting the data with a quadratic polynomial allowed us to estimate the number of units required to reach 90% of the maximum accuracy (Figure 4C1, n = 200; Figure 4C2, n = 120; Figure 4C3, n = 170). These results demonstrate that increasing the number of recorded M/Ts greatly benefits OB odor decoding, and that decoding performance approaches an optimal level when the number of recorded units reaches around 150. Whether different odors require different numbers of neurons for accurate decoding needs further investigation. To investigate this, we conducted an additional analysis involving odor-specific binary decoding tasks for all four odors (Figure S1; odor A vs. non-A, odor B vs. non-B, and so on). We found that at the OB level, there was no systematic difference in the number of M/Ts required for decoding different odors. The mechanisms behind this phenomenon remain an important topic for future research.

Furthermore, to investigate how the population of M/Ts represents odor information across different respiratory phases, we analyzed the classification weights of the 436 M/Ts. Considering that in the “instantaneous” and “cumulative spike count” coding schemes, each unit corresponds to one classification feature, whereas in the “concatenated” scheme, each unit corresponds to multiple features. We focused our analysis on the weights of the first two schemes for ease of interpretation (Figure S2). The results indicate that across different respiratory phases and various odor-pair (odor A vs. odor B, odor A vs. odor C …) classification tasks, the same subset of M/Ts consistently exhibited high weights. This demonstrates the existence of a key neuronal ensemble that plays a crucial role in odor representation, which also aligns with the sparse coding characteristic shown in Figure 1.

Odor decoding performance peaks approximately 80 ms after inhalation

Our previous results demonstrated that odor decoding performance in the OB is closely linked to the respiratory phase (Figure 4A1), reflecting the fact that inhalation facilitates the entry of odorant molecules into the nasal cavity and triggers olfactory processing. As a central olfactory structure, the OB is capable of analyzing the complex dynamics of inhaled chemical stimuli and encoding this information in real time, resulting in time-varying representations of odor perception.

To determine the time point at which odor perception reaches its peak after inhalation, we conducted an analysis that aligned neural responses to specific respiratory phases across multiple respiratory cycles. Specifically, for each 20 ms bin within a respiratory cycle, spikes occurring in the corresponding bin from all three respiratory cycles were aggregated and used as classification features (Figure 5A). For example, when assessing the decoding performance of the first bin (0–20 ms), we combined spikes from the first bin of all three respiratory cycles. This procedure was repeated for subsequent bins in a sequential manner (Figure 5A, purple).

Figure 5.

The odor-decoding performance reaches its maximum approximately 80 ms after inhalation

(A) A schematic diagram illustrating the classification performed based on the spike count of multiple inhalations. Purple, 1-bin spike counts per sniffing as features; lavender, 2-bin spike counts per sniffing as features; red, 3-bin spike counts per sniffing as features. The classification performance, utilizing all M/Ts (B1, n = 436 units), and a reduced population (B2, n = 120), exhibited similarity when employing the decoding schemes of “concatenated” bins (B3). (B4) The decoding accuracy reaches its peak approximately 80 ms after inhalation. One-way ANOVA, F_{(3, 32)} = 35.91, p < 0.0001; Tukey’s multiple comparisons test; 0–40 versus 40–80, p < 0.0001; 0–40 versus 80–120, p < 0.0001; 0–40 versus 120–160, p < 0.0001; 40–80 versus 80–120, p < 0.0001; 40–80 versus 120–160, p = 0.011; 80–120 versus 120–160, p = 0.54. The classification performance, utilizing all M/Ts (C1, n = 436), and a reduced population (C2, n = 170), exhibited similarity when employing the decoding schemes of “cumulative spike count” (C3). (C4) The decoding accuracy reaches its peak approximately 80 ms after inhalation. One-way ANOVA, F_{(3, 32)} = 29.42, p < 0.0001; Tukey’s multiple comparisons test; 0–40 versus 40–80, p < 0.0001; 0–40 versus 80–120, p < 0.0001; 0–40 versus 120–160, p < 0.0001; 40–80 versus 80–120, p = 0.0005; 40–80 versus 120–160, p = 0.18; 80–120 versus 120–160, p = 0.47. Data are presented as mean ± SEM. ∗p < 0.05, ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001. The black dashed lines represent 25% chance level.

The resulting real-time decoding performance curves showed a characteristic pattern: an initial rise followed by a decline, closely paralleling the pressure changes within the respiratory cycle (Figures 5B1, C1, purple). This trend persisted even when prior phase bins were cumulatively incorporated into the classification features (Figures 5B1 and C1). Repeating the analysis with reduced M/Ts population yielded consistent results (Figures 5B2, B3, C2 and C3).

By comparing odor decoding performance across different respiratory phases, we found that in awake, head-fixed mice, decoding accuracy—and thus the neural representation of odor perception—peaks at approximately 80 ms after inhalation onset (Figures 5B4 and C4).

Decoding performance of minimum distance classification is consistent with SVM results

To further demonstrate the dynamic odor decoding in OB M/Ts during single respiratory cycle, we performed principal component analysis on the population neuronal response matrix at individual time point, and then connected the reduced-dimensionality data points across time to construct the population trajectory representing the response to a given odor (Figure 6A). Prior to the onset of sniffing, the four odors were not distinguishable. As inhalation began, the trajectories representing each odor progressively diverged, reaching maximal separation during inhalation, and then nearly converged by the end of exhalation.

Figure 6.

Decoding performance of “instantaneous” firing rate using a minimum distance (MD) classifier

(A) The average response trajectories of M/Ts to each odor during a single sniff. The responses are projected onto the first three principal components. Each color represents an odor.

(B1 and B2) Decoding performance of “instantaneous” firing rate using an MD classifier exhibits a positive correlation with the number of recorded M/Ts (B2, linear regression, r = 0.0070, p = 0.0017, n = 45 mice). (B3) Distribution of maximum decoding accuracy across respiratory phases (n = 45 mice).

(C1 and C2) The maximum decoding accuracy of each project is related to the number of recorded M/Ts (C1, linear regression, r = 0.0028, p = 0.21, n = 7 projects; C2, unpaired t test, t₍₅₎ = 3.17, p = 0.025, n = 4 projects for small group, n = 3 projects for larger group).

(D) “Instantaneous”-based classification performance using a large population of M/Ts (n = 436 units).

(E1 and E2) Comparison of decoding performance of different decoder architectures and coding schemes over three respiration cycles (E1, Friedman test, p = 0.099, n = 45 mice; E2, one-way ANOVA, F_{(1.275, 7.650)} = 2.96, p = 0.12, n = 7 projects). Data are presented as mean ± SEM. ∗p < 0.05, n.s. not significant. The black dashed lines represent 25% chance level.

Based on this observation, we employed the Minimum Distance Classifier (MD) as an alternative approach for instantaneous odor decoding. This method classifies samples based on the geometric distance between population response vectors. The results obtained with MD classification were highly consistent with those from the SVM approach. At the level of individual mice, decoding accuracy showed a strong positive correlation with the number of recorded M/Ts (Figures 6B1 and B2), and approximately two-thirds of the mice exhibited peak decoding accuracy around 80 ms after inhalation onset (Figure 6B3, right). At the project level, groups with larger numbers of recorded units achieved significantly higher decoding accuracy compared to those with smaller populations (Figures 6C1 and C2). For the largest pooled population (436 units), the real-time decoding accuracy during a single sniffing cycle also exhibited a characteristic rise-and-fall pattern (Figure 6D).

Finally, we directly compared classification performance across different decoding algorithms (MD vs. SVM) and coding schemes. Across both the 45 mice and the seven project datasets, no statistically significant differences in overall accuracy were detected between decoders, likely due to substantial within-group variability (Figures 6E1 and E2). However, the “cumulative spike count” coding scheme consistently outperformed other schemes throughout the sniffing cycle (Figures 4B1–B3).

Discussion

How mitral and tufted (M/T) cells, the output neurons of the OB, encode odor information is a central question in OB function and a key issue in understanding how the olfactory system perceives smells.^3,4,10 Earlier studies have established respiration-coupled temporal coding in M/Ts spiking as an acknowledged principle for representing odor information in the olfactory system.^10,18 While these researches revealed the “what”—that temporal patterns within individual respiratory cycles carry odor information, key quantitative questions regarding “how many cells are required” and “how well it performs” remained underexplored, primarily due to typically limited sample sizes in single experiments.

This study integrates in vivo electrophysiological recordings and respiratory data from several independent projects conducted in our lab, aiming to quantitatively refine and extend this established knowledge. The findings reveal that, in awake mice, although the firing rates of M/Ts respond only weakly and sparsely to odor stimulation, their firing patterns within individual respiratory cycle can effectively decode odor information. Specifically, our analyses (1) determined the number of M/Ts required to achieve near-optimal odor decoding—when fewer than approximately 150 cells are recorded, increasing the number significantly enhances decoding performance; (2) pinpointed the peak of odor identity decoding at approximately 80 ms after inhalation onset; and (3) identified the “cumulative spike count” scheme as a superior strategy for temporal integration across respiratory cycles, providing new insights into how downstream brain regions might read OB signals.

In sensory systems, neural firing often conveys information about external stimuli. Neurons typically represent stimuli through various coding schemes, including firing rate, latency relative to stimulus onset, and temporal firing patterns.^10,18,19,20 As the first central relay of olfactory processing, the OB is thought to encode odor information with high fidelity.^4,5 M/Ts are the primary output neurons of the bulb. They receive direct input from peripheral olfactory sensory neurons and also transmit information to downstream olfactory cortex areas after being processed through the bulb’s internal circuitry.^21,22,23,24 Therefore, M/Ts activity is expected to reflect key odor attributes such as identity and concentration.

In anesthetized animals, M/Ts show marked increases in firing rates in response to odors,^7,25 making rate coding a viable strategy under those conditions. However, in awake animals, where spontaneous firing rates are higher, odor stimuli often elicit weak or even suppressed responses.^7,26,27 Consequently, firing rate alone may not effectively represent odor information. In our recordings from head-fixed, awake mice, only about one-third of M/Ts showed clear odor-evoked responses, consistent with previous observations.^18,26,27,28

Given the limitations of firing rate in awake conditions, could temporal firing patterns still support odor decoding? Since odor sampling depends on the animal’s respiration—odor molecules are drawn into the nasal cavity during inhalation—odor-evoked firing patterns within a single breathing cycle could provide informative temporal codes.^10,12,29 This hypothesis has been validated in freely moving rats, where the temporal distribution of firing within a respiratory cycle carries more reliable odor information than overall firing rate.¹⁸ Our study in mice further confirms this idea. Notably, we employed multiple decoding algorithms and provided detailed insights into decoding strategies based on respiration-locked firing patterns, such as the required number of neurons and temporal resolution for accurate decoding.

At the olfactory epithelium, olfactory sensory neurons expressing specific receptors encode thousands of odor molecules via population coding.³⁰ This information is relayed to the OB and integrated by M/Ts. Although this process involves convergence and transformation, a single M/T cell is generally insufficient to represent specific odors on its own—population coding across multiple M/Ts is essential.^31,32 Due to the constraints of conventional in vivo electrophysiological recording setups, only a few to a few dozen M/Ts can typically be recorded simultaneously in a single animal in most studies, including the individual projects pooled here.^11,26,33 Therefore, pooling data from multiple mice to construct a “virtual large neuronal population” is a commonly used and well-established method in the field for investigating population coding.^17,18 However, it has remained unclear how many M/Ts are needed to decode odor identity effectively.

By pooling data from multiple independent experiments, our study demonstrates that decoding performance improves with the number of M/Ts recorded. This provides a plausible explanation for the variability in decoding accuracy reported across different animals and studies—it likely depends on the number of M/Ts recorded. The consistency of this finding across 45 mice and seven different projects, which involved variations in experimenters, recording electrodes and minor experimental setups, greatly enhances the generalizability of our conclusions. The fact that the relationship between neuron number and decoding performance holds true both within individual animals and across pooled datasets strongly suggests that our findings reflect fundamental properties of the olfactory system, rather than artifacts arising from specific experimental configurations or particular animal subgroups. We also found that decoding performance plateaus when approximately 150 or more M/Ts are included, suggesting a practical benchmark for future studies of odor representation. Encouragingly, with ongoing advancements in electrophysiological recording and electrode fabrication, it is becoming increasingly feasible to record from more than 100 neurons simultaneously in a single animal.^34,35 Future studies that record 150 or more M/Ts in a single mouse will be able to explore odor-coding strategies at the individual level, eliminating potential confounds from inter-animal variability.

During the recording process, we recorded and analyzed a random population of M/Ts—we did not preselect specific functional subtypes but instead isolated as many units with good signal-to-noise ratio as possible from the mitral cell layer. This random sampling strategy was employed precisely to reflect the statistical properties of the OB’s overall output layer, rather than the properties of specific functional modules. Selecting only neurons of a particular response type would, conversely, introduce bias and overestimate their contribution under natural conditions.

In both our work and that of others, firing patterns have typically been analyzed using the respiratory cycle as a temporal framework.^10,18 However, the OB also exhibits other rhythmic activities—such as oscillations observed in local field potentials—that can serve as alternative time windows.^3,36,37 While theta oscillations are tightly coupled with respiration, beta and gamma oscillations may also act as temporal frameworks for encoding.^36,38,39 Previous studies suggest that firing events phase-locked to gamma oscillations might carry odor-specific information,¹¹ but more research is needed to clarify these mechanisms. In addition to temporal coding based on the firing patterns of M/Ts population, studies have also shown that different odorants elicit distinct yet overlapping odor maps.^40,41 However, in our study, electrodes were uniformly implanted into the mitral cell layer of the ventral OB, and there may be variations in the spatial overlap between electrode locations and the M/Ts population activated by different odors. This implies that electrode placement may have some influence on decoding performance. Future studies employing larger-scale recordings of M/Ts across broader OB regions would substantially address this issue.

Odor representation in the olfactory system can vary significantly depending on brain state—e.g., awake vs. anesthetized, asleep vs. awake, active vs. passive behavior, and with or without learning or past experience.^25,42,43 This study focuses on passive odor exposure in awake, head-fixed mice. How M/Ts represent odor information under active sniffing or task-related conditions remains largely unknown. Given that previous work has shown that active sampling state dynamically enhances odor representation in the OB,⁴⁴ and mice employing active sniffing strategies exhibit improved discrimination performance⁴⁵ and faster reaction times⁴⁶ in odor tasks, it is possible that different behavioral contexts may engage entirely distinct coding strategies. Future research is needed to examine whether the respiration-locked pattern coding observed here is also used under other behavioral or cognitive states.

This study focuses solely on the OB’s response to four single odor stimuli. However, investigating how the olfactory system encodes complex odor mixtures is a crucial aspect of natural olfactory scenes. Natural odors are complex mixtures of single-molecule components, each capable of binding to different olfactory receptors. Results from optical imaging and 2-photon microscopy⁴⁷ indicate that natural odors at their native concentrations can elicit dense representations in the OB, though this is influenced by the animal’s anesthetic state and odor concentration. This would likely lead to the activation of a larger and more diverse ensemble of M/Ts. To accurately resolve the identity of a complex mixture from this potentially noisier and more high-dimensional population activity, a larger sample of neurons might be necessary to achieve the same level of decoding accuracy as for single odors. Therefore, the number of ∼150 M/Ts we identified might represent a lower bound for robust decoding, with more complex stimuli potentially requiring larger populations. Furthermore, the additional challenge posed by mixtures lies not only in identifying their components but also in resolving their relative proportions. The precise temporal patterns of spikes within the M/Ts population may help distinguish mixtures that activate similar neuronal ensembles. Thus, we anticipate that temporal patterns would remain critical, and potentially even more critical, for the coding of complex mixtures.

In conclusion, our study combines large-scale pooled electrophysiological data with respiratory signals to systematically analyze how firing patterns locked to the respiratory cycle encode odor information in awake, head-fixed mice. These findings not only offer new insights into how the OB, via M/Ts activity, represents odor identity but also serve as an important reference for understanding how other sensory systems may use temporal pattern coding to decode environmental stimuli.

Limitations of the study

While this study demonstrates the temporal firing patterns and population dynamics underlying olfactory coding, it does not provide direct functional validation of the proposed coding mechanisms. Future experiments can comprehensively explore the behavioral significance of these sniff-related temporal features through optogenetic manipulation to confirm the causal role of these mechanisms in odor representation. Such experiments can be conducted in conjunction with electrophysiological recordings from downstream brain regions, such as the piriform cortex and lateral entorhinal cortex.

Furthermore, this study primarily focuses on quantifying the presence and temporal dynamics of odor identity information decodable from M/Ts population, without elucidating the precise format or pattern of the neural code (i.e., how specific neural activity patterns correspond to particular odors). We will address this question through in-depth investigation in future work.

Third, the scope of odorant stimuli in this study was limited to four representative monomolecular odors at a single concentration. Since natural odors are often complex mixtures, and odor perception can vary significantly with concentration, the odors used in this study do not encompass the full complexity of natural olfactory scenes. Therefore, our findings regarding the number of M/Ts required for decoding and the optimal temporal window may not fully generalize to the discrimination of complex mixtures or to different concentration gradients. Future studies employing a broader panel of odors, including natural blends and a range of concentrations, will be essential to elucidate how the OB population code adapts to represent more ethologically relevant stimuli.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Anan Li (anan.li@xzhmu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The dataset used in this study has been deposited at Mendeley data (https://data.mendeley.com/datasets/zdvxwzxdpz/draft?x=1) and are publicly available as of the date of publication.
• •The original codes used in the study are provided in GitHub: https://github.com/lishan911nju/In-vivo-electrophysiological-data-analysis.
• •Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (NSFC, 32471058 and 32271055 to A.L.).

Author contributions

S.L., P.W., and A.L. designed research; S.L. and P.W. performed research; S.L. and P.W. analyzed data; S.L. and A.L. wrote the manuscript.

Declaration of interests

The authors claim that there are no conflicts of interest.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Isoamyl Acetate	Sinopharm GroupChemical Reagent Co.	CAS No.:123-92-2
2-Heptanone	Sinopharm GroupChemical Reagent Co.	CAS No.:110-43-0
Benzaldehyde	Sinopharm GroupChemical Reagent Co.	CAS No.:100-52-7
Heptanoic Acid	Sinopharm GroupChemical Reagent Co.	CAS No.:111-14-8
Deposited data
All datas	Li et al., this article	Mendeley Data: https://data.mendeley.com/datasets/zdvxwzxdpz/draft?x=1
Experimental models: Organisms/strains
C57BL/6J mice	GemPharmatech	Strain Number: N000013
Software and algorithms
MATLAB R2020b	MathWorks	https://www.mathworks.com/
Custom Matlab analysis code	This article	GitHub: https://github.com/lishan911nju/In-vivo-electrophysiological-data-analysis
GraphPad Prism 8	GraphPad Software	https://www.graphpad.com/
Python 3.7.4	Python Software Foundation, PSF	https://www.python.org/

Experimental model and study participant details

Animals

This study utilized C57BL/6J male mice aged 8–16 weeks, which were group-housed under standard laboratory conditions (12 h light/dark cycle, constant temperature of 24±2°C, humidity 40%–70%, with food and water ad libitum). Following surgery, mice were individually housed and allowed to recover for at least one week prior to the experiments. All experimental protocols were conducted in accordance with the procedures submitted to and approved by the Xuzhou Medical University Institutional Animal Care and Use Committee (Approval No.: SYXK (SU) 2020-0048).

Method details

Surgery

Mice were briefly anesthetized with pentobarbital (90 mg/kg body weight, i.p.), then secured in a stereotaxic apparatus (RWD, Shenzhen, China). The scalp from the dorsal nasal region to the midpoint between the ears was removed to fully expose the skull surface. A hole was drilled above the right OB (AP, 4.0 mm; ML, 1.0 mm) for implantation of a microelectrode (4×4-channel, Kedou Brain-computer Technology, Suzhou, China) or tetrodes. Each tetrode consisted of four polyimide-coated nichrome wires (RO-800, Sandvik AB, Sandviken, Sweden) connected to a 16-channel electrode interface board (EIB-16, Neuralynx Inc.). For the remainder of our study, both the microelectrodes and the tetrodes were collectively referred to as “electrodes”. After connecting the reference electrodes to the grounding screw (1mm posterior to the bregma and 1 mm lateral to the midline), the electrodes were implanted into the mitral cell layer of the ventral OB with an average insertion depth of 1.8–2.5 mm.¹¹ Neuronal activity was monitored during electrode implantation, and the electrode was secured to the skull with dental acrylic once optimal placement was achieved. Finally, a custom aluminum headplate was affixed to the skull using dental acrylic to stabilize the head during recording.

To monitor respiratory activity, a hole was drilled in the dorsal recess of the mouse naris for cannula insertion. Following the completion of electrode implantation, a respiratory cannula was implanted into the nasal cavity, secured with adhesive, and stabilized with dental acrylic.

In vivo electrophysiological recordings and data preprocessing

In vivo electrophysiological recordings were performed after a minimum 1-week postoperative recovery period. The mouse was head-fixed and positioned on an air-suspended, freely rotating polystyrene ball. The recording electrodes were connected to the headstage of the electrophysiological acquisition system, allowing neural signals from the OB to be amplified by a 16-channel amplifier (Plexon DigiAmp, Plexon, Dallas, TX, USA; bandpass 1-5000 Hz, 2000× gain) and digitized at 40 kHz using the Plexon Omniplex recording system. Among various OB cell types, only M/Ts within the mitral cell layer generated spikes detectable by the electrodes. Spike sorting and unit identification were performed offline using Offline Sorter v4 software (Plexon, Dallas, TX, USA). Spike detection was performed using an amplitude threshold of −5 SD relative to the background noise. Subsequent separation of distinct units was carried out via principal component analysis. For automatically sorted units, those exhibiting unimodal frequency distribution curves and having less than 0.75% of total inter-spike intervals (ISIs) shorter than 1 ms were classified as single units. Finally, the identified single units were thoroughly evaluated, taking into account both the waveform shape and the number of spikes. M/Ts activity was quantified by either mean firing rate (MFR) or firing rate variability (ΔMFR = MFR-MFR_baseline; baseline, 0∼2 s before odor stimulation). Paired t-test or Wilcoxon sign-rank test was employed to compare baseline and odor-evoked firing rates for determining response significance. Based on the significance threshold of p < 0.05, odor-evoked responses were further subdivided into excitatory, inhibitory, or no response.

Sniffing measurement

Respiratory signals and in vivo electrophysiological data were recorded simultaneously in mice. The respiratory recording system primarily consisted of a pressure sensor (Model No. 24PCEFA6-G(EA), 0 psi–0.5 psi, Honeywell) and an amplifier (Plexon DigiAmp; 100× gain). The respiratory cannula transmitted the mechanical pressure generated by airflow to the sensor via a polyethylene tube, where it was converted into a voltage signal. The voltage signal was then amplified and sampled at 1 kHz through the Plexon Omniplex recording system.

Passive odor stimulation

The mice were stimulated with four neutral odorants: isoamyl acetate, 2-heptanone, benzaldehyde, and heptanoic acid (Sinopharm GroupChemical Reagent Co., Shanghai, China), which have been demonstrated to evoke stable neuronal responses in M/Ts.^22,43 All odorants were dissolved in mineral oil at 1% v/v dilution and delivered through tubing of the olfactometer system to a position 1 cm anterior to the mouse’s nose. During odor delivery, nitrogen gas flowed through the odorants at 100 mL/min, followed by a 1:20 dilution in the olfactometer. Additionally, charcoal-filtered air was continuously delivered at a constant flow rate of 1 L/min to minimize the impact of airflow variations during odor stimulation. Odor concentration stability was verified using a photoionization detector (PID; Aurora Scientific, Canada). As previously measured, the latency time of the odor delivery device was about 46.67 ms, and the rise time was about 128.73 ms.⁴⁸ Each odor was presented for 2 s in a pseudo-randomized order, followed by an inter-stimulus interval before the next odor presentation.

Odor decoding analysis

Odor decoding analysis was performed using population activity of M/Ts. Specifically, each odor trial served as a classification sample, with odor-evoked firing patterns across all recorded M/Ts constituting the sample features. Three coding schemes were implemented: for “instantaneous”-based coding scheme, the feature dimension = N neurons × 1(duration of time bin = 20 ms); for “concatenated”-based coding scheme, the feature dimension = N neurons × number of time bins (duration = 20 ms); for “cumulative spike count”-based coding scheme, the feature dimension = N neurons × 1 (time window is 0∼t ms). Decoding was performed in Python 3.7.4 using two distinct classifier architectures: support vector machine (SVM, linear kernel, 'C': 1.0, 'cache_size': 200, 'class_weight': None, 'coef0': 0.0, 'decision_function_shape': 'ovr', 'degree': 3, 'gamma': 'auto_deprecated', 'kernel': 'linear', 'max_iter': -1, 'probability': False, 'random_state': None, 'shrinking': True, 'tol': 0.001, 'verbose': False) and Minimum distance classifier (MD). Leave-one-out cross-validation (LOO-CV) was used to assess the model’s generalization ability, and the classification accuracy was defined as the proportion of correctly identified trials.

Quantification and statistical analysis

Statistical analysis was performed through MATLAB or GraphPad Prism software. The Anderson-Darling test was used to assessed whether the data conformed to the normal distribution. For paired samples, paired t-tests were used for normally distributed data, while Wilcoxon signed-rank tests were applied to non-normally distributed data. For three or more independent groups, one-way ANOVA with Tukey’s multiple comparisons test was used for normally distributed data, and the Kruskal–Wallis test was employed for non-normally distributed data. For three or more paired groups with non-normal distributions, the Friedman test was utilized. Data are expressed as mean ± SEM, with significance denoted as p < 0.05 (∗), p < 0.01 (∗∗), p < 0.001 (∗∗∗), p < 0.0001 (∗∗∗∗), and n.s. indicating non-significance.

Published: January 14, 2026