Stateful behavioral structure in an unstructured experimental paradigm.

A. Experimental setup. Mice were head-fixed or freely moving in a 40 by 15 cm arena while we recorded respiration and neuronal activity and captured video from below in visible light. B. The correlation structure of breathing and movement. Top, Histogram of instantaneous sniff frequencies of all mice (n = 4). Thick lines and shaded regions are mean and ±1 standard deviation, thin lines are individual mice. Blue: freely moving; black: head-fixed. Right , Histogram of instantaneous movement speeds, where the movement speed time series was sampled at each inhalation time. Center, 2D histogram of breathing frequency and movement speed. C. Long-timescale behavioral structure. Autocorrelations of sniff frequency (Left), movement speed (Right), and the cross correlation between sniff frequency and speed. Blue: freely moving; black: head-fixed. D. A three-state Hidden Markov Model (HMM) fit to the sniff frequency and movement speed time series captures the clustered correlation structure of breathing rhythm and movement. Colormaps show the instantaneous frequency and speed distributions of sniffs in each of three states: Orange: “rest”, blue: “grooming”, red: “exploration”. Right Overlay of the distributions from the three states. Overlap is indicated by color mixing and darkness (for colorbars, see Figure 1, supplemental video 2) E. The behavioral HMM captures the long-timescale states of breathing rhythms. Each dot indicates an inhalation time with its instantaneous frequency on the vertical axis. Black: head-fixed; other colors as in 1D.

A behavioral model captures the stateful structure of neuronal population activity in Olfactory Bulb.

A. Scatter plot of mean firing rates during the head-fixed and freely moving epochs of the recording sessions. Each dot indicates the firing rates of an individual unit (n=1680 units in all sessions; n=1274 in sessions with recordable sniff signals). B. Behavior, neuronal population activity, and similarity matrix from an individual session. Top, Each dot indicates an inhalation time with its instantaneous frequency on the vertical axis. Black: head-fixed; other colors as in 1D. Middle, Whole-session spike rates (5 s bins) of the neuronal population recorded in this session. Each row corresponds to an individual unit (n=58 total), with the color scale indicating the normalized firing rates. Each row is normalized separately between minimum and maximum. Bottom, Cosine distance matrix quantifies the similarity between the population activity pattern across time bins. C. Grand mean cosine distance matrix between states across mice (n=4). Each session’s cosine distance matrix is expressed in units of the number of standard deviations from a null distribution formed by circularly shifting the HMM state time series (see Methods). Positive values indicate greater similarity than expected from the null hypothesis of a “nonsense correlation”; negative indicates less similarity. D. Silhouette scores quantifying how well the behavioral states cluster the neuronal population activity patterns in all sessions (4 mice; 25 sessions). Scores are in units of the number of standard deviations from a null distribution formed by circularly shifting the HMM state time series as in 2C.

“Sniff fields” (SnFs) display how neurons track breathing rhythms.

A. Spike rasters from 4 units simultaneously recorded in the same session. Dots indicate spike times relative to inhalation. Each row shows two seconds of the recording centered at each inhalation time at time 0. Rows are sorted in descending order of sniff frequency B. Sniff field (SnF) plots from the same four units. Bottom left, Colormap indicates firing rates with respect to latency in the sniff cycle and instantaneous sniff frequency. Right, Sniff frequency profile of the SnF calculated by taking the max projection across the horizontal axis of the distribution. Top, Latency profile of the SnF calculated by taking the max projection across the vertical axis of the distribution.

Neuronal sniff field latency and frequency profiles fall into a small number of clusters across the population.

A. Right, SnF latency profiles of all units that were significantly predictable with a GLM fit to spike latency relative to inhalation (n=853/913 p<0.01, Sign rank test). Freely moving and head-fixed matrices are sorted the same. Left, units are segregated into two clusters by k-means clustering of the earth mover’s distance matrix quantifying the similarity of SnFs across units. Bottom, Within-cluster means for the two clusters. Green: putative tufted cells; Brown: putative mitral cells. Lines and shaded regions are within-cluster means ±1 standard deviation. B. Right, SnF frequency profiles of all units that were significantly with a GLM trained on instantaneous sniff frequency (n=638/913; p<0.01, Sign rank test). Freely moving and head-fixed matrices are both sorted according to selectivity index in the freely moving data, and differently than the matrices in 4A. Left, units are segregated into three types by k-means clustering of the earth mover’s distance matrix quantifying the similarity of SnFs across units. Bottom, Within-cluster means for the three clusters. Teal: low frequency units; Purple: medium frequency units; Blue: high frequency units. Lines and shaded regions are within-cluster means ±1 standard deviation.

Contribution of behavioral parameters to a predictive model of individual unit firing.

A. Both frequency and latency improve a model of individual unit firing. Top, Each dot indicates a unit with activity that was significantly predictable from a combined GLM based on both latency and frequency (p<0.01, Sign rank test). The contribution is defined as how much including a given parameter improves the model predictions on held-out data (see Methods). Lavender: units for which sniff frequency significantly improved the model prediction; Blue: units for which latency improved the model prediction; Lavender/blue: both parameters improve the model prediction. Marginal distributions of contributions from the two parameters are shown beside and above the scatter plot. Bottom, Relative contribution compares the improvement due to the two parameters. B. Movement speed minimally improves the predictions of a model incorporating sniff frequency and latency.

Place fields display allocentric location selectivity of olfactory bulb and hippocampal neurons.

Colormaps show occupancy-normalized firing rates as a function of location in the 15 x 40cm experimental arena parsed into a 12 by 5 grid and Gaussian smoothed by one bin width (see Methods). For consistency, OB and HPC colormaps are scaled between the 1st and 99th percentiles of each neuron’s firing rate in 10 s bins. Those values are displayed beside each unit’s colorbar. A. Four example OB units from each of four mice. Significance of Spatial Information (SSI) and p-values are defined relative to the circular shift null distributions. B. Four example HPC units from each of four mice. OB and HC recordings were performed in different animals.

Spatial selectivity of individual neurons and decoding of population activity from olfactory bulb and hippocampus.

A. Cumulative distributions of selectivity of OB and HPC units, quantified as the significance of spatial information (SSI), defined relative to circular shift null distributions (see Methods). B. Decoder model schematic. A classifier for each pair of spatial bins is trained on neuronal activity (firing rate in 5 s bins) and tested on held out data. The decoded spatial position of the mouse at a given time step is taken as the center of the bin that wins the most “votes”, defined as the bin that was predicted by the most pairwise classifiers. C. Decoder model performance of OB and HPC populations on real and shuffled controls. Decoding error is defined as the median distance between the decoded spatial position and the mouse’s actual position. Points are individual sessions, filled are p < 0.01 (sign-rank test).

Sniff fields do not explain place selectivity.

A. Top, Each dot indicates a unit that was significantly predictable from a combined GLM based on both sniff fields and place fields (p<0.01, Sign rank test). The contribution is defined as how much including a given parameter improves the model predictions on held-out data (see Methods). Lavender: units for which sniff parameters significantly improved the model prediction; Yellow: units for which place improved the model prediction; Lavender/yellow: both parameters improve the model prediction. Marginal distributions of contributions from the two parameters are shown beside and above the scatter plot. Bottom, Relative contribution compares the improvement of the two parameters.

Scent marks do not solely explain place selectivity.

In a subset of experiments, we rotated the floor 180 degrees midway through the experiment and compared the resulting place fields. A. Four example units each from OB of two mice. Colormaps are scaled between the 1st and 99th percentiles as in Fig 6. “Pre rotation” shows the place fields calculated from spiking and position time series before the rotation, “post-rotation” shows the place fields from the same units calculated from data after the floor rotation. B. Scatter plot of correlation between pre and post rotation place fields vs pre and virtually rotated post rotation (postR). C. Signed difference in correlations in B reveals that a majority of units’ place fields do not follow the scent marks while others do. D. Population decoding models trained on activity during the post rotation data and tested on post rotation data (Post/post), trained on pre rotation and tested on post rotation (Pre/post), and trained on pre rotation and tested on a 180 degree rotated control of post rotation data (Pre/postR).

A. Histogram of instantaneous sniff frequencies for each individual mouse and grand mean (n = 4). Thick lines and shaded regions are mean and ±1 standard deviation. Blue: freely moving; black: head-fixed. B. 2D histogram of breathing frequency and movement speed for each individual mouse and grand mean (n=4). C. Sniff rasters for three example sessions where each dot indicates an inhalation time with its instantaneous frequency on the vertical axis. Black: head-fixed; Colors based on movement speed during the freely-moving condition.

A. Bayesian Information Criterion (BIC) scores for HMMs with increasing numbers of hidden states. Scores drop substantially until three states. B. Cumulative histogram of inferred state durations across all sessions show that states typically last tens of seconds to minutes. C. Sniff Frequency histograms across inferred states for three individual sessions from three mice. D. Joint movement speed and sniff frequency scatter plots colored by state assignments for the same three sessions. E. Probability density estimates of joint movement speed and sniff frequency for the same three sessions.

Unit inclusion criteria

A. All spike waveforms for an example unit, their corresponding z-scored spike amplitudes, and mean waveform shape. To calculate amplitude on a spike by spike basis, we took the difference between the signal at the peak time and the trough time. B. Same as in A. for an example unit with positive leading waveform. C. Scatter of all cluster’s amplitude cutoff violations and refractory period violations. Green dashed lines show criteria for inclusion (amplitude cutoff violations < 10%, refractory period violations < 5%). D. As in C. for all clusters in sessions with simultaneous sniff recording.

Overlap between sniff latency and frequency clusters.

A. (Left) SnF latency profiles of all units sorted by latency (Right) SnF frequency profiles separately clustered within latency clusters show a diversity of frequency profiles exist within each functional latency cluster. Earthmovers distance matrices show clustered structure of profiles. B. As above, but sorted first by SnF frequency then latency clustered within frequency clusters.

Spatial distribution of behavioral state usage.

The spatial distribution of behavioral state usage for each OB mouse. Colormap overlays the usage of each state, normalized by sniffs, into a composite color as in Figure 1 (Figure 1 – Supplemental video 2).