Plume dynamics structure the spatiotemporal activity of glomerular networks in the mouse olfactory bulb

Although mice locate resources using turbulent airborne odor plumes, the stochasticity and intermittency of fluctuating plumes create challenges for interpreting odor cues in natural environments. Population activity within the olfactory bulb (OB), is thought to process this complex spatial and temporal information, but how plume dynamics impact odor representation in this early stage of the mouse olfactory system is not known. Limitations in odor detection technology have made it impossible to measure plume fluctuations while simultaneously recording from the mouse’s brain. Thus, previous studies have measured OB activity following controlled odor pulses of varying profiles or frequencies, but this approach only captures a subset of features found within olfactory plumes. Adequately sampling this feature space is difficult given a lack of knowledge regarding which features the brain extracts during exposure to natural olfactory scenes. Here we measured OB responses to naturally fluctuating odor plumes using a miniature, adapted odor sensor combined with wide-field GCaMP6f signaling from the dendrites of mitral and tufted (MT) cells imaged in olfactory glomeruli of head-fixed mice. We precisely tracked plume dynamics and imaged glomerular responses to this fluctuating input, while varying flow conditions across a range of ethologically-relevant values. We found that a consistent portion of MT activity in glomeruli follows odor concentration dynamics, and the strongest responding glomeruli are the best at following fluctuations within odor plumes. Further, the reliability and average response magnitude of glomerular populations of MT cells are affected by the flow condition in which the animal samples the plume, with the fidelity of plume following by MT cells increasing in conditions of higher flow velocity where odor dynamics result in intermittent whiffs of stronger concentration. Thus, the flow environment in which an animal encounters an odor has a large-scale impact on the temporal representation of an odor plume in the OB. Additionally, across flow conditions odor dynamics are a major driver of activity in many glomerular networks. Taken together, these data demonstrate that plume dynamics structure olfactory representations in the first stage of odor processing in the mouse olfactory system.


Introduction
Mice are adept at localizing odor sources ( (Gire et al., 2016); (Baker et al., 2018); (Liu 2 et al., 2019); (Gumaste et al., 2020)), but the spatiotemporal information in olfactory 3 environments that aids this search behavior is unknown. Odors travel in plumes which 4 pull odor away from its source in filaments that are broken and distorted as they travel in 5 air, creating complex odor environments. From the perspective of an olfactory searcher, 6 these intermittent filaments create stochastic odor encounters, or whiffs, such that odor 7 concentration dynamics fluctuate rapidly from moment to moment. Features of these 8 complex plume dynamics contain information regarding odor source location ( (Murlis 9 et al., 2000); (Celani et al., 2014)). For example, as a searcher encounters odors, the 10 frequency, strength, and timing of encounters provide complex cues about an odor source 11 ( (Vergassola et al., 2007); ; (Michaelis et al., 2020); (Atema, 1996). 12 A simple strategy such as averaging odor concentration across whiffs could eliminate 13 the complexity of an odor plume, allowing an animal to simply follow an increasing 14 odor concentration gradient to the odor source. However, an animal dependent on this 15 search strategy would operate at a timescale far slower than that observed in mice 16 engaged in olfactory-guided search (Gumaste et al., 2020). This suggests that rodents 17 most likely extract information from the complex spatiotemporal dynamics of olfactory 18 environments to support their efficient odor-guided search behavior. 19 The extraction of information from fluctuating odor plumes will necessarily be 20 impacted by the physics of odor transport. Factors such as wind speed and the Reynolds 21 number of the plume could impact early olfactory processing in mammals. Precise 22 olfactometers have been used to model certain features found in natural odor environments 23 such as fluctuating and intermittent odor concentration dynamics. Although this work 24 provides important insights, olfactometers do not capture the full complexity of the odor 25 environment. One problem is a lack of knowledge regarding which features of the plume 26 are relevant to olfactory search, constraining which features olfactometers have been 27 used to mimic. In addition, olfactometers create artificial plumes that decouple odor 28 concentration from features present in olfactory environments. This omits correlations 29 between concentration fluctuations and the surrounding air flow as well as small scale 30 details of odor transport like diffusion. Decoupling these factors creates challenges for 31 interpretation because it implicitly disrupts processing moderated by these features, 32 such as the impact of wind speed on the vibrissal system ( (Yu et al., 2016)) or feedback 33 regarding bilateral nasal sampling ( (Esquivelzeta Rabell et al., 2017); (Markopoulos 34 et al., 2012)). Directly observing how MT activity is impacted by the plume dynamics of 35 natural olfactory scenes will thus constrain hypotheses regarding which spatiotemporal 36 features of natural odor stimuli are conveyed by the brain. 37 We studied the response of MT cells in the OB to odor concentration dynamics 38 in awake mice as they processed natural olfactory scenes, i.e. odor plumes. We used 39 wide-field calcium imaging to measure MT activity, allowing us to study OB output 40 at the level of glomerular complexes on the dorsal surface of the OB. Simultaneous 41 recordings of the OB and plume dynamics show glomerular population activity follows 42 fluctuations of odor concentration during plume encounters. The reliability and following 43 2/31 behavior of glomerular responses were moderated by wind speed and the resulting 44 changes in plume structure. The fidelity of odor concentration tracking increased when 45 concentration dynamics were skewed, creating intermittent odor encounters across the 46 plume presentation. In addition, as the strength and reliability of odor-evoked activity 47 in MT cells increased, this activity more accurately followed plume dynamics. Together, 48 these data demonstrate for the first time that the rapid fluctuations present in natural 49 olfactory scenes significantly structure the activity of glomerular MT cell populations in 50 the mouse OB.

52
Olfactory Stimuli 53 Olfactory stimuli were released by an automated odor port within a 32" X 16" X 16" 54 acrylic wind tunnel where airspeed was controlled by a vacuum at the rear of the 55 wind tunnel, posterior to the animal's location. Concentration dynamics of olfactory 56 stimuli varied stochastically from trial to trial creating plumes with unique concentration 57 dynamics on each trial. Adjusting the velocity of wind flow allowed for variation in the 58 Reynolds number, resulting in characteristic changes of plume dynamics for different flow 59 levels (Supplemental Figure 1). Reynolds numbers were calculated using the mean flow 60 of each condition and the half height of the tunnel (8") as the head-fix setup is placed 61 on a stage that is elevated approximately 20cm from tunnel floor. Low, medium, and 62 high flow Reynolds numbers were 2400 ±1000, 8800 ±400, and 9800 ±200 respectively 63 (mean ± st dev). Absolute velocities of low, medium and high flow conditions were .40 64 ± 0.16, 1.31 ± 0.05, and 1.81 ± 0.17 and velocity fluctuations were 14 ± 3 fpm, 6 ± 2 65 fpm, and 5.3 ±1.5 fpm respectively.

66
A single session consisted of forty trials of odor presentation. Odor ports were located 67 upwind ∼ 13 cm anterior to the animal's nose, and each plume presentation had a 68 duration of 10 seconds. Odor release began approximately 10 seconds into the trial. 69 In order to avoid responses predicting the beginning of the plume, the exact time of 70 odor release was jittered by adjusting the duration of the 5 second intertrial interval 71 by a length of time drawn randomly from a uniform distribution, U (−2, 2). Random 72 clicking noise was used to control for the clicking sound of the port serving as a cue 73 for plume onset. Starting 5 seconds prior to plume onset, a number was drawn from a 74 uniform distribution, U (0, 1), for each camera frame, and if the number exceeded 0.95 a 75 clicking sound was produced. The odor used was a mixture of benzaldehyde and ethanol. 76 Ethanol concentration throughout each trial was measured by a modified, commercially 77 available ethanol sensor placed within 3.5-4mm from the mouse's right nostril.

79
Implantation of the imaging window was adapted from methodology detailed in Batista-80 Brito et al., (Batista-Brito et al., 2017). Mice were anesthetized with isoflurane for 81 surgery. 2 x 2.5mm or 2 x 3mm craniotomies were performed above the olfactory bulbs 82 and custom cut double windows were implanted. A customized stainless steel head plate 83 was glued directly on the skull posterior to the window, and two Invivo1 stainless steel 84 screws (McMaster-Carr) were placed posterior to the head plate. Metabond was then 85 added to cover all exposed skull and a thin layer built to cover the screws and the central 86 3/31 surface of the headplate. The position of the craniotomy was biased towards either the 87 left or right bulb.

88
In vivo imaging 89 Widefield fluorescent microscopy was used for awake, head-fixed imaging in Thy1-90 GCaMP6f-GP 5.11 (IMSR Cat JAX:024339, RRID:IMSR JAX:024339) mice to view 91 neural activity in the dorsal OB. Mice were between 11 weeks-13 months old when 92 imaged. 488 nm LED stimulation was used for the duration of the trial (30s), but was 93 absent during intertrial intervals (∼ 5 seconds) to avoid excessive bleaching. All mice 94 were imaged at 30Hz with 4x 0.13 NA objective (Nikon). Neural activity was recorded 95 using a Teledyne Photometrics Prime 95B sCMOS camera. For each session, mice were 96 head-fixed above a freely rotating, circular track, allowing mice to run at will during 97 imaging sessions.

98
In vitro OB slice imaging 99 To establish patterns of expression and signals obtained from the OB of Thy1-GCaMP6f-100 GP 5.11 animals, imaging experiments were conducted in OB slices. Horizontal OB slices 101 (300-400 m) were made following isoflurane anesthesia and decapitation. Olfactory bulbs 102 were rapidly removed and placed in oxygenated (95% O2, 5% CO2) ice-cold solution 103 containing the following (in mM): 83 NaCl, 2.5 KCl, 3.3 MgSO4, 1 NaH2PO4, 26.2 104 NaHCO3, 22 glucose, 72 sucrose, and 0.5 CaCl2. Olfactory bulbs were separated into 105 hemispheres with a razor blade and attached to a stage using adhesive glue applied to the 106 ventral surface of the tissue. Slices were cut using a vibrating microslicer (Leica VT1000S) 107 and were incubated in a holding chamber for 30 min at 32°C. Subsequently, the slices 108 were stored at room temperature. Slices were placed on a Scientifica SliceScope Pro 6000, 109 using near infrared imaging for slice placement and 488 nm LED illumination for imaging 110 activity and a QI825 Scientific CCD Camera (Q Imaging) for image acquisition. Imaging 111 was performed at 32-35°C. The base extracellular solution contained the following: 125 112 mM NaCl, 25 mM NaHCO3, 1.25 mM NaHPO4, 25 mM glucose, 3 mM KCl, 1 mM 113 MgCl2, and 2 mM CaCl2 (pH 7.3 and adjusted to 295 mOsm), and was oxygenated 114 (95% O2, 5% CO2). An elevated KCl solution (equimolar replacement of 50 mM NaCl 115 with KCl in the extracellular solution) locally applied through a borosilicate pipette 116 using a picospritzer 2 (Parker Instrumentation) was used to stimulate cells.

118
ImageJ was used to crop fields of view (FOVs) for data analysis. It was also used to 119 extract pixel averaged signal for hand-drawn region of interest (ROI) anaylsis.

120
Matlab 2019b was used to analyze data and plot figures. Data was aligned using 121 NoRMCorre software to perform piecewise rigid and non-rigid motion correction ( Pnev-122 matikakis and Giovannucci (2017)). This alignment corrected for both global frame 123 movement due to head jitter (rigid) and localized distortion due to brain movement 124 (non-rigid).
125 Figure 1. Plume presentations and head-fix setup for in-vivo recording experiments.a) All experiments conducted in 16" x 16" x 32" wind tunnel for quick clearing of odor presentations. The odor port (not pictured) was located ∼ 13 cms upwind of the animal's nose. b) (right) Graphic detailing experimental setup. (left) Ethanol odor concentration measured using a modified, commercially available ethanol sensor placed ∼ 4mm from the edge of the mouse's right nostril. c) Diagram depicting flow conditions (high, medium, or low) of the 40 trials within a single session. d) Example odor traces are depicted for each flow condition. e) Histograms of the odor concentration magnitude sampled across two examples trials show a change in skewness between low flow (right, blue) and high flow (left, red), with skewness increasing with increased airflow during plume presentations. f) Comparisons between the deconvolved sensor signal and a PID signal during a set of paired recordings show odor concentration dynamics of the deconvolution can recover dynamics observed in the PID recordings (r = 0.61, p < .001). An example from low flow (top) and high flow (bottom) are shown. g-h) For each trail, the skewness (F) and asymmetry (G) of the plume's distribution are calculated showing that high flow trials (orange) are separable from low (blue) for both PID and sensor recordings.
Precise tracking of plume dynamics 126 A miniaturized ethanol sensor modified from the commercially available metal oxide 127 (MOX) sensors (Tariq et al., 2019) was placed within 4mm of the lateral edge of the 128 mouse's right nostril to capture the odor concentration signal across plume presentations 129 for each trial. A single odor, a benzeldahyde-ethanol mixture, was used for each trial. 130 Odors released together travel together within plumes at sufficiently small scales because 131 dispersion dominates over diffusion ( (Celani et al., 2014); (Yeung and Pope, 1993)). 132 Therefore, the ethanol sensor measured the odor concentration of the benzeldahyde-133 ethanol mixture. The plume for each trial was released by an automated odor port at the 134 upwind end of the wind tunnel. The odor was released 10 seconds after the trial began 135 for a duration of 10 seconds. Each indicated flow condition was maintained throughout 136 the entire trial. Flow condition was set for each trial block 137 (figure 1C) by adjusting the strength of a vacuum exhaust at the downwind side of 138 the wind tunnel to one of three levels, low, medium, and high. A medium level flow 139 was presented for the first 10 trials, after which flow alternated by high and low flow 140 5/31 in blocks of 5 trials. Some sessions transitioned from medium to low flow initially and 141 others transitioned from medium to high flow.

143
Recent characterization of MOX sesnsors comparing their deconvolved signals to si-144 multaneously recorded photoionization detector (PID) signals have validated the use 145 of MOX sensors in capturing turbulent plume dynamics despite their slower recording 146 dynamics ( (Martinez et al., 2019); (Tariq et al., 2019)). For our experiments, ethanol 147 concentration throughout each trial was measured by a modified, commercially available 148 ethanol sensor placed within 3.5-4mm from the mouse's right nostril. A single session 149 consisted of 40 trials of odor presentation. Odor ports were located upwind ∼ 13 cm 150 anterior to the animal's nose, and each presentation had a duration of 10 seconds and 151 consisted of a mixture of benzaldehyde and ethanol.

152
Sensor signal was acquired at 100Hz and then low pass filtered at 30 Hz using a 153 Kaiser window. The signal, e, was then normalized within each trial using the mean 154 and standard deviation of the signal during the plume presentation. The signal was 155 deconvolved by adapting the deconvolution specified in Tariq et al., 2019. The kernel 156 was defined in the same manner, but instead of normalizing the range of the kernel, the 157 integral of the kernel is normalized. Thus, the kernel, k, is calculated as follows.
where t is an array with evenly spaced timestamps at the proper sampling rate for 160 the length of a single trial, τ decay = 0.0001, and τ rise = 0.4629. Both signals, e and k, 161 are then transformed into Fourier space using the Matlab Fourier transform function, 162 and the ethanol signal is deconvolved by dividingê byk. The inverse Fourier transform 163 of the resulting deconvolution is taken to obtain d. The deconvolved signal, d, is then 164 normalized within each trial using the mean and standard during the plume presentation. 165 The deconvolution optimizes the preservation of odor concentration dynamics across 166 trials, but does not preserve the absolute value of odor concentration.

167
To optimize the parameters for the deconvolution, a complete session of 40 plume 168 presentations was recorded with the usual ordering of flow condition blocks (figure 1C). 169 No mice were recorded during this session. Instead, a photoionization detector (PID) 170 was placed 4mm from the ethanol sensor in the same position where the animal is usually 171 head-fixed.

172
To optimize parameters, the PID signal, p was first downsampled to 100Hz to match 173 the sensor sampling rate. Next, the signal was normalized within each trial using the 174 mean and standard deviation of the signal during plume presentation. This normalized 175 signal was then compared to d deconvolved across a range of τ decay and τ rise parameter 176 values. The Kernel parameters were chosen by minimizing mean squared error between 177 the e and p signals averaged across all trials within the paired recording session. The 178 deconvolved ethanol sensor signal allows for the recovery of plume dynamics unique to 179 each trial (Supplemental Figure 2). It is significantly correlated with the PID signal as 180 measured during plume presentations (r = 0.61 , p < 0.001, both sampled at 100Hz ), 181 which is a 0.22 improvement from the correlation between the raw ethanol sensor and 182 6/31 PID signal (r = 0.39 , p < 0.001, both sampled at 100Hz ).

183
Finally, with the exception of Supplemental Figure 1, the deconvolved trace was 184 downsampled for figures and analyses to match the calcium trace (30Hz) by averaging 185 all samples taken across each camera frame.

186
Defining dynamic flow 187 Dynamic flow was calculated within each session. In the experimental setup, flow 188 condition is defined by wind speed. Intermittency was measured within a trial during 189 the middle 8 seconds of the 10 second plume presentation by calculating asymmetry or 190 by using the 3 rd moment of the sampled distribution. Low and High flow trials were 191 separable using either of these measures.

192
To determine if there was a main effect of flow condition on these stimulus properties, 193 skewness and asymmetry, a three-way analysis of variance (ANOVAs) test was conducted 194 for each parameter. For each ANOVA, a multiple comparisons using Tukey's pairwise 195 comparisons was performed to look for significant differences of the parameter between 196 flow where comparisons were considered significant for p < .01. These tests indicated that 197 both skewness and asymmetry varies significantly across flow condition (F (157) = 52.7, 198 p < .001 and F (157) = 54.9, p < .001 respectively). Multiple comparison tests examining 199 the differential effect of flow on these parameters showed that, for both, there was a 200 significant difference between low and medium flow and a significant difference between 201 low and high flow. No significant difference was found between medium and high flow. 202 Therefore, only low and high flow conditions were selected when examining the effect of 203 air flow on neural parameters.

204
Measuring glomerular responses to plume dynamics 205 Widefield imaging of the dorsal surface of the OB in Thy1-GCaMP6f-GP5.11 mice was 206 used to capture MT cell activity at the glomerular level (figure 2). Thy1 mice exhibit fast 207 kinetics and strong expression in MT cells within the OB (Dana et al., 2014). Global 208 MT activity is clustered into dendritic complexes known as glomeruli. Widespread 209 activity of secondary dendrites in the EPL of dorsal OB (figure 2A) causes diffuse 210 fluorescence across the imaging field. Therefore, CaImAn, a constrained nonnegative 211 matrix factorization(CNMF) algorithm, was used on each FOV to find regions of 212 interest(ROIs) and their activity traces ( (Pnevmatikakis et al., 2016)). Thus the spatial 213 decomposition of CNMF provided the ROI for each glomerulus, and the temporal 214 decomposition provided its corresponding denoised activity trace. The denoising of the 215 CNMF temporal decomposition helps to remove correlated signal between neighboring 216 glomeruli, accounts for calcium drift within recording sessions, and separates glomeruli 217 overlapping in the dorsal-ventral dimension. The change in the distribution of correlation 218 coefficients between ROIs before and after de-noising (Supplemental Figure 3) shows a 219 decorrelation of glomerular signals. A two-sample Kolmogorov-Smirnoff test shows the 220 distributions of correlation coefficients between the pixel averaged ROI traces, 0.87 ± 221 0.07 (mean ± st. dev), and the CNMF ROI traces, 0.36 ± 0.28 (mean ± st. dev), are 222 significantly different (D = 0.86, p < 0.001).

223
To protect against oversegmenting a single glomerulus into multilple ROIs, neighboring 224 ROIs whose baseline CNMF activtiy was correlated above 0.65 were selected as candidates 225 for ROI merging. The baseline period was examined as criteria for possible merging since 226 Figure 2. In-vivo recording of glomerular population response. a) In vitro image of MT cell activity in response to K+ puff. b) In vivo view of the dorsal olfactory bulb through an implanted cranial window. (left) Window activity averaged across a single trial. (right) Projected standard deviation for the same trial shows MT activity in the dorsal OB responsive to the odor presentation. c) Diagram depicting flow conditions (high, medium, or low) of the 40 trials within a single session. d) The deconvolved ethanol trace (blue) compared to the deconvolved response of each glomeruli (black) within the recorded FOV during a single low flow trial depicted by asterisk in (c). Red arrows indicate onset and offset of plume presentation. e) Same but for a single high flow trial from the same session also depicted by asterisk in (c).
glomeruli might have similar response profiles to the stimulus dynamics during plume 227 presentations. The baseline activity of the neighboring ROIs was then binarized using a 228 threshold of ±1 st dev. If the correlation of the binarized activity between neighboring 229 ROIs exceeded 0.65, the ROI with the lower mean activity (presumably encompassing 230 less of the glomerulus) was dropped from the analysis.

231
To validate the use of CNMF at the glomerular level, results were compared to a 232 hand drawn ROI analysis conducted on one of the fields of view (FOV 5) in ImageJ 233 (Supplemental Figure 5). Hand-drawn ROIs were selected after viewing footage and 234 reviewing standard deviation and maximum value projections of activity from the FOV 235 within each trial. The activity averaged from within hand drawn ROIs have higher 236 pairwise correlations than denoised CNMF activity traces. This mirrors what is seen 237 when pixel-averaged activity from within CNMF ROIs (without denoising) is compared 238 to the denoised traces. Thus, the denoising of CNMF increases the spatiotemporal 239 resolution of glomerular activtiy observed in the dorsal OB recordings in both types of 240 analyses. A two-sample Kolmogorov-Smirnoff test shows the distributions of correlation 241 coefficients between the deconvoled hand-drawn ROIs, 0.81 ± 0.08 (mean ± st. dev), 242 and the deconvoled CNMF traces, 0.1935 ± 0.3471 (mean ± st. dev), are significantly 243 different (D = 0.86, p < 0.001). Results of the hand-drawn ROI analysis (correlation and 244 power analyses) were qualitatively similar to those found using CNMF corroborating the 245 ability of glomerular networks to resolve odor concentration dynamics. Cross-correlations 246 show a relation between glomerular and ethanol signals during odor presentation with 247 all glomeruli having significant correlation with the plume during odor presentation as 248 compared to their respective null distributions from trial shuffled correlation anaylses. 249 In addition, a strong correlation between glomerular response power (0-5 Hz) and ability 250 to track odor concentration dynamics is also present in the hand-drawn ROIs (r = 0.80, 251 p =< 0.001). Thus, we find that CNMF captures the relationship between glomeruli 252 and plume dynamics while improving the resolution of glomerular network activity and 253 inter-glomerular temporal dynamics.

254
The CNMF activity traces from the identified glomeruli were baseline normalized 255 using the mean and standard deviation of a 5 second baseline activity period prior to 256 stimulus onset. Traces for each glomerulus were then deconvolved in the style of Stern 257 et al., 2020 to recover the average activity rate of each glomerulus. To find the optimal 258 penalty parameter, λ, for deconvolution, λ was optimized within each glomerulus. Then 259 the median of this optimized distribution was used as the λ in the deconvolution for all 260 glomeruli. After deconvolution, traces were standardized using the standard deviation of 261 the glomerulus' entire trace. Deconvolved signals were standardized in this way since 262 the florescence range of a glomerulus's response depends on the number of expressing 263 MT cells, the depth of the glomerulus from the dorsal surface, and other methodological 264 factors unrelated to the magnitude of the response.

265
Testing for responsive glomeruli 266 A glomerulus was considered to be responsive to odor if its deconvolved trace exceeded 267 threshold more time points than expected by chance during plume presentation (figure 7). 268 Since this preliminary measure does not rely on stimulus dynamics, it captures glomeruli 269 that respond to the plume even if their response is unrelated to odor concentration 270 dynamics or only present for part of the plume.

271
First, the deconvolved trace of a single glomerulus was split into two periods: baseline 272 9/31 activity and odor response. The baseline period is a 5 second period at the beginning of 273 each trial prior to plume onset. The odor response period is the time during which the 274 plume is present as well as one second immediately following the plume since inhibition 275 has been shown to induce excitatory rebound responses in tufted cells (Cavarretta et al., 276 2018). The signal is first baseline normalized by subtracting the mean and dividing by 277 the st dev of the baseline activity within each trial. Next, it is binarized, thresholding 278 for time points where activity exceeded the 95% confidence interval of the glomerulus's 279 original baseline activity (thresholded at ±1.96 baseline mean). In this way, each time 280 point that crossed the threshold was considered an event. Within each trial plume 281 presentation, if the number of events exceeded the null expectation (5% of the total 282 number of time points during plume presentation rounded up to the nearest integer), 283 the glomerulus was considered to be responsive to the plume during that trial. The 284 proportion of trials to which the glomerulus responded was calculated within three sets 285 of flow conditions (all flows, low flow, and high flow). The cumulative responsivity score 286 plotted in (figure 7A) is the sum of responsivity scores for all three flow conditions where 287 the contribution of each flow is plotted in different colors as a stacked bar graph. To understand the relation between stimulus and response time series, a preliminary 291 analysis was conducted by calculating the correlation coefficients between the two 292 signals for each glomerulus. In the future, more sophisticated techniques will be used to 293 establish how much of the neural representations can be explained by high-fidelity odor 294 concentration encoding.

295
Due to the stochastic nature of plume onset and offset times, the correlation is 296 only calculated for the middle 8 seconds of the 10 second plume so that onset and 297 offset dynamics are not included and the correlation measure represents the magnitude 298 of plume tracking during plume encounters. The cross-correlation coefficients of a 299 glomerulus r g is calculated between the ethanol e and calcium c deconvolutions during 300 plume presentations. For a single glomerulus, the correlation coefficient between the two 301 deconvolutions within a single trial n is calculated at all possible lags l. Using the xcorr() 302 function in Matlab to compute the coefficients, both signals are mean subtracted prior to 303 calculating the cross-correlations such that the correlation coefficients are synonymous 304 with calculating the pearson correlation coefficient between the two signals at each 305 respective lag value. 306 r g,n,l = corr(e n,l , c n,l ) The mean coefficient for each glomerulus, r g,l , is calculated by averaging across all 307 trials within the session (n=40 for 3 FOVs, and n=39 for 2 FOVs) at each possible lag. 308 The maximum coefficient mean is selected from all lags within a 500ms window w of the 309 neural activity following the ethanol signal.
This is considered to be a window of sufficient size to account for variable delays in 311 glomerular processing. The average time lag of r g was 130ms ± 100 (mean ± st. dev). 312

10/31
Within flow cross-correlations are calculated in the same manner but averaged only 313 across trials within the specified flow condition.

314
For plotting of tracking ability (figure 5A-C), r g is compared to a single trial shuffled 315 analysis using the same method as detailed above. The difference between the matched 316 and shuffled coefficients suggest correlations are not solely a result of plume structure, 317 but are driven by the temporal dynamics unique to each trial. Trials were shuffled within 318 each glomerulus by calculating the correlations between e n and c =n . In this way, any 319 relation dependent on the dynamics of the stochastic fluctuations within each plume 320 presentation is lost, but other statistical features of the plume presentation are preserved, 321 yielding a baseline value for the cross-correlation. Glomeruli are plotted in Figure 5 322 A-C if their correlation coefficient from the matched analysis exceeds ± 2 standard 323 deviations (st dev of the coefficient distribution from the shuffled analysis) of the shuffled 324 mean coefficient. Since correlation varies significantly within a glomerulus across flow 325 conditions, a glomerulus is considered to exceed the shuffled mean if it does in at least 326 one of the three defined conditions, all, low, or high flow.

327
Using the same shuffled correlation, a bootstrap analysis was conducted (10,000 328 iterations) creating a null distribution of the shuffled mean correlation coefficients to 329 test for significance (Supplemental Figure 7c). The mean correlations are compared to 330 their respective 95% confidence interval for the null distribution.

331
Comparison of correlation coefficients in the matched versus shuffled cross-correlations 332 does not naturally divide the glomeruli into two subpopulations, but rather the strength 333 of this relationship varies continuously across glomeruli. Therefore, instead of dividing 334 glomeruli into subpopulations of tracking versus non-tracking, our analyses consider 335 how the strength of odor concentration tracking compares to other properties of the 336 glomerulus and its response.

338
Measuring glomerular responses to plume dynamics 339 Using a commercially available, modified odor sensor combined with widefield calcium 340 imaging techniques in head-fixed mice, we reliably tracked plume dynamics and investi-341 gated glomerular response to this fluctuating input (figure 1). Imaging was conducted 342 in Thy1-GCaMP6f-GP5.11 mice which have fast kinetics and expression in mitral and 343 tufted (MT) cells within the olfactory bulb (OB) (Dana et al., 2014) (figure 2). Widefield 344 imaging of the dorsal surface of the OB allows for glomerular level resolution of the neural 345 response (figure 2B) (Fletcher et al., 2009). To explore a range of plume dynamics an 346 animal may encounter in its natural environment, we changed the airspeed in the wind 347 tunnel to create stochastic plumes with different odor concentration dynamics (figure 348 1D-H). Odor identity, concentration and volume released from the odor port remained 349 constant across all flow conditions, making the temporal structure of the plume the only 350 source of variation.  subjected to principal component analysis (PCA) and compared to the simultaneously 356 recorded plume dynamics. The FOV's were aligned prior to PCA, but no segmentation 357 or denoising was performed. To search for component activity responsive to plume 358 dynamics, the correlation between each principal component and the odor concentration 359 dynamics was calculated. There exists a high ranking component for each mouse that 360 correlates strongly with plume dynamics (figure 3B-D). Plotting the loading weights of 361 the maximally correlated component shows dense clusters of high variance resembling 362 partial spatial maps of glomerular activity. These findings demonstrate that MT 363 population activity recorded in the first relay of olfactory processing is correlated to 364 odor concentration dynamics during plume presentations. In order to establish whether 365 individual glomeruli are correlated to odor cues, we sought to segment the MT activity 366 into glomerular units to determine their respective contributions to the observed tracking 367 of plume fluctuations by population activity. . The mean deconvolved trace across trials was calculated 373 for each glomerulus during plume presentations. The mean was only considered during 374 the middle 8 seconds of the 10 second odor plume to concentrate on glomerular responses 375 to odor concentration dynamics during plume presentations and avoid responses to onset 376 or offset plume dynamics.

377
For each glomerulus, the average response mean (average of mean trace timepoints) 378 was also calculated within low and high conditions. A paired samples t-test found 379 Figure 4. The spatial and temporal decomposition of CNMF identifies glomeruli and denoises their traces. a) The white box outlines the FOV used for analysis as it relates to the larger recording window. The image shows the standard deviation projection of the aligned recording during a single odor presentation. b) Mean subtracted maximum projection of the same trial overlaid with ROIs from CNMF spatial decomposition shows segmentation of glomeruli for a single FOV using CNMF spatial decomposition. The spatial decomposition of the FOV results in 26 glomeruli (4 dropped units after merge analysis not pictured) as outlined and numbered. c) Shows the mean traces of each glomerulus's CNMF temporal decomposition within each flow condition (left to right : low, medium, high). Trials sorted by magnitude of normalized mean response during odor exposure. d) The deconvolved CNMF response of a single glomerulus (pink fill) to all low (grey) and high (black) flow trials across the recording sessions shows glomerular responses vary with the unique odor concentration dynamics of each plume. e) Deconvolution increases temporal accuracy of glomerular responses as shown by the mean deconvolved traces of the corresponding glomeruli depicted in (c). f) Sum of mean responses calculated within each flow condition. Mean responses (deconvolved) vary significantly between conditions (t(110) = 11.43, p < 0.001).

13/31
Figure 5. Glomerular population activity follows odor concentration dynamics across plume encounters. a) (left) The cross-correlation between the deconvolved ethanol trace and each glomerulus's deconvolved activity trace is calculated within each trial and then averaged across trials. Each row is a glomeruli and each time point represents the cross-correlation at the indicated lag. Glomeruli are sorted in order of decreasing magnitude of correlation coefficient (see methods). (right) Same but glomerular responses are trial shuffled so that the signals compared are not from the same trial. Glomeruli are sorted to match their corresponding unshuffled cross-correlation in the right panel. b) Scatterplot of the correlation coefficient of all glomeruli compared to their respective shuffled coefficient. Glomeruli plotted in (a) are marked in black and their coefficient exceeds their shuffled coefficient from a single trial shuffled comparison by 2 standard deviations. c) Cumulative scores for each glomerulus is the sum of their correlation coefficients calculated within low and high flow conditions. The cumulative correlation plot shows variation in a glomerulus's ability to detect changes in odor concentration dynamics varies significantly between conditions (t(110) = 12.81, p < 0.001), with most glomeruli having stronger correlation coefficients in high flow trials (*indicates glomeruli plotted in (a)). d) Binary cross-correlation. Top: Simultaneously recorded signals shown for two example glomeruli responding to the same example trial's odor plume. Odor and glomerular activity traces plotted with their respective thresholds (dotted, odor threshold: mean during plume presentation, neural threshold: ±2 st dev of baseline). Bottom: Resulting binarized traces plotted for each trial illustrate the magnitude of concurrent activity as events (stars) between the plume and the response of each glomerulus across the experimental session. that glomerular response means varied significantly across low and high flow conditions 380 (t(110) = 9.71, p < 0.001). Mean responses were higher in low flow conditions, during 381 which lower airspeed resulted in plume dynamics that were less intermittent, as shown 382 by lower skewness and asymmetry in the deconvolved odor signals of low flow trials as 383 compared to high flow trials ((F (157) = 52.7, p < .001 and F (157) = 54.9, p < .001 384 respectively)( figure 1E-H). In high flow conditions, increased intermittency produced 385 more brief, high concentration fluctuations followed by blanks, or periods without odor 386 signal. This decreased response mean was not due to a decrease in odor concentration 387 means though as plumes had higher concentration means during high flow trials (m = 202 388 a.u) as compared to low flow trials(m = 146 a.u) (t(116) = 3.30, p < 0.001). (As the 389 deconvolved traces are normalized using within trial averages, stimulus mean during high 390 and low flow trials are calculated using the raw sensor signal for relative comparison.) 391 Thus, the cumulative MT activity increased in low flow trials even as the cumulative 392 exposure to the stimulus decreased, suggesting the response of glomerular populations 393 are moderated by plume dynamics.

14/31
Correlation between stimulus and glomerular activity 395 To determine if plume dynamics could be moderating the glomerular population re-396 sponse, cross-correlation was used to quantify the relation between odor concentration 397 dynamics and simultaneously recorded glomerular activity (figure 5). Most glomeruli 398 significantly followed plume dynamics when correlation between neural activity and 399 odor activity was calculated across all trials (100/111), across low flow(97/111) trials, or 400 across high flow(100/111) trials. Significant tracking of the stochastic changes in odor 401 concentration across plume presentations is determined by comparing mean correlation 402 coefficients to a null distribution created using a trial shuffled bootstrap analysis (see 403 methods)(Supplemental Figure 7C). High correlation coefficients are not observed when 404 glomerular responses are trial shuffled and ethanol recordings are no longer compared 405 to the glomerular responses they elicited ( figure 5A-B). This shows that it is not the 406 statistics of stimulus presentations that drive this correlation, but rather the plume's 407 temporal dynamics unique to each trial.

408
Correlation coefficients increased from the null expectations by 0.13 ± 0.08 (mean ± 409 st dev) across all flow conditions, 0.08 ± 0.07 within low flow, and 0.16 ± 0.11 within 410 high flow. Glomeruli were significantly better at tracking plume dynamics in high flow 411 than they were in low (t(110) = 12.81, p < 0.001) with average correlation coefficients 412 increasing by 0.11 ± 0.09 (mean ± st dev). We wondered if this increase in correlation 413 could result from increased sparsity. Indeed correlation between two signals where the 414 values are constant or zero for most of the time is automatically high, even if the peaks 415 are entirely uncorrelated. If this was the case, the shuffled correlations in high flow 416 should be significantly higher than in low flow, but this is not observed when looking 417 at the confidence intervals for the null correlation coefficients computed within flow 418 conditions (supplementary figure7C). Thus, a large fraction of the glomerular population 419 follows fluctuations during plume encounters, and the degree of dynamic tracking is 420 moderated by plume dynamics, becoming stronger on average during plumes with higher 421 levels of intermittency (figure 5C).

422
Plume fluctuations structure glomerular network dynamics 423 Glomerular activity is moderated by changes in plume structure. We measured glomerular 424 responsivity and response power to see the effect of flow condition on these measures and 425 whether these measures were related to how well a glomerulus followed plume dynamics. 426 We found that there was a significant effect of flow condition on glomerular responsiv-427 ity. Responsivity is defined as the proportion of trials to which a glomerulus responded 428 to the odor (figure 7) (see methods). Glomeruli had significantly higher responsivity 429 during low flow trials(t = 12.1, p < 0.001), with repsonsivity scores increasing by 0.21 430 ± 0.18 (mean ± st dev) as compared to high flow. Therefore, glomeruli responded 431 to low flow trials more reliably than they responded to high flow trials. As noted 432 previously, the average correlation between plume dynamics and MT activity increased 433 in high flow conditions, so although glomerular responses became less reliable as airflow 434 increased, they became more correlated with plume dynamics. Thus, plume structure 435 changes glomerular responsiveness and response nature. Responsivity is a thresholded 436 measure that determines if a glomerulus is more active than expected by chance during 437 a plume presentation and does not capture the dynamics present in the response. The 438 strength of the dynamic activity of glomeruli was determined by measuring the change 439 in cumulative response power between baseline periods and plume presentations (see 440 Figure 6. Higher magnitude of glomerular response power (0 − 5Hz) is associated with higher correlation with plume dynamics. a) Short-time Fourier transforms of a single low flow trial and a sample of responding glomeruli show most response power of the glomeruli and odor signal is concentrated between 0 − 5Hz. Glomeruli are sorted by increasing correlation with plume dynamics. b) Same but for a single high flow trial from another example FOV. c) With both high and low flow conditions, correlation of the glomerulus with plume dynamics is plotted against its corresponding increase in power spectrum activity between 'odor off' (top) and 'odor on' (bottom) periods. Glomeruli with higher correlation coefficients have a stronger increase in response power during plume presentations (r = 0.74, p < 0.001). When calculated within flow, this relationship is significant within high flow (r = 0.73, p < 0.001), but not within low flow(r = 0.19, p = 0.05). The average repsonse across all glomerular is plotted (red/yellow) to represent the population response. Mean response power of the glomerular population is not significantly different between low (yellow dot) and high flow (red dot), except for when calculated amount glomeruli whose mean acitvity is in the 75th percentile (low = yellow circle, high average = red circle). Figure 7. Glomeruli that respond more reliably to plumes are more correlated with their dynamics. a) The cumulative responsivity is the sum of the responsivity scores calculated within each flow condition. Glomeruli are sorted by decreasing average correlation with plume dynamics. Glomeruli whose correlation coefficient exceeds its null confidence interval (see methods) are plotted in blue hues and the remainder of the glomeruli are plotted in grey hues. This shows the magnitude of odor concentration tracking is correlated with (r = 0.76, p < .001), but not strictly defined by, response reliability as glomeruli exist that respond strongly to odor presence but not to concentration dynamics. b) Within flow condition, repsonsivity is plotted against correlation with odor dynamics for each glomerulus (circles) and for the population average across all glomeruli (dots). Across glomeruli, responsivity is positively correlated with tracking ability as illustrated by the lines of best fit. For each glomerulus, their relative responsivity decreases significantly during high flow (dark blue) as compared to low flow (light blue) (t(110) = 12.1263, p < 0.0001). To represent the population response, the average responsivity across all glomeruli is plotted against average correlation with plume dynamics (dots) within both conditions illustrating how flow moderates these relationships. For a given glomerulus, higher flow predicts a decrease in its relative responsivity level (t(110) = 12.1263, p < 0.0001) and an increase in its relative tracking ability. methods). Fast Fourier transform was used to measure response power within 0-5 Hz, a 441 frequency range relative to the stimulus dynamics (figure 6). Response power (0-5 Hz) 442 increased significantly from baseline during plume presentations (t(110) = 20, p < 0.001) 443 by 5.7 ± 3.0 a.u. (mean ± st dev), a 448% increase (figure 6). On average 86.7% ± 4.9% 444 (mean ± st dev) of cumulative stimulus power for each session was within 0-5Hz. The 445 majority of cumulative response power for each glomerulus, 88.5% ± 7.9% (mean ± st 446 dev), was also found to be within this range. Thus, the majority of response power for 447 each glomerulus was measured to be within a relative response range of the stimulus 448 ( figure 6A-B). Across all glomeruli recorded,response power was not significantly different 449 between flow conditions. To examine the effect of flow conditions on the cells that most 450 strongly responded to the odor, we next analyzed only glomeruli whose mean response 451 was above the 75 th percentile. Within this group of glomeruli, response power did change 452 significantly between flow (t(27) = 5.52, p < 0.001), with stronger response power during 453 high flow conditions. This increase reflects the significant increase in stimulus power 454 (0-5 Hz) observed in high flow as compared to low flow trials (t(116) = 31, p < 0.001) 455 (Supplemental Figure 6). Thus, the response power of glomeruli with the strongest 456 signals was significantly affected by flow condition.

457
There exists a relationship between each of these two response features, responsivity 458 and response power, and how well a glomerulus follows plume dynamics. Across all 459 trials, glomeruli with higher responsivity to plume presentations were better at following 460 changes in odor concentration (r = 0.76, p < .001). Thus, the more reliably a glomerulus 461 responded to plume presentations, the more likely it was to follow changes in odor 462 concentration. This is not a perfect relationship as glomeruli that are highly responsive 463 to the plume but not its dynamics exist (figure 7), but a glomerulus with higher 464 responsivity is more likely to be correlated to plume dynamics than one with lower 465 responsivity. Also, as mentioned previously, a glomerulus's average responsivity level 466 is moderated by flow condition. Response power was also correlated with how well 467 the glomerulus followed changes in odor concentration, when averaged across all trials 468 , glomeruli with higher response power were significantly better at following plume 469 dynamics (r = 0.74, p < 0.001). When this relationship was examined within flow 470 conditions (figure 6C), high flow was significantly correlated (r = 0.73, p < 0.001), but 471 low flow was no longer significantly correlated (r = 0.19, p = 0.05) (All flow includes 472 medium flow, which is not analyzed in the between flow comparisons). In addition, for 473 a subset of strong responding glomeruli, this response power was moderated by flow 474 condition. These results suggest that both the reliability and the temporal pattern 475 of MT activity is significantly moderated by the odor concentration dynamics of the 476 incoming stimuli.

477
These results show that the spatiotemporal dynamics of plumes play a role in 478 structuring activity in the first olfactory relay of the mouse's brain during natural 479 olfactory processing.

481
Mice are adept at olfactory guided search despite the stochasticity and complexity of 482 odor plumes used in navigation. Spatiotemporal cues present in natural odor scenes are 483 thought to drive decision-making in olfactory search ( (Mafra-Neto and Cardé, 1994); 484 (Vickers, 2006)), (Pang et al., 2018), but how they moderate population activity in the 485 18/31 olfactory bulb (OB) is unknown. Releasing odor within a custom-built wind tunnel, we 486 were able to hold constant all properties of the odor stimulus and the animal's position 487 relative to the source and vary only the air velocity through which the plume traveled. 488 By using this approach we altered the Reynolds number of the flow and created plumes 489 with varying statistical structures and odor concentration dynamics. In this way, the 490 effect of plume dynamics on MT population activity could be examined using naturally 491 evolving odor plumes. Recording MT activity in mice expressing GCaMP6f, we show 492 that a significant fraction of glomerular populations of MT cells follow odor plume 493 dynamics. Additionally, the strength with which they do so is moderated by airflow, 494 such that increased flow velocity and turbulence (Reynolds number) results in increased 495 correlation of MT cell activity with plume dynamics. This work shows that plume 496 dynamics structure the activity of the OB, the first relay of olfactory coding in the 497 mouse's brain.

498
The recent history of an odor stimulus has been shown to be present in olfactory 499 encoding in both serial sampling of odor concentration in mice (Parabucki et al., 2019) 500 and tracking of odor concentration in invertebrates (Geffen et al., 2009), showing odor 501 concentration changes influence olfactory encoding. Although inter-sniff comparisons 502 in mice show that MT cells can detect the sign and magnitude of changes in odor 503 concentration (Parabucki et al., 2019), it is unknown whether they are able to resolve 504 the dynamics of natural plumes, which span across a range of temporal scales. If odor 505 concentration dynamics are resolved, computational work has shown that they are 506 informative for olfactory search ( (Baker et al., 2018); (Gumaste et al., 2020)). To 507 avoid the complexity of stochastic odor plumes, the averaging of odor concentration 508 dynamics could be an alternative strategy to navigate olfactory environments. Mean 509 odor concentration levels are moderated both by the distance from an odor source and 510 how close an animal is to the central stream of the plume (Crimaldi and Koseff, 2001). 511 While this measure is potentially informative, it does not by itself sufficiently inform 512 decision-making on the timescale observed in rodents (Gumaste et al., 2020). Therefore, 513 it is likely that the mouse relies upon spatiotemporal features of the plume for olfactory 514 search as information can be extracted from odor concentration dynamics (Baker et al.,515 2018).

516
Our study found a correlation between MT activity and odor concentration dynamics 517 during plume presentations. The temporal information conveyed by MT cells could 518 support a variety of navigation algorithms. For instance, two important dynamic features 519 are the length of odor encounters, whiffs, and the timing between odor encounters, blanks. 520 Whiff and blank duration are moderated by the distance between an animal and the odor 521 source. As an animal approaches an odor source, plume encounters become shorter and 522 more frequent ( (Wright and Thomson, 2005); (Celani et al., 2014)). Blank duration 523 has been shown to be particularly informative even when olfactory environments change. 524 Computational modeling of olfactory search in invertebrates ( (Park et al., 2016); (Rapp 525 and Nawrot, 2020)) as well as fluid dynamics modeling (Celani et al., 2014) shows that 526 the time between odor encounters, blank duration, is less sensitive to environmental 527 conditions such as plume velocity or potency of the odor source that are know to 528 affect interpretation of odor concentration dynamics ( (Connor et al., 2018); (Webster 529 and Weissburg, 2001)). Specifically, Park et al., found blank duration to be a more 530 efficient source of information for olfactory search than instantaneous tracking of odor 531 concentration. In our study, we observed that MT activity was more correlated with 532 plume dynamics in high flow trials than low flow trials. Odor concentration in low 533 19/31 flow trials was less skewed, meaning that these trials had lower intermittency and odor 534 concentration tended to fluctuate around a central value. Alternatively, high flow trials 535 were more skewed and were characterized by a whiff and blank structure. The fact that 536 correlations are higher in high flow suggests MT activity may be more responsive to whiff 537 and blank features as opposed to tracking fine fluctuations in odor concentration across 538 more constant plume encounters. Future studies exploring the effect of a broader range 539 of intermittency levels on MT activity during plume encounters could help determine 540 which spatiotemporal temporal features of intermittency are moderating MT responses 541 to plume dynamics.

542
The majority of glomerular activity responding to concentration dynamics could 543 be considered to be inefficient when the OB has to perform other tasks such as odor 544 identification and segmentation. Glomerular spatial maps, i.e. glomerular ensembles 545 consistently responding to an odor, are thought to be one of the primary means of odor 546 identification (Wachowiak and Shipley, 2006). Odors maps vary with concentration 547 ( (Xu et al., 2000); (Wachowiak and Cohen, 2001)), but are stable enough to reliably 548 encode odor identity (Belluscio and Katz, 2001). Although pulsed odors can be rapidly 549 discriminated (<200ms, or a single sniff, (Uchida and Mainen, 2003)), in a natural 550 olfactory environment where odors are intermittent ( (Celani et al., 2014); (Murlis 551 et al., 2000)) and mixed, identification becomes a much more complicated task especially 552 for identification of mixtures. Glomerular ensembles reliably responding across odor 553 encounters could aid odor discrimination in mixed odor environments. Spatial maps of 554 odor identification will overlap in natural olfactory scenes where an animal encounters 555 signals from multiple odor sources as it navigates a plume. Odors co-released travel 556 together (Celani et al., 2014), and therefore a mixture of odors emanating from the same 557 source will have correlated temporal dynamics in the plume and will thus be experienced 558 by a searcher as having correlated encounters across whiffs. This means the probability 559 of the signal from separate sources arriving together reliably across whiffs would be low 560 if odors are released from spatially separated sources. Grouping and demixing these 561 odor representations using the correlation, or lack thereof, in the odor concentration 562 dynamics could aid odor discrimination in complex environments (Hopfield, 1991).

563
Since our studies are not recorded at the individual cell level, the potential degree of 564 heterogeneous tuning to different features among MT cells within a single glomerulus 565 was not examined. It could be that observed correlations of glomerular MT populations 566 were a product of the collective activity of heterogeneously tuned MT cells within a 567 glomerulus, but MT responses have been shown to linearly sum odor inputs (Gupta 568 et al., 2015), which contradict the idea that they are directly tuned to different features 569 of plume dynamics. At the same time, this does not infer MT activity responding to 570 plume dynamics is homogeneous as the responsiveness (Adam et al., 2014) and the 571 response (Geramita and Urban, 2017) varies between MT cells across concentration levels 572 (Cleland and Borthakur, 2020). Future research across a variety of odor concentration 573 dynamic regimes and odor mixtures at both the cellular and population level are needed 574 to further investigate the degree to which bulbar responses are tuned to features of 575 odor concentration dynamics and how this tuning may impact optimal encoding of odor 576 information.

577
Our data show that MT activity in the OB of mice follows the temporal dynamics of 578 odor plumes. Additionally, we demonstrate that this effect is stronger under conditions 579 that generate larger Reynolds numbers. Following odor concentration dynamics within 580 plumes could enable MT cells to convey information useful for olfactory search. Following 581 20/31 the temporal dynamics of odor plumes may also be an efficient form of multiplexing 582 odor identity and source location for the first olfactory relay in mice. Although MT 583 activity responds to changes in odor concentration, the observed correlations do not 584 suggest perfect tracking at the level of individual glomeruli and indicate inter-individual 585 differences in the degree to which glomeruli follow plume dynamics. Future research 586 focusing on location encoding across a wide range of both intermittency regimes and 587 odor panels is needed to clarify the degree to which bulbar activity is tuned to features 588 of plume dynamics and how a balance between identity coding and concentration coding 589 is instrumental in supporting the wide variety of behaviors enabled by olfaction.  Figure S1. Deconvolved Ethanol signal and spectral decomposition from one session divided by flow type. a) The raw ethanol sensor output is shown for all trials within a single session. (red, orange, and blue are high, medium and low flow respectively). b) The deconvolution of the raw ethanol signal. c) The simultaneously recorded signal from a co-localized PID. Sensor deconvolution and PID are significantly correlated as calculated during plume presentations (r = .3089, p < .001). d) (top) The spectral decomposition calculated across all trials or within flow conditions for the raw sensor signal shown in (a). Only the middle 8 seconds of the 10 second plume presentation are analyzed to avoid onset or offset dynamics. Spectral decompositions are also plotted for the ethanol deconvolution (middle) shown in (b) and for the PID (bottom) shown in (c). Figure S2. Normalized autocorrelation of ethanol trace. a) The Normalized average autocorrelation of the deconvolved ethanol signal is shown for each of the 4 experimental sessions. Traces are calculated across all trials within the designated flow conditions. b) Same as (a) but trials are shuffled such that each cross-correlation is calculated between the deconvolved ethanol signal from two different trials. Figure S3. Pairwise correlations between glomerular ROIs are decorrelated by CNMF. a-e) Distributions of the pairwise linear correlation coefficients from all possible glomerular pairs within each FOV. Correlations are calculated on the temporal decomposition of CNMF ROI signals (without deconvolution) which include denoising and demixing (blue) or on the raw pixel values as averaged within the same ROI boundaries (red). Pairwise correlations are calculated between the two glomeruli during stimulus presentation. f) Scatterplot of each pairwise correlation shows CNMF de-noising decorrelates the signal for each pair of glomeruli. Pairwise correlations in grey are from ROI pairs that have been considered to originate from the same glomerulus after the merging threshold analysis (Supplemental Figure 4). Figure S4. Detecting oversegmentation of ROIs after CNMF decomposition. a) ROI segmentation of glomeruli for FOV 4 using CNMF spatial decomposition of activity across an entire session. ROIs circled in purple do not meet thresholds for independent ROIs (methods) and thus are considered to originate from the same glomerulus, while the ROI in magenta is considered to be a separate glomerulus. b) Raw fluorescent signal averaged across all pixels within each of the 3 ROIs (without denoising or de-mixing) is plotted for all trials in the session. Numbers above each plot refer to ROI numbering in panel (a). c) CNMF signal across all trials plotted for the corresponding ROIs in (b). Figure S5. Hand-drawn ROIs have qualitatively similar results but higher pairwise correlations. a) (left) St dev projection of a single FOV during the first trial of the session. (middle) ROIs from hand-drawn analysis are plotted over the projected st dev (left). ROIs determined using recordings and st dev projections averaged across trials within each flow condition. (right) Mean subtracted maximum projection of the same trial overlaid with ROIs from CNMF spatial decomposition. b) Distribution of hand-drawn ROI sizes from (a). c) Trial averaged linear pairwise correlation coefficients of ROI activity during plume presentations shows high correlations between ROIs. d) Glomerular correlation with odor dynamics is plotted against their corresponding change in repsonse power (0 − 5Hz) between 'odor off' (d) and 'odor on' periods showing glomeruli with higher correlation coefficients have higher activity power (r = 0.80, p =< 0.001). The dotted line plots the line of best fit using OLS regression. e) Glomerular activity is correlated with stimulus dynamics at varying lags (right) with all mean coefficients exceeding a 95% confidence interval of their shuffled mean coefficients. Each row is a glomerulus and each time point represents the mean correlation coefficient between odor and response deconvolutions at the indicated lag. Glomeruli are sorted in order of decreasing correlation. (right) Same but neural activity responses are shuffled so that the signals compared are not from the same trial. The glomeruli are sorted to match their corresponding unshuffled cross-correlation in the right panel. Figure S6. 0-5 Hz Power spectrums for odor and glomerular responses. a) Average binary cross-correlation (subplots same as shown in Fig. 6) for each glomerulus is plotted against the power spectrum of its activity during 'odor off' (top) and 'odor on' (bottom) periods. b) The power spectrums of the deconvolved odor (blue) and glomerular (red) traces are plotted for 8 example glomeruli (numbered accordingly in (a)) during 'odor off' and 'odor on' periods. Figure S7. Glomerular sizing and plume following behavior. a) Distribution of CNMF ROI sizes for all sessions (µm). b) ROI size (µm) plotted against the mean correlation coefficient between deconvolved glomerular and ethanol signals during odor presentation shows correlations between ROIs and plume dynamics are not driven exclusively by singular MT cell activity. Kendall's tau coefficient between size and odor concentration tracking (rτ = 0.1563, p = 0.202) shows that correlation to plume dynamics is not exclusive to or related to smaller sized ROIs (ones similarly sized to individual MT cells). This suggests response to plume dynamics across plume encounters is also occurring at the glomerular level. c) Mean correlation coefficients of glomeruli (red) are plotted against their respective bootstrapped 95% confidence interval of null mean coefficients (see methods for trial shuffled bootstrap analysis). Null confidence intervals are calculated across all flow conditions (black), within low (pink) or within high (blue) flow.