Introduction

Animals exhibit an impressive ability to change how sensory inputs map onto behavioral outputs. Understanding how animals learn to output different behaviors through experience is one of the fundamental problems in neuroscience. Over the last half century, the field has developed compelling frameworks to tackle this problem at both the algorithmic level (Rescorla-Wagner models, Q-learning models) (Rescorla, 1972; Sutton, 1988) and the mechanistic level (Hebbian learning, STDP, neuromodulation) (Dan and Poo, 2004). By comparison, we lack frameworks through which to understand how the brain might implement learning algorithms through the updating of synaptic weights. One strategy has been to identify neural correlates of latent features assumed to be required for these algorithms (e.g. dopamine as a neural substrate for reward-prediction-error) (Hollerman and Schultz, 1998; Schultz et al., 1997). These results, however, can often be difficult to interpret because reward related signals are found globally throughout the brain (Allen et al., 2019), and are likely multiplexed with signals about motor output and/or stimulus identity. We propose that a more powerful approach is one that compares 1) how the encoding of reward cues changes from one brain nucleus to its downstream target and 2) how much of the encoding can be explained by valence vs. other features such as identity or motor output. In this present work, we implement this comparative approach to the investigation of how encoding of olfactory reward cues is transformed between the olfactory tubercle (OT) and the ventral pallidum (VP) in the context of classical conditioning.

The OT, also known as the tubular striatum (Wesson, 2020), is a 3-layered striatal nucleus situated at the bottom of the forebrain. As with other striatal structures, the OT is composed primarily of Spiny Projection Neurons (SPN’s) which express either the Drd1 or Drd2 DA receptors (abbreviated as OTD1 and OTD2, respectively) (Tritsch and Sabatini, 2012). In addition to receiving a wide range of inputs from cortical and amygdalar areas (e.g. AI, OFC, BLA, PlCoA, Pir) (Zhang et al., 2017b), it receives dense DAergic input from the midbrain (Ikemoto, 2007) and direct input from the mitral and tufted cells of the olfactory bulb (Haberly and Price, 1977; Igarashi et al., 2012; Scott, 1981), a unimodal and primary sensory area. There is a range of experiments that suggest that the OT’s DAergic innervation is involved in reward processing. Coincident stimulation of the lateral olfactory tract and DAergic midbrain afferents supports LTP of excitatory current (Wieland et al., 2015) and rats self-administer cocaine, a DAergic drug, into the medial OT more vigorously than to any other striatal nuclei (Ikemoto, 2003). And while the OT neurons are known to respond to a wide range of odorants (Wesson and Wilson, 2010), pairing stimulation of midbrain DAergic neurons with an odor drives appetitive behavior towards the paired odor (Zhang et al., 2017a) and enhances the contrast of the odor-evoked activity (Oettl et al., 2020). Lastly, a number of recent publications report varying degrees of valence signals recorded from neurons in the OT (Gadziola et al., 2020, 2015; Martiros et al., 2022; Millman and Murthy, 2020; Oettl et al., 2020).

The most well-established target of the OT is the VP (Newman and Winans, 1980; Zahm and Heimer, 1987), a pallidal structure that lies immediately dorsal to the OT. In addition to OT input, VP receives strong input from the nucleus accumbens (Jones and Mogenson, 1980) and the subthalamic nucleus (Ricardo, 1980; Turner et al., 2001). More recently, it was reported that VP also receives inputs from several cortical and amygdalar areas that also project to the OT (e.g. Pir, BLA, OFC) (Stephenson-Jones et al., 2020). The VP contains GABAergic neurons, which respond to positive valence cues, and glutamatergic neurons, which respond to negative valence cues. Consistent with their responsiveness, the GABAergic and glutamatergic neurons drive real time place preference and avoidance, respectively (Faget et al., 2018). Though it is well-established that the VP plays a critical role in reward processing, there has been ongoing disagreement on what specific latent features are encoded by VP neurons. Interpretations have included valence (Ottenheimer et al., 2018, 2020b; Richard et al., 2016; Tachibana and Hikosaka, 2012), hedonics (Smith et al., 2009; Tindell et al., 2006), motivation (Faget et al., 2018; Fujimoto et al., 2019; Lederman et al., 2021; Tindell et al., 2005), and reward-prediction error (Ottenheimer et al., 2020a). This ongoing discussion highlights the need to adopt a more comparative approach outlined above.

Here, we investigated the transformation of learned association encoding between the OT and the VP. We began by refining our understanding of OT’s efferents to reveal that, contrary to a previous report, both OTD1 and OTD2 neurons project primarily to the ventrolateral portion of the VP and minimally elsewhere. Given this finding that VP may be the only robust output of the OT, we proposed that the OT to VP circuit is an ideal model system for examining how the encoding of reward cues is transformed between connected brain areas. Comparing the stimulus-evoked activity in OTD1, OTD2, and VP neurons with 2-photon Ca2+ imaging, we found that VP neurons encode reward-contingency in low-dimensional space with good generalizability. In contrast, activity in both OTD1 and OTD2 neurons was high-dimensional and primarily contained information about odor identity, though OTD1 neurons are modulated by reward. By examining the same neurons across multiple days of pairing, we propose a putative cellular mechanism for reward-cue responsiveness in VP wherein reward responsive VP neurons gradually become reward-cue responsive. Finally, using a novel classical conditioning paradigm, we provide evidence that non-overlapping sets of VP neurons contain information about the vigor of licking and reward-contingency, but not both.

Results

In order to compare odor-evoked activity in connected brain nuclei, we first characterized which specific subregions of the VP receive input from the anteromedial portion of the OT. While considerable effort has been made to unravel the anatomy and function of the NAc, much less attention has been directed at the OT. Though multiple studies have characterized its anatomical connectivity (Zahm and Heimer, 1987; Zhang et al., 2017b; Zhou et al., 2003), there is inconsistency regarding whether or not OT projects to areas other than the VP. We therefore aimed to clarify previously reported OT connectivity by independently conducting anterograde viral tracing experiments in OTD1 and OTD2 neurons of the anteromedial OT. To this end, we injected AAVDJ-hSyn-FLEX-mRuby-T2A-syn-eGFP in the anterior OT of Drd1-Cre (labels D1+ SPN’s) and Adora2a-Cre (labels D2+ SPN’s) animals (Fig1A, FigS1C-E). Because viral contamination of areas dorsal to the target site can lead to difficulties in interpretation of tracing data, we also injected the same virus to the AcbSh immediately dorsal to the OT for comparison. Consistent with past findings (Kupchik et al., 2015), we observed robust projections VP, LH, and VTA from AcbShD1 neurons and primarily VP projections AcbShD2 neurons (Fig1B-C). We also observed dense labeling of the VP in D1-Cre and A2A-Cre animals injected at the OT. Contrary to one report (Zhang et al., 2017b) but consistent with another (Zhou et al., 2003), we observed minimal labeling in LH and VTA, or anywhere else in the brain, for both OTD1 and OTD2 experiments (Fig1B-C, FigS1A-B), suggesting that neither OT subpopulation from the anteromedial OT projects strongly outside the VP. As previously reported (Groenewegen and Russchen, 1984). It is also notable that OT projections were restricted to the lateral portions of the VP.

OTD1 and OTD2 primarily project to the lateral portion of the VP.

(A) Schematic representation of Cre-dependent anterograde axonal AAV tracing experiments used to characterize outputs of OT neurons. Drd1+ and Drd2+ neurons were separately labeled by using Drd1-Cre and Adora2a-Cre mouse lines, respectively. (B) Representative images from OTD1 (top) vs. the AcbShD1 injection (bot). Target sites (far-left column) are stained with ⍺-tyrosine hydroxylase antibodies to visualize the boundary between VP and OT. (C) Quantifying the % of output regions with fluorescence (n=3-4). (D) Schematic representation of 2-color retrograde CTB tracing experiment used to confirm OT to VP connectivity. CTB::488 and CTB::543 were injected to the lateral and medial portion of the VP, respectively. (E) Representative images of CTB labeled neurons in the OT and Acb. (F) The number of labeled cells was quantified (n=4). (G) Schematic representation of retrograde CTB tracing experiment used to test OT to VTA connectivity. CTB::647 was injected in the VTA. (H) Representative image shows robust AcbSh and AcbC labeling but no OT labeling. (I) Quantification of labeling in different nuclei (n=3). Pairwise comparisons were done using the Student’s t-test. The p-values were corrected for FDR by Benjamini-Hocherg procedure. ***p<0.001, **p<0.01, *p<0.05. See Tables S1-S3 for detailed statistics.

To corroborate and more precisely describe the OT to VP projection, we conducted retrograde tracing by injecting CTB::488 and CTB::543 to the lateral and medial portion of caudal VP, respectively (Fig1D, FigS1F-G). We found strong labeling of soma by both CTB::488 and CTB::543 in the Acb, AI, and Pir (Fig1E-F). By comparison, we found predominantly CTB:488, but not CTB::543, labeling in OT soma, indicating OT neurons are more likely to project to the lateral portion of the VP than to the medial. Similarly, to corroborate the lack of OT to VTA projection, we injected CTB::647 into the VTA (Fig1G, FigS1H-I). Consistent with previous findings (Beier et al., 2015; Faget et al., 2016; Watabe-Uchida et al., 2012), we found dense labeling of soma in various areas of the striatum such as AcbSh, AcbC, and CPu (Fig1H-I). We also found some labeling of soma in some frontal cortical regions such as PrL, AI and IL cortices. In contrast, we found that hardly any OT neurons were labeled. The rare OT neurons that did have CTB labeling were exclusively localized to the dorsal most portion of layer III, closely bordering the VP. Taken together, we conclude that both D1 and D2 SPN’s of the anteromedial OT project primarily to the lateral portion of the VP and negligibly to other brain areas, including the VTA.

Once we had identified that OT has extremely constrained outputs to the lateral VP, we set out to comparatively characterize the encoding of reward cue in this striatopallidal circuit. Past analysis of valence encoding is confounded by not accounting for the difficult-to-avoid overlaps among identity, salience, and reward contingency. To address this, we carefully designed a 6-odor conditioning paradigm where these factors could be decoupled (Fig2A). During each trial, the animal is exposed to 1 of 6 odors for 2 seconds. At the end of odor delivery, the animal either receives: 2 µl of a 10% sucrose solution (S), 50 ms of airpuff at 70 psi (P) or nothing (X). 3 of the odors are ketones (hexanone, heptanone, octanone) and the rest are terpenes (terpinene, pinene, limonene), but the pairing contingencies are chosen such that each contingency group (S, P, or X) includes 1 ketone and 1 terpene. We reason that in a valence-encoding population, but not in an identity-encoding population, we should see that odor pairs of different reward-contingency (e.g. SK, a sucrose-paired ketone vs. PK, an airpuff-paired ketone) are more different than odor pairs of same reward-contingency (e.g. SK, a sucrose-paired ketone vs. ST, a sucrose-paired terpene). Additionally, because both sucrose-pairing and airpuff-pairing should make the associated odor more salient, we can disambiguate between increased discriminability due to salience vs. valence by comparing neural activity in response to sucrose-cues or airpuff-cues.

Head-fixed 2-photon Ca2+ imaging of OTD1, OTD2, or VP neurons during 6-odor conditioning paradigm.

(A) State-diagram of odor conditioning paradigm. Each trial begins with 2 seconds of odor delivery. Odors are chosen in pseudorandomized order such that the same odor is not repeated more than twice in a row. At the end of odor delivery, there is a variable delay (100-300ms), after which the animal is given either a 10% sucrose solution (SK and ST), a 70 psi airpuff (PK and PT), or nothing (XK and XT). Trials are separated by a variable intertrial interval (ITI; 12-18s). Schematic representation of (B) lens implant surgery and (C) headfix 2-photon microscopy setup. An example of spatial (D) and temporal (E) components extracted by CNMF from Drd1-Cre animal on day 3 of imaging. (D) The spatial footprints of 20 example neurons are shown on top of a maximum-correlation pixel image that was used to seed the factorization. The number displayed over each neuron matches the row number of the temporal components in (E). (F) An example raster plot (top) and averaged-across-trials trace (bottom) of the licking behavior recorded concurrently as (D) and (E). The timing of odor delivery is shown as shaded rectangles. The timing of US delivery is shown as arrowheads. (G) The mean total licks during each of the odors is shown averaged across all animals (n=17) after application of a moving-average filter with a window size of 10 trials. Red line marks the sucrose and airpuff contingency switch between day 3 and day 4. (H) Bar graph showing the licks during either sucrose cue expressed as a fraction of all licks during any odor. FWER-adjusted statistical significance for post hoc comparisons are shown as: ***p<0.001, **p<0.01, *p<0.05. See Tables S4-5 for detailed statistics.

To record the activity of the OT and VP neurons across multiple days of pairing, we injected C57BL/6 mice with AAV9-hSyn-jGCaMP7s-WPRE (lateral VP) and Drd1-Cre or Adora2a-Cre animals with AAV9-hSyn-FLEX-jGCaMP7s-WPRE (anteromedial OT) (Fig2B, FigS2A-F). Additionally, we implanted a 600µm Gradient Refractive Index (GRIN) lens 150µm dorsal to the virus injection site and cemented a head-fixation plate to the skull. 6-8 weeks after surgery, animals were water-restricted and habituated for 3-5 days in the head-fixation setup (Fig2C). We processed the acquired time-series images using Constrained Nonnegative Matrix Factorization (Pnevmatikakis et al., 2016) to obtain fluorescence traces from each putative neuron (Fig2D,E). In total, we recorded Ca2+ signals from 231 OTD2 neurons from 6 Adora2a-Cre animals (FigS3), 288 OTD1 neurons from 6 Drd1-Cre animals (FigS4) and 130 VP neurons from 5 C57BL6/J animals (FigS5).

After 3 days of odor-sucrose associations, the animals displayed anticipatory licking behavior primarily during sucrose-paired odors (Fig2F,G). Starting on day 4, the sucrose and airpuff contingencies were switched such that every odor had a reassigned contingency. By day 6, animals had adapted their anticipatory licking behavior to match the new sucrose-contingency (Fig2G). Quantification of the animal’s licking behavior showed that the accuracy of animals’ licks during odor increased across time and was not different across lens-placement groups (Fig2H, TableS4,S5; ANOVA: Fday=27.64, pday=2.29e-16, Flens location=2.30, plens location=0.11). These results show that the animals learn to associate S odors with reward in a flexible manner in our paradigm. Because we saw the strongest behavioral evidence that animals learned odor-sucrose associations by day 6, we focused our analysis on how reward cues are encoded on the last day of imaging. The animals also showed trends of behavioral changes in response to airpuff-cues, though they were not significant: during airpuff-cues, animals walked less and closed their eyes more than during other odors (FigS6D-G, TableS32-S36). These behavioral changes for aversive cues were less robust than that for reward association. However, animals show clear responses to the US indicating that they perceive the aversive stimulus.

OT and VP neurons showed heterogeneous responses to 6 odors across all 6 days of imaging (FigS3, FigS4, FigS5). To unbiasedly describe the difference between regions, we performed hierarchical clustering on the pooled trial-averaged responses to the 6 odors on the 6th day of imaging (Fig3A). We observed both inhibitory (clusters I, II) and excitatory (clusters III-VI) responses to odors as well as broad (clusters II, VI) and narrow (clusters IV, V) odor-tuning (Fig3B). Cluster I and cluster III most closely fit our description of putative valence-encoding neurons, i.e. neurons that had similar responses to 2 sucrose-cues (SK vs. ST) but different responses to a sucrose-cue and a puff-cue or control odor (SK vs. PK or XK). Although all clusters included neurons from all subpopulations, cluster I and cluster III, which showed larger responses to odors predicting sucrose, were enriched for VP neurons (Fig3C), leading us to hypothesize that individual neurons in the VP were more likely to be valence encoding neurons than in either OT subpopulation.

VP neurons encode reward-contingency more robustly than OTD1 or OTD2 neurons.

(A) Heatmap of odor-evoked activities in OTD1, OTD2, and VP neurons from day 6 of imaging. The fluorescence measurements from each neuron were averaged over trials, Z-scored, then pooled for hierarchical clustering. Neurons are grouped by similarity, with the dendrogram shown on the right and a raster plot on the left indicating which region a given neuron is from. Horizontal white lines demarcate the boundaries between the 6 clusters. Odor delivered at 0-2 seconds marked by vertical red lines and US delivery is marked by arrowheads. From left to right, the columns represent neural responses to sucrose-paired ketone and terpene, control ketone and terpene, and airpuff-paired ketone and terpene (SK, ST, XK, XT, PK, PT). (B) Average Z-scored activity of each cluster to each of the 6 odors on day 6 of imaging. Yellow bar indicates 2-seconds of odor exposure. (C) The distribution of clusters by population. (D) Percentage of total neurons that were significantly excited or inhibited by each odor (Bonferroni-adjusted FDR < 0.05) as a function of time relative to odor. Lines represent the mean across biological replicates and the shaded area reflects the mean ± SEM. (E) Bar graph showing % of neurons from each population that are responsive to both sucrose-paired odors in the same direction (left), responsive to only a single odor (middle), or responsive to at least 3 odors (right). Bars represent the mean across biological replicates and x’s mark individual animals. (F) Scatterplot comparing the magnitudes of SK responses (ΔΔSK) to ST responses (ΔΔST). The dotted line represents the hypothetical scenario where ΔΔSK = ΔΔST. For each population, the R2 value of the 2-d distribution compared to the ΔΔSK = ΔΔST line is reported. (G) Same as F but comparing ΔΔSK to ΔΔXK. (H) Lineplot showing the % of neurons from each population where the difference between ΔΔSK and ΔΔXK is lower than that between ΔΔSK and ΔΔST. (I) Bargraph showing % of neurons whose responses to {SK vs. XK} can be discriminated by a linear classifier with auROC>0.75. (J) Same as (I) but for {SK vs PK}. (K) Same as (I) but for {SK vs ST}. (L) Schematic representation of 4 possible categories for a joint-distribution of {SK vs. XK} and {SK vs. ST} auROC values. Identity-encoding neurons could be in any quadrant other than the bottom-left whereas valence-encoding neurons should be in the bottom-right quadrant. (M) Scatterplot of each neuron’s auROC value for {SK vs. XK} on the x-axis and {SK vs. ST} on the y-axis on days 1, 3 and 6 of imaging. (N) Stacked bar graph showing the distribution of neurons from each population that fall into each of the 4 quadrants across the 3 different imaging days. FWER-adjusted statistical significance for post hoc comparisons are shown as: ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05. See Tables S6-17 for detailed statistics.

To assess this hypothesis, we quantified the number of neurons that had statistically significant responses to each of the 6 odors on the last day of imaging. We found that more VP neurons were either excited (29.8±4.1%, 36.6±4.0% for SK, ST) or inhibited (24.5±3.0%, 29.4±3.8% for SK, ST) to either sucrose-paired odor than to control or puff-paired odors (7.6-11.1% excited, 8.1-12.9% inhibited) (Fig3D, FigS8A-B, TableS37). When compared across days, we found that the percentage of VP neurons that respond to both S odors increases from 6.1±2.2% on day 1 to 34.1±5.1% by the 6th day of imaging (Fig3E, TableS11). By comparison, the percentage of OT neurons that respond to both S odors in the same direction (i.e. excited by both S odors or inhibited by both S odors) did not increase through training. Furthermore, whereas OTD1 and OTD2 neurons were more likely to respond to a single odor than they were to respond to both S odors (12.6 vs 31.3% in OTD1, 11.8 vs 21.7% in OTD2), VP neurons were more likely to respond to both S odors than to a single odor (34.1 vs 23.3%).

Similarly, we found that the magnitude of trial-averaged odor responses in the VP were significantly higher for S odors than X or P odors on the last day of imaging (FigS9, TableS38). By comparison, neither sucrose-pairing nor airpuff-pairing had any impact on the magnitude of odor responses in OTD2 neurons on day 6. And though we did observe a significant effect of sucrose-pairing on response magnitudes in OTD1 neurons, both the effect size and significance were weaker than observed in VP. We propose that an ideal valence-encoding neuron should respond similarly to 2 odors of equal reward-contingency but disparate molecular structure, and we looked at the correlation between each neuron’s response to the sucrose-paired ketone (SK) and to the sucrose-paired terpene (ST). VP neurons had a high correlation between a neuron’s responses to SK and ST (Fig3F; R2=0.89). This similarity was much higher than between the sucrose-paired ketone (SK) and the control ketone (XK) despite the greater structural similarity between SK and XK (Fig3G; R2=0.33). In contrast, for both OTD1 and OTD2 neurons, there was a higher correlation between responses to similar molecular structure (Sk and XK,) than between responses to similar contingency (SK and ST) (OTD2: SK vs. ST R2=0.04, SK vs. XK R2=0.58; OTD1: SK vs. ST R2=0.13, SK vs. XK R2=0.40). Moreover, most VP neurons (76.5%), had a smaller absolute difference in the response magnitude to the 2 S odors (|SK-ST|) than the absolute difference between the sucrose-paired ketone and the control ketone (|SK-XK|) (Fig3H). By comparison, only half of OTD2 and OTD1 neurons showed smaller |SK-ST| than |SK-XK|, as would be expected if response magnitude to an odor did not depend on reward-contingency. This trend was not due to the fact that VP neurons were more likely to respond to both S odors than OT neurons were since it was consistent across various thresholds for odor response magnitude. This trend was consistent for other pairwise odor comparisons where one odor was a sucrose-cue and the other was not (e.g. SK vs. PT, FigS10A-B).

Finally, we reasoned that the activity of reward-contingency encoding neurons would support good decoding of odor pairs which have different valence but not of odor pairs that have the same valence. To do this, we trained binary logistic classifiers from each neuron’s response to all 15 odor pairs and quantified the area under their receiver operating characteristic (auROC). Because auROC values were non-normal with large spread, we quantified what percentage of neurons had an auROC of at least 0.75, halfway between ideal and at-chance decoding. We also note that all classifiers with auROC>0.75 showed bootstrapped p-values less than 10-3 (FigS10C-D). To assess whether neurons from each region were encoding valence, we compared a neuron’s {SK vs. XK} decoder performance (intervalence classification) against its {SK vs. ST} decoder performance (intravalence classification) (Fig3I-K, FigS10E-F). Across multiple days of imaging, we found that the percentage of neurons that support intervalence classification increased regardless of region but that this effect was markedly more pronounced among VP neurons than among OTD1 or OTD2 neurons (Fig3I-J, TableS12-S15, FigS10F, TableS39-S41). Intravalence classification, however, did not depend on days or region (Fig3K, TableS16-S17, FigS10F, TableS42-S44). By day 6, there were thrice as many VP neurons with good intervalence decoding than with intravalence decoding (51.8±5.0% vs. 14.4±5.8% for {SK vs XK} and {SK vs ST}, respectively). In contrast, a similar number of OT neurons displayed good intervalence decoding as did intravalence decoding (20.8% vs 19.9% of OTD1; 12.8% and 21.0% of OTD2 for {SK vs XK} and {SK vs ST}, respectively). The pattern of better intervalence decoding than intravalence decoding among VP neurons was observed across all 15 pairwise classifiers (FigS10H). Whereas 10.2% of all day 6 VP neurons had auROC>0.75 for {SK vs. ST}, 46.9-57.8% had auROC>0.75 for any classification between a sucrose-cue and a control odor or airpuff-cue. By comparison, there were few neurons with auROC>0.75 for any classification between a puff-cue and a control odor (2.3-10.9%), suggesting that negative valence is either not encoded in these VP neurons or the negative valence was not learned.

Plotting a neuron’s {SK vs. ST} auROC against its {SK vs. XK} auROC, we can categorize a neuron into the 4 categories (Fig3L). 1) a valence encoding neuron ({SK vs. ST}<0.75 and {SK vs. XK}>0.75), 2) an identity encoding neuron (both auROC>0.75), 3) an identity encoding neuron that does better with S odors ({SK vs. ST}>0.75 and {SK vs. XK}<0.75), and 4) an uninformative neuron (both auROC<0.75). According to this categorization, half of VP neurons were valence encoding by day 6, followed by OTD1 then OTD2 (Fig3M-N; 47.7, 16.2, 7.3% for VP, OTD1, and OTD2, respectively). The opposite was true for identity encoding. VP had a smaller percentage of identity encoding neurons than either OTD1 or OTD2 (14.8, 21.1, 22.9% for VP, OTD1, and OTD2, respectively). We note that these conclusions can also be replicated when analyzing multinomial regression (MNR) classifiers trained on single neuron activities (FigS11F-G, TableS50-S53). Namely, the rates of confusion between the 2 sucrose cues are highest in VP and lowest in OTD2 whereas the rates of confusion across all ketones (SK, XK, PK) are highest in OTD2 and lowest in VP. These single-neuron classifier analyses further indicate that VP neurons, more than either OTD2 or OTD1 neurons, were encoding reward contingency at the single neuron level. However, the most striking observation was that while only a subset (37.5%) of VP neurons had auROC<0.75 for both {SK vs. XK} and {SK vs. XK}, a majority of OTD2 and OTD1 neurons (69.7% and 62.6%, respectively) showed auROC<0.75 for both {SK vs. ST} or {SK vs. PK}. Thus, in comparison to the VP, most individual OT neurons have little discriminatory information about olfactory stimuli regardless of valence at the single-neuron level and may be better suited in a population code.

Our data indicated that valence encoding emerges in VP neurons over the course of learning. To explore the potential mechanisms at the cellular level, we compared the activity of a subset of neurons we could observe on both day 1 and day 3 (Fig4A-F). We noticed there were neurons that responded to the sucrose delivery on day 1 that responded to the sucrose cue on day 3 (Fig4C,F), reminiscent of models of Hebbian plasticity. When quantified, we found that 17.9, 20.9% of VP neurons were responsive to sucrose on day 1 and SK and ST on day 3, respectively (Fig4G). We specifically considered neurons that had the same direction of response (excitation or inhibition) to both cues on separate days. This figure was much lower among OT subpopulations (11.5, 8.2% for SK and ST in OTD1; 10, 2.5% for SK and ST in OTD2). Consistent with above observations, we also found that the odor responses to sucrose-cues were larger on day 3 than day 1 in 85% of tracked VP neurons, but only in 65% and 57% of OTD1 and OTD2 neurons, respectively (Fig4H-I, TableS18-S19). We did not see the same effect in VP neurons’ responses to control or puff-paired odors. Together, our data suggest that sucrose pairing causes sucrose-responsive VP neurons to increase their responses to the sucrose-predictive odors.

Sucrose responsive VP neurons become sucrose-cue responsive after pairing.

(A) The spatial footprints of 15 neurons from day 1 are outlined over a max-correlation projection image. (B) Heatmap of averaged-over-trials ΔF/F in response to 6 odors on day 1. Odor delivery period is shown with 2 red vertical lines and sucrose/airpuff timing is shown with downward arrowhead. (C) An example neuron’s responses on day 1 across 30 trials to 6 different odors. Individual trial traces are shown in light gray whereas the averaged-across trials trace is shown in black. Odor delivery period is depicted as shaded rectangles and US delivery is marked by arrowheads. (D-F) Same as (A-C), respectively, but for day 3. (G) Percentage of all tracked neurons that were both sucrose-responsive on day 1 and odor-responsive in the same direction on day 3. (H) Scatter plot of averaged-over-trials responses to SK or ST on day 1 (x-axis) and day 3 (y-axis). Each point is a neuron that was successfully matched from day 1 and day 3. Neurons from OTD2, OTD1, and VP are plotted as pink circles, blue crosses, and yellow squares, respectively. Neurons that have increased response magnitudes on day 3 would fall between the 2 dotted lines. (I) Violin plot showing the distributions of day 3 responsive magnitude – day 1 response magnitude. Black asterisks show statistical significance of pairwise comparisons and red asterisks show statistical significance for one-sample t-tests. Pairwise comparisons were done using the Student’s t-test. The p-values were corrected for FDR by Benjamini-Hocherg procedure. ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05. See Tables S18-19 for detailed statistics.

Olfactory brain areas are known to use population codes to encode sensory information, whereby single neurons have weak discriminatory information, but the activity of the population allows for an efficient encoding of high-dimensional data. To assess if there is discriminatory information about the odorants within the population-level activity, we compared the pairwise Euclidean distance of trial-averaged odor responses for all 15 odor pairs (Fig5A,B). We saw that, in general, the pairwise Euclidean distance for all odor pairs examined increases quickly after the onset of odor, reaches peak distance towards the end of the 2 second odor delivery, and slowly decays after odor ends (Fig5A). When examining the average pairwise distance during the last second of odor, there was a relatively unstructured distribution of pairwise distance in OTD2 odor-response such that ||SK-XK||, ||SK-PK||, ||SK-ST||, and ||XK-XT|| were all similar (Fig5B). By comparison, in VP populations, the distribution was structured such that intervalence pairwise comparisons between sucrose-paired and not sucrose-paired odors (e.g. ||SK-PK|| and ||SK-XK||) were larger than intravalence pairwise comparisons (e.g. ||SK-ST||, or ||XK-XT||). OTD1 populations showed an intermediate trend where most intravalence pairwise distances were smaller than intervalence pairwise distances with the exception of ||SK-ST||. Thus, at the population level VP representations appear to encode valence but not identity, whereas OT representations encode some valence information but appear to be better suited for identity encoding.

OT encodes odor identity in high-dimensional space and VP encodes reward-contingency in low-dimensional space.

(A) Average normalized pairwise Euclidean distance between odor-evoked population-level activity from day 6 of imaging shown as a function of time relative to odor delivery. Traces show the average value across biological replicates of the same population and the shaded areas represent the average ± SEM. (B) A heatmap of the average normalized pairwise distance during the odor delivery period. (C) Average CV accuracy of binary pairwise linear classifiers trained on population data plotted against time relative to odor delivery. (D). A heatmap of the average CV accuracy during the odor delivery period. (E) Schematic representation of generalized linear classification performance for an idealized valence encoder. Each row corresponds to the training odor-pair and each column corresponds to the testing odor-pair. For an idealized valence encoder, the decodability would generalize well across odor-pairs of the equal valence grouping outlined in red. Note that the elements along the diagonal are cases where training and testing odor-pairs are identical and do not reflect generalizability. (F) Heatmap representing the maximum generalized linear classification accuracy during odor delivery period averaged across biological replicates for each population. (G) Mean cross-validated linear classifier accuracy for S-cue vs. control or puff-cue classification and the generalized accuracy for S-cue vs. control or puff-cue classification after training on a different pair. Bar represents the mean across biological replicates and x’s mark accuracy values for individual animals. (H) Average PR normalized to n calculated after randomly subsampling an increasing number of neurons. (I) Average PR calculated after subsampling 15 neurons. (J) Average CV accuracy of linear classifiers trained on {SK vs. PK} plotted against number of principal components used for training. For each simultaneously imaged group of neurons, 15 neurons were subsampled and classifiers were trained on an increasing number of principal components. Thinner faded lines show mean accuracy across subsampling for individual animals. Markers represent the mean across biological replicates. Error bars indicate SEM across biological replicates. (K) Average CV accuracy of linear classifiers trained on {SK vs. ST}. (L) Comparison of the average accuracy of {SK vs. PK} classifiers trained on the 1st PC vs. {SK vs. ST} classifiers trained on all 15 PC’s. FWER-adjusted statistical significance for post hoc comparisons are shown as: ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05. See Tables S20-29 for detailed statistics.

In parallel, we also performed decoding analysis using linear classifiers to assess how reliably a given pair of odors could be decoded from population-level activity (Fig5C-D). To quantify this, we extracted the average ΔFi,k/F values for each trial i 𝜖 [1,m] and each neuron k 𝜖 [1,n]. The resulting matrix of size m x n was used to train a binary linear classifier with a logistic learner. For each classifier, we looked at the average accuracy across 5-fold cross-validation (CV accuracy). Classifiers were trained on simultaneously recorded populations (i.e. neurons from the same animal recorded on the same day) to capture biological variability. A total of 765 pairwise linear classifiers were trained (15 pairwise comparisons, 17 animals, and 3 days). When compared against 10,000 shuffles, 569 of these classifiers showed bootstrapped p-value less than 0.001 (FigS12A). Importantly, all classifiers with CV accuracy higher than 0.75 had p-value less than 0.001.

Linear classifiers trained on day 6 OTD2 population data had similar ranges of accuracy regardless of valence (Fig5D). For example, the intravalence classification {SK vs. ST} was more accurate (86.6±3.9%) than some and intervalence classifications (e.g. {SK vs. XK}, 72.8±5.4%) but less accurate than others (e.g. {ST vs. PK}, 88.2±3.1%). Classifiers trained on VP population activity, however, always showed more accurate intervalence decoding (range: 89.5-96.1%) than intravalence decoding ({SK vs. ST}, 79.9±6.1%). Additionally, whereas OTD2 population classifiers could decode the 2 control cues {XK vs. XT} at accuracy (85.8±4.2%) comparable to sucrose-cue vs. non-sucrose-cue, VP population classifiers were consistently less accurate (76.8±3.7%) at {XK vs. XT} than the aforementioned intervalence classifiers. This suggests that whereas OTD2 encodes odor identity agnostic to the valence, VP does not encode identity at all but rather encodes reward contingency or positive valence. OTD1 pairwise classification was a mixture of the other 2 regions: sucrose-cue vs. non-sucrose-cue classification was more accurate than most other pairwise classifications (range: 86.4-94.3%), but the {SK vs. ST} classification was comparably accurate (90.9±4.7%). This rules out the interpretation that OTD1 strictly encodes valence since the identity of 2 sucrose-cues can be decoded well.

To address the possibility that our results are due to the limitations of linear classification, we repeated the analysis using support vector machines (SVM’s) with a radial basis function kernel and found we could draw the same conclusions (FigS12E). Similarly, to verify our results are not epiphenomena of forcing the data into binary classification, we looked at population-level MNR classifiers trained on day 6 data. Importantly, we observe high confusion between 2 sucrose cues in MNR classifiers trained on VP data, but not those trained on OTD2 or OTD1 data (FigS12F), corroborating through an alternate analysis method that VP population activity encodes reward contingency whereas either OT subpopulations are better at encoding identity.

The fact that VP populations showed higher decoding for odor pairs of unequal sucrose-contingency provides strong evidence that VP encodes reward-contingency more than identity. Results from OT decoder analyses, however, are less intelligible: all 15 odor pairs, regardless of sucrose-contingency, could be decoded with above-chance success. Though this result is consistent with OT populations encoding identity rather than valence, it does not rule out the possibility that valence and identity are both encoded. In the context of cue-association, 2 cues of different valence cannot have the same identity, meaning that good decoding of {SK vs. PK} can be extracted from either valence encoding or identity encoding populations. To disambiguate these 2 possibilities, we looked at the generalizability of pairwise decoders. Briefly, linear classifiers were trained on each of the 15 possible odor pairs. Afterwards, the resulting classifier was tested on every other odor pair (Fig5E). We reasoned that if neural populations encode valence in addition to identity, classifiers trained on any odor pair of unequal sucrose-contingency should consistently perform above chance on a different odor pair of unequal sucrose-contingency (e.g. train on {SK vs. PK}, test on {ST vs. XT}). In other words, given valence encoding, {SK vs. PK} should be discriminable in a way that can also discriminate {ST vs. XT}. As expected, VP population decoders were consistently generalizable when trained on odor pairs of unequal sucrose-contingency then tested on other odor pairs of unequal sucrose-contingency (Fig5F,G). OTD2 population decoders, on the other hand, showed negligible generalizability across pairs of unequal sucrose-contingency. Similarly to other metrics of valence encoding, we found that OTD1 displayed a generalizability in between that of VP and OTD1, suggesting that OTD1 could encode some valence in addition to identity. However, we note that the VP population, on average, outperforms OTD1 at generalized valence decoding (95.0±2.0% vs 78.5±3.9%; TableS22-S23).

After performing these population-level analyses, we noticed a discrepancy: although single-neuron intervalence decoding was worse in OT than in VP (Fig3M-N), population-level intervalence decoding was comparable between either OT subpopulations and the VP (Fig5C-D). This led us to speculate that the encoding of odor information had a higher dimensionality in OT than in VP. To explicitly compare the dimensionality of VP and OT population activities, we looked at the extent to which the population vector is spread across multiple axes using principal component analysis (PCA). Dimensionality can further be quantified using the participation ratio (PR) of a population, which is the square of the sum of eigenvalues of its covariance matrix divided by the sum of the squares of its eigenvalues (Litwin-Kumar et al., 2017; Recanatesi et al., 2019). This value will have a range of 1 to n, where n is the total number of features. If a single principal component can describe all of the total population variance (i.e. the data is low-dimensional), the population will have PR equal to 1. Conversely, if every principal component equally describes nth of the total variance (i.e. the data is high-dimensional), the population will have PR equal to n. Because the number of total neurons recorded was different between OT and VP experiments, we first assessed if and how the normalized PR would vary with the number of total neurons through random sampling (Fig5H). After observing a consistent decrease in PR with increasing n, we compared the PR of OT and VP animals by repeatedly subsampling a fixed number of neurons (k=15) and found that VP animals had lower PR (PRVP =5.83±0.80) than either OTD2 (PRD2=9.61±0.37) or OTD1 (PRD1=9.24±0.44) animals after training (Fig5I, TableS24-S25). There was also a difference, however, in how valence information vs. identity information was encoded by VP populations. Though the first PC of each VP population was sufficient to train adequate {SK vs. PK} decoders (CV accuracyPC1 = 85.5±2.7%), all 15 PC’s were required for comparable {SK vs. ST} decoding (CV accuracyPC1:15 = 75.1±13.4%) (Fig5J-L, TableS26-S29). In either OT populations, the first PC did not support good decoding of either {SK vs. PK} or {SK vs. ST}. Together, our population-level analysis indicates that VP encodes valence, but not identity, in low-dimensional space, OTD2 encodes identity but not valence in high-dimensional space, and OTD1, has some valence information and encodes identity in high-dimensional space.

Analyses at the single-neuron and population levels showed that VP activity encodes reward contingency, rather than the identity, of the olfactory stimulus. However, due to the task design, the reward-contingency of a stimulus was highly correlated with the vigor of licking (Fig2F). This raised concerns that some neurons classified as robust reward-contingency encoders were potentially encoding motor-related information. Indeed, many VP neurons showed consistent increases in fluorescence time-locked to the onset of a licking-bout (FigS13A-B), and could be used to train distributed lag models to predict onset of licking bouts (FigS13C). Across all VP neurons, we observed a positive and significant correlation between a neuron’s valence decoding ability and licking decoding ability (FigS13D; slope=0.41, p=2.2x10- 10, R2=0.28). This motivated us to develop a new conditioning paradigm that could decouple reward-contingency of an odor cue from the behavioral output. Initially, we attempted to train animals on a symmetric Go/No-Go operant task where reward delivery was contingent on licking or withholding licks during odor. However, consistent with previous findings (Gubner et al., 2010), we found that animals struggled to learn the No-Go behavior in comparison to the Go behavior (data not shown). In an operant paradigm, this leads to a problematic difference in valence of Go/No-Go cues. Consequently, we opted to develop a classical conditioning paradigm whereby licks were encouraged/discouraged by physically moving the lick spout before odor presentation (Fig6A-B).

Separate VP populations encode reward-contingency and licking vigor.

(A) State diagram for odor pairing paradigm where lick spout is removed during the presentation of half of the odors. The paradigm is similar to one described in Fig2A with the following key differences: 1) the lick spout is moved away from the animal’s mouth during the presentation of half of the odors (Nhi, Nlo, NX). 2) sucrose is delivered after a longer variable delay (1.1-1.3s). 3) 2 of the odors have 100% sucrose contingency (Lhi, Nhi), 2 of the odors have 50% sucrose contingency (Llo, Nlo), and the other 2 have 0% sucrose contingency (LX, NX). (B) Schematic showing the timing of lick port movement relative to odor and sucrose delivery. (C) Licking behavior to 6 odors averaged across 30 trials from a representative animal. Duration of odor delivery is marked by the shaded rectangle and the average time of sucrose delivery is marked by the arrowhead. The time bin used for subsequent analysis (last 0.5s of odor and first 0.5s of delay) is outlined by square brackets (D) Average licks/s for each odor measured between the last 0.5s of odor and the first 0.5s of delay. Data were pooled from the day of highest difference between licks to Lhi and Nhi. (E) Heatmap of odor-evoked activity in VP neurons pooled from each animal’s day of highest difference between licks to Lhi and Nhi. Neurons are grouped according to the clustering dendrogram, shown on the right. Horizontal white lines demarcate the boundaries between the 3 clusters. Odor delivery is marked by vertical red lines. (F) Average Z-scored activity of each cluster to each of the 6 odors. Yellow bar indicates 2-seconds of odor exposure. (G) The percentage of single-neuron linear classifiers with auROC>0.75 as a function of time relative to odor delivery. Shaded area represents the SEM across biological replicates (n=5). (H) Heatmap of the percentage of pooled VP neurons with auROC>0.75 during the last 0.5s of odor and first 0.5s of delay. (I) Scatterplot comparing the auROC for {Lhi vs Nhi} (y-axis) and {Nhi vs. NX} (x-axis) for each neuron. The line of best fit is plotted as a dotted line, with the 95% confidence interval shaded in. (J) Same as (I) but comparing the auROC for {Lhi vs LX} (y-axis) and {Nhi vs. NX} (x-axis). (K) Scatterplot comparing regression models that explain each neuron’s activity on a given trial as a function of anticipatory licking or sucrose contingency. The values plotted are the loss in R2 in models without anticipatory licking (y-axis) or sucrose contingency (x-axis) when compared to a model with both variables and their interaction term. (L) CV accuracy for 5 different odor pairs as a function of time relative to odor delivery. (M) Heatmap of average pairwise CV accuracy trained on the last 0.5s of odor and the first 0.5s of delay. (N) Scatterplot of all pairwise classifier accuracies from all animals (y-axis) and the corresponding range-normalized average pairwise difference in anticipatory licking (x-axis). (O) Scatterplot of all pairwise classifier accuracies from all animals (y-axis) and the corresponding pairwise difference in reward-contingency (x-axis). (P) Scatterplot of all pairwise classifier accuracies (y-axis) and the adjusted combined model of ranged-normalized Δlick and Δreward-contingency (x-axis). FWER-adjusted statistical significance for post hoc comparisons are shown as: ***p<0.001, **p<0.01, *p<0.05, n.s. p>0.05. See Tables S30-31 for detailed statistics.

Briefly, headfixed animals were presented with 1 of 6 odors in pseudorandomized order. During the presentation of 3 of these odors, the lick spout was moved away from the mouse with a linear stepper motor. These odors are denoted as N odors (N for No-lick spout). During the presentation of the other 3 odors, the lick spout remained within licking distance of the mouse’s tongue. These odors are referred to as L odors (L for lick spout). 1 odor from each group served as a control odor that had 0% reward-contingency (LX, NX). The other 2 odors in each group were paired with sucrose at low (50%) or high (100%) probability (Llo, Lhi, Nlo, Nhi). We reasoned that this contingency could allow us to make pairwise comparisons where one odor has a higher value but lower anticipatory licking than the other (e.g. Nhi vs. Llo). To monitor anticipatory licking in the absence of the lick spout, we trained a distributed lag model (DLM) using features of the mouse’s face tracked using DeepLabcut (Mathis et al., 2018) (FigS13E-G). We chose to pool data across all mice from the day of the highest licking differential between Lhi and Nhi odors (FigS13H) to maximize the decoupling of value and motor output in our analysis. Anticipatory licking during Llo or Lhi began during the last second of the odor and increased gradually until sucrose delivery whereas licking during Nlo or Nhi was delayed by about one second (Fig6C). When quantified across animals on their days of highest lick differential, we found that mice consistently licked most during Lhi, followed by Llo and Nhi, then Nlo (Fig6D, TableS30-S31). Mice showed little to no licking during either control odors. Thus, this behavioral assay affords us the opportunity to assess the decoupled effects of reward-contingency and licking vigor on neural activity.

To begin to characterize the presence of reward-contingency and/or licking vigor encoding in the VP, we first pooled and clustered the neural activity taken from 5 animals on their days of highest lick-differential (Fig6E). When clustering VP neurons into 3 clusters, we found that one cluster (I) showed a largely similar inhibitory response to the 4 sucrose-paired odors, but not control odors-much like cluster (I) from the previous conditioning experiment (Fig3A-B, Fig6E-F). Another cluster (III) by comparison, showed a varied excitatory response to each of the 4 sucrose-paired odors, much like cluster III from the previous experiment. Cluster (III) neurons seemed to have a particularly strong response to Lhi for which there was most anticipatory licking. This led us to speculate the existence of both reward-contingency encoding and vigor encoding neurons in the VP.

To test this directly, we quantified single neuron decodability of odor pairs and examined how correlated decoding along the reward-contingency axis is to decoding along vigor axis (Fig6G-H). We reasoned that auROC values for {Lhi vs. Nhi} would be high for vigor encoding neurons but not value encoding neurons given these 2 odors have the same reward-contingency but disparate licking behaviors. Similarly, we reasoned that auROC values for {Nhi vs. NX} would be high for reward-contingency encoding neurons but not vigor encoding neurons given there is a large difference in value but small difference in licking between these 2 odors. First, we saw that while single neuron decodability along the reward-contingency axis (e.g. {Nhi vs NX}) was higher than along the lick axis (e.g. {Lhi vs Nhi}), there were more neurons that could decode {Nhi vs NX} than could decode {LX vs NX} at auROC>0.75 (Fig6G-H). Furthermore, we saw a lack of significant correlation between the single-neuron decodability of 2 odors that had similar licking but different reward-contingency ({Nhi vs NX}) and the decodability of 2 odors that had different licking but same reward-contingency ({Lhi vs. Nhi}) (Fig6I; slope=0.038, p=0.61, R2=0.0039). This decoupling suggests that reward-contingency and vigor information are both encoded in the VP but by different populations. As a control, we saw a significant correlation between 2 pairwise comparisons that both had high difference in reward-contingency ({Nhi vs NX} and {Lhi vs LX}) (Fig6J; slope=0.60, p=3.8x10-10, R2=0.30).

We also performed the converse experiment where the ΔΔF/Fbaseline values of each neuron were linearly fitted to either 1) the reward contingency, 2) the anticipatory licking or 3) both values and the interaction term. Then, we compared the ΔR2 when either variable was omitted in the model and plotted the ΔR2-valence against the ΔR2-licking (Fig6K). We reasoned that, if a typical VP neuron’s activity could be well-explained by either reward-contingency or vigor but not both, we would see points along either x or y-intercepts. On the other hand, if a typical neuron’s activity could be well-explained by a linear combination of the 2 variables, we would see data fall along a line of positive slope. We found that most neurons tended to have large ΔR2-valence or large ΔR2-licking values but not both, supporting the idea that 2 largely non-overlapping sets of VP neurons encode reward-contingency or vigor but not both.

Lastly, we trained linear classifiers of pairwise odor comparisons using population-level activity to assess if both reward-contingency and vigor information were present in the population-level activity. Consistent with single-neuron decoder analysis, we found that {Lhi vs LX} and {Nhi vs NX} could both be decoded better than {LX vs NX} (Fig6L-M). Because we train each classifier using simultaneously recorded neural activity (i.e. from a single animal), we had a total of 75 classifiers (15 pairwise classifiers for 5 animals). The cross validated accuracies of these classifiers were then fitted to a linear model of pairwise differences in either 1) reward-contingency, 2) anticipatory licking, or 3) both. If the population VP activity encodes either reward-contingency or vigor but not both, we expect to see one of the single-variable models outperform the other greatly. But if the population VP activity encodes both variables, we expect the multivariable model would outperform either single model. We found that Δlicking has a weak and not significant relationship with pairwise CV accuracy (Fig6N; slope=0.076, p=0.079, R2=0.042). By comparison Δreward-contingency (or P(S), as in probability of sucrose delivery following odor) had a larger and statistically significant correlation with pairwise accuracy (Fig6O; slope=0.15, p=3.7x10-4, R2=0.16). The combined model, however, showed larger coefficients and larger R2 than either single variable model, suggesting an additive effect of both features on CV accuracy (Fig6P; accuracy = 0.18Δlick + 0.24ΔP(S) - 0.22Δlick*ΔP(S) +0.65, R2=0.23). Thus, we conclude that both reward-contingency and licking vigor are encoded in the population-level activity of VP neurons.

Discussion

Our anatomical investigations demonstrate that the primary output of the OT is to the VP, with minimal connections to the VTA. Given its constrained connectivity, we propose that the OT to VP circuit is an ideal model system for examining how the encoding of reward cues is transformed across brain circuits. Utilizing comparative longitudinal imaging, we found that VP, but not OTD2, robustly encodes the sucrose-contingency of odors. Although our analyses revealed that sucrose-contingency influences odor-evoked responses in OTD1 neurons more so than in OTD2 neurons, other evidence suggests valence encoding is not the appropriate framework for interpreting OTD1 activity. Specifically, information about sucrose-contingency in OTD1 resides in a high-dimensional space and generalizes poorly, whereas VP encodes reward-contingency robustly in a low-dimensional and generalizable manner. Thus, we suggest that the changes in OTD1 activity are more likely to reflect increased contrast of identity or an intermediate and multiplexed encoding of valence and identity. Finally, using a novel classical conditioning paradigm, we assigned motor-related signals and expected-value signals to non-overlapping VP subpopulations.

Some of our findings were unexpected. For example, we found no evidence that either OTD1 or OTD2 have significant extrapallidal outputs. This is in contrast to a previous study which reported that OTD1 neurons, and to a lesser extent, OTD2 neurons, project to the LH and VTA (Zhang et al., 2017b). It is possible that other parts of the OT have extrapallidal outputs, as we only performed anterograde tracing from the anteromedial portion. It is also possible that at least some of the VTA labeling Zhang and colleagues observed from anterograde viral tracing experiments could be due to backflow of the tracer virus in nuclei immediately dorsal to the OT (e.g. AcbSh). As a critical control, we provide evidence that retrograde tracing from VTA robustly labels AcbSh neurons but hardly any OT neurons. And the few VTA projecting OT neurons we did observe were restricted to the distal portions of layer III bordering the VP. Consistent with this, quantification of OT afferents is glaringly absent from 2 independent characterizations of brainwide inputs onto VTA (Beier et al., 2015; Faget et al., 2016). In contrast, OT has been reported to be one of the most prominent inputs to both GAD2+ and Vglut2+ VP neurons (Stephenson-Jones et al., 2020). It is difficult, however, to completely rule out the existence of OT 𝑡𝑜 midbrain projections due to the limitations of our experiments: we primarily targeted layer II in the anteromedial portion of the OT for anterograde tracing and only tested the VTA with retrograde tracers. More posterior and/or lateral portions of the OT could have extrapallidal outputs posterior to the VTA. Despite these caveats, the evidence suggests that Drd1+ neurons in the anteromedial portion of the OT have little extrapallidal projections when compared to the AcbSh.

Though we found little difference in the output patterns of OTD1 and OTD2 neurons, we observed differences in how these 2 subpopulations encode odor valence. Consistent with a previous report (Martiros et al., 2022), we found that OTD1 activity, more than OTD2 activity, is modulated by reward contingency. For example, OTD1 neurons, but not OTD2 neurons, were more likely to respond to sucrose-paired odors than other odors. And the magnitude of responses in OTD1 but not OTD2 neurons were significantly larger to sucrose-paired odors than to other odors. We refrain, however, from concluding that the primary feature encoded in OTD1 neurons is valence or reward contingency, for the following reasons. First, the above-mentioned effects of sucrose-contingency on neural activity are much stronger for VP than for OTD1. Additionally, whereas more than 50% of VP neurons could be categorized as reward-contingency encoders, this figure was less than 20% for OTD1. Lastly, population-level decoders trained on odor pairs of different valence can generalize in the case of VP populations, but not OTD1 populations. While we acknowledge that there is poor standardization when it comes to defining valence encoding, it is unlikely that discrepancies between our conclusions and those of Martiros et al. stem from differences in interpretation alone. Comparative examination of our analyses reveals clear dissimilarities in the effect-size of shared metrics (e.g. % odor responsive). Given the high Z-resolution afforded by 2-photon microscopy, it is probable that we recorded from different layers of the OT, which should not be assumed to have identical physiology. We note that the lens placements in our experiments are considerably more ventral than those reported in Martiros et al. It is possible that these neurons are recorded from layer III of the OT whereas the majority of the neurons in the present study are recorded from layer II. A direct comparison of layer II and layer III OT neurons and their valence encoding could prove useful in understanding the discrepancies between the two studies. It is also possible that some of the neurons recorded in Martiros et al. could be from the rostral portion of the VP which lies immediately dorsal to layer III of the OT. Although Adora2a and Drd1 are not expressed as mRNA in the VP, the BAC-transgenic lines used for both the present work and work by Martiros et al. labels neurons in the VP.

Our comparison of OT and VP is reminiscent of previous comparisons made between value encoding in VP and NAc (Ottenheimer et al., 2018; Richard et al., 2016). These publications showed that VP encodes incentive value more robustly than the NAc. Given that OT and NAc share many anatomical, physiological, and molecular traits, it is tempting to speculate that the encoding schemes, too, would be similar between the 2 areas. Optogenetic activation of OTD1 supports RTPP (Murata et al., 2019), as does activation of D1 or D2 neurons of the NAc (Soares-Cunha et al., 2020). While we acknowledge stimulation experiments provide unique insights that cannot be obtained from recordings alone, we note that SPN’s have extensive inhibitory collaterals and exhibit high-dimensional activity. Given these peculiarities of the striatum, we predict that bulk stimulation leads to activity patterns well outside the physiologically relevant range and that this warrants conservative extrapolations regarding OT SPN’s endogenous role.

An exciting conclusion from our work is that, within the context of our conditioning paradigm, the dimensionality of neural activity was much lower in VP than in OT. Furthermore, the dimensionality of the imaged subpopulations were anti-correlated with the robustness of sucrose-contingency encoding: OTD2 displayed the highest dimensionality and lowest valence encoding whereas VP displayed the lowest dimensionality and highest valence encoding. As discussed elegantly by others (Chu et al., 2016; Shannon, 1948), there is generally a tradeoff between the efficiency of a neural population (i.e. its total information capacity) and the robustness of its encoding scheme (i.e. redundancy of encoding). Consistently, it is likely that VP neurons display such robust encoding of valence, in large part, due to the loss of odor identity information. By comparison, OT populations may be able to encode information about the large olfactory identity space due to their high dimensionality. We speculate that the extensive inhibitory collaterals among SPN’s play a role in enforcing the high dimensionality of OT activity. Though it is entirely unknown what anatomical or physiological strategies are used to reduce VP dimensionality, we consider this an important piece of the puzzle in understanding VP computations.

We saw little evidence of negative valence neurons in any of the 3 populations that were imaged. This was surprising given previous reports of negative valence neurons in the VP (Stephenson-Jones et al., 2020). We consider 2 potential explanations for this discrepancy. First, it is possible that our conditioning paradigm was not sufficiently aversive for the animals. Although our behavioral evidence for aversive association is significant, it is less robust than sucrose association raising the possibility that the learning was insufficient. This could be due to the fact that we targeted the airpuff to the animal’s hindquarters rather than to the face. But we note that in a previous report, airpuff delivery to the snout and to hindquarters elicited similar ingress response in a burrowing assay (Fink et al., 2019). Additionally, we observed clear unconditioned responses to the airpuff itself. Another possibility is that, while negative valence neurons do exist in the VP, as has been reported, they were outside of our field-of-view. Previous work in the VP supports positive and negative valence as being encoded by Vgat+ and Vglut2+ neurons, respectively (Faget et al., 2018; Stephenson-Jones et al., 2020). Most Vglut2+ neurons are found in the dorsomedial portion of the VP, whereas our lenses were specifically targeted to the ventrolateral portion where we found the most OT afferents. Given this distinction, our results are not inconsistent with previous reports of negative valence neurons in the VP.

In this work, we present evidence that may appear to contradict previous anatomical and physiological characterizations of the OT. We find that the anteromedial portion of the OT sends high-dimensional information about odor identity primarily to the VP and not the VTA. By directly comparing OT and VP population-level activity in the same paradigm, we bridge together, for the first time, the fields of OT and VP. This provides valuable context which not only helps us evaluate past conclusions about valence encoding in the OT but also consider the implications of the stimulus-evoked activity in the OT. This comparative approach leads us to conclude that the OT has relatively little valence information. However, our findings are not generally inconsistent with what has been observed in previous studies. We do find reward modulation in the OTD1 population, however, we do not find valence encoding single neurons and the population vector does not generalize between two rewarded odors as it does in the VP. Therefore we propose that representation in the OT reflects either an intermediate representation of reward-contingency or a contrast modulation to reflect the contingency.

Speculation

It is interesting to note the discrepancy between the anatomical organization of dorsal striatum (DS) vs. ventral striatum (VS): SPN’s of the DS project exclusively to either the substantia nigra pars reticulata (SNr) of the midbrain (Drd1+) or the exterior portion of the globus pallidus (GPe) (Drd2+), but Drd1+ neurons in the VS (Acb) project to both the VTA of the midbrain and the ventral pallidum (Kupchik et al., 2015). The OT appears to have further limited output divergence, whereby both OTD1 and OTD2 neurons project primarily to the VP. This may reflect at a gradient of anatomical connectivity where the most dorsal Drd1+ SPN’s project primarily to the midbrain and the most ventral Drd1+ SPN’s (i.e. OTD1 neurons) project primarily to the pallidum. Functionally, the lack of evidence for OTD1 to midbrain connectivity challenges the dichotomy of direct vs. indirect pathways in the ventral basal ganglia. In this model, DA orchestrates motor initiation by oppositely modulating Drd1+ and Drd2+ SPN’s, which have differential downstream targets. Given the lack of clear differences in OTD1 and OTD2 projections, we think this canonical model of basal ganglia connectivity inadequately explains the functional consequences of DA modulation in the OT.

In our work, we described key differences in how reward cues are encoded in 2 synaptically connected nuclei. But what insights can we infer about the role of OT on shaping VP activity through this comparison? The most salient observation of VP activity is the large and widespread excitatory responses to sucrose-cues. Though the effect size is smaller, OTD1 neurons also showed larger excitatory responses to sucrose-cue when compared to other odors. Given that these neurons are GABAergic and their primary target are the VP neurons, it is difficult to explain how these 2 responses are related. We consider 3 possible explanations for this paradox. First, in addition to large excitatory responses that were specific to the sucrose-cues, we also observed inhibitory responses that were specific to the sucrose-cues. It is possible that the excitatory VP activity during sucrose-cue presentation is driven mainly by the numerous excitatory afferents (Pir, BLA, etc.) while the inhibitory VP activity is driven mainly by OTD1 and OTD2 afferents. In a second model, there could be mechanisms downstream of somatic activity that could explain the discrepancy. For example, though brief optical stimulation of D2 neurons in Acb leads to a decrease of VP activity, prolonged activation causes an increase in VP activity via the δ-opioid receptor (Soares-Cunha et al., 2020). Our experiments do not provide any information on how neuropeptide release from OT neurons is different during presentation of sucrose-cue vs. control odor. Similarly, we cannot measure if and how positively valent stimuli change the input-output-function of OT neurons. Previous reports have found that Drd2 agonism in Acb neurons leads to a decrease in collateral inhibition through a presynaptic mechanism (Dobbs et al., 2016). Given that more DA is expected to be released during presentation of sucrose-cues, it is plausible that the probability of GABA release from OT boutons onto VP dendrites is affected. In a third and perhaps the most parsimonious model, endogenous OT activity does not contribute significantly to explaining the bulk excitatory activity in VP. This goes against the prevailing working model in Acb to VP circuit which assumes that Acb excitation leads to VTA disinhibition by inhibiting the VP. And while there is evidence supporting from bulk stimulation of D1 or D2 neurons in the Acb (Soares-Cunha et al., 2020), under endogenous conditions, both Acb neurons and VP neurons are excited in response to reward-cues (Lederman et al., 2021; Ottenheimer et al., 2018). Furthermore, given that GABAergic synapses from SPN’s to VP neurons is likely dendritic (Bolam et al., 1986), we think it is unlikely that OT to VP drives large-scale shunting of action potential in the presence of excitatory drive from other areas known to respond preferentially to reward cues such as the BLA (Beyeler et al., 2018) or the OFC (Wang et al., 2020). We consequently propose an alternate framework in which the mechanistic role of the OT in this circuit is to provide spatiotemporally precise inhibition to coordinate the integration of excitatory inputs onto VP. This form of inhibition could gate which excitatory synapses go through Hebbian potentiation vs. anti-Hebbian depression. Under such a framework, OT would function as a high-dimensional filter for VP neurons to adaptively scale its various excitatory afferents.

Methods

Stereotaxic Surgery

All procedures were approved by the UCSD Institutional Animal Care and Use Committee. Animals were anesthetized with isoflurane (3% for induction, 1.5-2.0% afterward) and placed in a stereotaxic frame (Kopf Model 1900). Mouse blood oxygenation, heart rate and breathing were monitored throughout surgery, and body temperature was regulated using a heating pad (Physio Suite, Kent Scientific). A small craniotomy above the injection site was made using standard aseptic technique. Virus was injected with needles pulled from capillary glass (3-000-203-G/X, Drummond Scientific) at a flow rate of 2nl/s using a micropump (Nanoject III, Drummond Scientific). For OT anterograde tracing experiments, 50µl of AAV9-phSyn1-FLEX-tdTomato-T2A-SypEGFP-WPRE diluted to 1012 vg/ml was injected into the rostral portion of the medial OT (AP: 1.6mm, ML: -1.0mm, DV: -5.375mm) in Drd1-Cre or Adora2a-Cre mice.

For VP retrograde tracing experiments, 100 nl of Cholera Toxin Subunit B CF 488A (Biotium) was injected at into the caudal portion of the ventrolateral VP (AP: 0.75mm, ML: -1.4mm, DV: - 5.4mm) and 100 nl of Cholera Toxin Subunit B CF 543 (Biotium) was injected into the dorsomedial VP (AP: 0.75mm, ML: -1.0mm, DV: -5.35mm) in C57BL6/J mice. For VTA retrograde tracing experiments, 100 nl of Cholera Toxin Subunit B CF 647 (Biotium) was injected to the rostral portion of the VTA (AP: -3.1mm, ML: 0.8mm, DV: -4.5mm) in C57BL6/J mice. CTB injections were done at 1 mg/ml dilution in PBS. In some cases, tracers were injected bilaterally and each hemisphere was analyzed independently. Following each injection, the injection needle was left at the injection site for 10 minutes then slowly withdrawn.

For imaging experiments, the skull was prepared with OptiBondTM XTR primer and adhesive (KaVo Kerr) prior to the craniotomy. After performing a craniotomy 800 um in diameter centered around the virus injection site, a 27G blunt needle was used to aspirate 1.5 mm below the brain surface. For OT imaging experiments, 500 ul of AAV9-syn-FLEX-jGCaMP7s-WPRE (Addgene viral prep #104491-AAV9) was diluted to 1012 vg/ml and injected into the left and rostral portion of the medial OT in D1-Cre or A2A-Cre mice. For VP imaging experiments, 300 ul of AAV9-syn-jGCaMP7s-WPRE (Addgene viral prep #104487-AAV9) was diluted to 1012 vg/ml and injected into the left and caudal portion of the ventrolateral VP in C57BL6/J mice. Following the viral injection, a head-plate (Model 4, Neurotar) was secured to the mouse’s skull using light-curing glue (Tetric Evoflow, Ivoclar Group). At least 30 minutes after viral injection, a 600um GRIN lens (NA, ∼1.9 pitch, GrinTech) was sterilized with Peridox-RTU then slowly lowered at a rate of 500 um/min into the craniotomy until it was 200 um dorsal to the injection coordinate. The lens was adhered to the surface of the skull using Tetric Evoflow. We then placed a hollow threaded post (AE825ES, Thorlabs) to act as a housing for the lens and adhered it using Tetric Evoflow. Any part of the skull that was still visible was covered using dental cement (Lang Dental). Finally, the housing was covered with a Nylon cap nut (94922A325, McMaster-Carr) screwed onto the thread post to protect the lens in between imaging. Animals were left on the heating pad until they fully recovered from anesthesia.

Histology

Mice were administered ketamine (100 mg/kg) and xylazine (10 mg/kg) and euthanized by transcardial perfusion with 10 ml of cold PBS followed by 10 ml of cold 4% paraformaldehyde in PBS. Brains were extracted and left in a 4% PFA solution in PBS overnight. 50 um coronal sections were cut on a vibratome (VT1000, Leica). A subset of tissue was labeled using the following simplified staining protocol. First, brain sections were incubated for 48 hours at 4°C in the primary antibody diluted in PBST (0.3% Triton-X in PBS). Brain sections were then washed 3 times for 15 minutes in PBST before and after incubating for 2 hours at room temperature in the secondary antibody diluted in PBST. The antibodies used in this study and their dilutions are: Rb ⍺-substance P (1:1,000 dilution; 20064, Immunostar), Rb ⍺-TH (1:1,000 dilution; AB152, Millipore), Dk ⍺-Rb Alexa FluorTM 488 (1:2,000 dilution; A-210206, Thermo Fisher Scientific), Dk ⍺-Rb Alexa FluorTM 647 (1:2,000 dilution; A-31573, Thermo Fisher Scientific). Slices were mounted using Fluoromount with a DAPI counterstain (SouthernBiotech) and imaged on an Olympus BX61 VS120 Virtual Slide Scanner and 10x objective (Olympus). Brains were harvested 21-30 days or 5-7 days after surgery for anterograde and retrograde tracing experiments, respectively. Brains injected for Ca2+ imaging were harvested within a week of the last imaging session.

For anterograde tracing quantification, 4-6 slices containing each of the brain regions of interest (VP, LH, and VTA) were analyzed per animal. To quantify the relative abundance of OT axons in a given brain region, boundaries for the region were drawn on ImageJ Fiji (National Institutes of Health) with reference to the Paxinos and Franklin Mouse Brain Atlas. Afterwards, the percentage of the 16-bit pixels within the boundary that had intensity above 200 was quantified. For retrograde tracing experiments, cells were counted manually every 4th slice.

Behavior

Mice were water restricted to reach 85-90% of their initial body weight and given access to water for 5 minutes a day in order to maintain desired weight. Prior to imaging, mice were habituated to the head fixation device (Neurotar) and treadmill for 3-5 days, 15-30 minutes per session. The treadmill parts were 3D printed using a LCD printer (X1-N, EPAX) from publicly available designs (Jackson et al., 2018). During habituation, mice were provided 10% sucrose from the water spout. Walking and licking behaviors were measured using a quadrature encoder (HEDR-5420-es214, Broadcom) and a capacitance sensor (1129_1, Phidgets), respectively. A video feed of the animal’s face was also recorded using a camera (acA1300-30um, Basler) with a 8-50mm zoom lens (C2308ZM50, Arducam) at 20 Hz with infrared illumination (VQ2121, Lorex Technology).

Odor was delivered to the mouse using a custom-built olfactometer. Compressed medical air was split into 2 gas-mass flow controllers (GFC17, Aalborg). One flow controller directed a constant rate of 1.5 L/min to a hollowed out teflon cylinder. The other flow regulator was connected to a 3-way solenoid valve (LHDB1223418H, The Lee Co.). Prior to odor delivery, the 3-way valve directs clean air at 0.5 L/min to the teflon cylinder. During odor delivery, the 3-way valve directs air to an odor manifold, which consists of an array of 2-way solenoid valves (LHDB1242115H, The Lee Co.), each connected to a different odor bottle. Depending on the trial type, the appropriate 2-way valve opens, directing 0.5 L/min of air flow through the odor bottle containing a kimwipe blotted with 50 ul of diluted odor. All odors were diluted in mineral oil (M5310, Sigma-Aldrich) to 1.5 mmHg. The kinetics and consistency of odor delivery were characterized for 30 trials of terpinene delivery using a miniature Photoionization Detector (mPID) (Aurora Scientific, Inc).

During classical conditioning, animals were exposed to the following odors for 2 seconds: 3-hexanone, 3-heptanone, 3-octanone, ⍺-terpinene, ⍺-pinene, and (R)-(+)-limonene (all odors were purchased from Sigma with the highest available purity). In days 1-3 of training, each of the 6 odors and associated outcomes were provided 30 times with 12-18 seconds of inter-trial interval. Hexanone and terpinene were not associated with any outcome, heptanone and pinene were associated with 2 ul of 10% sucrose, and octanone and limonene were associated with a 70 psi airpuff delivered to their hindquarters. Sucrose or airpuff was delivered 100-300 ms after the end of odor delivery. Trials were organized into 30 blocks, each of which consisted of 1 trial of each of the 6 odors in randomized order. In days 4-6 of training, the outcome contingencies were switched such that heptanone and limonene were not associated with any outcome, octanone and terpinene were associated with 2 ul of 10% sucrose, and hexanone and pinene were associated with 70 psi airpuff.

In the lick-no-lick paradigm, trials were also structured into 30 blocks, each of which consisted of 1 trial of each of the 6 odors in randomized order. Hexanone and terpinene were not associated with any outcome, heptanone and pinene were paired with 2 ul of 10% sucrose at 50% chance, and octanone and limonene were paired with 2 ul of 10% sucrose at 100% chance. 200 ms prior to the onset of 3 of the odors (terpinene, octanone, and limonene), the lick spout was retracted 30 mm away from the animal’s mouth using a linear stepper motor (BE073-1, Befenybay) and driver (A4988, BIQU). The lick spout would return to its original position 100 ms prior to the earliest possible time of sucrose delivery.

DeepLabCut

DeepLabCut2.3.3 with Tensorflow 2.12 was used to track 4 points on the periphery of the eye during 2-photon Ca2+ imaging. The mini-batch k-means clustering method was used to extract a total of 100 frames (20 frames from 5 animals). These frames were labeled and used to train a Deep Neural Network (DNN) model for 100,000 iterations. After the first training session, 20 outlier frames were picked up from each video and added to the training data for a second training session. The area of the eye at a given time point was estimated as an ellipse. For the lick-no-lick paradigm, we used DeepLabCut to track the tip of the tongue, the corner of the mouth, the upper lip and the lower lip. To record licking in the absence of the lick spout, we trained a linear classifier using logistic regression of the following metrics: 1) the confidence score for the tip of the tongue, 2) the confidence score for the corner of the mouth and 3) the Euclidean distance between the upper and lower lip. Data collected from the capacitive lick sensor was used as ground truth for the classifier.

2-photon Ca2+ imaging in head-fixed, behaving mice

Mice were habituated to the head-fixation setup for 3 days beginning 8-10 weeks after surgery. Ca2+ imaging data was acquired using an Olympus FV-MPE-RS Multiphoton microscope with Spectra Physics MaiTai HPDS laser, tuned to 920 nm with 100 fs pulse width at 80 MHz. Each 128x128 pixel scan was acquired with a 20x air objective (LCPLN20XIR, Olympus), using a Galvo-Galvo scanner at 5Hz. Stimulus delivery and behavioral measurements were controlled through a custom software written in LabVIEW (National Instruments) and operated through a DAQ (USB-6008, National Instruments). Each imaging session lasted between 30-45 minutes and was synchronized with the stimulus delivery software through a TTL pulse. The imaging depth was manually adjusted to closely match that of the first imaging day such that we recorded from overlapping populations across days of imaging. Animals were excluded from analysis if a) histology showed that either the GRIN lens or the jGCaMP7s virus was mistargeted or b) the motion during imaging was too severe for successful motion-correction. 2 animals were excluded due to mistargeting and 2 animals were excluded due to excessive motion.

Image Processing

Ca2+ imaging data were first motion-corrected using the non-rigid motion correction algorithm NoRMCorre (Pnevmatikakis and Giovannucci, 2017). Afterwards, neural traces were extracted from the motion-corrected data using constrained nonnegative matrix factorization (CNMF) (Giovannucci et al., 2019; Pnevmatikakis et al., 2016). Briefly, this algorithm estimates a spatial matrix (analogous to the idea of ROI’s in manual processing methods) and a temporal matrix whose products equal the motion-corrected spatiotemporal fluorescence data. Spatial components identified by CNMF were inspected by eye to ensure they were not artifacts. A Gaussian Mixture Model (GMM) was used to estimate the baseline fluorescence of each neuron. To account for potential low-frequency drift in the baseline, the GMM was applied along a moving window of 2,500 frames (500 seconds). The fluorescence of each neuron at each time point t was then normalized to the moving baseline to calculate ΔF/F = Ft - Fbaseline/Fbaseline. For analysis comparing the activity of the same neuron across multiple, spatial components from two different imaging days were matched manually. All subsequent analyses were performed using custom code written in MATLAB (R2022b).

Hierarchical clustering of pooled averaged responses

ΔF/F in response to all 6 odors on day 6 were averaged across trials then Z-scored. The resulting trial-average values from the following timebins were averaged across time: 1) the first second during each odor, 2) the last second during each odor, and 3) the first second after each odor. The resulting 18-element vectors were sorted into 6 clusters after agglomerative hierarchical clustering using euclidean distance and ward linkage.

Responsiveness criteria

To determine how many neurons were responsive to a given odor, we compared ΔF/F at each frame during the 2 second odor period against a pooled distribution of ΔF/F values from the 2-seconds prior to odor onset using a Wilcoxon rank sum test. The resulting p-values were evaluated with Holm-Bonferroni correction to ensure that familywise error rate (FWER) was below 0.05. We then calculated the percentage of responsive neurons for each animal to show the mean and the standard error as a function of time. We also counted the number of neurons that were significantly responsive for at least 4 frames during the odor period to report the total percentage of responsive neurons during odor.

Single neuron logistic classifiers

To test how reliably a single neuron’s fluorescence could discriminate between 2 odors, we assessed the performance of binary logistic classifiers trained on a single neuron’s responses to 2 odors. For each neuron and odor pair, we averaged the ΔF/F during the last second of the odor exposure for each trial then Z-scored across all trials. The resulting 60-element vector was used to train a linear classifier using logistic regression. The receiver operator characteristic (ROC) was evaluated for each single neuron pairwise classifier and the area under the curve (AUC) reported. To test if a given pairwise classifier performed significantly better than chance, we compared the accuracy of each classifier against a distribution of 10,000 classifiers trained on shuffled labels.

Normalized ΔΔF/F correlations

To compare the average response of a neuron to each odor, the trial-averaged ΔF/F during the last second of odor exposure from each trial was averaged and then subtracted from the trial-averaged ΔF/F during the 2 seconds prior to odor delivery. This ΔΔF/F value was scaled to the largest positive ΔΔF/F value of each neuron for all odors. To assess the similarity of the average response to a given pair of odors 𝑖 and 𝑗, we looked at the null linear model in which all neurons respond identically to both odors, i.e. ΔΔFj/F = ΔΔFi/F. To assess how well this describes the data, we report the R2 value of the fit.

Pairwise euclidean distance

To quantify the differences among population-level responses to the 6 odors, we quantified the pairwise Euclidean distance between the trajectories of odor responses. First, we subtracted the ΔF/F values during the 2 seconds prior to odor delivery from each frame then averaged these values across trials for each odor. The pairwise Euclidean distance at each frame was computed for each odor pair and normalized to the maximum pairwise distance measured in all odor pairs at any time bin. These calculations were carried out separately for each animal and then averaged across biological replicates to report the mean and the standard error.

Population pairwise classifiers

To assess the discriminability of odor responses in high-dimensional space, we measured the accuracy of binary classifiers for a given odor pair. At each time point relative to odor delivery, we pooled ΔF/F values from all trials during which either odor was presented. These values were then normalized and used to train a linear classifier using either a logistic regression or a Support Vector Machine (SVM). The accuracy of the classifier was evaluated via 5-fold cross-validation. To test if a given pairwise decoder performed significantly better than chance, we compared the accuracy of each classifier against a distribution of 10,000 classifiers trained on shuffled labels. All classifiers were trained on populations of neurons simultaneously recorded from individual mice. The resulting cross-validated accuracies were averaged across biological replicates to report the mean and the standard error.

Dimensionality analysis

To quantify the dimensionality of each simultaneously recorded neural population, we calculated its participation ratio (PR). First, we performed principal component analysis of the whole dataset using the singular value decomposition algorithm. The PR was calculated as the square of the sum of the eigenvalues of the covariance matrix divided by the sum of the square of its eigenvalues (Litwin-Kumar et al., 2017; Recanatesi et al., 2019). To account for the differences in number of recorded neurons across individuals, we bootstrapped the PR by randomly sampling n neurons from each dataset 1,000 times and reported the average PR value.

Statistical analysis

For simple pairwise comparisons, we used Student’s t-tests or, when appropriate, Wilcoxon rank sum tests with Benjamini Hochberg correction to adjust for false discovery rate (FDR). For post hoc comparisons following ANOVA’s, we used Tukey’s honestly significant difference test which adjusts for family-wise error rate (FWER). For linear mixed-effects models with individual animals as random effect, we used the MATLAB fitlme function with maximum likelihood estimation algorithm and Quasi-Newton optimization.

Author contributions

D.L. and C.M.R. conceived of the project, participated in its development, and wrote the manuscript. L.L. assisted with anatomy, histology and behavioral analysis. D.L. performed all imaging experiments and analyzed the data.

Acknowledgements

We thank members of Root lab for discussions, M. Aoi for discussions on data analysis, and T. Komiyama and for comments on the manuscript. This research was supported by grants from the NIH (R00DC014516, R01DC018313), and C.M.R. was a Hellman Fellow.