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 three-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 OT_D1 and OT_D2, 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; Gadziola et al., 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). Although 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; Ottenheimer et al., 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 OT_D1 and OT_D2 neurons of the anteromedial OT 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 anteromedial 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 OT_D1, OT_D2, and VP neurons with two-photon Ca²⁺ imaging, we found that VP neurons encode reward-contingency in low-dimensional space with good generalizability. In contrast, activity in both OT_D1 and OT_D2 neurons was high-dimensional and primarily contained information about odor identity, although OT_D1 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.

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. Although multiple studies have characterized its anatomical connectivity (In ’t Zandt et al., 2019, 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 OT_D1 and OT_D2 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 (Figure 1A, Figure 1—figure supplement 1C–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 AcbSh_D1 neurons and primarily VP projections AcbSh_D2 neurons (Figure 1B–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 OT_D1 and OT_D2 experiments (Figure 1B–C, Figure 1—figure supplement 1A, 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.

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 (Figure 1D, Figure 1—figure supplement 1F–G). We found strong labeling of soma by both CTB::488 and CTB::543 in the Acb, AI, and Pir (Figure 1E–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 (Figure 1G, Figure 1—figure supplement 1H). 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 (Figure 1H–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 neurons within any part of the OT 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 the anteromedial 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 (Figure 2A). During each trial, the animal is exposed to 1 of 6 odors for 2 s. At the end of odor delivery, the animal either receives: 2 µl of a 10% sucrose solution (S), 50ms 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. S_K, a sucrose-paired ketone vs. P_K, an airpuff-paired ketone) are more different than odor pairs of same reward-contingency (e.g. S_K, a sucrose-paired ketone vs. S_T, 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.

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; Figure 2B, Figure 2—figure supplement 1A-F). Additionally, we implanted a 600 µm Gradient Refractive Index (GRIN) lens 150 µm dorsal to the virus injection site and cemented a head-fixation plate to the skull. Six to eight weeks after surgery, animals were water-restricted and habituated for 3–5 days in the head-fixation setup (Figure 2C). We processed the acquired time-series images using Constrained Nonnegative Matrix Factorization (Pnevmatikakis et al., 2016) to obtain fluorescence traces from each putative neuron (Figure 2D and E). In total, we recorded Ca²⁺ signals from 231 OT_D2 neurons from 6 Adora2a-Cre animals (Figure 2, Figure 2—figure supplement 2), 288 OT_D1 neurons from 6 Drd1-Cre animals (Figure 2, Figure 3) and 130 VP neurons from 5 C57BL6/J animals (Figures 2 and 3, Figure 4).

After 3 days of odor-sucrose associations, the animals displayed anticipatory licking behavior primarily during sucrose-paired odors (Figure 2F and 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 (Figure 2G). 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 (Figure 2H, Appendix 1—tables 4 and 5; ANOVA: F_day = 27.64, p_day = 2.29e-16, F_{lens location}=2.30, p_{lens 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 (Figures 2—4, Figure 5D-G, Appendix 1—tables 32–36). 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 (Figure 2—figure supplement 2, Figure 2—figure supplements 3 and 4). 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 (Figure 3A). 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 (Figure 3B). Cluster I and cluster III most closely fit our description of putative valence-encoding neurons, that is neurons that had similar responses to 2 sucrose-cues (S_K vs. S_T) but different responses to a sucrose-cue and a puff-cue or control odor (S_K vs. P_K or X_K). 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 (Figure 3C), leading us to hypothesize that individual neurons in the VP were more likely to be valence encoding neurons than in either anteromedial OT subpopulation.

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 S_K, S_T) or inhibited (24.5 ± 3.0%, 29.4 ± 3.8% for S_K, S_T) to either sucrose-paired odor than to control or puff-paired odors (7.6–11.1% excited, 8.1–12.9% inhibited; Figure 3D, Figure 2—figure supplement 7A). For statistical comparisons see (Figure 2—figure supplement 7B, Appendix 1—table 37). 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 1–34.1 ± 5.1% by the 6th day of imaging (Figure 3E,Appendix 1—table 11). 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 OT_D1 and OT_D2 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 OT_D1, 11.8 vs 21.7% in OT_D2), 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 (Figure 2, Figure 2—figure supplement 8, Appendix 1—table 38). By comparison, neither sucrose-pairing nor airpuff-pairing had any impact on the magnitude of odor responses in OT_D2 neurons on day 6. And though we did observe a significant effect of sucrose-pairing on response magnitudes in OT_D1 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 two 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 (S_K) and to the sucrose-paired terpene (S_T). VP neurons had a high correlation between a neuron’s responses to S_K and S_T (Figure 3F; R²=0.89). This similarity was much higher than between the sucrose-paired ketone (S_K) and the control ketone (X_K) despite the greater structural similarity between S_K and X_K (Figure 3G; R²=0.33). In contrast, for both OT_D1 and OT_D2 neurons, there was a higher correlation between responses to similar molecular structure (S_k and X_K,) than between responses to similar contingency (S_K and S_T) (OT_D2: S_K vs. S_T R²=0.04, S_K vs. X_K R²=0.58; OT_D1: S_K vs. S_T R²=0.13, S_K vs. X_K R²=0.40). Moreover, most VP neurons (76.5%), had a smaller absolute difference in the response magnitude to the 2 S odors (|S_K-S_T|) than the absolute difference between the sucrose-paired ketone and the control ketone (|S_K-X_K|) (Figure 3H). By comparison, only half of OT_D2 and OT_D1 neurons showed smaller |S_K-S_T| than |S_K-X_K|, 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 the 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. S_K vs. P_T, Figure 3—figure supplement 1A, 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 (Figure 3—figure supplement 1C and D). To assess whether neurons from each region were encoding valence, we compared a neuron’s {S_K vs. X_K} decoder performance (intervalence classification) against its {S_K vs. S_T} decoder performance (intravalence classification) (Figure 3I–K, Figure 3—figure supplement 1E and 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 OT_D1 or OT_D2 neurons (Figure 3I–J, Appendix 1—tables 12–15, Figure 3—figure supplement 1F, Appendix 1—tables 39–41). Intravalence classification, however, did not depend on days or region (Figure 3K, Appendix 1—tables 16 and 17, Figure 3—figure supplement 1F, Appendix 1—tables 42–44). 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 {S_K vs X_K} and {S_K vs S_T}, respectively). In contrast, a similar number of OT neurons displayed good intervalence decoding as did intravalence decoding (20.8% vs 19.9% of OT_D1; 12.8% and 21.0% of OT_D2 for {S_K vs X_K} and {S_K vs S_T}, respectively). The pattern of better intervalence decoding than intravalence decoding among VP neurons was observed across all 15 pairwise classifiers (Figure 3—figure supplement 1H). Whereas 10.2% of all day 6 VP neurons had auROC >0.75 for {S_K vs. S_T}, 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 {S_K vs. S_T} auROC against its {S_K vs. X_K} auROC, we can categorize a neuron into the 4 categories (Figure 3L). (1) a valence encoding neuron ({S_K vs. S_T}<0.75 and {S_K vs. X_K}>0.75), (2) an identity encoding neuron (both auROC >0.75), (3) an identity encoding neuron that does better with S odors ({S_K vs. S_T}>0.75 and {S_K vs. X_K}<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 OT_D1 then OT_D2 (Figure 3M and N; 47.7, 16.2, 7.3% for VP, OT_D1, and OT_D2, respectively). The opposite was true for identity encoding. VP had a smaller percentage of identity encoding neurons than either OT_D1 or OT_D2 (14.8, 21.1, 22.9% for VP, OT_D1, and OT_D2, respectively). We note that these conclusions can also be replicated when analyzing multinomial regression (MNR) classifiers trained on single neuron activities Figure 3—figure supplement 2F and G, Appendix 1—tables 50–53. Namely, the rates of confusion between the 2 sucrose cues are highest in VP and lowest in OT_D2 whereas the rates of confusion across all ketones (S_K, X_K, P_K) are highest in OT_D2 and lowest in VP. These single-neuron classifier analyses further indicate that VP neurons, more than either OT_D2 or OT_D1 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 {S_K vs. X_K} and {S_K vs. X_K}, a majority of OT_D2 and OT_D1 neurons (69.7% and 62.6%, respectively) showed auROC <0.75 for both {S_K vs. S_T} or {S_K vs. P_K}. Thus, in comparison to the VP, most individual anteromedial 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 (Figure 4A–F). We noticed there were neurons that responded to the sucrose delivery on day 1 that responded to the sucrose cue on day 3 (Figure 4C and 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 S_K and S_T on day 3, respectively (Figure 4G). 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 S_K and S_T in OT_D1; 10, 2.5% for S_K and S_T in OT_D2). 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 OT_D1 and OT_D2 neurons, respectively (Figure 4H–I, Appendix 1—tables 18 and 19). 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.

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 (Figure 5A and 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 s odor delivery, and slowly decays after odor ends (Figure 5A). When examining the average pairwise distance during the last second of odor, there was a relatively unstructured distribution of pairwise distance in OT_D2 odor-response such that ||S_K-X_K||, ||S_K-P_K||, ||S_K-S_T||, and ||X_K-X_T|| were all similar (Figure 5B). By comparison, in VP populations, the distribution was structured such that intervalence pairwise comparisons between sucrose-paired and not sucrose-paired odors (e.g. ||S_K-P_K|| and ||S_K-X_K||) were larger than intravalence pairwise comparisons (e.g. ||S_K-S_T||, or ||X_K-X_T||). OT_D1 populations showed an intermediate trend where most intravalence pairwise distances were smaller than intervalence pairwise distances with the exception of ||S_K-S_T||. Thus, at the population level VP representations appear to encode valence but not identity, whereas the anteromedial OT representations encode some valence information but appear to be better suited for identity encoding.

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 (Figure 5C–D). To quantify this, we extracted the average ΔF_i,k/F values for each trial i∈ [1,m] and each neuron k∈ [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 fivefold 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 (Figure 5—figure supplement 1A). Importantly, all classifiers with CV accuracy higher than 0.75 had p-value less than 0.001.

Linear classifiers trained on day 6 OT_D2 population data had similar ranges of accuracy regardless of valence (Figure 5D). For example, the intravalence classification {S_K vs. S_T} was more accurate (86.6 ± 3.9%) than some and intervalence classifications (e.g. {S_K vs. X_K}, 72.8 ± 5.4%) but less accurate than others (e.g. {S_T vs. P_K}, 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 ({S_K vs. S_T}, 79.9 ± 6.1%). Additionally, whereas OT_D2 population classifiers could decode the 2 control cues {X_K vs. X_T} 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 {X_K vs. X_T} than the aforementioned intervalence classifiers. This suggests that whereas OT_D2 encodes odor identity agnostic to the valence, VP does not encode identity at all but rather encodes reward contingency or positive valence. OT_D1 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 {S_K vs. S_T} classification was comparably accurate (90.9 ± 4.7%). This rules out the interpretation that OT_D1 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 (SVMs) with a radial basis function kernel and found we could draw the same conclusions (Figure 5—figure supplement 1E). 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 OT_D2 or OT_D1 data (Figure 5—figure supplement 1F), corroborating through an alternate analysis method that VP population activity encodes reward contingency whereas either anteromedial 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. Although 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, two cues of different valence cannot have the same identity, meaning that good decoding of {S_K vs. P_K} can be extracted from either valence encoding or identity encoding populations. To disambiguate these two 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 (Figure 5E). 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 {S_K vs. P_K}, test on {S_T vs. X_T}). In other words, given valence encoding, {S_K vs. P_K} should be discriminable in a way that can also discriminate {S_T vs. X_T}. 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 (Figure 5F and G). OT_D2 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 OT_D1 displayed a generalizability in between that of VP and OT_D1, suggesting that OT_D1 could encode some valence in addition to identity. However, we note that the VP population, on average, outperforms OT_D1 at generalized valence decoding (95.0 ± 2.0% vs 78.5±3.9%; Appendix 1—tables 22 and 23).

After performing these population-level analyses, we noticed a discrepancy: although single-neuron intervalence decoding was worse in anteromedial OT than in VP (Figure 3M–N), population-level intervalence decoding was comparable between either OT subpopulations and the VP (Figure 5C–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 n^th 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 (Figure 5H). 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 (PR_VP = 5.83 ± 0.80) than either OT_D2 (PR_D2=9.61 ± 0.37) or OT_D1 (PR_D1=9.24 ± 0.44) animals after training (Figure 5I, Appendix 1—tables 24 and 25). 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 {S_K vs. P_K} decoders (CV accuracy_PC1=85.5 ± 2.7%), all 15 PCs were required for comparable {S_K vs. S_T} decoding (CV accuracy_PC1:15=75.1 ± 13.4%) (Figure 5J–L, Appendix 1—tables 26–29). In either OT populations, the first PC did not support good decoding of either {S_K vs. P_K} or {S_K vs. S_T}. Together, our population-level analysis indicates that VP encodes valence, but not identity, in low-dimensional space, OT_D2 encodes identity but not valence in high-dimensional space, and OT_D1, 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 (Figure 2F). 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 (Figure 6—figure supplement 1A, B), and could be used to train distributed lag models to predict onset of licking bouts (Figure 6—figure supplement 1C). Across all VP neurons, we observed a positive and significant correlation between a neuron’s valence decoding ability and licking decoding ability (Figure 6—figure supplement 1D; slope = 0.41, p=2.2 × 10^–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 (Figure 6).

Briefly, headfixed animals were presented with one of six odors in pseudorandomized order. During the presentation of three 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 three 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). One odor from each group served as a control odor that had 0% reward-contingency (L_X, N_X). The other two odors in each group were paired with sucrose at low (50%) or high (100%) probability (L_lo, L_hi, N_lo, N_hi). 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. N_hi vs. L_lo).

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; Figure 6—figure supplement 1E–G). We chose to pool data across all mice from the day of the highest licking differential between L_hi and N_hi odors (Figure 6—figure supplement 1H) to maximize the decoupling of value and motor output in our analysis. Anticipatory licking during L_lo or L_hi began during the last second of the odor and increased gradually until sucrose delivery whereas licking during N_lo or N_hi was delayed by about one second (Figure 6C). When quantified across animals on their days of highest lick differential, we found that mice consistently licked most during L_hi, followed by L_lo and N_hi, then N_lo (Figure 6D, Appendix 1—tables 30 and 31). 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 five animals on their days of highest lick-differential (Figure 6E). When clustering VP neurons into three 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 (Figure 3A–B, Figure 6E–F). Another cluster (III) by comparison, showed a varied excitatory response to each of the four sucrose-paired odors, much like cluster III from the previous experiment. Cluster (III) neurons seemed to have a particularly strong response to L_hi 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 (Figure 6G–H). We reasoned that auROC values for {L_hi vs. N_hi} would be high for vigor encoding neurons but not value encoding neurons given these two odors have the same reward-contingency but disparate licking behaviors. Similarly, we reasoned that auROC values for {N_hi vs. N_X} 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 two odors. First, we saw that while single neuron decodability along the reward-contingency axis (e.g. {N_hi vs N_X}) was higher than along the lick axis (e.g. {L_hi vs N_hi}), there were more neurons that could decode {N_hi vs N_X} than could decode {L_X vs N_X} at auROC >0.75 (Figure 6G–H). Furthermore, we saw a lack of significant correlation between the single-neuron decodability of two odors that had similar licking but different reward-contingency ({N_hi vs N_X}) and the decodability of two odors that had different licking but same reward-contingency ({L_hi vs. N_hi}) (Figure 6I; 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 two pairwise comparisons that both had high difference in reward-contingency ({N_hi vs N_X} and {L_hi vs L_X}) (Figure 6J; slope = 0.60, p=3.8 × 10^–10, R²=0.30).

We also performed the converse experiment where the ΔΔF/F_baseline 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 (Figure 6K). 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 two 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 two 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 {L_hi vs L_X} and {N_hi vs N_X} could both be decoded better than {L_X vs N_X} (Figure 6L, 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 (Figure 6N; 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 (Figure 6O; slope = 0.15, p=3.7 × 10^–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 (Figure 6P; 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.

Our anatomical investigations demonstrate that the primary output of the anteromedial OT is to the VP, with minimal connections to the VTA. Given its constrained connectivity, we propose that the anteromedial 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 OT_D2, robustly encodes the sucrose-contingency of odors. Although our analyses revealed that sucrose-contingency influences odor-evoked responses in OT_D1 neurons more so than in OT_D2 neurons, other evidence suggests valence encoding is not the appropriate framework for interpreting OT_D1 activity. Specifically, information about sucrose-contingency in OT_D1 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 anteromedial OT_D1 activity are more likely to reflect increased contrast of identity or an intermediate encoding of valence that also encodes 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 OT_D1 or OT_D2 have significant extrapallidal outputs. This is in contrast to a previous study which reported that OT_D1 neurons, and to a lesser extent, OT_D2 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 neurons from any part of the OT. 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 OTmidbrain 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.

Although we found little difference in the output patterns of anteromedial OT_D1 and OT_D2 neurons, we observed differences in how these two subpopulations encode odor valence. Consistent with a previous report (Martiros et al., 2022), we found that OT_D1 activity, more than OT_D2 activity, is modulated by reward contingency. For example, OT_D1 neurons, but not OT_D2 neurons, were more likely to respond to sucrose-paired odors than other odors. And the magnitude of responses in OT_D1 but not OT_D2 neurons were significantly larger to sucrose-paired odors than to other odors. We refrain, however, from concluding that the primary feature encoded in OT_D1 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 OT_D1. Additionally, whereas more than 50% of VP neurons could be categorized as reward-contingency encoders, this figure was less than 20% for OT_D1. Lastly, population-level decoders trained on odor pairs of different valence can generalize in the case of VP populations, but not OT_D1 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. label 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 two areas. Optogenetic activation of OT_D1 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 interesting 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: OT_D2 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 two 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 anteromedial 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 OT_D1 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 anteromedial OT reflects either an intermediate representation of reward-contingency or a contrast modulation to reflect the contingency.