Midbrain endocannabinoids actuate dopamine-based action selection

SUMMARY

Quickly adapting behavior to changing environments is key for survival. Brain dopamine signals reinforce actions that procure rewards and avoid adverse outcomes. However, the molecular mechanisms within the circuits that actuate these signals at subsecond timescales remain unexplored. We show that midbrain endocannabinoid mobilization is the conditio sine qua non for behavioral invigoration evoked by external cues. Reward- and punishment-predictive signals require intact retrograde release of the endocannabinoid 2-arachidonoylglycerol (2-AG) from ventral tegmental area dopamine neurons. Moreover, we show that genetic deletion of presynaptic cannabinoid type-1 receptors (CB1Rs), the target for 2-AG, recapitulates behavioral and dopaminergic deficits. Exploiting causal inference tools to interpret continuous brain signals, we describe that 2-AG/CB1R communication is essential for striatal dopamine release to signal salient cues and propel conditioned responding. These findings reveal an endocannabinoid-mediated disinhibition mechanism exploited by dopamine neurons to orchestrate striatal dopamine release events, enabling adaptive action selection informed by cues.

Graphical Abstract

In brief

Dopamine signals guide adaptive responses to reward- and threat-predictive cues, yet the mechanisms shaping these rapid dynamics remain unclear. Luján et al. identify midbrain endocannabinoid signaling onto pallidal inputs as a temporally precise disinhibitory mechanism that enables striatal dopamine release to support adaptive action selection in response to salient cues.

INTRODUCTION

The nervous system guides behavior through computations that leverage environmental cues¹ to secure rewards and avoid adverse outcomes.^2,3 The brain’s dopamine system is a key mediator of these computations, motivating behavior following unexpected events and promoting learning.^4,5 Building upon prior voltammetric evidence, genetically encoded real-time dopamine biosensors^6,7 now reveal how subsecond dopamine fluctuations regulate ongoing actions.⁸ For instance, striatal dopamine fluctuations act as continuous teaching signals, conferring spontaneous behavior its temporal structure.^9-13 However, the precise molecular mechanisms by which continuous dopamine computations integrate ongoing exteroceptive information inherited from parallel processing inputs at short temporal scales are yet to be explored.

Prior evidence suggests a critical role for the midbrain endocannabinoid (eCB) system in shaping dopamine cell excitability and striatal release events.¹⁴ In the ventral tegmental area (VTA), G_i/o protein-coupled CB1Rs are found in afferent terminals impinging upon dopamine neurons.¹⁵ Activation of CB1R by the ligand 2-arachidonoylglycerol (2-AG) preferentially results in the inhibition of GABAergic terminals,¹⁶ thereby disinhibiting VTA dopamine neurons.¹⁷ We previously showed that midbrain 2-AG mobilization is required for operant performance in appetitive and aversive tasks.^18-20 However, it is currently unknown whether this mechanism participates in dopamine-mediated subsecond structuring of motivated behavior and whether CB1R signal transduction plays a similar role as 2-AG mobilization in enabling salient information processing.

Here, we manipulate a candidate cannabinoceptive circuit from the basal forebrain that innervates the VTA and shapes dopamine-based continuous encoding of cued behavior. Using integrated genetic, photometric, and causal inference computational tools, we find that VTA dopamine cells require 2-AG release to modulate incoming basal forebrain afferents. These findings reveal a temporally precise dopamine-eCB circuit interaction required for fundamental adaptive functions involved in the proactive avoidance of threats and the rapid pursuit of rewards.

RESULTS

Threat avoidance requires mobilization of the endocannabinoid 2-AG by VTA dopamine neurons

Previous studies indicate that eCB signaling in dopamine neurons is essential for their canonical role in appetitive reward processing.^18,19,21 However, whether eCB mobilization from dopamine neurons participates more broadly in goal-directed behaviors, including negative reinforcement and punishment, remains unknown (but see Wenzel et al.²⁰). To specifically test how 2-AG mobilization from VTA dopamine neurons impacts learning about aversive outcomes, we implemented the conditioned lever-pressing active avoidance paradigm in DGLa^f/f mice that we previously published using rats.^19,20 In these animals, we expressed cre recombinase under control of the tyrosine hydroxylase (TH) promoter, a marker of dopamine neurons, bilaterally across the entire VTA (Figure 1A). Targeting dopamine cells using the AAV9-rTH-cre vector has been repeatedly validated before, yielding penetration ratios of ~60%, while showing negligible leakage.^22-26 Here, cre recombination with loxP sites flanking the dagla gene prevented the ability of VTA dopamine neurons to synthesize 2-AG.¹⁸ Using fluorescence in situ hybridization, we confirmed confined expression of cre in the VTA (Figure 1B). As we previously reported,¹⁸ cre co-localization with th transcripts resulted in specific deletion of dagla mRNA (Figure 1C). To control for virus- and genotype-related confounds, we locally transduced rTH-cre in wild-type (WT) animals and artificial cerebrospinal fluid (aCSF) in DGLa^f/f mice, respectively. This strategy was used to abrogate eCB-mediated retrograde signaling from VTA dopamine neurons to incoming CB1R-expressing afferents (Figure 1D).

Figure 1. Cell-type-specific deletion of DGLa in VTA dopamine neurons hampers operant active avoidance performance.

(A) loxP structure surrounding the DGLa (dagla) gene in DGLa^f/f mice.

B) AAV9-rTH-cre construct was bilaterally transduced in the VTA of male and female WT and DGLa^f/f mice. (Right) Midbrain fluorescence in situ hybridization section showing intra-VTA-specific infection and co-localization of th and dagla transcripts (scale bar, 150 μm).

(C) cre recombinase in the VTA is exclusively found on th+ cells (white arrows), but not th− cells (scale bar, 20 μm). (Right) Dagla expression is knocked down after transduction of cre in th+ cells (white arrows) but not cre− cells (empty arrows) (scale bar, 5 μm).

(D) Schematic representation illustrating the resulting disruption in endocannabinoid retrograde signaling from VTA dopamine cells in DGLa^TH cKO mice. CTRL group included WT/rTH-cre mice as well as DGLa^f/f aCSF controls.

(E–G) Schematic illustration of the operant active avoidance task. Redundant multisensory cues (tone + house lights and white noise + cue light) were used to avoid potential genotypic sensory confounds.

(H) Knocking down DGLa^TH did not alter daily (left) or average (right) escape responses throughout active avoidance training.

(I and J) Active avoidance is severely impaired in DGLa^TH cKO mice in terms of total avoidance responses (I) (unpaired t test, **p = 2.1 · 10⁻⁵) and avoidance rates (% avoided responses) (J) (unpaired t test, ***p = 2 · 10⁻⁴).

For all panels, CTRL, n = 15; DGLa^TH cKO, n = 8. Data are presented as mean ± SEM.

We next subjected the animals to the operant active avoidance training paradigm (Figures 1E-1G; STAR Methods). Due to their similar behavioral phenotypes (see Table S1 for detailed statistics), both WT (receiving AAV9-rTH-cre) and SHAM DGLa^f/f mice were consolidated into a single control group (CTRL) and compared against DGLa^TH conditional knockout (cKO) mice. Analysis of escape performance revealed no significant effect of VTA DGLa^TH deletion (Figure 1H). However, comparison of avoidance rates between groups showed that DGLa^TH cKO mice were impaired in avoiding an incoming threat in response to warning stimuli (Figures 1I and 1J). This result expands previous pharmacological evidence²⁷ ascertaining that 2-AG release specifically from VTA dopamine neurons is critical for avoidance. Furthermore, these findings also give rise to two important questions regarding the upstream and downstream components of this signaling pathway, which we address next.

Momentary uncoupling of NAc dopamine release dynamics and active avoidance in DGLa^TH cKO mice

Our group has shown that phasic mesolimbic dopamine signaling causally drives active avoidance learning.²⁰ To test the downstream consequences of downregulated midbrain 2-AG mobilization, we measured dopamine release in the medial portion of the nucleus accumbens (NAc) core during operant active avoidance (Figures 2A and 2B). Our fiber photometry recordings reveal learning-dependent encoding of response strategy (escape vs. avoidance, Figures 2C and 2D). Early in training (sessions 1–4), warning signal (WS) presentations in both trial types elicited dopamine decreases, followed by increases during the safety period (Figure 2C, representative traces). In later sessions (8–12), as avoidance stabilized, dopamine responses at WS onset diverged by strategy: decreases accompanied escape, whereas increases predicted avoidance (Figure 2D, representative traces). Figure 2E compiles GrabDA_2m traces from all CTRL and DGLa^TH cKO trials, averaged across early (1–4), mid (4–8), and late (8–12) training sessions.

Figure 2. Loss of NAc dopamine threat (WS) encoding in DGLa-deficient mice.

(A) Schematic representation of VTA DGLa^TH knockdown and GrabDA_2m expression in the NAc. (Right) Immunohistochemical section illustrating GrabDA_2m (GFP) staining and optic fiber placement in the core section of the NAc (NAcc).

(B) Schematic illustration of the operant active avoidance task.

(C and D) Representative CTRL NAc GrabDA_2m traces associated with escape and avoidance responses across early (1–4) and late (8–12) training sessions.

(E) Group-averaged NAc GrabDA_2m trials time-locked to threat cue presentation (CS+) in early, mid, and late training sessions. CTRL, n = 4 (349 trials); DGLa^TH cKO, n = 6 (190 trials). Data are shown as mean ± SEM.

(F) Graphical illustration of the FLMM procedure. At every trial time point, GrabDA_2m Z scores, Y_i(s), are pooled from all trials and correlated against genotype (β₁), trial outcome (β₂), and their interaction (β₃).

(G) Group-averaged GrabDA_2m trial dataset (late training) used for FLMM.

(H–J) Coefficient estimates from FLMM analysis of the trial outcome × genotype association with dopamine. The plots are aligned to warning signal (WS) onset. Trial periods in which the light shaded area does not contain 0 indicate pointwise significant coefficients. Dark shaded areas denote jointly significant coefficient periods. Joint confidence intervals (CI) are additionally corrected for multiple comparisons and inform about significant time intervals, instead of significant time bins (pointwise CI). Vertical black lines surrounded by gray shaded areas represent average avoidance latency ±SEM. CTRL, n = 4 (349 trials); DGLa^TH cKO, n = 6 (190 trials). Horizontal bars indicate time intervals selected for two-way ANOVA corroboration (Satterthwaite method; interaction, ***p = 0.0005).

To capitalize on the high temporal resolution afforded by GrabDA_2m, we refrained from using summary measures, as these can mask or bias the characterization of rapid, trial-by-trial event-locked dopamine dynamics. To do so, we modeled NAc dopamine release using a functional linear mixed model (FLMM). FLMM enables hypothesis testing of multiple effects and its interaction, at every trial time point, allowing for comparison of the timing and magnitude of signal-behavior subsecond correlations, while accounting for between-animal differences.²⁸ This allowed us to examine the contribution of VTA 2-AG mobilization to moment-to-moment orchestration of avoidance by dopamine. In this experiment, we modeled late training NAc GrabDA_2m transients as a function of genotype (CTRL = 0, cKO = 1; β₁), trial outcome (escape = 0, avoidance = 1, β₂), and its interaction (β₃) at every trial time point (Figures 2F and 2G). The resulting plots (Figures 2H-2J) provide joint and pointwise 95% confidence intervals (CIs). Each time interval for which the joint 95% CI does not contain 0 denotes a significant effect of genotype or trial outcome on the corresponding dopamine release transient throughout that time window (Figure 2F). We, therefore, hypothesized that there would be significant differences in mean dopamine levels between escape and avoidance trials during threat cue presentation (0–3 s, relative to cue onset). More importantly, we posited that this mean difference would dissipate when knocking down DGLa^TH. First, FLMM indicated no differences in mean signal values between genotypes (CTRL vs. cKO, β₁) on escape trial signals after controlling for animal-specific random effects (Figure 2H), suggesting equivalent dopamine dynamics during unconditioned responding. Second, we find that there are significantly higher dopamine levels during WS presentation on avoidance trials compared to escape trials among CTRL mice (Figure 2I, β₂ 95% joint CI lower bound>0 from 0.85 s). Finally, Figure 2J shows that there was a significant interaction between trial outcome and genotype during WS presentation (95% joint CI upper bound <0 from 1.04 s), an effect confirmed by a two-way ANOVA (Satterthwaite method). Since the interaction was negative, ${\widehat{\beta }}_3(s)<0$, this indicates that the difference in mean signal between avoided and escaped trials is significantly smaller in cKO compared to CTRL mice. Importantly, dopamine release and DGLa^TH status were uncoupled before WS onset and the coupling only became significant just before avoiding (average latency to avoid = 1.09 ± 0.03 s). In addition to trial outcome, shorter latencies were also significantly associated with higher dopamine amplitudes during the WS, an effect significantly reduced in DGLa^TH cKO mice (Figure S1). Finally, FLMM also identifies a significant interaction following threat avoidance (95% joint CI lower bound >0 from 5.09 s) (Figure 2J). In comparison, CTRL animals did not show such increases in dopamine release after avoidance, as this outcome was fully expected. Altogether, these fiber photometry results highlight a prominent role for VTA dopamine neuron mobilization of 2-AG in the generation of downstream NAc release events associated with moment-to-moment threat prediction. We show that such dopamine events are functionally coupled with active avoidance probability and, in line with our previous causal manipulations,²⁰ can, therefore, explain the behavioral impairments seen in DGLa^TH cKO mice.

Using FLMM, we observe that in WT mice there is a linear relationship between high/low dopamine release at cue presentation and avoidance probability, which is missing in eCB-deficient mice. However, cue-evoked DGLa^TH cKO dopamine release could still be related to avoidance selection through an unknown signaling mechanism, different from that used by WT mice. To rule out this possibility and confirm that dopamine signaling after DGLa^TH deletion does not account for active avoidance performance whatsoever, we trained random forest classifiers on GrabDA_2m traces to decode binary avoidance/escape states using a sliding-window decoding approach (Figure 3A; STAR Methods). Based on our hypothesis, we predicted that classifiers trained on CTRL traces would outperform cKO classifiers only during WS presentation (0–3 s, relative to cue onset). Notably, the earliest time window at which CTRL classifiers outperformed cKO classifiers included GrabDA_2m training data from only 0.4 s to 0.45 s after WS onset (Figure 3B). By the last point in this time window, only 1.8% of trials had resulted in avoidance (average latency = 1.09 ± 0.03 s), suggesting that the decoding ability of CTRL-trained classifiers at this moment is not merely a by-product of completed actions. Predictably, CTRL classifier accuracy peaked later on (2–3 s after WS onset), likely reflecting that most animals had avoided by that point. Finally, the cKO classifiers failed to show any temporal variation in predictive performance across the trial (Figure 3B), highlighting that there is a complete, temporally discrete disconnection between cue-evoked dopamine release and threat avoidance behavior when VTA dopamine cells cannot synthesize 2-AG. Next, we sought to determine what role midbrain CB1Rs—the pre-synaptic component of this retrograde mechanism—play in the orchestration of active avoidance.

Figure 3. Time-specific features of NAc dopamine release no longer convey information about trial outcomes in DGLa-deficient mice.

(A) Every 50 ms, GrabDA_2m Z scores pulled from CTRL or cKO trials are used to train an independent random forest classifier (100 trees). Features from this time window are then used to predict the escape/avoidance outcome of the trial (10-fold cross-validation). This decoding strategy is repeated every 50 ms, from −2 to 6 s relative to WS onset. CTRL-trained decoders are only used to predict CTRL trial outcomes, and vice versa. To compare when differences in accuracy performance emerge within the trial time window, permutation tests (n = 200) are conducted at every 50-ms time interval.

(B) CTRL- and cKO-trained accuracy results across the trial, aligned to WS onset. Horizontal purple bars denote time intervals in which CTRL-trained decoders outperformed cKO models (false discovery rate corrected *p < 0.05, q-threshold = 0.01). Vertical black lines surrounded by gray shaded areas represent average avoidance latency ±SEM. CTRL, n = 4 (349 trials); DGLa^TH cKO, n = 6 (190 trials).

Characterizing a basal forebrain cannabinoceptive input to VTA dopamine cells

Pharmacological studies have suggested that midbrain CB1Rs balance conditioned and unconditioned responding during learning.²⁹ However, to date, there is no in vivo evidence as to what cannabinoceptive inputs modulate dopamine neuron activity in the midbrain. A promising candidate is the basal forebrain. It contains one of the highest densities of CB1R mRNA seen in the mouse brain³⁰ and heavily innervates the VTA.³¹ In fact, the ventral pallidum (VP), a basal ganglia/forebrain nexus for reinforcement learning,³² is one of the top VTA projectors in the entire brain,^33,34 homogenously innervating all subregions and driving avoidance behavior.³⁵ Owing to this evidence, we tested if VTA^TH-projecting basal forebrain cells express CB1Rs and whether they modulate dopamine neuronal output.

To achieve this, we implemented a cre-dependent rabies-virus-based transsynaptic retrograde tracing strategy,^34,36 combined with fluorescence in situ hybridization of Cnr1 mRNA, to quantify CB1R expression in monosynaptic inputs to VTA dopamine neurons (STAR Methods). To enable transsynaptic retrograde transfer, the rabies virus cognate receptor (TVA), as well as the virus envelope glycoprotein (RG), was expressed in all VTA subregions of DAT-cre mice (Figure 4A). TVA and RG were first expressed under the control of cre recombinase, thus exclusively infecting dopamine neurons (starter cells) (Figure S2). After 14 days, a pseudotyped rabies virus (SAD-eGFP(EnvA)) was expressed in the VTA (Figure 4A). We then multiplexed widefield imaging of SAD-eGFP(EnvA) with fluorescence in situ hybridization of Cnr1 mRNA, focusing on the basal forebrain (Figures 4B and 4C). There, Cnr1 transcripts co-localized with GFP+ cell bodies, confirming that the basal forebrain sends CB1R-enriched monosynaptic projections to VTA dopamine neurons. Compared to the surrounding regions, the VP stood out as the primary, but not sole, cannabinoceptive input to VTA dopamine cells, with higher staining in its medial portion (Figure 4D). Other basal forebrain structures such as the medial, lateral, and magnocellular preoptic nuclei, as well as the diagonal band of Broca, also innervated VTA dopamine cells while expressing CB1Rs, although to a much lesser extent (Figures 4D and 4E).^37-39

Figure 4. Basal forebrain inputs to VTA dopamine cells are under the control of presynaptic CB1Rs.

(A) Experimental design illustrating monosynaptic retrograde labeling of VTA dopamine cell inputs using SADΔG-eGFP(EnvA) in DAT-cre mice.

(B) Widefield imaging of retrogradely traced basal forebrain inputs (GFP) to VTA dopamine neurons (scale bar, 500 μm).

(C) Fluorescence in situ hybridization of cnr1 transcripts in the basal forebrain used to determine co-localization with rabies-traced inputs to the VTA.

(D) Percentage of basal forebrain VTA dopamine projectors expressing cnr1 (one-way ANOVA followed by Holm-Šidák’s post hoc, *p < 0.05 vs. VP) (n = 4).

(E) Basal forebrain localization of all VTA dopamine projectors expressing cnr1.

(F) Visualization of VP coronal sections using in situ hybridization reveals that cre+ cell bodies co-express vGAT and CB1R (white arrows), but not vGlut (yellow arrows). Some VP cells express vGlut and CB1R, but these do not co-express cre.

(G) (Top) Experimental setup for testing the effects of the CB1R agonist MDMB on IPSCs evoked by photostimulation of ChR2-expressing VP axon terminals synapsing on VTA DA neurons (oIPSC, n = 5) or by electrical stimulation near the dopamine (DA) neuron recording electrode (evIPSC, n = 9). (Bottom) Examples of averaged oIPSCs and evIPSCs collected simultaneously during baseline (gray traces) and 20 min after beginning application of MDMB. Traces in left panel show the effect of the GABA_A receptor antagonist gabazine (10 μM) on averaged VP axon-activated oIPSCs.

(H) Mean time course of experiments showing inhibition of oIPSCs and evIPSCs by MDMB. The agonist significantly invariably inhibited evIPSCs and oIPSCs. (Inset) Summary histograms of time course IPSC experiments (Bonferroni’s, *p < 0.05 vs. baseline).

(I) Mean oIPSC time courses showing a reduction of the inhibitory effect of MDMB in brain slices pre-incubated with the neutral CB1R antagonist NESS-0327. Untreated oIPSC time course (blue dots) re-plotted for comparison. (Inset) Summary histograms of time course oIPSC experiments (Bonferroni’s, *p = 0.049, ^n.s. p > 0.99 vs. baseline) (untreated, n = 5; +NESS-0327, n = 7). Data shown as mean ± SEM.

The loss of dopamine canonical function ascribed here to 2-AG-deficient mice suggests that pre-synaptic CB1Rs are preferentially located on inhibitory terminals (but see Faget et al.³⁷). We, therefore, sought to discern whether these cannabinoceptive connections were primarily GABAergic or glutamatergic. Mice received a retrograde cre construct injection covering the entire VTA (Figure 4F). By overlapping vesicular markers of GABAergic (vGAT1) and glutamatergic (vGlut2) in VP cells using in situ hybridization, we found that Cnr1 mRNA was exclusively expressed on VP cre+/vGAT1+ neurons (Figure 4F).⁴⁰ This aligns with studies revealing that VP neurons robustly inhibit dopamine via GABA release,³⁸ thereby decreasing VTA neuron spontaneous firing.^39,41

Next, we corroborated these results with functional circuit mapping, by injecting ChR2 into the VP and recording light- and electrically evoked inhibitory post-synaptic currents (oIPSCs and evIPSCs, respectively) in VTA dopamine neurons using patch-clamp electrophysiology (Figure 4G). oIPSCs from VP terminals onto VTA dopamine cells displayed short latencies (2.8 ± 0.08 ms) and were insensitive to DNQX (10 μM), confirming their monosynaptic GABAergic nature (Figure S3). Conversely, oIPSCs were prevented by bath application of the GABAA antagonist gabazine (10 μM) (Figure 4G). Consistent with the in situ hybridization results, VP→VTA^DA oIPSCs were inhibited by bath application of a CB1R agonist (MDMB, 100 nM) (Figure 4H). Importantly, CB1R activation reduced oIPSCs and evIPSCs to a similar extent (Figure 4H, inset), suggesting that the VP represents the primary pathway through which eCB signaling from dopamine cells modulates GABAergic afferents in the midbrain. Accordingly, pre-application of the neutral CB1R antagonist NESS-0327 significantly reduced the inhibition of MDMB on oIPSCs (Figure 4I). These experiments identify VP CB1Rs as a substrate for retrograde signaling from VTA dopamine neurons, which (along with other yet-to-be-identified cannabinoceptive inputs) participate in a feedback circuit mechanism capable of sustaining dopamine cell activation and its downstream functions.

CB1Rs from basal forebrain neurons projecting to the VTA enable cued dopamine release and active avoidance

Using an intersectional gene-delivery approach,⁴² we downregulated CB1R expression from basal forebrain neurons projecting to the VTA (Figure 5A, left and central panels). Within the basal forebrain, injections were targeted to the VP due to its relative enrichment in CB1R-expressing VTA-projecting neurons and its established involvement in reinforcement. By injecting Flp-dependent cre recombinase throughout the entire VP (AAV9-EF1α-fDIO-cre) and a retrograde flipase (AAVrg-EF1α-mCherry-Flpo) in the VTA of CB1^f/f mice (Figure 5B), a circuit-specific knockout of Cnr1 mRNA was achieved in VTA-projecting (cre+) VP neurons (Figure 5C). The resulting subjects, hereafter referred to as ^VP→VTACB1R cKO mice, were then trained on the same operant active avoidance task (Figure 5D) as above. Mirroring 2-AG-deficient mice, ^VP→VTACB1R cKO mice displayed normal unconditioned stimulus (US)-evoked escape behavior (Figure 5E), but impaired cue-elicited active avoidance (Figure 5F).

Figure 5. Midbrain CB1Rs are required for operant active avoidance and predictive encoding of threat.

(A) Intersectional gene strategy used in CB1^f/f mice to selectively halt expression of CB1R in basal forebrain cells projecting to the VTA.

(B) Immunohistochemical sections showing expression of Flpo (mCherry) and cre recombinase centered in the VTA (scale bar, 300 μm) and the basal forebrain (scale bar, 500 μm), respectively.

(C) In situ hybridization coronal forebrain section demonstrating lack of cnr1 expression in cre+ VTA-projecting VP cells (white arrows), but not cre− cells (empty arrows).

(D) Schematic representation of the operant active avoidance task.

(E) Unaffected daily (left) and average (right) escape responding in ^VP→VTACB1R compared to CTRL mice.

(F) Impaired daily (left) and maximum (right) active avoidance performance in ^VP→VTACB1R mice (unpaired t test, ***p = 7.28 · 10⁻⁵). CTRL, n = 12; ^VP→VTACB1R cKO, n = 11.

(G) Viral transduction strategy used to generate ^VP→VTACB1R cKO mice and probe NAc dopamine fluctuations using fiber photometry. (Bottom) Coronal brain section exemplifying GrabDA_2m (GFP) expression in the NAc core and placement of optic fiber.

(H and I) Group-averaged NAc GrabDA_2m trial signals time-locked to threat cue presentation (CS+) followed by escape (dashed line) or avoidance responses (solid line) in CTRL and ^VP→VTACB1R mice.

(J) For clarity, avoidance-related NAc dopamine signals from both groups are shown together. CTRL, n = 7 (419 trials); ^VP→VTACB1R cKO, n = 5 (295 trials). Data are shown as mean ± SEM.

(K) Graphical illustration of the FLMM procedure using genotype (β₁; CTRL = 0, cKO = 1), trial outcome (β₂; escape = 0, avoidance = 1), and their interaction (β₃) as covariates for the GrabDA_2m signal, Y_i(s).

(L–N) Coefficient estimates from FLMM analysis of the trial outcome × genotype association with dopamine. Horizontal bars indicate time intervals selected for two-way ANOVA confirmation (Satterthwaite method, interaction, ***p = 0.0007). Vertical black lines surrounded by gray shaded areas represent average avoidance latency ±SEM. CTRL, n = 4 (620 trials); ^VP→VTACB1R cKO, n = 4 (904 trials).

In this cohort, mice received AAV9-GrabDA_2m injections and optic fiber implants in the medial portion of the NAc core (Figure 5G). WS-aligned dopamine fluctuations were averaged across sessions 8–12, and traces were grouped based on genotype (CTRL vs. ^VP→VTACB1R cKO) and trial outcome (escape vs. avoidance). As described above, avoidance in CTRL mice was accompanied by a sustained increase in dopamine release during WS presentation (0–3 s, relative to cue onset), which was not present in escape trials (Figure 5H). In contrast, ^VP→VTACB1R cKO GrabDA_2m signals during escape and avoidance trials followed comparable dynamics, both with sharp downward deflections ~0.5 s after WS onset (Figure 5I). A graphical comparison between CTRL and ^VP→VTACB1R cKO reveals diametrically opposite dopamine dynamics during avoidance (Figure 5J), pointing to impaired threat-cue salience attribution and important behavioral implications.

To test whether ^VP→VTACB1R deletion disrupted NAc core dopamine release continuous shaping of negative reinforcement, we again fit an FLMM (see Figures 5K-5N) to estimate differences in mean dopamine levels across trial outcome types (escape = 0, avoidance = 1, β₂) and ^VP→VTACB1R status (CTRL = 0, cKO = 1; β₁). The FLMM model identified no significant differences in mean GrabDA_2m fluorescence levels between cKO and CTRL mice on escape trials (Figure 5L; β₁). As expected, there were higher mean dopamine fluorescence levels on avoidance trials than escape trials among CTRL mice during WS (Figure 5M; β₂ 95% joint CI lower bound>0 from 0.5 s). Finally, the significant interaction term shows that the difference between avoidance-escape signals is smaller in cKO mice during WS, compared to WT (Figure 5N; β₃, 95% joint CI upper bound<0 from 0.66 s). We corroborated this by conducting a two-way ANOVA test collapsing all fluorescence Z values between 0.66 and 1.76 s. In summary, we report that removing CB1Rs from one of the top cannabinoceptive projections to the VTA, originating in the basal forebrain and centered in the VP, critically impairs moment-to-moment generation of negative reinforcement by NAc core dopamine release events.

Basal forebrain CB1Rs in the VTA are required for positively reinforced conditioned responses

To investigate the role of CB1Rs on appetitive tasks, we trained ^VP→VTACB1R cKO mice on a food-reinforced, variable time-out (VTO) task (Figures 6A and 6B), designed to parse out rapid conditioned responses to reward-predictive stimuli⁴³ (STAR Methods). In this trial-based task, the latency to press is interpreted as a proxy for the cue’s association with forthcoming reward. Both groups learned to obtain most rewards within 8 sessions, with ^VP→VTACB1R cKO mice displaying slower learning rates (Figure 6C). As hypothesized, deletion of midbrain CB1Rs significantly delayed conditioned responding, even after both groups matched their active lever pressing (days 5–8; Bonferroni’s, **p < 0.01, *p < 0.05) (Figure 6D). In addition, cue- and reward-associated NAc core GrabDA_2m transients were monitored during VTO training and data from sessions 6–8 were used for group comparison purposes (Figure 6E). Next, we applied FLMM to investigate the functional relevance of the cue-evoked dopamine differences observed (Figures 6F-6H). In particular, we tested how ^VP→VTACB1R status (subject genotype) interacted with NAc core dopamine amplitudes at every trial time point to explain reaction times. Here, we modeled reaction time as the negative of response latency (i.e., −1·latency). Results indicate that, among CTRL mice, faster reaction times were associated with higher GrabDA_2m amplitudes only during cue onset (β₁; 95% joint CI lower bound >0 around 0.35 s) (Figure 6F). There were no significant differences between the genotypes β₂ (genotype) in mean dopamine levels (Figure 6G) among trials with average reaction times, even though the mean signal of both groups increased in response to task stimuli (i.e., both groups showed fluorescence increases). Remarkably, the subsecond relationship between dopamine amplitude changes and reaction times during cue onset was significantly lower among cKO than CTRL mice (β₃; 95% joint CI upper bound <0 from 0.31 s) (Figure 6H). Of note, we find the opposite relationship during post-reward trial time points: the reaction time-dopamine association was significantly larger among cKO than CTRL mice (β₃; 95% joint CI lower bound >0 from 6.01 s after cue onset). Figures 6I-6K depict the effects highlighted by the FLMM modeling, separating trials into fast and slow latencies.

Figure 6. ^VP→VTACB1R promote cue-evoked reward seeking, dopamine predictive encoding, and effortful motivation.

(A) Intersectional gene strategy used in CB1^f/f mice to selectively halt expression of CB1R in pre-synaptic terminals innervating the VTA from the basal forebrain while monitoring dopamine release dynamics in the NAc using fiber photometry.

(B) Schematic representation of the appetitive VTO task used. Redundant multisensory cues (tone + cue light and white noise + house light) were used to avoid potential genotypic sensory confounds.

(C) Impaired conditioned responding in ^VP→VTACB1R mice resulted in fewer food rewards achieved across VTO training (CTRL vs. cKO, Bonferroni’s; *p < 0.05, **p < 0.01, ***p < 0.001).

(D) ^VP→VTACB1R mice cKO mice displayed longer response latency to press across the first seven VTO training sessions (CTRL vs. cKO, Bonferroni’s; **p < 0.001, *p < 0.05). CTRL, n = 13; ^VP→VTACB1R cKO, n = 10. Data are shown as mean ± SEM.

(E) Group-averaged NAc GrabDA_2m trials used for FLMM analysis. Inset shows average dopamine response amplitudes during cue presentation (0–5 s) (unpaired t test, *p = 0.028).

(F) Increases in dopamine release right after cue onset are significantly associated with reaction times (β₁, -latency).

(G) NAc dopamine dynamics were similar across genotypes (β₂) after controlling for animal-specific random effects.

(H) The association between dopamine release and reaction times is significantly smaller in cKO compared to CTRL mice (β₃) (Satterthwaite method; interaction, *p = 0.0105, **p = 0.007). CTRL, n = 7 (418 trials); ^VP→VTACB1R cKO, n = 5 (295 trials).

(I) When trials are separated by reaction time (fast vs. slow; median split), NAc GrabDA_2m traces across show larger cue-evoked transients preceding short-latency trials.

(J and K) In CTRL mice, fast trials are preceded by enhanced cue-evoked dopamine release, whereas this relationship is markedly blunted in ^VP→VTACB1R cKO mice.

(L) Daily active lever presses during PR training (CTRL vs. cKO, Bonferroni’s; **p < 0.001).

(M and N) ^VP→VTACB1R cKO earned less rewards during PR responding than CTRL mice (unpaired t test, **p = 1.3 · 10⁻⁵) despite similar inactive lever responding. CTRL, n = 13; ^VP→VTACB1R cKO, n = 10.

(O) Sucrose preference test using two-bottle choice for 1% sucrose solution vs. water. CTRL, n = 9; ^VP→VTACB1R cKO, n = 7.

We have previously described that deficits in dopamine-based associative learning are often linked to impairments in effortful motivation—the ability to exert high effort (i.e., lever presses) to obtain a positive reinforcer, which may have implications for psychiatric disorders associated with abulia.^18,40,43,44 To test this possibility, CTRL and ^VP→VTACB1R cKO mice were trained under a progressive ratio (PR) schedule of reinforcement. Despite similar performance at the beginning of training (days 1–2), CTRL mice quickly outperformed ^VP→VTACB1R cKO mice in lever presses (Figure 6L). On average, ^VP→VTACB1R cKO earned less rewards compared to CTRL mice (Figure 6M) despite showing similar inactive lever presses (Figure 6N), motor coordination skills (rotarod), and body weights (Figure S4). Finally, to rule out potential confounding factors in sucrose pellet-directed behaviors, animals were tested for non-specific effects on sucrose preference. ^VP→VTACB1R cKO exhibited comparable homecage, non-contingent sucrose preference (Figure 6O). Overall, these experiments outline a crucial role of eCB signaling between VTA dopamine neurons and VP afferents not only for active avoidance behaviors but also for food-driven positive reinforcement, confirming that this feedback mechanism operates independently of stimulus valence and sensory modality.

Optostimulation of inhibitory VP→VTA terminals recapitulates ^VP→VTACB1R genetic deletion

Our electrophysiological recordings (Figure 4) indicate that one of the ways by which midbrain CB1R signaling can limit inhibitory synaptic drive onto VTA dopamine cells is by downregulating basal forebrain-VP inputs. We, therefore, hypothesize that cue-aligned optogenetic activation of inhibitory pallidal VTA terminals in WT mice will mimic the reduction of conditioned responding seen in eCB-deficient mice. To test this idea, we expressed ChR2 under the control of vGAT (AAV2/9-hVGAT-ChR2) in the entire VP of WT mice and implanted optic fibers in the parabrachial pigmented area of the VTA (Figures 7A-7D). We then trained ^vGATChR2 mice in the operant active avoidance task, pairing optogenetic stimulation (473 nm, 10 mW, 40 Hz) with every WS (3 s, from WS onset) for 7 sessions (Figure 7E). Compared to ^vGATChR2 mice receiving no optostimulation (laser OFF), activation of inhibitory VP terminals in the VTA substantially impaired operant active avoidance learning (Figure 7F). During the next 5 sessions, no optostimulation was delivered, rescuing active avoidance behavior in the group previously exposed to optostimulation (Figures 7F and 7G). In the next experimental phase (sessions 13–15), threat-paired optostimulation (3 s, from WS onset) was randomly delivered within subject in 50% of the trials (Figure 7H). We found that the average probability of avoiding was significantly reduced in trials coupled with laser activation (Figure 7H, right panel). In the next 3 sessions (Figure 7I), we shortened the laser activation to match the time interval (0.5–1.8 s, from WS onset) at which genotype and trial outcome showed significant coupling with GrabDA_2m NAc dynamics (β₃, from Figure 2J). Targeting this reduced temporal domain was sufficient to significantly hamper avoidance probability (Figure 7I, right panel).

Figure 7. Transient activation of inhibitory VP→VTA projections phenocopies ^VP→VTACB1R cKO learning impairments.

(A) Gene delivery and optostimulation strategy used to activate inhibitory VP terminals in the VTA using AAV2/9-hVGAT-ChR2 (^vGATChR2).

(B) Representative coronal brain section showing ChR2 (mCherry) expression in the VP, overlaid with substance P immunostaining (scale bar, 200 μm).

(C) Zoomed VP section showing colocalized expression of ChR2 (mCherry)—driven by an hVGAT promoter—and the GABAergic neuron marker GAD1 (scale bar, 10 μm).

(D) Representative VTA coronal brain section showing VP^vGATChR2 fibers (mCherry) innervating TH+ dopamine cells and fiber tract alignment (scale bar, 200 μm).

(E) Schematic representation of positively and negatively reinforced operant tasks (VTO and operant active avoidance, respectively) used to parse out conditioned responding in ^vGATChR2 mice.

(F) During operant active avoidance training (sessions 1–7), half of ^vGATChR2 mice received VTA laser stimulations time-locked to each warning signal (WS) presentation, leading to a significant decrease in the percentage of avoided trials per session (Bonferroni’s, **p = 0.001, *p = 0.015 vs. ^vGATChR2–Laser OFF, same session). Upon withdrawal of laser stimulation (sessions 8–12), previously stimulated ^vGATChR2 mice exhibited a progressive improvement in avoidance performance, reaching levels comparable to controls by session 12 (Bonferroni’s, ^n.s.p = 0.23, *p = 0.033 vs. ^vGATChR2–Laser OFF, same session).

(G) Maximum rates of active avoidance responding in the presence (sessions 1–7) and absence (sessions 8–12) of WS-paired laser stimulation (Bonferroni’s, **p < 0.01). ^vGATChR2–Laser OFF, n = 9; ^vGATChR2–Laser ON, n = 9.

(H) In subsequent sessions (13–15), all ^vGATChR2 mice received WS-paired (0–3 s, from cue onset) optostimulation in 50% of the trials, resulting in a decrease in active avoidance performance (paired t test, ***p = 7.5 · 10⁻⁵).

(I) A shortened laser activation (0.5–1.8 s, from WS onset)—targeting the greatest effect of ^VP→VTACB1R deletion on GrabDA_2m dynamics when avoiding—is sufficient to impair conditioned avoidance behavior (paired t test, ***p = 0.001). Line plot insets illustrate the outcome × genotype interaction effect previously described by FLMM. ^vGATChR2, n = 17.

(J) Alternatively, ^vGATChR2 mice were trained on the positively reinforced VTO task used to monitor conditioned response latencies. Animals receiving cue-paired optostimulation in each trial of this task (sessions 1–10) showed impaired learning of food-predictive cues (Bonferroni’s, **p < 0.01, *p < 0.05). Upon withdrawal of laser stimulation (sessions 11–15), previously stimulated ^vGATChR2 mice exhibited an improvement in response latencies, reaching levels comparable to controls by session 15 (Bonferroni’s, ^n.s.p = 0.22, *p < 0.05, **p < 0.01 vs. ^vGATChR2–Laser OFF, same session).

(K) Response latency on sessions 10 and 15 as a function of laser stimulation (Bonferroni’s, ***p < 0.001). ^vGATChR2–Laser OFF, n = 9; ^vGATChR2–Laser ON, n = 10.

(L) Increased conditioned response latencies by within-subject activation of inhibitory VP→VTA projections during CS+ presentation (sessions 16–19; paired t test, ***p = 1.8 · 10⁻⁵.

(M) A shortened laser activation (0–2 s, from CS+ onset)—targeting the greatest effect of ^VP→VTACB1R deletion on GrabDA_2m dynamics at short latencies—is sufficient to impair conditioned avoidance behavior (paired t test, *p = 0.011). ^vGATChR2, n = 18. Data shown as mean ± SEM.

We next investigated whether optogenetic activation of GABAergic VP terminals in the midbrain was sufficient to disrupt conditioned responding to appetitive cues (Figure 7E). Animals expressing ChR2 in VP GABAergic terminals were implanted with optic fibers in the VTA, and light stimulation was paired with every cue presentation (2 s from CS+ onset) during the VTO appetitive task. Compared to ^vGATChR2 mice without optostimulation, activation of VP GABAergic terminals in the VTA significantly suppressed rapid, conditioned lever pressing in response to reward-predictive cues (Figure 7J, sessions 1–10). When optostimulation was withdrawn for the following five sessions, conditioned response latencies rapidly recovered (Figures 7J and 7K). During the next 4 sessions, conditioned responding under ^vGATChR2 optostimulation was tested in a within-subject design by which every animal had a 50% chance of receiving light activation during CS+ (Figure 7L, left panel). Under these conditions, VP terminal activation in the VTA continued to significantly impair rapid conditioned responding (Figure 7L, right panel). Finally, we tested a shorter light stimulation window (0–2 s following CS+ onset)—mimicking the time window in which NAc dopamine and ^VP→VTACB1R caused faster reaction times (β3, interaction, Figure 6I) (Figure 7M, left panel). Crucially, this brief stimulation was sufficient to disrupt rapid cue-directed behavior (Figure 7M, right panel), demonstrating that transient activation of inhibitory VP→VTA projections suppresses reward-predictive responding in an exquisitely time-defined manner.

DISCUSSION

Organisms maximize their survival by rapidly mobilizing goal-directed behavioral resources in response to environmental stimuli predicting outcomes of differing valence. This adaptive process relies on the continuous updating of stimulus-action-outcome associations through subsecond dopamine release across the mesocorticolimbic system. Here, we uncover an eCB actuator favoring moment-to-moment coupling between accumbal dopamine release and cue-triggered behavior. We describe how rapid 2-AG/CB1R VTA signaling disinhibits dopamine cells by silencing incoming cannabinoceptive inputs from the basal forebrain, mostly the VP, which oppose conditioned responding. Supporting this view, we report (1) a loss of operant active avoidance responding in DGLa^TH cKO mice lacking the ability to mobilize 2-AG from VTA dopamine neurons, (2) a breakdown of positively and negatively reinforced cued responding in ^VP→VTACB1R cKO mice, and (3) a transient uncoupling between cue-triggered NAc dopamine release and conditioned responding in eCB-deficient mice.

Unlike appetitive cues, the basic properties of dopamine signals evoked by threat-predictive stimuli remain a subject of ongoing debate (reviewed in Lopez and Lerner⁴⁵). In the striatum, these signals are highly heterogeneous, often displaying opposite polarities depending on subtle anatomical and task variations between studies.^46-53 While it is proposed that such phasic fluctuations reinforce proactive strategies of fear avoidance,^54,55 evidence has remained elusive given the conflicting signals reported.⁵⁶ To resolve this conundrum, we adapted a statistical framework (FLMM)²⁸ that allowed us to look beyond summarized differences in amplitude of fluorescence signals and explore instead how time-resolved features of these transients evolve as escape/avoidance action selection takes place. Our findings show that higher dopamine release invariably predicts higher avoidance probability. Moreover, we also uncover the temporal structure of this relationship: dopamine fluctuations become transiently aligned with avoidance behavior within 400–800 ms of WS onset and dissipate shortly after cue offset. Crucially, this dopamine coupling precedes the avoidance response itself, hinting at the causal structure of the decision-making process guiding fear-related action selection. These results corroborate previous causal findings from our laboratory showing that optoactivation of VTA TH+ neurons drives avoidance behavior and NAc dopamine release.²⁰

Applying FLMM modeling also allowed us to assess how other factors (in this case, 2-AG mobilization in DGLa^TH cKO and WT mice) interacted with dopamine dynamics to explain avoidance probability. This linked 2-AG release from VTA dopamine neurons with the transitory coupling between dopamine and avoidance during cue presentation. Previous research from our group suggested this possibility, highlighting that pharmacological blockade of DGLa in the VTA abrogated active avoidance learning.²⁰ Here, we expand this finding by revealing that 2-AG synthesis by VTA dopamine cells is required for signaled operant avoidance only within a narrow temporal window, from WS identification through <600 ms after its offset. This suggests that midbrain eCBs are more than mere facilitators of dopamine cell firing⁵⁷; they temporally dictate (actuate) the polarity of dopamine release events and the information carried out downstream by them.

Multiple mechanisms can explain this pivotal role. However, rather than surveying all possibilities, we concentrated on one strong candidate. We report that basal forebrain projections to VTA TH+ cells contain high levels of CB1R expression³⁴ with VP neurons showing the highest levels within this region. The VP provides some of the strongest inhibitory and excitatory inputs to VTA dopamine neurons across the entire brain.^34,58,59 In line with this anatomical arrangement, cued active avoidance, and its rapid coupling with NAc dopamine dynamics, was robustly impaired in ^VP→VTACB1R cKO mice. This is consistent with the increase in escape responding observed after intra-VTA injections of rimonabant,²⁰ intra-NAc administrations of dopamine receptor antagonists,⁶⁰ and dopamine cell lesions.⁶¹ We build on this evidence by showing, ex vivo, that suppression of inhibition in the VTA by eCBs is fully attributable to the pool of CB1Rs located on VP axons—explaining the phenotypic similarity between ^VP→VTACB1R and DGLa^TH cKO mice, as well as the lack of tonic eCB response on GABA release in DGLa^TH cKO mice.⁶² Of note, the intersectional gene-delivery strategy employed did not allow us to rule out, in vivo, the participation of cannabinoceptive glutamatergic VP→VTA terminals. Nevertheless, considering that Gi-coupled CB1Rs are inhibitory,⁶³ the VP^GABA→VTA^TH pathway represents the most parsimonious possibility for our findings: disinhibition of a glutamatergic input would be expected to enhance, not diminish, dopamine output. While retrograde 2-AG/CB1R signaling has classically been associated with relatively slow kinetics, recent multiphoton imaging studies indicate that eCB-mediated depolarization-induced suppression of inhibition can occur on subsecond timescales.^64-66 One proposed mechanism enabling such rapid dynamics is the presence of a pre-existing presynaptic pool of extracellular microvesicles containing 2-AG.⁶⁷ Such mechanism would permit a rapid sequence of events by which VTA dopamine neuron burst firing drives 2-AG synthesis and release sufficiently quickly to activate CB1Rs located in inhibitory afferences (i.e., the VP) within ~400 ms of WS perception. Alternatively, dopamine neuron 2-AG release might operate more tonically, persistently dampening the inhibitory drive exerted by VP projections (among others). Such mechanism would favor fast excitatory afferents arising from sensory-processing structures, including the pedunculopontine tegmental nucleus,^68-70 dorsal raphe,⁵⁸ or superior colliculus.^71,72

In a broader context, the strong participation of the VP-originating eCB pathway that we uncover here refines our view of this structure’s complex and nuanced role in motivated behavior and learning. Historically viewed as an output relay of the basal ganglia circuitry,⁷³ the VP is now increasingly recognized as a core hub for generating predictive neural representations.^32,74-76 In our functional circuit mapping testing, VP GABAergic inputs modulated VTA dopamine neurons as expected: providing a strong source of synaptic inhibition. Accordingly, anticipatory responding was compromised in vivo by optostimulation of VP→VTA ^vGATChR2 terminals. Previous evidence reveals, however, a more complex picture of the VP, acting either as a driver or limiter of positively and negatively conditioned behaviors.⁷⁷ For instance, somatic optostimulation of VP GABA neurons curtailed active avoidance⁷⁸ but facilitated cued reward seeking.^78,79 VP glutamatergic cells, in contrast, are more generally implicated in responding to aversive outcomes and constraining reward seeking.^78-81 More relevant to our discussion, pharmacological lesions of VP→VTA projections disinhibited NAc dopamine release,⁸² similar to what we report here. To reconcile these seemingly contrasting lines of evidence, it is of paramount importance to consider that stress exposure, such as in our shock-reinforced task, strengthens the inhibitory influence of VP GABA cells onto VTA dopamine neurons.⁸³ We postulate that VP afferents carry the precise time-resolved information to integrate CB1R/2-AG signaling emanating from the VTA as an evolutionarily conserved coping mechanism enabling dopamine neurons to overcome the inhibitory drive linked to innate, non-specific defensive behaviors (i.e., escape)⁸⁴ and to support anticipatory avoidance responses. It is important to emphasize that the optostimulation strategy followed here proves the sufficiency of VP^vGAT→VTA terminals to suppress cued avoidance and sucrose seeking. However, these experiments do not establish that 2AG/CB1R signaling onto those terminals is itself sufficient to account for the observed behavioral effects.

Despite the striking effects of VP→VTA eCBs in dopaminergic encoding of threat predictive cues, our results suggest that errorlike signals remained intact in ^VP→VTACB1R and DGLa^TH cKO mice. That is, eCB-deficient mice displayed larger dopamine transients in avoided trials than control animals, implying that the outcome was perceived as unexpected/better than expected. However, this information did not reach downstream action selection processes required to minimize forthcoming punishments, especially in response to WSs. We, nonetheless, observed that ^VP→VTACB1R constituted a value-complete mechanism, contributing equally to operant associations involving appetitive and aversive outcomes. This is consistent with current and prior observations of impaired conditioned responding in DGLa^TH cKO mice trained to quickly secure a food reward¹⁸ or to avoid a foot shock (Figure 1).

Limitations of the study

The study of circuit-specific modulatory effects of eCB signaling is still in its early stages. For this reason, to ensure robust effects in the absence of prior reports, here we prioritized broad structure coverage and penetrance over subregional specificity. This approach limited our ability to resolve region- or subpopulation-specific contributions within the VTA and VP. While all VTA neurons express DGLa,⁸⁵ it may serve distinct functions across molecularly defined dopamine subpopulations distributed across the parabrachial pigmented area and paranigral, interfacial, rostral linear, and parainterfascicular nuclei.^86-89 Future studies should confine genetic eCB manipulations to defined VTA subregions to dissect these potentially divergent roles. Similarly, dopamine dynamics were monitored in a single subregion of the NAc core, guided by prior evidence.^20,90 Therefore, this design did not address possible differences in dopamine signaling within NAc subregions, including the shell, where recent photometric studies suggest markedly distinct dynamics during negative reinforcement.^49,45,90,91 Targeted recordings across NAc subdomains will be necessary to determine how eCB signaling differentially shapes dopamine release in these circuits.

Collectively, these data outline a multi-transmitter brain circuitry that fine-tunes anticipatory behaviors in response to threat- and reward-predictive cues. Phasic NAc dopamine fluctuations govern the selection and vigor of cued goal-directed responses, while VP→VTA GABA projections limit dopamine cell output, biasing behavior toward innate, unconditioned responses. Bridging these circuit nodes, cue-evoked retrograde 2-AG signaling from VTA dopamine neurons to CB1Rs located on pre-synaptic terminals transiently filters out inhibitory tone onto dopamine neurons, enabling anticipatory action selection. eCBs in the midbrain, therefore, are precise temporal filters (actuators) shaping continuous dopamine neuron computations. This insight not only broadens our understanding of eCB function but also highlights their potential as targets for treating conditions marked by maladaptive proactive coping strategies—such as anxiety disorders⁹²—or excessive motivational drive, including substance use disorders.⁹³

STAR METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

All experimental procedures conformed to the NIH Guide for the Care and Use of Laboratory Animals. Ethical approval was granted by the Institutional Animal Care and Use Committee (IACUC) at the University of Maryland, Baltimore (Protocol #00000054).

For all experiments, we used either wild-type (WT) C57BL6/J mice or conditional knockout (cKO) mice on a C57BL6/J background. Male and female mice were subjected to surgery procedures at 80-day of age and were 120-days old at the start of the behavioral procedures. DGLa^f/f mice (Dagla gene targeted) were obtained from Vanderbilt University⁹⁴ and maintained in-house by homozygote x homozygote breeding, as previously established.¹⁸ CB1^f/f (RRID:IMSR_JAX:036107, Cnr1 gene targeted) and DAT-cre (RRID:IMSR_JAX:006660, Slc6a3 gene targeted) mice were purchased from Jackson Laboratories and bred by homozygote x homozygote crossings. Animals were housed in a temperature- and humidity-controlled environment (24°C; 40–50% humidity) on a 12-h light/dark cycle (lights on at 07:00 h). All experiments were conducted during the light phase. The number of mice used in each experiment is reported in Table S1 and the corresponding figure legends. Sex was included as a factor in all parametric analyses. When no significant sex differences were detected, data from males and females were pooled, and subsequent analyses were conducted without including sex as a biological variable.

METHOD DETAILS

Stereotaxic surgeries

For all stereotaxic procedures, mice were anesthetized with isoflurane in O2 (4% induction and 1–2% maintenance, 2 L/min). Following skull exposure, a craniotomy (0.5 mm) was performed, and adeno-associated viruses (AAVs) were delivered into the target brain region using a microinjector pump. Viral injections (0.1 μL/min) used graduated pipettes (Drummond Scientific Company), broken back to a tip diameter of ~20 μm. Viral preparations were diluted and injected at 2-6×10¹² gc/ml. As previously validated,¹⁸ conditional DGLa knockout in dopamine neurons was induced by bilateral injections of AAV9-rTH-PI-Cre-SV40 (400-nL/side; Addgene plasmid #107788, packaged by UMB Vector Core) into the VTA (−3.3 AP, ±0.5 ML, −4.0 DV, mm relative to bregma) of DGLa^f/f mice. Control subjects (CTRL) for this manipulation were (a) WT mice injected with AAV9-rTH-PI-Cre-SV40 into the VTA as well as (b) SHAM DGLa^f/f mice receiving aCSF in the same structure. ^VP→VTACB1R cKO mice were generated by injecting AAV9-EF1α-fDIO-cre (RRID:Addgene_121675) bilaterally (500-nL/side) into the VP (+0.26 AP, ±1.5 ML, −4.55 DV, mm relative to bregma) and AAVrg-EF1α-mCherry-Flpo (RRID:Addgene_55634) bilaterally (500-nL/side) into the VTA of CB1^f/f mice. CTRL CB1^f/f mice received AAVrg-EF1α-mCherry-Flpo in the VTA and aCSF (500-nL/side) in the VP. To express GrabDA_2m, we unilaterally injected 500-nL of AAV9-hsyn-GrabDA4.4 (RRID:Addgene_140553) into the NAc (+1.2 AP, ±1.1 ML, −3.7 DV, mm relative to bregma). A 400-μm fiber-optic cannula (Doric, #MFC_400/430-0.48_4.0mm_ZF1.25_FLT) was placed 0.1-mm above the injection site. To generate ^vGATChR2 mice, WT animals were transduced bilaterally (400-nL/side) with AAV2/9-VGAT1-hChR2(H134R)-mCherry-WPRE (BrainVTA, #PT-0643) in the VP. In the same surgical procedure, a 200-μm fiber-optic cannula (Doric, #MFC_200/245-0.37_5mm_MF1.25_FLT) was implanted, unilaterally, into the VTA (−3.16 AP, ±0.99 ML, −4.24 DV; 7° ML inclination). For the functional circuit mapping slice experiments, bilateral injections (400-nL/side) of AAV5-hSyn-ChR2(H134R)-eYFP (Addgene viral prep #26973-AAV5) in the VP were conducted All fiber-optic implants were embedded within zirconia ferrules secured to the skull using dental cement containing charcoal powder. No skull-penetrating screws were used. For the rabies-virus tracing, we first unilaterally injected 30-nL of rAAV5-EF1a-FLEX-TVA-mCherry (UNC Vector Core) and rAAV8-CA-FLEX-RG (UNC Vector Core) in the VTA (−3.3 AP, ±1 ML, −4.5 DV; 10° ML inclination) of DAT-cre mice, using a 1:1 mixture cocktail. 14 days later, the same mice received 750-nL of the pseudotyped rabies virus SAD-eGFP(EnvA) (UNC Vector Core) 0.1mm above the VTA DV coordinate. After all stereotaxic procedures, mice were gently removed from the stereotaxic apparatus and placed on a heat pad in their home cages. They were singly housed and monitored daily for wound healing, with a recovery period of 4–6 weeks prior to experimentation.

Operant apparatus

All behavioral experiments were conducted in sound-attenuating chambers (21.6 × 17.6 × 14 cm; Med Associates) equipped with footshock grids, a pellet delivery receptacle, two retractable levers, cue lights positioned above each lever, a house light, and speakers for tone and white noise delivery. Behavioral programs were controlled by MedPC software.

Operant active avoidance

The operant active avoidance procedure was adapted to mice from our previous rat study.²⁰ Before active avoidance training, mice were shaped to press a lever to terminate (escape) a footshock (0.09-mA) in three daily 15-min sessions. At the start of each shaping session, subjects were presented with a lever, white noise (70-dB), and a cue light paired with continuous footshock. A lever press resulted in the immediate termination of footshock and the initiation of a 20-s “safety” period, signaled by a distinct set of multi-sensory cues: lever retraction, cue light dimming, house light illumination, white noise cessation, and tone presentation (2.5-kHz, 70-dB). Subjects were gradually shaped toward the lever by the experimenter until successful acquisition of the escape behavior.

After shaping, subjects underwent daily 30-min active avoidance training sessions. At the onset of each trial, the response lever was extended, and a warning signal (WS; cue light + white noise) was presented. A lever press during the 3-s window preceding footshock onset was recorded as an avoidance response and triggered the retraction of the lever, dimming of the cue light, cessation of white noise, and the onset of a 20-s safety period. During this period, the house light was illuminated, a tone was presented, and no footshock was delivered. If animals failed to respond within the 3-s avoidance window, recurring footshocks (0.5-s, 0.12 to 0.18-mA, delivered every 2-s) were administered until a lever press occurred. A response during this period was classified as an escape response, which immediately terminated footshock and initiated the 20-s safety period. Responses on the inactive lever were recorded but had no programmed consequences. Mice were trained for at least 12 sessions and a maximum of 18 sessions. Data from each subject’s final 12 training sessions were used for graphing and statistical analyses. When recording GrabDA_2m transients, animals remained untethered throughout the shaping phase and were connected to the patch cord during each operant active avoidance training session.

Variable time-out (VTO)

As previously stablished,¹⁸ mice first underwent three acclimation sessions in which sweetened grain pellets (Bio-Serv, F05684) pellets (sucrose) were non-contingently delivered under a VTO schedule (30 trials/session) with inter-trial intervals ranging from 30 to 60-s (average 45-s). Subsequently, mice were trained under an operant VTO reinforcement schedule. In this task, each trial began with illumination of a cue light above the active lever and presentation of a 2.5-kHz tone for 5-s prior to lever extension. Presses on the active lever resulted in immediate delivery of a sucrose pellet, if the response occurred within 120-s of lever extension. Failure to meet the response requirement led to lever retraction, cue light offset, and trial omission. Responses on the inactive lever were recorded but had no programmed consequences. Each session consisted of 30 trials, and 8 sessions were conducted in total. When fiber photometry recordings were performed, mice were tethered on every VTO session but remained untethered on the acclimatation sessions. Animals were food restricted to 90% of their initial body weights throughout the duration of the appetitive operant training.

Progressive ratio (PR)

A subset of VTO-trained mice (Figure 6) progressed to a PR schedule of reinforcement. This schedule was used to assess the effort mice were willing to exert to obtain a sucrose pellet reward. On each successive trial, the response requirement (number of lever presses) increased in a near-logarithmic fashion, following the equation:

$$\mathit{Response}\mathit{requirement}=\lceil 5\times e^{0.2\times \mathit{reward}\mathit{number}}-5\rceil$$ (Equation 1)

The response ratios for the first sixteen trials were: 1, 2, 4, 6, 9, 12, 15, 20, 25, 32, 40, 50, 62, 77, 95, and 118. At the start of each session, both active and inactive levers were extended, and both the house light and the cue light above the active lever were illuminated. Responses on the inactive lever were recorded but had no programmed consequences. Upon completion of the required number of responses, a sucrose pellet was delivered, followed by a 10-s time-out period during which the levers were retracted, lights were turned off, and a 2.5-kHz tone was played. PR sessions ended if no reward was obtained within 20-min. Mice underwent seven PR sessions in total.

Sucrose preference test

To discard changes in innate sucrose palatability across genotypes, a two-bottle sucrose preference test was conducted in the home cage of singly-housed mice. Animals were first habituated for 4 days to two identical bottles, both containing tap water. During the 4-day test phase, mice were given access to one bottle of tap water and one bottle of 1% sucrose solution. Bottles were refreshed daily, and their positions were alternated each day to control for side preference. Fluid consumption was recorded every 24-h, and sucrose preference was calculated as the average percentage of sucrose solution consumed over total fluid intake across the 4 days: [sucrose intake/(sucrose + water intake)] × 100.

Fiber photometry

Fiber photometry recordings were performed using two fiber-coupled LEDs (465-nm: Lx465; 405-nm: Lx405, Tucker-Davis Technologies) as excitation sources. Light from each LED was directed through a fluorescence minicube (FMC4, Doric Lenses) and delivered to the brain via a 2-m, 400-μm core optical fiber (Doric Lenses) connected to the implanted optic cannula. LED power was measured at the tip of the patch cord and adjusted to 5-μW (405-nm) and 10-μW (465-nm). Emitted GrabDA_2m fluorescence was collected through the same fiber, passed back through the minicube, and detected by a photoreceiver (LxPS1, Tucker-Davis Technologies). The resulting current signal was converted to voltage using a real-time processor (RZ10x, Tucker-Davis Technologies) and filtered using a sixth-order low-pass Butterworth filter (6-Hz cutoff) to eliminate high-frequency noise. TTL signals from the operant chamber were sampled at 6-kHz and recorded using Synapse software (Tucker-Davis Technologies).

GrabDA_2m signals were processed using the MATLAB script developed by Barker et al.⁹⁵ Noise-related fluctuations in fluorescence were corrected by using the 405 nm isosbestic control signal. The 405 nm trace was scaled and regressed onto the 465 nm ligand-sensitive signal to generate a noise model. This predicted noise waveform was then subtracted from the raw GrabDA_2mrecordings, effectively removing movement artifacts, photobleaching, and fiber-bending noise. Peri-event time histograms (PETH) were constructed using 10-ms GrabDA_2m bins surrounding the event of interest. Each PETH bin was z-scored by subtracting the mean fluorescence over a baseline period and dividing by the s.d. observed across the baseline periods (n = number of trials). Baseline periods were defined as −4s to −2s (relative to WS onset) for GrabDA_2m active avoidance PETHs, −2s to −1s (relative CS + onset) for GrabDA_2m VTO PETHs.⁵

In vivo optogenetics

Optogenetics was used to provide causal evidence for the opposing contribution of VP→VTA GABAergic terminals on conditioned approach and avoidance (Figure 7). For the active avoidance experiments, all WT mice received AAV2/9-hVGAT-ChR2 injections and were trained under the same conditions described above, being tethered only after completing the shaping sessions. Half of the animals were randomly assigned to the ‘Laser ON’ condition and, during the first 7 sessions, every WS led to delivery of a train of light stimulation. Light was delivered by a diode-pumped solid-state laser (473-nm, 10-mW, 40-Hz) coupled to a 200-μm, 0.37 NA fiber-optic cannula (Doric). To prevent prolonged light exposure during trials with slow responding, each trial was capped at a maximum duration of 20-s. Mice in the ‘Laser OFF’ condition were tethered but did not receive laser stimulation during the WS. From sessions 8 to 12, all animals were tested while tethered, but no optogenetic stimulation was delivered to either group. During sessions 13 to 15, all animals received light stimulation on 50% of randomly selected trials. Finally, during sessions 16–18, the length of the random light stimulation train was shortened to the first 1.3-s after WS onset.

Similarly, all WT animals trained on the appetitive conditioning task received AAV2/9-hVGAT-ChR2 injections and were trained under the same VTO conditions described above. For the first 10 sessions, half of the mice were randomly assigned to the ‘Laser ON’ condition and received light stimulation (473 nm, 10 mW, 40 Hz) for the entire duration of the food-predictive CS + cue. During sessions 11 to 15, no optostimulation was delivered to any animals. In sessions 16 to 18, all mice received light stimulation on 50% of randomly selected trials. Finally, in sessions 20 and 21, the random light stimulation train was shortened to 2-s, time-locked to CS + onset.

Slice electrophysiology

Four weeks after AAV5-hSyn-ChR2(H134R)-eYFP transductions, WT mice were deeply anesthetized with isoflurane and decapitated. The brains were then extracted and transferred to ice-cold N-methyl-D-glucamine (NMDG) cutting solution (in mM: NMDG, 93; KCl,2.5; NaH2PO4,1.2; NaHCO3, 30; HEPES, 20; Glucose, 25; Ascorbic acid, 5; Sodium pyruvate, 3; MgCl2, 10; CaCl2, 0.5). The tissue was blocked and then glued to the stage of a vibrating tissue slicer (Leica VT1200S, Leica Biosystems, Wetzler, Germany). Horizontal sections (220-μm; 2–3 per animal) were collected and transferred to an oxygenated (95% O2/5% CO2) holding chamber filled with HEPES-containing artificial cerebrospinal fluid (aCSF) (in mM: NaCl, 109; KCl, 4.5; NaH2PO4,1.2; NaHCO3, 35; HEPES, 20; Glucose, 11; Ascorbic acid, 0.4; MgCl2,1; CaCl2,2.5) maintained at 35°C for 10-min, followed by incubation at room temperature for at least 30-min prior to recording.

For recordings, slices were transferred to a chamber (RC-26; Warner Instruments, Hamden, CT) mounted to a fixed stage on a vibration isolation table (TMC Vibration Control, Peabody, MA). Slices were continuously perfused (2-mL/min) with oxygenated aCSF (in mM: NaCl, 126; KCl, 3; NaH2PO4, 1.2; NaHCO3, 26; Glucose, 11; MgCl2, 1.5; CaCl2, 2.4) using a peristaltic pump (Cole-Parmer, Vernon Hills, IL), and warmed to 30°C–32°C using an in-line solution heater (Warner Instruments). Recordings of neurons were performed in the lateral VTA, medial to the terminal nucleus of the accessory optic track (MT) and anterior to the third cranial nerve. Dopamine neurons were identified in the lateral VTA using electrophysiological criteria in cell-attached mode. Only cells demonstrating regular pacemaker firing (>3-Hz) and action potential widths >2.5-ms were chosen for further recording.⁹⁶ Fluorescence from enhanced yellow fluorescent protein (eYFP) that was co-expressed with ChR2 was visualized using an epifluorescence-equipped upright microscope (BX51WI, Olympus, Tokyo, Japan). The microscope was also equipped with differential interference contrast (DIC) optics, and a 900nm infrared light source to facilitate patch clamping of VTA neurons.

Whole-cell recording electrodes were fabricated using borosilicate pipette glass (Sutter Instruments, Novato, CA, 1.5-mm O.D. × 0.86-mm i.d.) and a horizontal puller (P-97; Sutter Instruments). They were filled with a potassium-based internal solution consisting of (in mM): K-gluconate, 140; KCl, 5; HEPES, 10; EGTA, 0.2; MgCl2, 2; Mg-ATP, 4; Na2-GTP, 0.3; Na2-phosphocreatine, 10), ~295–305 mOsm, neutralized to a pH of 7.2 using potassium hydroxide. Electrode resistances were 3–7-MΩ when filled with this solution. Whole-cell patch clamp recordings were performed using an Axon Instruments 700B Multi-Clamp amplifier (Molecular Devices, San Jose, CA). Unless otherwise indicated, neurons were voltage-clamped at −50-mV. Electrically evoked IPSCs were recorded in the presence of DNQX (10-μM) to block AMPA-mediated currents. For optogenetic stimulation, an LED driver (ThorLabs DC4104, Newton, NJ) was used to deliver 470-nm light through the microscope objective via a liquid light guide connected to the epifluorescence port. Electrically-evoked IPSCs were elicited by a bipolar stimulating electrode (FHC, Bowdoin ME) positioned 100–150 μm from the recorded neuron. A single current pulse (0.2-ms, 150 to 250-μA) was delivered via an isolated constant current stimulating unit (Digitimer DS-3, Hertfordshire UK). Signals were collected using WinLTP software (WinLTP Ltd, Bristol, UK) and an A/D board (National Instruments, Austin, TX, PCI-6251) housed in a personal computer. Hyperpolarizing voltage steps (−10-mV) were delivered via the recording electrode every 30-s to monitor whole-cell access and series resistance. Cells with an access resistance change of more than 30% were excluded from analysis. All drugs were dissolved at their final concentration in aCSF. MDMB-PICA is a high affinity synthetic cannabinoid agonist (Ki = 0.38-nM)⁹⁷ that has been demonstrated to potently inhibit synaptic transmission in the hippocampus via CB1Rs.⁹⁸ We chose MDMB-PICA over the widely adopted agonist WIN55,212-2 because of its greater affinity for CB1R and lack of off-target effects (inhibition of voltage-dependent Ca2+ channels) at concentrations typically used in brain slice experiments.^99-101 We therefore evaluated the actions of a 100-nM bath-application on both electrically evoked and ChR2-elicited IPSCs in VTA neurons. Pretreatment with the neutral CB1 antagonist NESS-0327 (Ki = 350-fM)¹⁰² (30-min, 2 μM) was used to block the effect of MDMB-PICA.

Rabies-virus-based transsynaptic retrograde tracing

Retrograde tracing

For retrograde monosynaptic tracing experiments, we used a two-step strategy that involved injection of rAAV5-EF1a-FLEX-TVA-mCherry and rAAV8-CA-FLEX-RG in the VTA of DAT-cre (+/+) mice. The AAV helper viruses contained the re-dependent mCherry tag, the cre-dependent rabies glycoprotein (RG) expression cassette and the tumor virus A (TVA) protein, which enabled cell type-specific infection of DATcre+ neurons. Mice were allowed to recover for 14 days and then received an injection of the pseudotyped rabies virus SAD-eGFP(EnvA) (1.72×10⁸ gc/ml) into the same site. Brain tissue samples were harvested 4 days later.

Multiplexed in situ hybridization and immunolabeling

Coronal free-floating sections (12-μm thickness) from 4 mice were processed as previously described.¹⁰³ Sections were incubated for 10-min in phosphate buffer (PB) containing 0.5% Triton X-100, followed by two 5-min rinses in PB. They were then treated with 0.2-N HCl for 10-min, rinsed again (2 × 5-min in PB), and acetylated in 0.25% acetic anhydride dissolved in 0.1-M triethanolamine (pH 8.0) for 10-min. After two additional 5-min rinses in PB, sections were fixed in 4% PFA for 10-min. Before hybridization, sections were incubated for 2-h at 55°C in hybridization buffer containing 50% formamide, 10% dextran sulfate, 5 × Denhardt’s solution, 0.62-M NaCl, 50-mM DTT, 10-mM EDTA, 20-mM PIPES(pH 6.8), 0.2% SDS, 250-μg/mL salmon sperm DNA, and 250-μg/mL tRNA. Hybridization was carried out for 16-h at 55°C using [35S]- and [32P]-labeled single-stranded antisense or sense probes against mouse Cnr1 (nucleotides 104–1561, Accession # NM_024394.2), at 107 cpm/ml. Post-hybridization, sections were incubated with 4-μg/mL RNase A for 1-h at 37°C, washed in 1 × SSC/50% formamide at 55°C for 1-h, followed by 0.1 × SSC at 68°C for 1-h. After SSC washes, sections were rinsed in PB and incubated for 1-h in PB containing 4% BSA and 0.3% Triton X-100. Immunolabeling was then performed by overnight incubation at 4°C with a sheep anti-TH monoclonal antibody (Cat. No. AB1542; Abcam; 1:500 dilution). Sections were rinsed 3 × 10-min in PB and processed using an ABC kit (Vector Laboratories, Burlingame, CA). For detection, sections were incubated for 1-h at room temperature in biotinylated secondary antibody (Cat. No. 713-155-003, Jackson; 1:200 dilution), rinsed in PB, and incubated for another hour with avidin–biotin–HRP complex (Figure 4) or donkey anti-sheep IgG (Cat. No. 713-155-003, Jackson; 1:50 dilution) for Figure S2. The peroxidase reaction was developed using 0.05% DAB and 0.003% H2O2. Free-floating sections were mounted onto coated slides, dipped in Ilford K.5 nuclear track emulsion (Polysciences, Inc., Warrington; 1:1 dilution in double-distilled water), and stored in the dark at 4°C for 4 weeks before development.

Immunohistochemistry (IHC)

Mice were anesthetized with isoflurane (5%) and transcardially perfused with 4% paraformaldehyde (PFA) in 0.1-M PBS (pH 7.4). After perfusion, brains were post-fixed in PFA overnight at 4°C. Coronal brain sections (40-μm thickness) were prepared using a vibratome (Leica). For all IHC procedures, sections were incubated for 30-min in PB containing 3% normal donkey serum (Jackson 017-000-121). GrabDA_2m immunolabeling was performed via a 2-h incubation with an anti-GFP rabbit antibody (Thermo Fisher Scientific, Cat. No. A-11122, RRID AB_221569; 1:200), followed by labeling with Alexa Fluor Plus 555-conjugated goat anti-rabbit secondary antibody (ThermoFisher, A-32732, RRID AB_2633281; 1:2000, 45 min). Cre recombinase immunostaining was achieved incubating VP sections with an anti-cre primary rabbit antibody (D7L7L)XP conjugated with Alexa Fluor 488 (Cell Signaling Technology, #15036; 1:100) overnight at 4°C. Substance P and GAD1 proteins were labeled using rabbit anti-substance P (ImmunoStar, #20065, RRID: AB_572266, 1:250) and chicken anti-GAD67/GAD1 (Abcam, #ab75712, RRID: AB_1310248, 1:200) antibodies, which were incubated overnight at 4°C. Secondary antibodies were goat anti-chicken 488 Alexa Fluor (ThermoFisher, A-11039, AB_2534096, 1:1000) and donkey anti-rabbit Alexa Fluor 647 (Jackson, #711607003, RRID: AB_2340626, 1:200) Finally, VTA dopamine cells were imaged using a mouse anti-TH (ImmunoStar, #22941, RRID:AB_572268, 1:1000), incubated overnight, and a donkey anti-mouse Alexa 647 secondary antibody (Jackson, #706545148, 1:1000).

RNAscope in situ hybridization (ISH)

ISH was used to determine the cell type-specific expression of Cre recombinase, Dagla, Th, Cnr1, Slc32a1 (vGAT) and Slc17a6 (vGLUT) mRNA transcripts in a cell type-specific manner. After deep anesthesia, mouse brains were rapidly extracted and snapfrozen at −80°C for coronal sectioning (14-μm thick) using a Leica microtome. The resulting brain sections were mounted onto glass slides (Fisher Scientific) and stored at −80°C until ISH was performed. RNAscope probes were obtained from ACDbio: Mm-CRE-C1 (Cre; cat. no. 312281), Mm-Slc17a6-C2 (vGLUT; cat. no. 319171-C2), Mm-Slc32a1-C3 (vGAT; cat. no. 319191-C3), Mm-Th-C2 (Th; cat. no. 317621-C2) and Mm-Dagla-C3 (Dagla; cat. no. 478821-C3). RNAscope assays were carried out in accordance with the manufacturer’s instructions. Stained sections were coverslipped with DAPI-containing mounting medium (Fluoroshield ab104139; Abcam) and imaged on an Olympus Fluoview confocal microscope at 20× magnification.

QUANTIFICATION AND STATISTICAL ANALYSIS

Functional linear mixed models (FLMM)

We used FLMM to evaluate the influence of genotype and behavioral outcome on GrabDA_2m fluorescent changes at every time-point of the trial. FLMM provides a suitable statistical framework for revealing transitory time-varying signal changes that would have been otherwise obscured by analyzing only summary measures (e.g., average z-scores during the WS). GrabDA_2m signals were analyzed using the R CRAN package fastFMM developed in Loewinger et al.²⁸ (see https://github.com/gloewing/photometry_FLMM for an analysis guide and installation instructions). This approach can be described conceptually as fitting a linear mixed model to the photometry signal at every trial time point, smoothing the estimates across trial time points, and then adjusting for the multiple comparisons by constructing a joint confidence interval.

First, trial-by-trial z-scored GrabDA_2m traces are pooled across trials and animals into one dataset and jointly analyzed. Denoting Y_ij(s) as the photometry value for animal i on trial j at trial time point s, we model the average signal as:

$$\mathbb{E}\left[{\mathrm{Y}}_{ij}(s)\mid {\mathbb{X}}_i,{\mathbb{Z}}_i,{\gamma }_i(s)\right]={\mathbb{X}}_i\beta (s)+{\mathbb{Z}}_i{\gamma }_i(s)$$ (Equation 2)

where $\mathbb{X}$ and $\mathbb{Z}$ are the design matrices for the fixed-effects (the independent variables tested; i.e., genotype) and random-effects (between- and/or within-subjects variance) regression coefficients, respectively. In all cases, random effects (γ_i) were set as animal-specific random intercepts (without random slopes). Using animal-specific random intercepts allowed the mean photometry signal (Z score) to vary between mice when all covariates in the model were equal to 0 (e.g., genotype ‘CTRL’ or outcome ‘escape’). To avoid biasing our model selection, alternative models including random slope effects were tested, but those yielded inferior fits (AIC/BIC values) and were therefore discarded.

yielded estimates for the regression coefficients of the independent variables tested, together with pointwise and joint 95% CIs at every trial time-point. When interpreting FLMM plots, β₁(s), β₂(s) and β₃(s) values should be read as the average GrabDA_2m change in z-scores at trial-timepoint s associated with a one unit increase in covariate x, while holding other covariates constant. Intervals during which 95% CIs did not include 0 were deemed significant (*p < 0.05). To account for multiple comparisons when inspecting coefficients across the entire trial, we report and plot a corrected 95% joint CI. All of our interpretations were based on the joint 95% CI, but pointwise CIs were used to select the time intervals for ANOVA testing using the Satterthwaite method.¹⁰⁴

To test the association between NAc dopamine release, genotype and behavioral outcome (escape or avoidance), as seen in Figures 2H-2J and 5L-5N, we use the formula syntax: photometry ~ outcome * genotype + (1 ∣ID). Because these covariates are binary, the fixed effects are interpreted as mean differences in the photometry signal between conditions at each trial timepoint (akin to a repeated-measures ANOVA at each timepoint). This fits the model:

$${\mathrm{Y}}_{ij}(s)={\beta }_0(s)+{\gamma }_{0,i}(s)+{\beta }_1(s){\mathit{outcome}}_{i,j}+{\beta }_2(s){\mathit{genotype}}_i+{\beta }_3(s){\mathit{outcome}}_{i,j}\times {\mathit{genotype}}_i+{ϵ}_{i,j}(s)$$ (Equation 3)

The coefficients are interpreted.

• β₀(s): The mean signal at trial timepoint s among CTRL group mice on escape trials.

• β₁(s): The mean difference in the signal at trial timepoint s between escaped and avoided trials among CTRL mice.

• β₂(s): The mean difference in the signal at trial timepoint s between CTRL and cKO mice on escape trials.

• β₃(s): The interaction term (β₃), interpreted as a difference in differences. For instance, β₃(s) < 0 denotes that the difference in the mean signal at trial timepoint s between avoided and escaped and trials is smaller in cKO compared to CTRL mice.

Similarly, we tested the association between the GrabDA_2m signal, genotype and response latency using the formula syntax: photometry ~ reaction time * genotype + (1 ∣ID), as shown in Figures 6G-6I and Figures S1C-S1E. This yielded the model:

$${\mathrm{Y}}_{ij}(s)={\beta }_0(s)+{\gamma }_{0,i}(s)+{\beta }_1(s)reactio{n}_{i,j}+{\beta }_2(s){\mathit{genotype}}_i+{\beta }_3(s){\mathit{reaction}}_{i,j}\times {\mathit{genotype}}_i+{ϵ}_{i,j}(s)$$ (Equation 4)

Where.

• β₀(s): The mean signal at trial timepoint s among CTRL group mice when responding at average latencies.

• β₁(s): The mean difference in the signal at trial timepoint s for a one second change in reaction time among CTRL mice.

• β₂(s): The mean difference in the signal at trial timepoint s between CTRL and cKO mice for trials when responding at average latencies.

• β₃(s): The interaction term β₃ interpreted as a difference in differences. For instance, β₃(s) < 0 denotes that the difference in the mean signal associated with a one unit change in reaction time is smaller in cKO compared to CTRL mice.

For these analyses, we calculated reaction times as reaction_i,j = 1 – latency_i,j, such that higher reaction times corresponded to higher dopamine release. Reaction time values were then centered to a mean of zero to improve interpretability of the model coefficient estimates.

Random forest classifiers

To track the emergence of predictive information in subsecond dopamine signals, we implemented a sliding-window decoding approach using random forest classifiers. GrabDA_2m fluorescence z-scores from individual trials were segmented into 50-ms time windows, sampled every 50-ms, spanning a −2 to 6-s peri-event time interval (relative to WS onset). For each time bin, a distinct classifier was trained using dopamine activity restricted to that interval and sampled across trials. This resulted in a series of temporally-specific models, each independently predicting trial outcome (escape vs. avoidance) based solely on information present within its respective 50-ms window. All classifiers were trained using 10-fold cross-validation: trials were partitioned into 10-folds, and the model was trained on 9-folds and tested on the remaining fold, cycling through all partitions. This procedure was repeated independently for each time window and each group (CTRL vs. cKO). Note that for each fold, the training and test datasets contained trials from all animals within that group. Finally, random forest classifiers were built using 100 decision trees per forest, and average accuracy was computed as the proportion of correctly predicted trials across all test folds.

To test whether the decoding accuracy difference between two model families (CTRL vs. DGLa^TH cKO) were statistically significant, we performed two-sided permutation tests at every time window. For each window, the null hypothesis was generated by randomly permuting group labels across trials 200 times, obtaining a shuffled distribution of differences in accuracies. This null distribution was compared against the actual distribution of accuracy differences observed between CTRL- and cKO-trained models on that trial window. p-values were calculated as the proportion of shuffled accuracy differences than the actual difference observed between CTRL-trained and cKO trained models. To account for multiple comparisons, p-values were corrected using the Benjamini–Hochberg false discovery rate (FDR) procedure (q-threshold = 0.01). FDR-corrected significant p-values indicated time windows in which CTRL GrabDA_2m traces contained more information to predict avoidance vs. escape outcomes than cKO GrabDA_2m traces. Of note, decoding accuracy peaked late into the WS period (2 to 3-s after its onset), when most animals had already avoided. Nonetheless, the earliest time interval at which CTRL GrabDA_2m traces outperformed cKO datasets was 0.4 to 0.45-s after WS onset. At this moment into the trial, only 1.8% of trials had resulted in avoidance. We therefore reason that the decoding ability of CTRL-trained classifiers at 0.4-s is not a byproduct of completed actions.

Statistics and reproducibility

Behavioral and dopamine data were analyzed using one- or two-way repeated measures (RM) ANOVAs, or unpaired t tests, with Tukey post-hoc corrections applied where multiple comparisons were conducted. For all parametric tests, Welch’s correction was used in instances where homoscedasticity was violated. In such cases, Dunnett’s post hoc tests were employed. Statistical significance was determined at p < 0.05 using two-tailed tests. Analyses were conducted using Prism 10.3 (GraphPad), MATLAB 2024b, and RStudio 2024. Detailed statistical results and sample sizes are provided in Table S1. No formal statistical methods were used to predetermine sample sizes. Comparable expression patterns and targeting accuracy were consistently observed across at least three independent biological replicates. Experimenters were blinded to genotype during all experimental procedures.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2026.117298.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-GFP	Thermo Fisher Scientific	Cat# A-11122; RRID AB_221569
Alexa Fluor Plus 555-conjugated goat anti-rabbit	Thermo Fisher Scientific	Cat# A-32732; RRID AB_2633281;
Rabbit anti-cre (D7L7L)XP conjugated with Alexa Fluor 488	Cell Signaling Technology	Cat# 15036
Rabbit anti-substance P	ImmunoStar	Cat# 20065; RRID: AB_572266
Chicken anti-GAD67/GAD1	Abcam	Cat# ab75712; RRID: AB_1310248
Goat anti-chicken 488 Alexa Fluor	ThermoFisher	Cat# A-11039; AB_2534096
Donkey anti-rabbit Alexa Fluor 647	The Jackson Laboratory	Cat# 711607003; RRID: AB_2340626
Mouse anti-TH	ImmunoStar	Cat# 22941; RRID:AB_572268
Donkey anti-mouse Alexa 647	The Jackson Laboratory	Cat# 706545148
Bacterial and virus strains
AAV9-rTH-PI-Cre-SV40	Addgene	plasmid #107788
AAV9-EF1α-fDIO-cre	Addgene	RRID:Addgene_121675
AAVrg-EF1α-mCherry-Flpo	Addgene	RRID:Addgene_55634
AAV5-hSyn-ChR2(H134R)-eYFP	Addgene	viral prep #26973-AAV5
SAD-eGFP(EnvA)	UNC Vector Core	N/A
rAAV8-CA-FLEX-RG	UNC Vector Core	N/A
rAAV5-EF1a-FLEX-TVA-mCherry	UNC Vector Core	N/A
AAV9-hsyn-GrabDA4.4	Addgene	RRID:Addgene_140553
AAV2/9-VGAT1-hChR2(H134R)-mCherry-WPRE	BrainVTA	Cat# PT-0643
Critical commercial assays
RNAscope™ Multiplex Fluorescent Assay	ACDBio	Cat# 323110
Deposited data
R script required to generate the random forest decoder in Figure 3		https://doi.org/10.5281/zenodo.19154695
Experimental models: Organisms/strains
Mouse: DGLaf/f	Vanderbilt University	NM_198114
Mouse: CB1f/f	The Jackson Laboratory	RRID:IMSR_JAX:036107
Mouse: DAT-cre	The Jackson Laboratory	RRID:IMSR_JAX:006660
Oligonucleotides
Mm-CRE-C1	ACDBio	Cat# 312281
Mm-Slc17a6-C2	ACDBio	Cat# 319171-C2
Mm-Slc32a1-C3	ACDBio	Cat# 319191-C3
Mm-Th-C2	ACDBio	Cat# 317621-C2
Mm-Dagla-C3	ACDBio	Cat# 478821-C3
Software and algorithms
MATLAB 2024b	MathWorks	https://www.mathworks.com/products/matlab.html
Prism 10.3	GraphPad	https://www.graphpad.com/
RStudio 2024	Posit	https://posit.co/download/rstudio-desktop/

Highlights.

• Dopamine neuron 2-arachidonoylglycerol is required for active avoidance

• Cue-evoked striatal dopamine release enables active avoidance

• CB1 receptors are found on inhibitory basal forebrain projections to the midbrain

• Endocannabinoid signaling couples dopamine release dynamics to action selection