Coordinated yet asymmetric striatal neuromodulatory dynamics encode associative learning

This study demonstrates that dopamine and acetylcholine in the anterior dorsolateral striatum form a hierarchically coupled system where fast, event-locked dopamine signals act as a directional temporal scaffold to organize sustained acetylcholine fluctuations, thereby encoding associative learning states through coordinated low-dimensional dynamics rather than generic motor output.

The Gist

Imagine your brain's learning center (the striatum) as a busy, high-tech construction site. To build new habits and learn from rewards, this site needs two main foremen working together: Dopamine (DA) and Acetylcholine (ACh).

For a long time, scientists thought these two foremen worked independently, like two separate construction crews just happening to be in the same building. This new paper, however, reveals they are actually a highly coordinated, albeit slightly uneven, team. They work together to turn a "random guess" into a "learned skill," but they do it in very different ways.

Here is the story of their partnership, broken down into simple concepts:

1. The Two Foremen with Different Styles

Think of Dopamine as the Flashy Architect.

• Style: Fast, sharp, and precise.
• Job: When a mouse hears a sound that predicts a treat, Dopamine fires a quick, bright "flash" of signal. It's like a camera shutter snapping a photo of the exact moment a reward is coming.
• Change over time: At first, it flashes for everything (even random noises). But as the mouse learns, the Architect gets smarter. It only flashes when the specific "correct" sound happens. If the sound doesn't lead to a treat, the Architect stops flashing.

Think of Acetylcholine as the Steady Site Manager.

• Style: Slow, broad, and sustained.
• Job: Instead of a quick flash, Acetylcholine creates a long, steady "hum" or background noise that changes shape over time. It's like the manager walking the site, checking the mood, and adjusting the overall energy level.
• Change over time: As learning happens, this "hum" gets more complex. It doesn't just sit there; it starts to rise and fall in specific patterns that match the learning process.

2. The "Low-Dimensional Manifold": The Secret Dance Floor

The researchers didn't just look at how loud the signals were (amplitude); they looked at the shape and timing of the signals.

Imagine the signals from Dopamine and Acetylcholine as two dancers.

• Before learning: They are dancing randomly, bumping into each other, with no rhythm.
• After learning: They find a secret dance floor (what the scientists call a "low-dimensional manifold"). On this floor, their movements are perfectly synchronized. Even if you can't see the individual steps clearly, the pattern of their dance tells you exactly what the mouse is thinking: "I know this cue means a treat is coming!"

The Big Discovery: You can predict whether a mouse has learned a task just by watching this "dance pattern." If the dance is chaotic, the mouse is guessing. If the dance is coordinated, the mouse has learned.

3. Who Leads the Dance? (The Asymmetry)

Here is the most surprising part: Dopamine leads, and Acetylcholine follows.

The researchers used a statistical tool called "Granger Causality" (think of it as a predictive crystal ball) to see who influences whom.

• The Result: The crystal ball showed that Dopamine's movements predict what Acetylcholine will do next. But Acetylcholine's movements do not predict what Dopamine will do.
• The Metaphor: Imagine a conductor (Dopamine) waving a baton. The orchestra (Acetylcholine) hears the baton and adjusts their playing accordingly. The orchestra doesn't tell the conductor when to wave the baton; the conductor sets the tempo, and the orchestra follows.

This means Dopamine acts as the temporal scaffold—a rigid, reliable frame of time that organizes the more flexible, wobbly movements of Acetylcholine.

4. Learning vs. Just Moving

The study also looked at what happens when the mouse is just licking a water spout randomly (not during a learning task).

• Random Licking: The signals are just about "how hard" the mouse is licking (amplitude). It's like a volume knob being turned up and down. There is no complex dance pattern here.
• Learned Licking: When the mouse is anticipating a reward, the signals form that complex, coordinated dance pattern again.

The Takeaway: The brain uses this special "dance pattern" specifically for learning and memory, not just for moving muscles.

Summary: The New Picture of Learning

Before this paper, we thought Dopamine and Acetylcholine were two separate radio stations broadcasting different messages.

This paper shows they are actually a single, hierarchical system:

1. Dopamine is the Director. It provides the precise timing and the "aha!" moment when a prediction is correct.
2. Acetylcholine is the Stage Crew. It adjusts the lighting, the mood, and the background energy to support the Director's vision.
3. Together, they create a synchronized performance that allows the brain to turn a simple cue into a learned habit.

If you want to understand how we learn, you can't just look at one of them. You have to watch how they dance together, with Dopamine leading the way.

Technical Summary

1. Problem Statement

The striatum is a critical hub for transforming cortical inputs into flexible motor actions, a process heavily reliant on the interplay between dopamine (DA) and acetylcholine (ACh). While DA is known to encode reward prediction errors (RPE) and ACh acts as a gain controller, the principles governing their in vivo coordination during adaptive learning remain unclear.

• Limitations of Prior Work: Previous studies relied on indirect approaches (e.g., loss-of-function ablations suggesting independence) or constrained analyses (e.g., trial-averaged responses or brief recording windows). These methods obscured trial-by-trial temporal heterogeneity, cross-channel covariance, and the dynamic, directional structure of DA-ACh interactions essential for encoding complex learning states.
• Core Question: How do DA and ACh coordinate their activity temporally and structurally to jointly organize striatal computation during the trajectory of associative learning?

2. Methodology

The authors employed a multi-modal approach combining longitudinal in vivo recording with advanced computational modeling:

• Experimental Model:
  • Subjects: Cre-negative mice (to avoid confounding effects of Cre-recombinase) expressing dual fluorescent sensors in the anterior dorsolateral striatum (aDLS).
  • Sensors: Green GRAB-ACh4h (acetylcholine) and Red GRAB-DA2m (dopamine).
  • Technique: Longitudinal dual-color fiber photometry across 15 sessions of Pavlovian cue-reward conditioning.
  • Groups:
    • Paired Group ($n=13$ for photometry): Cue (1.5s LED) consistently followed by reward (4μL sucrose).
    • Unpaired Group ($n=10$ for photometry): Cues and rewards delivered randomly with no contingency.
• Behavioral Analysis:
  • Lick Clustering: Trials were clustered based on inter-lick intervals (ILI) around cue onset to distinguish Learned Trials ($t_{cL}$) (anticipatory, time-locked licking) from Not-Learned Trials ($t_{cNL}$).
  • Burst Analysis: Spontaneous licking during waiting periods was categorized into "short" ($\le$ 4 licks) and "long" ($>$ 4 licks) bursts.
• Computational & Statistical Framework:
  • Functional Regression: Used Generalized Additive Mixed-Effects Models (GAMM) and functional regression (pffr) to model temporal signal evolution across learning stages (Cue-Only, Early, Late).
  • Low-Dimensional Manifold Analysis: Applied Principal Component Analysis (PCA) and k-means clustering on bootstrapped trial-wise traces to extract temporal "motifs."
  • Decoding: Trained logistic regression decoders using motif similarity scores to classify learned vs. unlearned states and burst sizes.
  • Causality: Performed bivariate Granger Causality (GC) analysis with time-reversal controls and sliding-window analysis to determine directional influence between DA and ACh.

3. Key Contributions

1. Dual-Color Longitudinal Recording: First simultaneous, trial-by-trial tracking of DA and ACh dynamics throughout the full trajectory of associative learning in the aDLS.
2. Manifold-Based Decoding: Demonstrated that low-dimensional temporal motifs (covariance structure) are superior to scalar amplitude measures for decoding learned behavioral states.
3. Temporal Asymmetry: Established a robust, unidirectional causal hierarchy where DA acts as a predictive scaffold for ACh dynamics.
4. Dissociation of Motor vs. Cognitive Encoding: Showed that while motor output (lick bursts) is encoded via amplitude scaling, associative learning states are encoded via coordinated low-dimensional reorganization.

4. Key Results

A. Behavioral and Signal Divergence

• Behavior: Paired mice developed stereotyped, anticipatory licking ($t_{cL}$) over sessions, while unpaired mice did not.
• DA Dynamics: Fast, phasic responses (<500 ms). In paired mice, cue-evoked DA transients remained stable; in unpaired mice, they collapsed (novelty response faded without contingency).
• ACh Dynamics: Broad, sustained modulation. Paired mice developed a structured, biphasic ACh response (early pause followed by a post-cue peak) that was absent in unpaired controls.

B. Low-Dimensional Motifs and Decoding

• Motif Extraction: PCA revealed that ACh signals had a more stereotyped temporal organization (higher variance explained by PC1-2) compared to DA.
• Decoding Performance:
  • A decoder trained on motif similarity scores (combining cue/reward periods from both DA and ACh) achieved AUC = 0.76 in distinguishing learned vs. unlearned trials.
  • Generalization: The decoder generalized well across mice (Leave-One-Mouse-Out) but failed to classify unpaired trials as "learned," confirming the manifold is specific to associative contingency.
  • Feature Importance: ACh motifs provided the strongest within-group decoding power, while DA cues provided unique predictive value. The divergence between paired and unpaired groups was driven primarily by the absence of stabilized DA cue motifs in unpaired mice.

C. Spontaneous Lick Reorganization

• Burst Structure: Learning reorganized spontaneous licking, shifting from diffuse bursts to structured, prolonged bouts.
• Neuromodulatory Signatures:
  • DA: Showed a conserved, training-independent suppression at burst initiation (training history did not alter this).
  • ACh: Showed a training-dependent amplification. Long bursts in paired mice elicited significantly larger ACh elevations than in unpaired mice, indicating ACh encodes motor vigor in a state-dependent manner.

D. Granger Causality: Directional Asymmetry

• Directionality: DA reliably predicts subsequent ACh fluctuations ($DA \rightarrow ACh$), but ACh does not predict DA ($ACh \nrightarrow DA$).
• Magnitude: The predictive gain (Wald F-statistic) for $DA \rightarrow ACh$ was significantly higher than the reverse.
• Temporal Specificity: Significant coupling was concentrated in the 0–500 ms lag range. Time-reversal controls abolished this effect, confirming it is not due to shared slow drifts.
• Stability: The $DA \rightarrow ACh$ influence was stable across 65% of sliding windows, whereas the reverse was intermittent (~12%).

5. Significance and Conclusion

This study reframes the relationship between striatal dopamine and acetylcholine from independent channels to a hierarchically coupled system:

• DA as the Scaffold: Dopamine acts as a directional, temporal scaffold that organizes ACh dynamics. It provides the consistent, learning-dependent temporal structure required to define the "associative manifold."
• ACh as the Flexible Executor: Acetylcholine undergoes broader, more variable reorganization, acting as a gain controller that adapts to the learned state scaffolded by DA.
• Computational Insight: Learning is not merely a change in average signal amplitude but a reorganization of low-dimensional covariance structure. This manifold organization preferentially encodes learned cognitive states rather than generic motor outputs.

These findings resolve previous contradictions regarding DA-ACh independence by showing that while they may operate independently in loss-of-function contexts, they are temporally asymmetric and functionally coupled during active learning, with DA driving the predictive structure of the cholinergic system. This framework offers a new lens for understanding striatal plasticity in health and disease.