The Periaqueductal Gray Selectively Supports Reversal Learning During a Flexible Discrimination Task in Mice

Using whole-brain fMRI in behaving mice combined with reinforcement learning modeling, this study reveals that the periaqueductal gray (PAG) is selectively recruited during reversal learning to suppress previously rewarded actions and update behavior, highlighting a novel role for this midbrain structure in cognitive flexibility beyond its traditional association with threat processing.

The Gist

Imagine your brain is a highly sophisticated navigation system, like a GPS in a car. Usually, this GPS works great: it tells you that "Turn Left at the big oak tree" leads to a delicious pizza (a reward). You learn this rule, and you drive there happily every day.

But what happens when the pizza place closes down and a new one opens up on the right? Your old GPS is now stuck on autopilot, trying to take you to the closed shop. To get the new pizza, you have to unlearn the old rule and learn a new one. This ability to change your mind when the rules of the game change is called cognitive flexibility.

For a long time, scientists thought this "re-routing" was handled mostly by the brain's "reward center" (the part that gets excited about pizza). But this new study, using a high-tech MRI scanner on mice, discovered a surprise player in this process: a tiny, ancient part of the brain called the Periaqueductal Gray (PAG).

Here is the story of what they found, broken down simply:

1. The Experiment: The "Sniff and Lick" Game

The researchers set up a game for mice inside an MRI machine.

• The Setup: Mice were trained to sit still while smelling two different scents.
• The Rule (Acquisition): Scent A meant "Lick the tube to get a water drop!" (Reward). Scent B meant "Don't lick, or you get nothing."
• The Twist (Reversal): Once the mice mastered the game, the researchers flipped the script. Suddenly, Scent A meant "Don't lick," and Scent B meant "Lick for water!"

The mice had to figure out that the old rules were broken and learn the new ones quickly.

2. The Old Theory vs. The New Discovery

Scientists already knew that the Nucleus Accumbens (let's call it the "Happy Button") lights up when a mouse does the right thing and gets a reward. It's like a cheerleader shouting, "Yes! Do that again!"

However, when the rules changed, the researchers noticed something strange happening in the PAG.

• The PAG's Job: Think of the PAG as the brain's "Brake Pedal" or "Stop Sign."
• What it did: When the mouse successfully stopped itself from licking at the wrong scent (the "No-Go" signal), the PAG lit up like a Christmas tree.
• The Twist: The PAG didn't care about the reward; it cared about inhibition. It was the part of the brain saying, "Wait! Don't do that old thing anymore! Stop!"

3. The "Double Dissociation" (The Perfect Team)

The study found a beautiful teamwork dynamic between the two brain regions during the rule change:

• The Happy Button (Nucleus Accumbens): Shouted "GO!" when the mouse licked at the new correct scent.
• The Brake Pedal (PAG): Shouted "STOP!" when the mouse resisted licking at the old wrong scent.

It's like driving a car: You need the gas pedal (Happy Button) to move forward toward the new goal, but you desperately need the brakes (PAG) to stop yourself from crashing into the old, familiar path.

4. Why This Matters

Here is the most surprising part: The PAG only showed up when the rules changed. When the mice were just learning the game for the first time, the PAG was quiet. It only woke up when the mice had to unlearn something they already knew.

This suggests that the PAG isn't just for dealing with pain or fear (which is what we used to think it did). Instead, it's a crucial part of our ability to be flexible. It helps us suppress our habits when the world changes, even if no one is punishing us for being wrong—we just have to realize the old way doesn't work anymore.

The Big Picture

This study is like finding a new, essential gear in a watch that everyone thought was already perfect. It shows that our ability to adapt isn't just about chasing rewards; it's equally about having a specialized "brake system" in our brain that helps us let go of the past.

By using MRI to watch the whole mouse brain at once, the researchers found that cognitive flexibility is a team sport involving both the "Go" signals and the "Stop" signals working in perfect harmony. Without that "Stop" signal from the PAG, we would all be stuck driving in circles, trying to find a pizza place that closed years ago.

Technical Summary

1. Problem Statement

Flexible, goal-directed behavior requires the ability to update value representations when environmental contingencies change (cognitive flexibility). While reinforcement learning (RL) mechanisms are well-studied, particularly the role of the dopaminergic ventral tegmental area (VTA) to ventral striatum pathway, the full distributed brain network supporting these processes remains unclear.

• Limitations of Current Methods: Human fMRI studies provide whole-brain insights but lack causal control and cellular resolution. Rodent studies often rely on resting-state fMRI or invasive single-unit recordings, which lack the systems-level perspective needed to map distributed networks.
• Knowledge Gap: The specific neural circuits supporting the transition from initial acquisition to reversal learning (switching rules) in rodents, particularly the involvement of non-canonical regions like the midbrain Periaqueductal Gray (PAG), are not fully characterized. The PAG is traditionally linked to aversion and threat processing, but its role in appetitive value-based decision-making and cognitive flexibility is underexplored.

2. Methodology

The study employed a multimodal approach combining high-resolution task-based fMRI in behaving mice with computational reinforcement learning modeling.

• Subjects & Setup:

  • Subjects: 12 male C57Bl/6J mice.
  • Task: A go/no-go odor discrimination task. Mice learned to lick for a rewarded odor (S+) and withhold licking for an unrewarded odor (S-).
  • Phases:
    1. Acquisition: Mice learned the initial contingencies (n=12).
    2. Reversal: A subset (n=6) underwent a rule reversal where S+ and S- contingencies were switched.
  • Hardware: A custom, MRI-compatible behavioral platform allowed for precise odor delivery (1s duration), water reward control (~2.5 µl), and simultaneous high-resolution recording of sniffing and licking behavior.
  • Imaging: 9.4 Tesla MRI scanner. Event-related fMRI using Spin-Echo EPI sequences (TR=2.5s, voxel size 200×200×300 µm³).

• Computational Modeling:

  • Algorithm: A model-free Q-learning algorithm was used to estimate trial-by-trial state-action values ($Q$) for licking vs. not licking.
  • Parameters: Learning rate ($\alpha$), intercept ($\beta_0$), and inverse temperature ($\beta_1$) were estimated using hierarchical Bayesian inference (Stan).
  • Regressor: The Decision Variable ($\Delta V = Q_{lick} - Q_{no-lick}$) was used as a parametric modulator in the General Linear Model (GLM) to identify brain regions tracking value updates.

• Data Analysis:

  • Whole-Brain Analysis: Group-level parametric (n=12) and non-parametric permutation tests (n=6) were performed to map BOLD correlates of the decision variable.
  • Region of Interest (ROI): Finite Impulse Response (FIR) analyses were conducted on the Nucleus Accumbens (ACB) and PAG to dissect responses to specific behavioral outcomes: Hit, Miss, False Alarm (FA), and Correct Rejection (CR).

3. Key Contributions

1. Methodological Advancement: Demonstrated the feasibility of using task-based fMRI with computational modeling in awake, behaving mice to dissect dynamic learning processes at a mesoscale resolution.
2. Discovery of PAG Role: Identified the Periaqueductal Gray (PAG) as a critical, previously unrecognized node in the circuit for reversal learning, specifically in the suppression of previously rewarded actions.
3. Functional Dissociation: Revealed a "double dissociation" between the Nucleus Accumbens (ACB) and PAG during reversal learning, showing they encode opposing signals (approach vs. suppression).
4. Appetitive Context: Showed that PAG recruitment occurs in an appetitive task without explicit punishment, challenging the view that PAG is solely an aversive/threat-processing center.

4. Key Results

• Behavioral Performance: Mice successfully learned the initial discrimination (reaching >80% accuracy) and rapidly adapted to the rule reversal (reaching criterion within ~2 sessions). Reaction times decreased as proficiency increased.
• Acquisition Phase (n=12):
  • Whole-brain analysis correlated the decision variable with activity in canonical reward and sensory regions: Nucleus Accumbens (ACB), Dorsomedial Striatum (DMS), Orbitofrontal Cortex (ORB), and olfactory/gustatory areas.
  • PAG was NOT significantly engaged during initial acquisition, even when performance levels were comparable to the reversal phase.
• Reversal Phase (n=6):
  • In addition to striatal and cortical regions, the PAG showed significant positive correlation with the decision variable during reversal.
  • Opposing Activity Patterns (ROI Analysis):
    • Nucleus Accumbens (ACB): Showed strong activation for Hits (licking to S+) and inactivation for Correct Rejections. This aligns with its role in positive prediction errors and approach behavior.
    • Periaqueductal Gray (PAG): Showed the opposite pattern: strong activation for Correct Rejections (withholding licking to S-) and inactivation for Hits.
  • Specificity: PAG activation was specific to the reversal phase. It did not track the decision variable during the acquisition phase, indicating its recruitment is tied to the updating of rules rather than general value learning.

5. Significance

• Redefining PAG Function: The study expands the functional repertoire of the PAG beyond defensive behaviors and aversion. It positions the PAG as a key midbrain hub for cognitive flexibility, specifically mediating the suppression of outdated actions when contingencies change, even in the absence of explicit punishment.
• Circuit Mechanism: The findings suggest a complementary coding scheme where the striatum (ACB) drives the exploitation of current rewards, while the PAG drives the inhibition of previously reinforced but now invalid actions. This opponency facilitates rapid behavioral adaptation.
• Translational Potential: By validating task-fMRI in rodents combined with computational modeling, the study provides a powerful framework for bridging the gap between human systems neuroscience and rodent cellular/molecular mechanisms. This is crucial for understanding neuropsychiatric disorders characterized by impaired behavioral flexibility (e.g., OCD, addiction, depression).
• Theoretical Implication: The results support the hypothesis that the PAG acts as a general mediator of adaptive response suppression, potentially biasing dopaminergic signaling to update value representations when the environment becomes volatile.