Distinct involvements of the subthalamic nucleus subpopulations in reward-biased decision-making in monkeys

This study reveals that distinct subpopulations within the monkey subthalamic nucleus differentially encode sensory evidence, reward expectations, and decision-related dynamics during complex reward-biased decision-making, thereby delineating their specialized roles in forming and evaluating choices.

The Gist

Imagine your brain is a bustling city, and the Subthalamic Nucleus (STN) is a highly sophisticated, multi-purpose traffic control tower. For years, scientists knew this tower helped manage the flow of traffic (movement) and prevented drivers from running red lights (impulsivity). But they weren't sure exactly how it managed the complex decisions we make every day, like choosing between a quick, small snack or waiting for a bigger, tastier meal.

This paper is like a team of detectives who finally got a clear look inside the control tower while it was handling a very tricky traffic jam. Here is what they found, broken down into simple stories:

The Big Experiment: The "Juice Bar" Game

The researchers trained two monkeys to play a game. They had to watch a cloud of moving dots and decide which way the dots were flowing.

• The Twist: Sometimes, choosing "Left" gave a tiny drop of juice, while "Right" gave a huge splash. Other times, it was the opposite.
• The Goal: The monkeys had to balance the evidence (which way are the dots moving?) with the reward (which choice gets me the big juice?).

While the monkeys played, the scientists recorded the electrical signals of 156 individual neurons inside the STN. They wanted to see if these neurons were just "traffic cops" or if they were also "rewards managers."

The Discovery: It's Not One Big Crowd; It's Three Different Teams

Previously, scientists thought the STN was a bit of a jumbled mess. But this study found that the STN isn't one big group of identical workers. Instead, it's made up of three distinct teams (subpopulations), each with a very different job description, even though they all sit in the same building.

Think of it like a restaurant kitchen:

1. The "Speed Bump" Team (Cluster 1):

   • What they do: These neurons fire up early in the decision process.
   • The Analogy: Imagine a construction crew putting up a "Slow Down" sign or a speed bump on the road. Their job is to tell the brain, "Hey, don't rush! The evidence is shaky, so let's raise the bar before we commit to a choice."
   • The Result: They control the decision threshold. They decide how much proof you need before you make a move. If the reward is huge, they might lower the bar; if the evidence is weak, they raise it.

2. The "Bias" Team (Cluster 2):

   • What they do: These neurons are the most picky about which choice you make. They fire strongly when the monkey chooses a specific direction.
   • The Analogy: Imagine a salesperson who really wants you to buy a specific product. They don't just wait for you to decide; they actively tilt the scales in their favor.
   • The Result: They handle the reward bias. If the "Right" choice has a big reward, this team whispers to the brain, "Hey, the evidence for 'Right' is actually stronger than it looks!" They help the brain weigh the evidence differently based on how good the prize is.

3. The "Cleanup Crew" Team (Cluster 3):

   • What they do: These neurons fire late, right around the time the monkey makes the eye movement (saccade).
   • The Analogy: Think of them as the stagehands who come in right before the curtain falls to make sure the lights are on and the props are ready. They handle the final "non-decision" stuff, like the time it takes to actually move your eye after you've made the choice.
   • The Result: They manage the reaction time and the final motor execution, ensuring the decision translates into action smoothly.

The "Mixing Bowl" Surprise

Here is the most fascinating part: These three teams are completely mixed up.
If you looked at a map of the STN, you wouldn't see Team 1 on the left and Team 2 on the right. They are intermingled like a bowl of M&Ms. A neuron from the "Speed Bump" team might be sitting right next to a neuron from the "Bias" team.

This explains why previous experiments (like electrical stimulation) were confusing. If you zap the STN with electricity, you accidentally zap all three teams at once, making it impossible to tell which one was doing what. This study used a new method (listening to individual neurons) to finally sort them out.

Why Does This Matter?

This research changes how we understand the brain's decision-making process:

• It's not a single switch: Making a decision isn't just one part of the brain flipping a switch. It's a coordinated effort between different specialized teams.
• We are complex: Our brains don't just look at facts (the dots); they also look at feelings and rewards (the juice). The STN is the place where these two things get mixed together.
• Better Treatments: Since this area is often targeted for Deep Brain Stimulation (DBS) to treat Parkinson's disease, understanding that there are different "teams" inside could help doctors tune their treatments more precisely. Maybe they can target the "Speed Bump" team to help with impulsivity without messing up the "Bias" team's ability to make smart choices.

In a nutshell: The STN is a busy, mixed-up control tower where three different types of workers collaborate. One team sets the rules for how much proof you need, one team tilts the scales based on the reward, and the third team gets you moving once the decision is made. Together, they help us navigate a world full of choices and temptations.

Technical Summary

1. Problem Statement

The subthalamic nucleus (STN) of the basal ganglia is well-known for its role in motor control and is a primary target for deep brain stimulation in Parkinson's disease. However, its role in cognitive functions, specifically complex decision-making involving the integration of sensory evidence and reward expectations, remains poorly understood.

• Gap in Knowledge: Previous studies identified that the STN contains functionally heterogeneous subpopulations of neurons. However, it was unclear how these specific subpopulations contribute to the computational mechanisms of decision-making (e.g., evidence accumulation, decision bound adjustment, and non-decision processes) when decisions are biased by reward.
• Limitation of Previous Methods: Prior attempts to distinguish these roles using electrical microstimulation were unsuccessful because stimulation likely affected intermingled subpopulations with distinct functions simultaneously.

2. Methodology

The authors employed a combination of single-unit electrophysiology, behavioral modeling, and advanced statistical clustering in two rhesus monkeys.

• Behavioral Task:
  • Monkeys performed a random-dot motion discrimination task.
  • Asymmetric-Reward Condition: Unlike previous equal-reward tasks, this version manipulated reward context. In alternating blocks, one choice direction was paired with a large reward and the other with a small reward.
  • Variables: Motion coherence (evidence strength), motion direction, and reward context (large vs. small reward for specific choices) were randomized.
• Data Collection:
  • Recorded from 156 STN neurons (93 from Monkey C, 63 from Monkey F).
  • Only neurons with robust task-related modulation were selected for deep analysis (87 neurons met a z-score threshold >1.5 for activation or suppression).
• Computational Modeling:
  • Drift-Diffusion Model (DDM): Behavioral data (Choice and Reaction Time) were fitted to a DDM to extract latent decision parameters:
    • Decision Variable (DV) components: Evidence scaling ($k$), evidence bias ($m_e$), starting point bias ($z$).
    • Bound components: Initial bound height ($a$), bound collapse rate ($B_{collapse}$), and delay ($B_d$).
    • Non-decision time: Sensory/motor delays ($t_0$).
  • Pseudo-session Analysis: Trials were split based on median firing rates to correlate neural activity with specific DDM parameters.
• Statistical Analysis:
  • Clustering: Neurons were grouped into subpopulations using k-means and hierarchical linkage clustering based on high-dimensional firing rate vectors and regression coefficients.
  • Regression: Multiple linear regression was used to link neural activity to decision factors (choice, coherence, reward context, and their interactions).
  • Evaluation Signals: Partial Spearman correlations were calculated to separate neural encoding of choice accuracy vs. reward expectation.

3. Key Contributions

1. Functional Subpopulations: The study definitively identifies three distinct functional subpopulations of STN neurons that are anatomically intermingled but computationally distinct.
2. Integration of Reward and Evidence: It demonstrates that STN neurons do not just process sensory evidence; they integrate reward expectations to bias decision formation.
3. Mapping to Computational Models: The study maps specific STN subpopulations to distinct components of the Drift-Diffusion framework (bound dynamics, evidence bias, and non-decision times), refining existing theoretical models of basal ganglia function.
4. Decision Evaluation: It reveals that different STN subpopulations encode post-decision evaluation signals (accuracy vs. reward expectation) differently.

4. Key Results

A. Neural Sensitivity to Reward and Evidence

• A substantial fraction of STN neurons (~40%) showed joint modulation by motion coherence and reward context/size.
• Neurons were sensitive to:
  • Choice: Selective activation for contralateral or ipsilateral choices.
  • Reward Context: Modulation based on which choice offered the larger reward.
  • Reward Size: Modulation based on the magnitude of the expected reward.
• These patterns were comparable to the caudate nucleus but distinct from the Frontal Eye Field (FEF), suggesting the basal ganglia are more directly involved in integrating visual and reward information.

B. Three Distinct Subpopulations

Clustering analysis revealed three stable groups with unique activity patterns and computational roles:

• Cluster 1 (Early Modulation / Bound Control):

  • Activity: Early activation during motion viewing that returns to baseline before the saccade.
  • Computational Role: Strongly associated with decision bound dynamics. Higher activity correlated with lower initial bound heights ($a$) and slower bound collapse ($B_{collapse}$).
  • Interpretation: Likely receives input via the hyperdirect pathway (cortex $\to$ STN) to set global decision thresholds independent of specific reward contexts.

• Cluster 2 (Choice Selective / Evidence Bias):

  • Activity: Strong choice selectivity (contralateral preference), particularly around saccade onset.
  • Computational Role: Associated with evidence bias ($m_e$). Activity correlated with a reward-dependent scaling of evidence, driving faster/more frequent choices toward the high-reward option.
  • Interpretation: Likely receives input via the indirect pathway (Caudate $\to$ GPe $\to$ STN) to implement context-dependent biases in the decision variable.

• Cluster 3 (Late Modulation / Non-Decision Processes):

  • Activity: Late-onset activation, emerging closer to the time of choice.
  • Computational Role: Associated with non-decision times ($t_0$), reflecting extra sensory or motor processing delays.
  • Interpretation: Likely involved in visual/motor processing independent of the core decision formation, potentially via indirect pathway inputs.

C. Decision Evaluation Signals

• All three clusters encoded post-decision signals, but with different distributions:
  • Cluster 1: Tended to encode a negative choice accuracy signal and positive reward expectation signals.
  • Cluster 2: Encodings depended heavily on the specific choice made.
  • Cluster 3: Tended to encode reward expectation signals more strongly than choice accuracy.

D. Anatomical Organization

• The three subpopulations were anatomically intermingled. There was no systematic spatial segregation (e.g., dorsal vs. ventral) distinguishing the clusters, explaining why previous stimulation studies (which affect large volumes) failed to isolate specific functions.

5. Significance

• Refining Basal Ganglia Theory: The study moves beyond the view of the STN as a monolithic "brake" on movement. It proposes a nuanced model where distinct, intermingled subpopulations perform specific computations: setting decision bounds (Cluster 1), biasing evidence accumulation based on value (Cluster 2), and managing processing delays (Cluster 3).
• Clinical Implications: Understanding these distinct roles is crucial for optimizing Deep Brain Stimulation (DBS) for Parkinson's disease. It suggests that DBS might inadvertently disrupt specific cognitive computations (like reward-based decision making) if it affects these intermingled subpopulations indiscriminately.
• Methodological Advance: By combining single-unit recording with DDM parameter fitting and clustering, the authors successfully disentangled the functions of intermingled neurons without relying on perturbation techniques that lack spatial specificity.
• Future Directions: The findings suggest that future research should focus on the specific connectivity (hyperdirect vs. indirect pathways) and molecular markers of these subpopulations to enable precise targeting for therapeutic interventions.