Developmental and genetic modulation of evidence integration dynamics in zebrafish sensorimotor decision-making

By combining high-throughput behavioral assays with drift-diffusion modeling in larval zebrafish, this study reveals that evidence integration dynamics mature during development and are selectively disrupted by mutations associated with human epilepsy and schizophrenia, demonstrating a scalable approach to studying the algorithmic basis of sensorimotor decision-making in health and disease.

The Gist

The Big Picture: How Fish Make Decisions

Imagine you are walking through a crowded room, trying to figure out which way to go. You hear whispers, see people moving, and feel the flow of the crowd. You don't decide instantly; you gather these tiny clues over a few seconds, weigh them in your mind, and then finally take a step.

This paper is about how baby zebrafish do the exact same thing. They swim around in a tank, looking at moving dots on a screen (like a digital crowd), and they have to decide which way to swim. The scientists wanted to understand the "mental math" the fish are doing to make that choice.

The Problem: We Can't Read Fish Minds

The tricky part is that we can't ask a fish, "Hey, how confident are you in that decision?" or "How much noise is in your brain right now?" We can only see the final result: Did the fish swim the right way? How long did it take to decide?

To solve this, the scientists used a Drift-Diffusion Model. Think of this model as a mental scale or a bucket.

• The Bucket: The fish has a bucket in its brain where it pours "evidence" (clues from the moving dots).
• The Drift: If the dots are moving clearly to the right, the water (evidence) pours in fast. If the dots are chaotic, the water trickles in slowly.
• The Leak: Sometimes, the bucket has a hole. If the hole is big, the fish forgets the clues it gathered a second ago. If the hole is small (or even negative, meaning the bucket fills itself up!), the fish holds onto those clues tightly.
• The Threshold: Once the bucket is full, the fish makes a decision and swims.

The Innovation: A "Smart" Detective

In the past, figuring out the size of the bucket, the speed of the pour, and the size of the leak was like trying to guess the recipe of a soup just by tasting the final dish. Scientists had to guess and check manually, which was slow and often inaccurate.

This team built a digital detective (using something called Bayesian Optimization).

• They fed the computer thousands of hours of fish swimming data.
• The computer played a game of "Guess the Recipe." It simulated millions of different fish brains with different bucket sizes and leak rates.
• It kept adjusting the recipe until the simulated fish behaved exactly like the real fish.
• Once it matched, the computer revealed the "secret recipe" (the specific brain parameters) for that individual fish.

The Discoveries

1. Fish Get "Smarter" (and Stickier) as They Grow

The scientists tested baby fish at different ages (from 5 days old to 9 days old).

• The Baby Fish (5 days): Their mental buckets had big holes. They forgot clues quickly. They were easily distracted and didn't hold onto a decision for long.
• The Older Fish (7-9 days): As they grew, their buckets changed. They developed a "self-reinforcing" mechanism. Imagine the bucket has a little pump that adds more water every time it gets a clue. This makes the decision "stickier." Once the fish starts leaning toward a direction, it holds that thought firmly, making the decision process more persistent and stable.

Analogy: Think of a toddler vs. a teenager. A toddler might forget why they wanted a cookie the moment they see a toy. A teenager can hold onto a goal (like studying for a test) even when distractions pop up. The fish brains matured to be more like the teenager.

2. The "Broken" Brains (Genetic Mutations)

The scientists then looked at fish with genetic mutations linked to human diseases like epilepsy and schizophrenia.

• These fish had a broken "pump" in their mental buckets.
• Even though they were the same age as the "smart" older fish, their brains couldn't hold onto the clues. The "leak" was too big, or the "self-reinforcing" pump was broken.
• The Result: These fish were slower to decide and less consistent. They couldn't build up a strong internal feeling of "I know which way to go."

Why This Matters

This isn't just about fish. It's a new way to look at how brains work.

• For Science: It proves that we can use simple math models to understand complex brain circuits without needing to open the skull.
• For Medicine: Because these fish have genes similar to humans, this method could be a super-fast way to test drugs. If a drug fixes the "leaky bucket" in a mutant fish, it might help fix similar decision-making problems in humans with epilepsy or schizophrenia.

The Takeaway

The scientists built a high-tech "mind-reading" tool for fish. They discovered that as fish grow up, their brains learn to hold onto thoughts better, making them better decision-makers. When the genes that build this "thought-holding" mechanism are broken, the fish struggle to decide. This gives us a powerful new way to study how our own brains make choices and what happens when they go wrong.

Technical Summary

1. Problem Statement

Animals rely on temporal integration of sensory information to make decisions, a process often modeled by Drift-Diffusion Models (DDM). While DDMs are powerful theoretical frameworks for describing decision-making across species, several challenges limit their application in high-throughput biological research:

• Latent Variable Inference: DDM parameters (e.g., diffusion, drift, leak) are hidden variables that must be inferred indirectly from behavioral data.
• Stochasticity and Data Scarcity: The stochastic nature of these models makes automatic parameter inference difficult, especially when trial numbers are limited.
• Manual Tuning: Previous approaches often relied on manual parameter tuning or complex multi-objective fitting, which produced non-unique estimates and hindered systematic analysis of individual variability.
• Biological Linkage: There is a gap in understanding how specific developmental stages or genetic mutations (linked to human diseases like epilepsy and schizophrenia) modulate the algorithmic implementation of evidence integration.

The authors aim to develop a scalable, automated pipeline to infer latent DDM parameters from high-throughput behavioral data in larval zebrafish, thereby linking behavioral algorithms to developmental and genetic factors.

2. Methodology

A. Experimental Setup

• Subject: Larval zebrafish (Danio rerio) at various developmental stages (5–9 days post-fertilization, dpf) and specific genetic mutants.
• Behavioral Paradigm: A random-dot-motion (RDM) task adapted for freely swimming larvae.
  • Stimulus: A kinematogram where a fraction of dots moves coherently (0%, 25%, 50%, 100%) while the rest move randomly.
  • Real-time Tracking: The stimulus direction is locked to the fish's orientation in real-time.
  • Readout: Swimming events are binarized into "correct" (turning with motion) or "incorrect" (turning against motion). The inter-swim interval serves as a proxy for response delay.
• Genetic Models:
  • scn1lab mutants (heterozygous): Linked to neuronal excitability, Dravet syndrome, and epilepsy.
  • disc1 mutants (heterozygous and homozygous): Linked to schizophrenia and neurodevelopmental scaffolding.

B. Computational Framework

• Model: A simplified, classical Drift-Diffusion Model with five free parameters:
  1. Diffusion ($\sigma$): Noise in the integration process.
  2. Drift ($\mu$): Weight of sensory input.
  3. Leak ($\lambda$): Decay of accumulated evidence (negative values imply self-reinforcement).
  4. Reset ($r$): State of the integrator after a decision.
  5. Delay ($\delta$): Refractory period.
  • Note: The decision boundary ($B$) was fixed at 1 as it is mathematically redundant with other parameters.
• Optimization Algorithm: Bayesian Optimization was employed to automatically fit the model to experimental data.
  • Loss Function: A modified Kullback-Leibler (KL) divergence ($D_{KL}^*$) was used. Unlike standard KL, this metric weights histogram bins proportionally to their height, prioritizing frequently occurring events (shorter inter-swim intervals) during fitting.
  • Validation: The pipeline was first validated on artificially generated datasets with known parameters to ensure repeatability, robustness to noise, and minimal data requirements (approx. 900s per coherence level).

3. Key Contributions

1. Automated Inference Pipeline: Development of a robust, Bayesian-optimization-based method to automatically extract latent DDM parameters from high-throughput behavioral data without manual tuning.
2. Individual-Level Analysis: Demonstration that significant inter-individual variability exists in sensorimotor integration strategies, necessitating analysis at the single-animal level rather than population averages.
3. Developmental Mapping: Quantification of how evidence integration algorithms change during early development (5–9 dpf).
4. Genetic Phenotyping: Application of the model to disease-associated mutants to reveal specific algorithmic deficits linked to scn1lab and disc1 genes.

4. Key Results

A. Model Validation and Individual Variability

• The Bayesian optimization pipeline reliably recovered known parameters from synthetic data and converged to near-zero loss on real experimental data.
• Analysis of 39 wild-type larvae (5 dpf) revealed substantial inter-animal variability in latent parameters, particularly in the leak ($\lambda$) and diffusion ($\sigma$) parameters, despite similar behavioral outputs (accuracy and response time).
• A significant portion of the population exhibited negative leak values, indicating self-reinforcing integration dynamics (priming-like behavior) where accumulated evidence reinforces itself rather than decaying.

B. Developmental Modulation (5–9 dpf)

• Leak Parameter ($\lambda$): The most significant developmental change was observed in the leak parameter. At 5 dpf, the population distribution was centered near zero. By 6 dpf and onwards, the distribution shifted significantly toward negative values.
  • Interpretation: As larvae mature, their integrators become more self-reinforcing and persistent, suggesting the development of working-memory-like properties or recurrent neural feedback.
• Drift Parameter ($\mu$): Remained stable, suggesting the ability to extract momentary sensory evidence does not change significantly during this early window.
• Diffusion ($\sigma$): Slightly increased with age, potentially reflecting increased behavioral flexibility or sampling of additional cues.

C. Genetic Modulation (Disease Models)

• scn1lab Mutants (Epilepsy/Excitability):
  • Heterozygous mutants showed a significantly less negative leak compared to controls.
  • Interpretation: The mutation impairs the self-reinforcing, persistent dynamics of the integrator, leading to "leakier" internal states and potentially shorter persistence of evidence. This aligns with the gene's role in neuronal excitability.
• disc1 Mutants (Schizophrenia/Development):
  • Homozygous mutants (disc1–/–) also exhibited less negative leak values compared to controls.
  • Interpretation: Disruption of neurodevelopmental scaffolding proteins similarly disrupts the recurrent network dynamics required for stable evidence accumulation.
• Behavioral Correlates: These algorithmic changes (reduced self-reinforcement) manifested behaviorally as longer inter-swim intervals (slower decisions) with only minor reductions in accuracy.

5. Significance and Implications

• Scalable Phenotyping: This approach provides a scalable, automated method to generate experimentally testable hypotheses about the "algorithmic" implementation of decision-making in health and disease. It bridges the gap between behavioral observations and underlying neural circuit mechanisms.
• Neural Circuit Insights: The finding that the leak parameter is modulated by development and disease suggests that recurrent positive feedback within the integrator (likely in the anterior hindbrain) is a critical, tunable component of sensorimotor decision-making.
• Disease Modeling: The study demonstrates that mutations linked to human neuropsychiatric disorders (scn1lab, disc1) do not just cause gross behavioral deficits but specifically alter the dynamics of evidence integration (reducing persistence). This offers a new quantitative framework for screening drugs or genetic modifiers that restore normal integration dynamics.
• Generalizability: The framework relies only on event timing and correctness labels, making it applicable to other species (insects, rodents, primates) and decision-making paradigms.

In conclusion, the paper establishes that larval zebrafish employ self-reinforcing evidence integration dynamics that mature during development and are selectively disrupted by disease-associated mutations. The proposed Bayesian optimization pipeline offers a powerful tool for dissecting the computational logic of the brain in health and disease.