Biomimetic brain model matches animal learning and reveals hidden neural patterns

A multi-institutional team has created a computational model of the brain that learns a visual categorization task with the same erratic progress as lab animals, despite never being trained on animal data. The model's accuracy led researchers to discover a previously unnoticed group of neurons whose activity predicts errors, revealing a potential mechanism for exploring alternative strategies when rules change.

The model, developed by researchers at Dartmouth College, MIT, and the State University of New York at Stony Brook, represents a significant step toward "biomimetic" modeling—systems that closely replicate biological principles rather than simply mimicking behavior. When presented with the same task of categorizing dot patterns, the model produced neural activity and behavioral results that matched animal data almost exactly, according to a study published in Nature Communications.

"It's just producing new simulated plots of brain activity that then only afterward are being compared to the lab animals. The fact that they match up as strikingly as they do is kind of shocking," says Richard Granger, a professor of psychological and brain sciences at Dartmouth and senior author of the study.

Building from biological principles

Unlike many computational models that focus on either microscopic details or large-scale architecture, the biomimetic model incorporates both. Dartmouth postdoc Anand Pathak created the model to obey constraints observed in real brains, such as how neurons synchronize through broader rhythms.

The model's architecture uses "primitives"—small circuits of a few neurons each that connect based on electrical and chemical principles. For example, within the model's version of the cortex, one primitive design has excitatory neurons that receive visual input via glutamate-affected synapses. These neurons then densely connect with inhibitory neurons in a "winner-take-all" architecture that regulates information processing, a structure found in real brains.

At a larger scale, the model encompasses four brain regions needed for basic learning and memory tasks: a cortex, a brainstem, a striatum, and a "tonically active neuron" (TAN) structure that injects variability through acetylcholine bursts. This multi-scale approach ensures the model captures both the "trees" (individual circuits) and the "forest" (regional interactions).

Learning with naturalistic dynamics

When the model performed the visual categorization task, it demonstrated learning dynamics remarkably similar to animals. As learning progressed, the cortex and striatum became more synchronized in the "beta" frequency band of brain rhythms—a pattern commonly observed in animal research. This increased synchrony correlated with times when the model made correct category judgments.

The TAN structure played a crucial role in learning. Initially, it ensured variability in how the model acted on visual input, allowing exploration of different actions and their outcomes. As learning continued, cortex and striatum circuits strengthened connections that suppressed the TAN, enabling the model to act with increasing consistency.

"The idea is to make a platform for biomimetic modeling of the brain so you can have a more efficient way of discovering, developing, and improving neurotherapeutics," says Earl K. Miller, Picower Professor in The Picower Institute for Learning and Memory at MIT and a co-author of the study. "Drug development and efficacy testing, for example, can happen earlier in the process, on our platform, before the risk and expense of clinical trials."

Revealing "incongruent" neurons

The model's most surprising contribution came when it revealed a group of neurons—about 20 percent of the population—whose activity appeared highly predictive of error. When these "incongruent" neurons influenced circuits, the model would make the wrong category judgment.

Initially, the team assumed this was a quirk of the model. "Only then did we go back to the data we already had, sure that this couldn't be in there because somebody would have said something about it, but it was in there, and it just had never been noticed or analyzed," Granger explains.

Miller suggests these counterintuitive cells might serve an important purpose: while learning task rules is valuable, what happens when those rules change? Trying out alternatives from time to time enables a brain to stumble upon newly emerging conditions. This hypothesis aligns with recent evidence from another Picower Institute lab showing that humans and other animals sometimes explore alternatives even when they know the current rules.

From research to biotech applications

The team has been expanding the model to handle greater variety in tasks and circumstances. They've added more brain regions and new neuromodulatory chemicals, and begun testing how interventions like drugs affect its dynamics.

The research has commercial implications through Neuroblox.ai, a company founded by Miller, Granger, and other team members. Lilianne R. Mujica-Parodi, a biomedical engineering professor at Stony Brook and lead principal investigator for the Neuroblox Project, serves as CEO. The company aims to develop the models' biotech applications for more efficient neurotherapeutic discovery and testing.

The model's ability to generate predictions that lead to biological discoveries demonstrates the value of biomimetic approaches. Rather than simply reproducing known behaviors, these models can reveal hidden patterns in existing data and suggest new hypotheses about brain function.

As the team continues refining the model, its applications may extend beyond basic research to clinical settings, potentially offering a platform for testing interventions before expensive and risky clinical trials. The approach represents a bridge between computational neuroscience and practical neurotherapeutic development.

Paper: "Biomimetic model of corticostriatal micro-assemblies discovers a neural code" (Nature Communications)

Related resources:

• Miller Lab at MIT
• The Picower Institute for Learning and Memory
• Department of Brain and Cognitive Sciences at MIT
• Neuroblox.ai (company developing biomimetic modeling platform)