Summary: A biologically grounded computational model, constructed to mirror real neural circuits and never trained on animal data, learned a visual categorization task in the same way lab animals did—matching their accuracy, variability, and the underlying neural rhythms. By combining fine-scale synaptic rules with a multi-region architecture spanning cortex, striatum, brainstem, and acetylcholine-modulated elements, the model reproduced hallmark patterns of learning, including strengthened beta-band synchrony between regions during correct decisions.
The model also exposed a set of neurons dubbed “incongruent neurons” whose activity predicted upcoming errors. Researchers reexamined their animal recordings and found the same signal had been present but previously unnoticed. This biomimetic platform offers a robust blueprint for exploring circuit changes in disease and for in-silico testing of therapeutic interventions, creating a promising avenue for next-generation neurotherapeutics development.
Key Facts
- Biology-first design: The model incorporates realistic neuronal connectivity rules, neurotransmitter dynamics, and multi-region brain architecture to emulate biological computation.
- Emergent realism: Without training on biological datasets, the model spontaneously produced learning behavior, beta synchrony, and decision patterns that match animal experiments.
- Hidden signals exposed: The discovery of “incongruent neurons” identified an error-predictive signal that had been overlooked in previous analyses of experimental data.
Source: Picower Institute at MIT
A new computational model that closely follows brain biology and physiology not only learned a simple visual category task as effectively as lab animals, but also revealed unexpected neural activity that researchers had missed in their experimental data, report scientists from Dartmouth College, MIT, and the State University of New York at Stony Brook.
Crucially, the model achieved these results without any prior exposure to animal data. It was assembled from first principles to faithfully represent how neurons connect, communicate electrically and chemically, and organize across regions to produce cognition and behavior.
When the team tasked the model with the same visual categorization challenge used in their animal studies—viewing dot patterns and judging which of two categories they belonged to—the model generated neural activity and behavioral learning curves strikingly similar to those observed in the animals, including comparable trial-to-trial variability and overall learning dynamics.
“The model produces simulated brain activity that we can then compare to animal recordings. That the match is so close is striking,” said Richard Granger, Professor of Psychological and Brain Sciences at Dartmouth and senior author of the study in Nature Communications.
The model is not only intended to illuminate healthy brain function. As co-author Earl K. Miller, Picower Professor at MIT, explained, the platform is being developed to test how circuit dynamics shift in disease and to evaluate interventions that might restore normal function.
Miller, Granger, and colleagues have founded the company Neuroblox.ai to translate these biomimetic models into biotech applications. Lilianne R. Mujica-Parodi, a biomedical engineering professor at Stony Brook and Lead Principal Investigator for the Neuroblox project, serves as CEO.
“The goal is to build a biomimetic modeling platform that speeds discovery and development of neurotherapeutics,” Miller said. “By running drug efficacy and mechanistic tests in silico, researchers can reduce early-stage risk and cost before moving to clinical trials.”
Building a biomimetic model
Dartmouth postdoctoral researcher Anand Pathak led the model’s development. Unlike many models that emphasize either micro-scale physiology or macro-scale architecture, this design integrates both. It includes detailed rules for how individual neuron pairs connect and communicate, and it embeds larger-scale organization across regions influenced by neuromodulators such as acetylcholine.
The team iterated the design to ensure it respects biological constraints observed in real brains, such as how neuronal populations synchronize to broader rhythms. “We didn’t want to lose the tree, and we didn’t want to lose the forest,” Pathak said.
At the micro-scale, the model uses small circuit motifs—called primitives—that mimic how a few neurons interact through known electrical and chemical mechanisms to perform basic computations. For example, in the model’s cortical modules, excitatory neurons receive visual input via glutamatergic synapses and engage in dense interactions with inhibitory neurons to implement winner-take-all competition, a motif common in biological cortex.
At the macro-scale, the simulation includes four primary regions necessary for this learning task: cortex, striatum, brainstem, and a tonically active neuron (TAN) structure that injects stochastic modulation through bursts of acetylcholine. Early in training, TAN-mediated variability encourages exploration of different actions; as learning progresses, cortico-striatal circuits strengthen and suppress TAN influence, producing more consistent, learned responses.
As learning unfolded, the model reproduced a hallmark observation from experimental work: increased synchrony between cortex and striatum in the beta frequency band. This enhanced beta-band coherence correlated with correct categorizations in both the model and animal recordings.
Revealing “incongruent” neurons
Unexpectedly, the model highlighted a subpopulation of neurons—about 20 percent—whose activity reliably predicted forthcoming errors. These “incongruent neurons,” when they drove downstream circuits, increased the likelihood of incorrect category judgments. Initially viewed as a modeling quirk, the team reanalyzed existing animal datasets and confirmed the same signal was present there, though it had gone unreported.
Miller suggests these counterintuitive cells could have adaptive value. Occasional deviations from the learned rule may enable the brain to explore alternatives when environmental contingencies shift, supporting flexibility when task structure changes.
Following the study, the team has continued to expand the model—adding additional brain regions, incorporating other neuromodulators, and beginning systematic tests of pharmacological and other interventions to observe their effects on network dynamics.
In addition to Granger, Miller, Pathak, and Mujica-Parodi, the paper’s authors include Scott Brincat, Haris Organtzidis, Helmut Strey, and Evan Antzoulatos.
Funding: The research was supported by the Baszucki Brain Research Fund (United States), the Office of Naval Research, and the Freedom Together Foundation.
Key Questions Answered:
A: The model learned the visual category task with nearly identical progress patterns, neural activity signatures, and learning dynamics—even though it was never trained on biological data.
A: It revealed a population of “incongruent neurons” that reliably predicted errors. Reanalysis of animal recordings confirmed the same pattern had been present but previously overlooked.
A: The platform directly links mechanistic, spiking-level phenomena with high-level decision and reinforcement signals, enabling simulation of disease states and preclinical testing of neurotherapeutics before expensive clinical trials.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was added by staff.
About this AI, learning, and neuroscience research news
Author: David Orenstein
Source: Picower Institute at MIT
Contact: David Orenstein – Picower Institute at MIT
Image: The image is credited to Neuroscience News
Original Research: Open access. “Biomimetic model of corticostriatal micro-assemblies discovers a neural code” by Richard Granger et al. Nature Communications. DOI: 10.1038/s41467-025-67076-x
Abstract
Biomimetic model of corticostriatal micro-assemblies discovers a neural code
Computational models have advanced neuroscience, but linking low-level physiological activity (spikes, field potentials) and biochemistry (transmitters and receptors) directly to high-level cognitive functions (decision-making, working memory) and disorders remains challenging.
Here, the authors present a mechanistically detailed, multi-scale model that generates realistic simulated physiology from which extended neural and cognitive phenomena emerge. The model directly produces spiking activity, field signals, phase synchronies, and synaptic plasticity, giving rise to working memory, decisions, and categorization behavior.
These emergent behaviors were validated against extensive macaque experimental data, despite the model receiving no prior training on that data. Furthermore, the simulation uncovered a previously unrecognized neural code—“incongruent neurons”—that specifically predicts upcoming errors, a finding later confirmed in empirical recordings.
This biomimetic model thus provides a direct, predictive link between decision and reinforcement signals of computational interest and spiking and field codes of neurobiological importance, offering a powerful tool for probing brain function and testing therapeutic strategies.