In a breakthrough that could revolutionize neuroscience, researchers have harnessed the power of artificial intelligence to create a highly accurate “digital twin” of the mouse brain. This advanced AI model can predict how neurons respond to entirely new visual stimuli—something no previous model has accomplished with such precision.
Recently published in Nature, the study—led by scientists from Baylor College of Medicine, Stanford University, and the Allen Institute for Brain Science—introduces a sophisticated artificial neural network called a “foundation model.”
The model replicates brain activity and mirrors the intricate structural details of neural circuits, offering a powerful new tool for exploring the brain’s inner workings.
Much like how ChatGPT and other large language models transformed natural language processing, this brain model could reshape how we study perception, behavior, or even consciousness.
“If you build a model of the brain and it’s very accurate, that means you can do a lot more experiments,” Stanford professor of ophthalmology and senior author of the study, Dr. Andreas Tolias, explained in a press release. “We’re trying to open the black box, so to speak, to understand the brain at the level of individual neurons or populations of neurons and how they work together to encode information.”
A Brain Digital Twin That Thinks Ahead
For decades, decoding the brain’s language has been one of science’s most enduring puzzles. Traditional neural network models, often trained to respond to specific data sets like object recognition tasks or motion detection, performed well but only within their comfort zones. These models struggle once unfamiliar data is introduced, such as new types of images or stimuli.
Inspired by the power of foundation models in AI—large-scale models trained on massive datasets that generalize remarkably well to new domains—neuroscientists set out to create something similar for the brain.
They recorded real-time neural responses in 14 awake mice as the animals watched natural videos and interacted with their environment. “It’s very hard to sample a realistic movie for mice because nobody makes Hollywood movies for mice,” Dr. Tolias said. “[However] mice like movement, which strongly activates their visual system, so we showed them movies that have a lot of action.”
These recordings captured visual stimuli, behavior (like eye movement and pupil dilation), and contextual variables across six visual brain areas. Over 900 minutes of data from the mice watching clips of action movies, such as Mad Max, were fed into an artificial neural network (ANN).
The result was a four-part model: a “perspective” module to account for how a mouse’s eye sees a stimulus, a “modulation” module to interpret behavioral signals, a “core” that processes the bulk of the sensory data, and a “readout” that translates that data into predictions of neural activity. Trained on just eight mice, the model proved startlingly robust—accurately predicting brain responses in new mice and across novel stimulus types it had never seen before.
The crown jewel of this ANN is what researchers call the “foundation core.” Once trained, this shared internal representation of visual processing could be “transferred” to new individual mice with minimal additional data—sometimes as little as four minutes of new recordings. This dramatically reduced the amount of real-world data needed to accurately model a new brain.
Compared to previous leading models, the new foundation model achieved a 25–46% improvement in predictive accuracy, even in notoriously complex higher-order visual areas of the brain. These performance gains represent a blueprint for developing increasingly accurate models of the brain that can adapt quickly to new data without starting from scratch.
Remarkably, the model didn’t stop at predicting how neurons would fire. It also inferred their physical characteristics—like where they sit in the brain and what kind of neuron they are.
In another major leap for neuroscience, researchers with the MICrONS project recently unveiled the most detailed functional map of the brain to date, offering unprecedented insight into how neurons connect, interact, and communicate. This feat was once considered “impossible.”
Using data from the MICrONS project, researchers applied their model to more than 70,000 neurons. The model accurately predicted anatomical cell types, dendritic structures, and even synaptic connectivity patterns despite never having been trained on anatomical data.
This suggests that the “functional barcodes” generated by the model—essentially, how a neuron processes visual information—can be used as fingerprints to reveal a cell’s type and structure.
Unlimited Potential, Zero Invasiveness with digital twins
A key advantage of creating a digital twin of the brain is its scalability. Because the model can simulate neural responses inside a computer, researchers can now run unlimited experiments that would be time-consuming, invasive, or even impossible to perform on living brains.
For example, researchers tested how digital brains responded to classical vision experiments—like moving dots, Gabor patches, or flashing lights—and found that the digital neurons exhibited the same orientation or spatial selectivity as their real counterparts. In one test, simulated neurons matched the actual preferred orientation of live neurons within just 4 degrees, an extraordinary degree of precision.
Toward a Unified Model of the Brain
By leveraging similarities between brains, rather than treating each one as a separate puzzle, researchers can now model shared cognitive features while still accounting for individual quirks.
Additionally, what makes this recent breakthrough stand out isn’t just the technical prowess—it’s the philosophical leap. Foundation models, by their nature, don’t just memorize; they internalize general principles. The hope is that a similar approach in neuroscience could eventually uncover a universal set of rules that governs how all brains process information.
By uncovering a shared set of rules that govern how neurons encode, process, and transmit information, scientists could be on the verge of establishing foundational laws of neuroscience, akin to the laws of physics that govern the natural world.
See Also
Harvard Scientists Say New Research “Sets the Stage” For Helping Humans Regrow Lost Limbs
Just as equations like Newton’s laws or Einstein’s theories of relativity provide a framework for understanding motion and energy across the universe, a universal neural code could offer a unifying theory for how brains function across species, individuals, and contexts.
Such a discovery would revolutionize our understanding of cognition and behavior and lay the groundwork for predictive, standardized models of brain activity, ushering in a new era of neuroscience.
“Our present foundation model is just the beginning, as it only models parts of the mouse visual system under passive viewing conditions,” researchers wrote. “As we accumulate more diverse multimodal data—encompassing sensory inputs, behaviors, and neural activity across various scales, modalities, and species, foundation neuroscience models will enable us to decipher the neural code of natural intelligence, providing unprecedented insights into the fundamental principles of the brain.”
Next Steps and Broader Implications
The success of this model raises significant questions about the future of neuroscience—and the ethical and philosophical implications of simulating brains with such fidelity.
Could similar models one day simulate the human brain? Could they lead to better brain-computer interfaces, more accurate diagnoses for neurological diseases, or even a deeper understanding of consciousness?
The researchers caution that we’re not there yet. Their digital brain twin currently applies only to mice’s visual cortex under passive viewing conditions. But the groundwork has been laid for a much broader project.
Ultimately, in the same way that large language models unlocked new capabilities in artificial intelligence, this digital twin of the brain could unlock new frontiers in biology, medicine, and cognitive science. And in doing so, it brings us a step closer to answering the enduring mysteries of the brain.
“In many ways, the seed of intelligence is the ability to generalize robustly,” Dr. Tolias said. “The ultimate goal — the ‘holy grail’ — is to generalize to scenarios outside your training distribution.”
“Eventually, I believe it will be possible to build digital twins of at least parts of the human brain. This is just the tip of the iceberg.”
Tim McMillan is a retired law enforcement executive, investigative reporter and co-founder of The Debrief. His writing typically focuses on defense, national security, the Intelligence Community and topics related to psychology. You can follow Tim on Twitter:@LtTimMcMillan. Tim can be reached by email:tim@thedebrief.orgor through encrypted email:LtTimMcMillan@protonmail.com
