Just like how a pilot could train using a flight simulator, scientists may soon have the ability to conduct experiments on a detailed simulation of a mouse’s brain. In recent research from Stanford Medicine with partners, experts utilized an AI model to create a “virtual replica” of the section of the mouse brain responsible for interpreting visual data.
The digital replica was developed using extensive data sets of brain activity recorded from the visual cortices of actual mice while they were viewing film segments. This model can subsequently forecast how tens of thousands of nerve cells will react to fresh video content and pictures.
Virtual replicas might simplify and expedite the examination of the brain’s internal mechanisms.
“If constructing an exceptionally precise brain model becomes possible, then numerous additional experiments could be conducted,” stated Dr. Andreas Tolias, who serves as a Stanford Medicine professor of ophthalmology and is also the lead author of this study. study published April 9 in Nature The most promising ones can subsequently be tested in an actual brain.
Eric Wang, Ph.D., who is a medical student at Baylor College of Medicine, authored the main part of the study.
Beyond the training distribution
In contrast to earlier AI models of the visual cortex that were limited to simulating responses based solely on the types of stimuli encountered during their training, this updated model has the capability to forecast how neurons will react to various novel visual inputs. Additionally, it can infer specific characteristics of individual neurons.
The latest version represents a type known as a foundation model, which belongs to a recently emerged category of artificial intelligence systems designed to absorb information from extensive databases. These models have the ability to utilize this acquired knowledge for tackling novel tasks and different kinds of data—a capability referred to by experts as “extending beyond the scope of their initial training.” (An easily recognizable instance of such a foundational model would be ChatGPT, which harnesses massive volumes of textual content to grasp and produce fresh text accordingly.)
Tolias stated that, in numerous aspects, the foundation of intelligence lies in the capacity for strong generalization. The primary objective, the ultimate aim, is to apply this generalization beyond the scope of the training data encountered.
Mouse movies
To develop the new AI model, the researchers initially captured the brain activity of actual mice as they viewed films designed for humans. These movies were intended to mimic what the mice might encounter in their natural environment.
"It’s extremely difficult to find a suitable film for mice since no one produces Hollywood movies specifically tailored for them," Tolias explained. However, action films were close enough.
Mice possess low-resolution eyesight — akin to how humans perceive things at the edges of their vision — implying that they primarily notice motion instead of intricate details or colors. "Since mice respond well to movement, which greatly stimulates their visual senses, we presented them with videos filled with dynamic actions," explained Tolias.
During numerous brief viewing sessions, scientists logged over 900 minutes of brain activity data from eight mice observing segments of high-energy films like "Mad Max." The cameras tracked their ocular movements and behaviors simultaneously.
The scientists utilized combined data to develop a primary model, which could subsequently be tailored into a digital replica of an individual mouse through further training.
Accurate predictions
The digital twins managed to accurately mimic the neural responses of their living equivalents when exposed to different types of visual inputs like videos and still pictures. According to Tolias, the extensive amount of compiled training information played a crucial role in achieving this accuracy. He noted, "Their precision can be attributed to being trained using massive datasets."
Although these new models were trained exclusively on neural activity, they have the potential to apply to various other kinds of data as well.
The specific mouse’s digital twin could forecast the anatomic positions and types of cells for thousands of neurons within the visual cortex along with their interconnections.
The researchers confirmed these predictions using high-resolution electron microscope images of the mouse's visual cortex, as part of an extensive effort to meticulously chart the architecture and functionality of the mouse visual cortex with unparalleled precision. results of that project , known as MICrONS , were released at the same time in Nature .
Opening the black box
Since a digital twin remains functional far beyond the life span of a physical mouse, researchers have the potential to conduct an effectively limitless series of tests using what amounts to the same virtual creature. What might normally require years could be accomplished within mere hours, with countless trials running concurrently. This accelerates investigations into how our brains handle data and uncovering the foundational concepts of intelligence.
Tolias mentioned, "We aim to unlock the mysteries of the black box, essentially delving into an understanding of the brain at the neuronal level—both individual neurons and groups of them—and explore how these components collaborate to store data."
Actually, the latest models have begun providing fresh perspectives. Additionally, they relate closely to other findings. study , was also released concurrently in Nature In their study, researchers employed a digital twin to uncover how neurons within the visual cortex select other neurons for forming connections.
Researchers were aware that analogous neurons often establish connections, much like individuals developing friendships. The digital twin helped identify which specific similarities played a crucial role. It showed that neurons have a preference for connecting with those that react to identical stimuli—for instance, recognizing the color blue—rather than just responding to the same region within their field of vision.
"It’s akin to choosing friends based on shared interests rather than proximity,” Tolias stated. “This allowed us to uncover a more exact principle governing the organization of the brain.”
The research team intends to broaden their modeling efforts to encompass additional regions of the brain as well as various species like animals and primates known for their sophisticated cognitive skills.
I think eventually we'll be able to construct digital replicas of at least portions of the human brain," Tolias stated. "We've only just scratched the surface.
The researchers from the University of Göttingen and the Allen Institute for Brain Science were involved in this research.
More information: Eric Y. Wang et al., A foundation model of neural activity forecasts responses to novel stimulus categories, Nature (2025). DOI: 10.1038/s41586-025-08829-y
Furnished by Stanford University Medical Center
This tale was initially released on Medical Xpress . Subscribe to our newsletter For the most recent science and technology news updates.