Different Brains share a common neural grammar

What does the brain do when it is not receiving any specific external stimuli and is not engaged in an active task? Far from being silent, neurons exhibit a rich and structured dynamic. This background activity is not mere noise: it reflects the architecture of neural circuits, determines how the brain processes information, and contributes to the formation of the internal states of living beings.

In humans, functional magnetic resonance imaging has already shown that the brain exhibits coherent patterns of activity at rest. These “default mode networks” are used to study development, aging, and certain psychiatric disorders. However, this imaging remains indirect and relatively coarse. Indeed, it does not allow us to track the activity of individual neurons or to precisely identify the circuits that produce these dynamics. At the other extreme, much simpler organisms, such as the C. elegans worm, allow us to record the activity of nearly all neurons and link it to a highly stereotyped anatomy: their nervous system consists of a small number of neurons, each of which can be uniquely identified within each individual.  The situation is quite different in vertebrates: two brains of the same species share a similar overall organization, yet no single neuron has an identifiable counterpart from one individual to another. Consequently, understanding the functional properties common to multiple individuals becomes particularly challenging. 

This research was carried out in the following CNRS laboratories:

• Laboratoire Jean Perrin (LJP, CNRS / Sorbonne Université)

• Institut de physique théorique (IPhT, CEA/CNRS)

• Laboratoire de physique de l'ENS (LPENS, CNRS / ENS-PSL / Sorbonne Université / Université Paris Cité)

This is precisely what a team of researchers from three French laboratories set out to achieve by using zebrafish larvae as a vertebrate model. This small freshwater fish is particularly well-suited for such a study: its brain is compact and transparent, allowing researchers to record the activity of virtually the entire brain (tens of thousands of neurons) at the cellular level. The immediate challenge posed by such precision is to compare this brain activity across individuals in order to determine what is common to the species as a whole, beyond the very high variability that exists among different individuals.

To do this, the researchers used probabilistic models inspired by statistical physics, known as Restricted Boltzmann Machines (RBMs). These models “learn” to represent spontaneous brain activity based on collective patterns, which are groups of neurons that tend to fire together. These patterns can be viewed as the building blocks of brain activity, or as the words of a neural vocabulary. Together, they form a low-dimensional “representation” of brain activity.  The researchers behind the study developed a variant of the model, called LaRBM (Latent-aligned Restricted Boltzmann Machine), which forces multiple models trained on different fish to use the same representation. Each individual thus retains its own neurons, but the patterns learned by the model can be “aligned” across animals. This approach enables a conceptually powerful operation: “translating” a snapshot of one fish’s brain into the brain of another. Obviously, this does not involve transferring actual neural activity. The model takes a pattern of activity observed in one animal, encodes it in the shared representation space, and then calculates which pattern of activity would be most plausible in the brain of another animal. The configurations obtained in this way retain a spatial and statistical organization that is plausible for the recipient brain. And this is true even though the two individuals did not exhibit the same activity when their respective brains were recorded.

This result suggests that the spontaneous brain activity of zebrafish larvae is not merely a collection of fluctuations in individual neurons. It is constructed from a shared set of functional patterns, stable enough to be identifiable across brains but flexible enough to adapt to interindividual anatomical variability. To use a linguistic analogy, this approach reveals a common grammar within brain activity, despite the variations existing among speakers. 

The significance of this work extends beyond the zebrafish. It provides a quantitative framework for comparing brain activity across individuals without requiring an exact correspondence between neurons. This is an important development for experimental neurobiology: it could make it possible to quantify how a brain deviates from the normal functional repertoire during development, due to a genetic mutation, as a result of treatment, or in models of neurological disorders. The study also demonstrates the specific contribution of statistical physics to the analysis of large-scale neural recordings. Whereas a purely descriptive approach would simply accumulate activity maps that are difficult to compare, the model constructs a common, interpretable representation. It does more than simply summarize the data: it allows researchers to test whether a pattern of activity is likely, to reconstruct it, and to compare it across brains. This study thus opens up a new avenue for high-resolution comparative neuroscience, comparing not only anatomies but also the dynamic grammars through which brains organize their activity. It is published in the Proceedings of the National Academy of Sciences.