How much of a nervous-system model does behaviour identify?
This study demonstrates that while behavioral data can constrain specific directions in the parameter space of whole-organism nervous system models, the vast majority of connectome-scale parameters remain undetermined, with the extent of identifiability depending on the diversity and functional demands of the elicited behaviors.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Scientists have long sought to build digital twins of living creatures, computer programs that mimic the way a worm crawls or a fly walks. These simulations are built from thousands of tiny parts: the strength of the connections between nerve cells, the speed at which electrical signals travel, and the chemical reactions that keep the cells alive. The hope has been that if a computer model can perfectly reproduce an animal's movement, then the numbers inside that model must match the real biology of the animal. If the virtual worm moves just like the real one, the logic goes, the virtual worm must be built with the same settings as the real one. But a growing body of work suggests this might not be true. It is possible for many different sets of numbers to produce the exact same movement, meaning that watching an animal move might not be enough to figure out exactly how its nervous system is wired.
A new study by Chenxi He at the University of Cambridge puts this question to the test on a massive scale. Instead of looking at a single nerve cell or a small group of them, the researcher examined whole-organism simulators of a nematode worm and a fruit fly. These models contain thousands of parameters that define how the nervous system works. By systematically tweaking every single number in these models and watching how the behavior changed, the study mapped out exactly which parts of the model are locked in by the animal's actions and which parts are free to vary. The results show that behavior is a very poor guide to the full details of a nervous system. While the animal's movements do pin down a few specific combinations of settings, they leave the vast majority of the model's internal numbers completely unconstrained.
The study focused first on a detailed model of the nematode C. elegans, a tiny worm whose entire nervous system wiring diagram is known. This model contains 3,076 different connection strengths between its nerve cells. The researcher ran thousands of simulations, changing one connection at a time to see if the worm's crawling pattern shifted. The findings were striking: out of those 3,076 connections, the worm's behavior only determined eight specific directions in the space of all possible settings. Even when the researcher looked at a slightly looser standard of measurement, the number rose to only 47. This means that 99% of the connection strengths in the model could be changed without the worm noticing a difference in how it moved. The behavior of the worm acts like a narrow spotlight, illuminating a tiny, specific slice of the model while leaving the rest in darkness.
This pattern held true even when the study looked at a different animal and a different part of the nervous system. In a model of the fruit fly's visual system, which processes what the fly sees, the researcher found that out of 330 adjustable settings, behavior only constrained 14. In both the worm and the fly, the parts of the model that were tightly controlled by behavior were the connections that carry the main signals for movement or vision. The parts that were left free to vary were often the specific chemical currents inside the cells. For example, in the worm model, the behavior was largely insensitive to two specific types of calcium channels, meaning their strength could be altered significantly without changing how the worm walked. This suggests that the nervous system has a great deal of flexibility; it can achieve the same result with many different internal settings.
The research also explored whether watching an animal do more things would help pin down more details. The team tested the worm model with different types of movement, such as crawling in a straight line versus moving toward food. They found that adding more of the same kind of movement, like walking faster or slower, did not reveal much new information. The model's settings that controlled these similar actions were already locked in by the first movement. However, when the animal performed a functionally different task, such as navigating toward a smell instead of just walking, the model revealed a new set of constrained settings. The worm's behavior in a chemical gradient locked down different connections than its behavior in a simple walk. Yet, even with these different tasks, a large portion of the model remained free. The study showed that to learn more about the nervous system, one needs to observe a wide variety of distinct behaviors, not just longer recordings of the same action.
To ensure these findings were not just a quirk of the computer models, the researcher tested the same methods on simpler cell models where the correct answers were already known. In these cases, the method worked perfectly: when the model was small and the behavior was simple, the correct settings were recovered. But when the model was complex, like the famous MAPK signaling pathway, the simulation could match the behavior exactly while using completely wrong numbers for the internal parts. This confirmed that a perfect match to behavior does not guarantee that the underlying biology has been correctly identified. The study also compared its results to real biological data from other scientists, such as measurements of how much nerve cells vary between individuals of the same species. The patterns of what varies and what stays the same in real worms and flies matched the patterns found in the computer models, suggesting that this "sloppiness" is a real feature of biology, not just a flaw in the simulations.
The ultimate conclusion is that behavior identifies a specific, low-dimensional space within a nervous system model, but it leaves the vast majority of that model's space open. The parts of the system that are constrained are the ones essential for the specific task at hand, while the free parts represent directions in which biology can vary without consequence. This means that building a digital organism that behaves like a real one does not mean we have discovered the real organism's exact settings. Instead, we have found a working combination of settings that happens to produce the right movement. The study reframes this uncertainty not as a failure of measurement, but as a feature of life: the nervous system is robust enough to tolerate a wide range of internal changes as long as the final output, the behavior, remains effective. The researcher concludes that to truly understand a nervous system, we must look beyond just the behavior it produces and recognize that the path to that behavior is paved with many possible routes.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.