Connectome-constrained Drosophila neural control and reward-modulated plasticity in a simulated Maxwell-demon particle system

Imagine a tiny, invisible agent standing between two rooms filled with gas molecules. One room is hot, the other cool. The agent's job is to watch every molecule as it approaches the door between them. If a fast, energetic molecule comes from the cool side, the agent opens the door to let it into the hot room. If a slow, sluggish molecule approaches from the hot side, the agent keeps the door shut. Over time, this sorting action makes the hot room hotter and the cool room cooler, seemingly creating energy from nothing. This thought experiment, known as Maxwell's demon, has challenged physicists for over a century. It forces us to ask how information and physical reality interact: does the act of measuring and sorting molecules carry a hidden energy cost? While modern physics has resolved the paradox by showing that the information processing itself generates heat, the question remains how a biological brain, evolved to navigate a world of smells and light, would perform such a task if placed in this role.

A new study by Yash Sharma at Queen's University Belfast explores this question by building a digital laboratory where a simulated fly brain attempts to act as this sorting agent. The researchers did not use a real fly, but rather a computer model of the entire adult fruit fly brain, mapped down to the level of individual connections between neurons. They connected this biological blueprint to a simulation of a gas-filled chamber with a sliding door. The goal was not to prove that a fly could actually sort gas molecules in the real world, but to create a closed-loop system where they could watch, second by second, how a brain constrained by real biological wiring reacts to a physical environment and learns from the results. By recording every decision the digital brain made, every time a molecule crossed the door, and every tiny adjustment the brain's connections made in response to success or failure, the team created a detailed map of how a biologically realistic controller interacts with the laws of thermodynamics.

The experiment was set up as a relay race between two different controllers. For the first part of each run, a simple, pre-programmed computer script took the role of the demon. This script knew exactly what to do: it opened the door only for the "right" molecules to maintain a temperature difference between the two chambers. For the next part of the run, the control was handed over to the fly-brain model. This model received the same information about the gas temperature and the speed of the molecules, but it had to decide when to open the door based on its own internal activity. The researchers ran this simulation eight times, with sessions lasting between five minutes and twenty minutes. In every session, the two controllers alternated every two and a half minutes, allowing the team to see how the fly-brain model performed against the perfect, pre-programmed script under identical conditions.

The results showed a clear distinction between the two controllers. When the simple script held the door, it successfully maintained a temperature difference between the two chambers, keeping the gas sorted as intended. However, whenever the fly-brain model took control, the temperature difference quickly faded, and the two sides of the chamber returned to a state of balance. The fly-brain model did not learn to sort the molecules effectively during these short sessions. Instead of keeping the hot side hot and the cool side cool, it allowed the gas to mix, failing to sustain the separation that the demon task requires. The data showed that the temperature difference under the script's control was significantly larger than when the fly brain was in charge, indicating that the biological controller was not yet capable of performing this specific thermodynamic sorting job.

Despite the failure to sort the gas, the experiment revealed that the fly-brain model was actively learning and changing. The researchers tracked the strength of the connections between the neurons, known as synapses, throughout the simulation. They found that with every decision the brain made, these connections were being tweaked. The total amount of change in these connections grew steadily in every single run, showing that the network was responding to the rewards and penalties it received for its actions. The type of reward the brain received mattered, too. When the researchers used a steady, predictable way of giving rewards, the brain made "good" decisions about 48 percent of the time. When the rewards were random and unpredictable, the brain's success rate dropped to about 34 percent. This suggests that the biological wiring was sensitive to the pattern of feedback, even if it had not yet learned a reliable strategy for the task.

The study also provided a rare look inside the "mind" of the controller while it worked. The researchers recorded the activity of specific regions within the fly-brain model, known as neuropils, which are clusters of neurons that handle specific tasks like smell or vision. They found that certain areas, particularly those involved in processing smell and visual information, were the most active during the experiment. These activity patterns were consistent across all the runs, showing that the biological structure of the brain dictated which parts were engaged, regardless of whether the controller was successful at the task. The researchers could see exactly which parts of the brain were firing and how the activity shifted as the simulation progressed, offering a complete view of the sensorimotor loop from the moment a molecule approached the door to the moment the brain adjusted its internal connections.

It is important to understand what this study did not prove. The author is careful to state that they have not demonstrated that a fly brain can actually act as a Maxwell's demon. In fact, the results suggest the opposite: in this specific setup, the biological controller was a much weaker sorter than the simple script. The study does not claim that the fly brain has learned a new skill or that it has solved the thermodynamic puzzle. Instead, the value of the work lies in the platform itself. The researchers have built a verified system where a biologically constrained network can be tested against a physical environment, with every single event logged and available for analysis. This allows scientists to measure exactly how a brain built from real wiring diagrams behaves when asked to control a physical system, something that was previously difficult to do with such precision.

The researchers acknowledge several limits to their current findings. For instance, the simulation included an air jet that added energy to the system, and this jet was used differently by the two controllers, which may have influenced the temperature results. They also note that they only ran one session for each condition, so the results are descriptive rather than statistically proven. The temperatures in the simulation are measured in arbitrary units, not real-world degrees, and the "rewards" given to the brain were based on a simple rule rather than a complex thermodynamic calculation. These factors mean the study is a proof of concept for the experimental platform, not a final answer on how fly brains handle thermodynamics.

Ultimately, this work opens a door for future research. By coupling a realistic model of a fly brain to a simulated physical world, the researchers have created a tool that can be used to test how different learning rules or different brain structures might perform in the future. The next steps would involve running more sessions to confirm the patterns, testing the system with the air jet turned off to isolate the brain's control, and perhaps even running the simulation on specialized hardware designed to mimic the brain's structure. For now, the study stands as a detailed record of a digital fly brain trying, and mostly failing, to sort gas molecules, while simultaneously showing that the brain was actively changing its own wiring in response to the challenge. It is a demonstration of how we can begin to measure the interaction between biological complexity and physical laws, one simulated molecule at a time.