AI as a Biological Primitive Enables Targeted Inversion Across Disease, Population, and Neuronal Systems

Most of the time, when we use artificial intelligence to understand biology, we ask it to look forward. We feed a computer a picture of a patient's current health or a snapshot of a virus spreading through a city, and we ask the machine to predict what will happen next. Will that person's blood sugar rise tomorrow? Will the number of infections increase next month? This forward-looking approach has become the standard for modern medicine and public health. It is a powerful way to forecast risk, but it leaves a crucial question unanswered. A doctor or a policy maker rarely needs to know only what will happen; they need to know what to do to make something different happen. They need to know exactly how much insulin to give, or how strict a lockdown should be, to move a system toward a specific, desired outcome.

This is the difference between watching a river flow and knowing which rocks to move to change its course. For decades, scientists have tried to build "digital twins"—virtual copies of patients or populations that mimic real life so closely they can be tested with different treatments. However, these digital twins often fail because they are built on averages. They describe how a typical person reacts, not how a specific individual reacts to a specific action. If a model has never seen a particular drug in its training data, it cannot reliably tell you what that drug will do for a specific patient, even if it gives a confident-sounding answer. The result is a system that might predict well but cannot safely guide action.

A new study by Maurice Antony Ewing proposes a different way to think about these computer models. Instead of building a twin that mimics a person, the author treats a trained AI model as a "biological primitive." This is a simpler concept: the model is not a replica of a person's entire history, but a learned record of how a system actually responded to the things done to it in the past. If the data shows that a specific amount of exercise lowered blood sugar in a group of people, the model learns that specific relationship. The researcher then turns the model around. Instead of asking, "What happens if I do this?", the model is asked, "What amount of this action will get us to that goal?" This process, called targeted inversion, searches for the precise levels of intervention needed to move a system toward a target state, but only using the levers that the data actually supports.

To test if this idea works in the real world, the study applied the method to three very different biological systems: the individual physiology of people with type 1 diabetes, the population dynamics of the COVID-19 pandemic, and the microscopic responses of neurons in the brain. The researchers started by training standard AI models to predict future states in each of these areas. For the diabetes data, the models learned to predict future blood glucose levels based on past insulin, food, and activity. For the pandemic data, they learned to predict future case numbers based on policies like lockdowns and testing rates. For the brain data, they learned how neurons fired in response to specific visual stimuli. These forward models were highly accurate, proving they had learned the underlying patterns of the data.

Once the models were trained, the researchers asked them to work backward. They set a target—for example, a safe blood sugar level for a diabetic patient or a lower number of infections for a city—and asked the model to find the specific combination of actions that would get there. In the diabetes group, which involved 400 high-risk cases, the method was remarkably effective. When the researchers asked the model to find the right amount of physical activity alone, it reduced the gap between the patient's current state and the target by 76.5 percent. When they asked for the right amount of insulin alone, the gap shrank by 60.6 percent. When they allowed the model to combine both activity and insulin, but kept the suggestions within safe, clinical limits, the gap was reduced by 87.8 percent. If they removed those safety limits and let the model search freely, it found solutions that reduced the gap by 96.4 percent. In almost every single case, the model found a way to move the system closer to the goal.

The results were slightly more modest but still consistent when applied to the larger scale of the COVID-19 pandemic. Here, the goal was to reduce the burden of future cases. When the model searched for the right level of lockdown measures, it reduced the projected gap by 17.9 percent. Testing alone helped by 5.8 percent, and vaccination by 3.8 percent. However, when the model combined these three strategies, it reduced the gap by 25.1 percent. This difference highlights a key finding: while a single action can help, the complex dynamics of a pandemic are shaped by the combination of many factors working together. The method showed that population-level problems require a mix of interventions rather than a single silver bullet.

Perhaps the most surprising part of the study was how much the researchers could achieve without searching through every possible combination of actions. In the real world, there are thousands of potential treatments or policies, but many of them are not supported by the data available. The study tested a "compact" approach, where the model was restricted to searching only through a small, curated list of interventions that are known to be plausible and are actually recorded in the data. For the diabetes cases, this compact list recovered 79.0 percent of the possible improvement toward the target. The full, unrestricted search only added another 2.2 percent of improvement. For the pandemic data, the compact list recovered 85.9 percent of the improvement, with the full search adding just 0.5 percent more. This suggests that we do not need to guess wildly or search through millions of impossible options to find effective solutions. We only need to look at the actions we know exist and have data for.

The study also demonstrated a critical safety feature of this approach: the model refuses to invent solutions it cannot support. In the diabetes group, the researchers tried to ask the model to find solutions involving certain medications, such as metformin or GLP-1 drugs. Because the original data for these patients did not include records of those specific drugs, the model simply skipped them. It did not hallucinate a result or pretend it knew what would happen. It returned a flat answer, indicating that it could not target those specific levers. This is a stark contrast to many current AI systems, which might confidently suggest a treatment even if they have never seen data on it. The "primitive" approach is grounded in what the data actually contains, making it a safer tool for decision-making.

The same logic was applied to the brain. Using data from the Allen Brain Observatory, which contains over 1.2 million observations of how neurons respond to light, the researchers built a primitive that learned how specific cells react to stimuli. They were able to ask the model what level of stimulus would move a neuron toward a specific firing pattern. This showed that the method works not just for whole people or entire populations, but for the smallest units of biology as well. The neuron, the patient, and the city are all treated as systems that respond to inputs, and the model learns the map of those responses.

The author is careful to state that this is a method of inquiry, not a final prescription. The results come from looking back at historical data, not from running new clinical trials. The model identifies the levels of intervention that would move a system toward a target based on past behavior, but it does not prove that those interventions will work in a new, real-world setting. It is a way to narrow down the possibilities and find the most promising paths to explore. The study argues that by treating AI as a biological primitive that can be interrogated in reverse, scientists can move beyond simple prediction. Instead of just asking what will happen, they can ask what level of action is needed to make something happen, provided they stick to the levers that the data supports. This shift from guessing the future to targeting the present offers a more grounded, practical way to use artificial intelligence in medicine and public health.