AI as a Biological Primitive: Individual Biological Simulation Without Population-Scale Reference Cohorts for Personalized Care and Drug Development

In the quest to treat complex diseases, medicine has long relied on a strategy of averages. Doctors and researchers study large groups of people to find patterns, then apply those general rules to individual patients. This approach works well when everyone responds similarly, but it falters when a person's biology is unique or when a standard treatment fails. To bridge this gap, scientists have begun developing "digital twins"—virtual models of a specific person that can be tested with different treatments before anything is tried on the real body. Traditionally, however, these models are built by taking a broad population's data and tweaking it to fit one person. They start with the many and try to narrow down to the one.

A new line of research challenges this direction. Instead of starting with a crowd and shrinking it down, this approach asks if a single person's own detailed history is enough to build a working model of themselves. If a doctor has enough records of how a patient's body reacts to food, medicine, and daily life, can those records alone create a reliable simulation? This question moves the focus from finding a patient who fits a population model to building a model that fits the patient's own unique history. It suggests that with enough observation, a person does not need to be compared to others to be understood; their own past actions can become the laboratory for their future care.

Maurice Antony Ewing, a researcher at Conquer Medical Health, tested this idea using a dataset of twelve people living with type 1 diabetes. These individuals were monitored closely for eight weeks, generating nearly 200,000 snapshots of their blood sugar levels, insulin usage, and daily activities. The goal was to see if the computer could learn a model for each person using only their own data, without leaning on the average responses of the other eleven people or any larger group. The researchers created what they call an "artificial biological primitive." Think of this not as a static profile, but as a living, breathing digital version of the patient that has learned its own rules of behavior from its own history.

The team first tested whether this self-built model could identify viable paths toward a healthy state. They took moments when a patient's blood sugar was dangerously high and asked the model: "If you were to take a specific action, like a dose of insulin or a bit of exercise, could this move your blood sugar toward a target?" The model successfully identified interventions that would reduce the gap between the current high state and a healthy target by nearly 97% in the most flexible tests. This proved that the model could map the specific relationships between actions and outcomes for that individual, generating candidate interventions rather than asserting causal treatment effects.

However, the more critical test was to see if this individual model was truly different from a model built on the group average. The researchers compared the advice given by the individual's own history against the advice given by a model trained on all twelve people combined. The results were striking: the two models rarely agreed. In fact, they suggested the exact same treatment plan in only about one-quarter of the cases. The shape of the response—the way the body reacted to a change—was so unique to each person that the group average was a poor guide for the individual. This finding rules out the idea that a patient can be accurately represented simply by adjusting a population model; the individual's own history contains a geometry of response that the group simply does not see.

To ensure these digital models were not just guessing, the researchers looked at what happened in the real world when these people actually took insulin or exercised. They measured how much the person's blood sugar trajectory changed after a real intervention compared to times when nothing happened. The data showed that real treatments caused a significant, measurable shift in the body's path, a "deformation" that was nearly twice as large as the natural wiggles seen during stable periods. This confirmed that the body's history holds a clear signal of how it reacts to change, providing the raw material needed to build a reliable simulation.

The study then moved to a practical application: using the digital twin to test new ideas before trying them on the person. The computer proposed thousands of potential changes to treatment, such as adjusting insulin doses or timing. The system then checked these proposals against the person's own past. If the person had never tried a specific adjustment, or if their history showed that a similar move had previously led to bad results, the system flagged it as risky or unsupported. Remarkably, about 78% of the proposed changes had a precedent in the person's own history, and of those, nearly two-thirds had worked well before. This created a safety net where the computer could say, "We have tried this before, and it worked," or "We have tried this, and it failed," without ever needing to experiment on the patient again.

Crucially, the model was not static. As the person's condition changed from hour to hour, the best treatment suggestion changed with it. When the researchers asked the model for advice at different times or with different goals, it gave different answers in more than two-thirds of the cases. This proves that the model is not just assigning a person to a fixed category, like "Type A diabetic," but is acting as a dynamic experimental partner that evolves with the patient.

The researchers also tested if this method worked beyond diabetes. They applied the same logic to population-level data on government policies during the pandemic and to data on how individual brain cells respond to visual stimuli. In both cases, the system could learn from the specific history of that system—whether a country or a neuron—and simulate how it would respond to new inputs. This suggests the approach is not limited to one type of biology but is a general way of using deep observation to understand how any system behaves.

The study concludes that for personalized care and drug development, the most powerful tool may not be a larger database of other people, but a deeper record of the individual. By treating a person's own history as the primary source of truth, doctors and researchers can build a virtual laboratory that is specific to that person. This allows them to screen treatments, reject those that history says will fail, and prioritize those that history says will work, all before a single pill is swallowed or a single injection is given. The work does not claim to have solved the problem of disease, but it offers a new way to think about it: that the answer to a patient's unique problem might be found entirely within their own past, waiting to be read by a machine that knows how to listen.