Neuroscientists build computer models of brain parts to study its inner workings. Together with experiments, these models are valuable tools. They are mainly based on averages of many humans or non-human animals, so the conclusions from the model have a better chance of generalizing beyond the individual. However, a model can also be fit to an individual person, which is done to generate models of patients with brain disorders such as epilepsy. These models can be used to support treatments because they can simulate the effects of taking medicine or having surgery.

How can a model that accurately represents a patient’s brain be created? Some aspects can be measured. Magnetic resonance imaging (MRI), for example, can measure the size of brain regions and the thickness of the connections between them. But that is where modern, direct measurements in humans end, leaving us with no way to estimate the synaptic connections between brain areas. Functional MRI (fMRI) tells us about the correlations between brain areas but not about their causal connections. Even electrode recordings that are sometimes done inside the brains of patients only measure correlations. However, the correlations result from the causal connections of the patient’s brain. That is where simulation-based inference comes in.

The patient’s brain generates the measured data. A model likewise generates data when we simulate it, and that data can either be very similar to the observed data or very different. Whether it is different or similar depends on the simulation parameters we define. Simulation-based inference uses the data from the simulator to find the likely parameters that could generate the measured data. It is simulation-based because it uses a lot of data from the simulator. What is inferred is the simulator parameters that are likely to have generated that data, which is similar to the measured data. That is precisely what we need to find useful models.

A big issue remains: many different sets of causal parameters could generate the same correlational data in a complex brain model. How do we know which parameters are faithful to the patient’s brain? The bad news is that for most patients, we will likely never know exactly how their brain works. If we did, we would not need to simulate at all. The good news is that simulation-based inference calculates the probability across the entire plausible parameter range. In other words, it makes the uncertainty explicit, and we can incorporate it into our judgment about the model. Furthermore, as our knowledge, our simulators, our hardware, and our data improve, the uncertainty decreases. This can make simulators more and more valuable over time.

But how useful are personalized brain simulators in today’s medicine? For most brain disorders, there is little practical evidence. In epilepsy care, however, a clinical trial called EPINOV (Improving EPilepsy Surgery Management and progNOsis Using Virtual Epileptic Patient Software) is ongoing. In that trial, neurosurgeons access a personalized brain model to decide which part of a patient’s brain to remove during surgery. Removing parts of the brain is a treatment option for patients who have tried several anti-epileptic drugs to stop seizures. If none of the drugs has stopped the seizures, neurosurgery might be attempted, and if the parts of the brain that cause the seizure are removed, patients can become seizure-free. Neurosurgeons, therefore, need to decide which parts to remove to stop seizures while minimizing side effects. The personalized brain model is fit to MRI data and electrical recordings inside the brain. Besides the personalized brain model, neurosurgeons use all the data and tools they usually use to plan the surgery. Some will have access to the personalized brain model, while others do not. This randomized controlled trial design will be the most substantial test of the usefulness of virtual brain models.

In summary, simulation-based inference integrates data into a model that would otherwise be impossible to integrate. This can make the models more accurate, and studies to determine the effectiveness of these models are ongoing.