Deep brain stimulation for Parkinson’s disease works well for some patients and poorly for others. The clinical question before surgery isn’t whether DBS works in general — it’s whether it will work for this specific brain. Du and colleagues build a computational answer: a virtual brain model that predicts individual treatment response before the intervention happens.

The architecture has two stages. First, a generative foundation model trained on 2,707 subjects learns the universal dynamics of brain connectivity — what functional patterns look like across the population. Then, the model is personalized for each patient, generating a virtual brain that reproduces that individual’s functional connectivity with correlation r = 0.935 to empirical measurements.

The virtual brain doesn’t just mimic the patient’s current state. It simulates the effect of stimulation — predicting which brain regions will respond and how the network dynamics will shift. The prediction of clinical outcomes outperforms existing methods because it captures the full network context: not just where the electrodes sit but how the stimulation propagates through a specific patient’s connectivity architecture.

The foundation-to-personalization pipeline is the structural contribution. A model trained on thousands of brains learns what’s universal. A personalized instantiation captures what’s individual. The prediction lives in the gap between the two — how this particular brain deviates from the population norm, and how those deviations interact with the stimulation protocol. The universal model provides the dynamics. The individual data provides the initial conditions. The prediction requires both.