Digital twin brain generates personalized behavior predictions from connectomes, paving the way for individualized psychiatry

A groundbreaking digital twin brain framework capable of translating an individual's brain connectome into high-fidelity predictions of multitask behavior has been developed by researchers at the National Center of Neurology and Psychiatry, Japan, and Tohoku University. The work, published in BME Frontiers, offers a transformative approach to personalized psychiatric care by simulating how unique neurobiology drives complex cognitive and affective functions.

Personalized psychiatry has long sought models that can predict an individual's behavioral and neural responses across multiple functional domains—such as emotional processing and cognitive control—from their distinct brain structure. Existing approaches, however, struggle to bridge the gap between static connectomes and dynamic, multitask behavior, limiting their clinical utility.

To overcome this challenge, the research team designed a novel two-component architecture. A hypernetwork uses an individual's resting-state functional connectome to generate personalized parameters for a main recurrent neural network. This main network then simulates participant-specific behavioral choices, reaction times, and blood-oxygen-level-dependent (BOLD) signals across tasks probing both affective and cognitive domains.

Validation studies on 228 participants—including individuals with psychiatric disorders and healthy controls—demonstrate remarkable predictive capabilities. The system achieves over 90% accuracy in predicting behavioral choices across diverse tasks, with reaction time correlations exceeding r > 0.85. More impressively, it predicts BOLD signal patterns with r = 0.84 accuracy, validated through group-level GLM analysis. These metrics confirm the system's ability to capture complex neurobehavioral dynamics with exceptional fidelity.

Crucially, by leveraging the end-to-end architecture linking connectomes directly to behavior, the team used gradient backpropagation to identify connectome manipulations that selectively modulate targeted functions. In silico interventions successfully altered amygdala response strength (an affective indicator) and processing speed (a cognitive indicator), revealing realistic interindividual variability in simulated treatment effects arising from each person's baseline connectome.

While the study's sample size and task scope present limitations, the digital twin framework's ability to learn flexibly from sensory inputs and behavioral outputs opens avenues for modeling real-life dynamics. Future work integrating microscopic (e.g., molecular) scales and larger datasets could extend simulations to pharmacological interventions.

This research represents a significant step toward mechanistic, individualized psychiatry, offering a powerful platform for understanding pathophysiology and designing personalized therapies.

END