Personalized psychiatry demands digital twins that translate individual brain connectomes into predictions of multidomain functions (e.g., emotion, cognition) and treatment responses. We present a digital twin brain framework where a Hypernetwork, using resting-state connectomes, generates parameters for a Main recurrent network. This enables the Main network to jointly simulate behavioral and blood-oxygen-level-dependent (BOLD) time series across multiple tasks. Validated in a transdiagnostic cohort (n=228), our model demonstrated high-fidelity prediction of behavioral choices (90% accuracy), reaction times (r0.85), and distributed BOLD activation (r=0.84) during affective and cognitive tasks, capturing individual neurobehavioral dynamics. Leveraging the model's differentiability, we identified gradient-based connectome manipulation targets (e.g., limbic, parietal, prefrontal) to selectively modulate affective or cognitive functions. Crucially, in-silico trials simulating these manipulations reproduced realistic inter-individual variability in treatment effects (effect size range 2 SD). This establishes our digital twin brain as a quantitative, mechanistic platform for simulating and designing personalized psychiatric interventions.
Takahashi et al. (2025) studied this question.