Mechanistic corticostriatal circuit model predicts learning-dependent fMRI dynamics and individual reward bias in...

Circuit-level computational models can do more than explain existing data; they can generate novel hypotheses and capture individual differences in human populations. We demonstrate this using a biophysical corticostriatal model, transforming simulated neuronal activity from local field potential (LFP) into functional magnetic resonance imaging (fMRI) signals via the balloon model, and finally generating behavioral outcomes. We then validate the results against those obtained from human subjects. The model generates a counterintuitive yet testable prediction. Although prefrontal-striatal coherence demonstrates an increase at the neural LFP scale during learning, the same circuit-passed through the hemodynamic transform--predicts a decrease in prefrontal-striatal BOLD correlation, driven by category-selective representations that emerge in cortex but not striatum. This prediction is confirmed in fMRI data optimized for single-subject-level detection sensitivity. The model further enables single-subject fitting of a behavioral outcome, classifying individuals into those differentially responding to positive versus negative reward bias. Established biomarkers, including activation measured by the amplitude of low-frequency fluctuations and dopaminergic effects on hemodynamic latency, are also conserved across LFP and fMRI scales. These findings reposition circuit models as generative tools for human neuroscience, capable of producing mechanistically grounded hypotheses and parsing individual variation in ways inaccessible to data-driven approaches alone.