A REINFORCEMENT LEARNING-BASED NEURAL CIRCUIT MODEL OF SCHIZOPHRENIA: PREFRONTAL EXCITATION/INHIBITION IMBALANCE DEGRADES AND REVERSES FEEDBACK ALIGNMENT IN THE BASAL GANGLIA-MIDBRAIN CIRCUIT

Schizophrenia has positive, negative, and cognitive symptoms, characterized by impairments in credit assignment, motivational control, and working memory, respectively. Regarding the causes, dopamine and glutamate abnormalities in the cortico-basal ganglia (BG)-midbrain circuits and prefrontal excitation/inhibition imbalance have been suggested. We present a model, which could explain such diverse symptoms and causes, by elaborating reinforcement learning-based models of the cortico-BG-midbrain circuits. Standard models describe that state value can be learned through reward-prediction-error (RPE) encoded by dopamine, but cannot explain (i) formation of context-dependent appropriate state representation and (ii) reward-specific motivational control. For (i), our recent work (Tsurumi et al. 2025 eLife) showed that training of prefrontal recurrent network by RPE enabled learning of appropriate state representation through Feedback-Alignment (FA) (Lillicrap et al. 2016 NatCommun: a bio-plausible alternative to the Back-Propagation algorithm) of cortico-BG-midbrain connections. For (ii), Millidge and colleagues (2024 PLoSComputBiol) showed that training of BG-midbrain by multi-dimensional RPE enabled reward-specific motivational control though BG-midbrain FA. We integrate these two models and show that learning of appropriate state representation and reward-specific motivational control can be simultaneously achieved through 'double FAs'. Crucially, prefrontal excitation/inhibition balance is a key for successful learning: excessive excitation induces aberrant persistent activity, which degrades FA in BG-midbrain and even reverses it into anti-alignment. This impairs reward-specific motivational control, and further, credit assignment, explaining the schizophrenia's negative and positive symptoms, respectively, while aberrant persistent activity itself could explain cognitive (working memory) impairments. Our model (Morita & Kumar, 2025 bioRxiv) could potentially explain reported altered brain activations in patients.