Topological basal ganglia model with dopamine-modulated spike-timing-dependent plasticity reproduces reinforcement learning, discriminatory learning, and neuropsychiatric disorders
The basal ganglia (BG) are central to action selection and reinforcement learning, yet how the topological organization of the BG circuit with dopamine (DA) D1- and D2-receptors shape learning remains unclear. We present a topologically organized spiking model of macaque BG with cortico-striatal inputs organized into competing channels, D1/D2 medium spiny neurons (MSNs), three-factor DA-modulated STDP for cortical synapses, asymmetric intra-striatal collaterals, and partially overlapping direct/indirect pathways. We validate resting activity and action selection, then study conditioning and generalization-discrimination learning using DA bursts (CS+) and dips (CS). Two structural determinants emerged. First, pathway overlap ({lambda}) trades off selection efficiency and learning speed: higher overlap degrades GPi-based selection efficiency during conditioning, yet accelerates convergence during discrimination by strengthening D2 influence on GPi. Second, lateral inhibition from MSN-D2 to MSNs ({kappa}) helps constrain competing actions but is not sufficient alone; robust discrimination requires DA-dip-dependent up-modulation of D2 collateral efficacy (), which speeds and, at low overlap, enables convergence. Simulations under Parkinsonian and schizophrenia-like settings showed different deficits. A hypodopaminergic "Parkinsonian" STDP regime (D1 LTP loss, D2 LTD loss) impaired conditioning and failed to enhance discrimination. In contrast, attenuated D2 plasticity during DA dips (modeling methamphetamine-induced changes/schizophrenia-related dysregulation) selectively disrupted discrimination while sparing conditioning. Finally, we demonstrate efficient scaling on the Fugaku supercomputer to rodent and non-human primate-relevant sizes, supporting large-scale, biologically grounded BG simulations. Together, the results highlight how pathway overlap and D2 collateral dynamics jointly regulate the speed and reliability of discrimination learning, and how specific DA perturbations map to distinct learning impairments.