Computational physiology of the Basal Ganglia, movement disorders, and their therapy

Classical models of the basal ganglia (BG) depict them as closed cortex-BG-cortex loops. The motor cortex activity is driven by the opposite effects of dopamine on the excitability of striatal projection neurons expressing D1 and D2 dopamine receptors and the subsequent change in discharge rate of neurons in the direct and indirect BG pathways. More modern computational BG models depict them as an actor/critic machine learning network. The BG main axis (actor) connects the cortical networks encoding the current state of the subject to the BG input stages (striatum) and continues through BG downstream structures to the cortical and subcortical motor centers. Dopamine modulates the coupling between the state and motor encoding networks by the modulation of the efficacy of the cortico-striatal synapses. Here, we present a novel computational model of the BG network that combines the main features of both classical and modern BG models. The BG networks are built as actor/critics network. The dimensionality reduction networks of the BG main axis (actor) connect the thalamocortical networks encoding the current state of the subject to the BG input stages (striatum and the subthalamic nucleus, STN). The information then flows through the central nucleus of the basal ganglia (the external segment of the globus pallidus, GPe) to the BG output stages (internal segment of the globus pallidus and the substantia nigra reticulata, GPi and SNr, respectively) that innervate the cortical and subcortical (brainstem) motor centers. The main computational goal of the BG is multi-objective optimization of behavior (e.g., to maximize future cumulative gains and minimize costs). The competitive networks of the BG main axis flexibly extract relevant features for ongoing and future actions from the current state of the thalamocortical activity. The BG critics (neuromodulators) include the dopaminergic, cholinergic, serotonergic, and histaminergic projections to the striatum. These BG critics differentially modulate the excitability of striatal neurons and the efficacy of the cortico-striatal synapses. Modulation of striatal and BG excitability enables instantaneous optimal trade-offs between exploratory (gambling) and exploitative (greedy) behaviors. Adjustment of the cortico-striatal synaptic efficacy empowers long-term learning of optimal behavioral policy (state-to-action associations). Degeneration of midbrain dopaminergic neurons and other BG neuromodulators (e.g., in Parkinson’s disease) leads to abnormal competitive dynamics and synchronous oscillatory discharge of the neurons in the BG main axis. Because the BG networks are the default connection between the neural networks encoding state and actions, the other neural networks (e.g., cortico-cortical networks) cannot compensate for the abnormal BG activity. Therapy of BG-related movement disorders can be achieved by either dopamine replacement therapy (DRT) or by functional inactivation of the BG main axis, as achieved by deep brain stimulation (DBS) paradigms. Functional inactivation of the BG main axis enables compensation by other neuronal networks and restoration of close-to-normal state-to-action coupling. Future DBS therapies might be improved by mimicking the BG multi-objective optimization paradigms. These therapies should aim at restoring normal BG activity, motor behavior, and quality of life by the provision of more precise (in time and space) functional inactivation of the abnormal activity in the BG networks.