LATENT SUBSTRATES OF COMPULSIVITY AND BEHAVIOURAL FLEXIBILITY

Abstract

Parallel decisional systems constantly compete for control of our day-to-day choices. Recent computational models propose two distinct forms of reinforcement learning, model-based and model free, which drive goal-directed and habitual behaviours, respectively. Although most of us seem to effortlessly alternate between the two systems, several pathological disorders have been associated with their imbalance. Thus the need to further characterize such systems is clear.To examine the neural networks of these parallel processes, we employ a two-step sequential learning task used previously to show simultaneous engagement of model-based and model-free learning in healthy volunteers [1] and their imbalance in pathological states [2]. We separately acquired fMRI data from 66 healthy volunteers during rest using a novel multi-echo planar sequence that greatly boosts signal-to-noise ratios compared to traditional single-echo sequences thus allowing higher spatial resolution of subcortical structures [3]. We quantified the connectivity between regions of interest (ROI’s) by calculating Pearson correlations coefficients (R) between a priori hypothesized regions (henceforth referred to as ‘connectivity’). These connectivity values were then correlated with behavioural measures. Higher model-based learning was positively correlated with medial orbitofrontal cortex and ventral striatum connectivity (R=0.32, p=0.01). Model-based scores also positively correlated with connectivity between dorsolateral prefrontal cortex and ventral striatum (R=0.27, p=0.03). In contrast, model-free learning was positively correlated with posterior putamen and supplementary motor area connectivity (R=0.266, p=0.033); and negatively with medial orbitofrontal cortex and left amygdala connectivity (R=-0.255, p=0.042). Model-free learning was also positively correlated with right ventral striatum and amygdala connectivity (R=0.274, p=0.028). We finally show that cognitive flexibility in the form of attention set shifting is associated with functional connectivity between dorsolateral prefrontal cortex and ventral striatum (R=-0.298, p=0.021).These findings suggest a neural network for model-based learning involving: integration of instrumental performance (ventral striatum); flexible, computationally-driven updating of outcome value based on changing internal motivational states and external feedback (medial orbitofrontal cortex); and computation of the relative advantage of alternative action values (dorsolateral prefrontal cortex). The current findings provide a more integrative network model of how previously implicated cortical regions organize to drive goal-directed model-based learning. Whereas most previous findings have focused on the neural correlates of model-based learning, we highlight a motoric frontal-striatal network associated with model-free learning. Together, we illustrate the sensitivity of the multi-echo sequence for the distinguishment of such elusive behavioural control systems.