The mesolimbic system and the loss of higher order network features in schizophrenia when learning without reward

Abstract

Introduction: Schizophrenia is characterized by a loss of network features between cognition and reward sub-circuits (notably involving the mesolimbic system), and this loss may explain deficits in learning and cognition. Learning in schizophrenia has typically been studied with tasks that include reward related contingencies, but recent theoretical models have argued that a loss of network features should be seen even when learning without reward. We tested this model using a learning paradigm that required participants to learn without reward or feedback. We used a novel method for capturing higher order network features, to demonstrate that the mesolimbic system is heavily implicated in the loss of network features in schizophrenia, even when learning without reward.

Methods: fMRI data (Siemens Verio 3T) were acquired in a group of schizophrenia patients and controls (n=78; 46 SCZ, 18 ≤ Age ≤ 50) while participants engaged in associative learning without reward-related contingencies. The task was divided into task-active conditions for encoding (of associations) and cued-retrieval (where the cue was to be used to retrieve the associated memoranda). No feedback was provided during retrieval. From the fMRI time series data, network features were defined as follows: First, for each condition of the task, we estimated 2^nd order undirected functional connectivity for each participant (uFC, based on zero lag correlations between all pairs of regions). These conventional 2^nd order features represent the task/condition evoked synchronization of activity between pairs of brain regions. Next, in each of the patient and control groups, the statistical relationship between all possible pairs of 2^nd order features were computed. These higher order features represent the consistency between all possible pairs of 2^nd order features in that group and embed within them the contributions of individual regions to such group structure.

Results: From the identified inter-group differences (SCZ ≠ HC) in higher order features, we quantified the respective contributions of individual brain regions. Two principal effects emerged: 1) SCZ were characterized by a massive loss of higher order features during multiple task conditions (encoding and retrieval of associations). 2) Nodes in the mesolimbic system were over-represented in the loss of higher order features in SCZ, and notably so during retrieval.

Discussion: Our analytical goals were linked to a recent circuit-based integrative model which argued that synergy between learning and reward circuits is lost in schizophrenia. The model's notable prediction was that such a loss would be observed even when patients learned without reward. Our results provide substantial support for these predictions where we observed a loss of network features between the brain's sub-circuits for a) learning (including the hippocampus and prefrontal cortex) and b) reward processing (specifically constituents of the mesolimbic system that included the ventral tegmental area and the nucleus accumbens. Our findings motivate a renewed appraisal of the relationship between reward and cognition in schizophrenia and we discuss their relevance for putative behavioral interventions.

Keywords: cognition; functional MRI; learning; memory; mesolimbic system; schizophrenia.

Figure 1

Task paradigm. The figure schematically depicts the deployed paradigm. Participants viewed illustrations of nine common objects presented in sequential random order in their associated locations during Encoding (3 s per object; 27 s total, successive presentations of “Key” and “Tree” are depicted). After a brief rest period (27 s), locations were cued in random order. Following each cue, participants were required to verbally name the associated object (Retrieval; 3 s per object; 27 s total). A brief post-retrieval consolidation period followed (27 s). The entire cycle went through eight iterations.

Figure 2

Methods outlined. The pipeline used to estimate higher order features is depicted. (A) In an initial step, the full second order functional connectivity matrix was formed for all sixteen nodes in the network (schematically depicted for four nodes, A-D). The matrix was derived from time series for each node, with each participant contributing matrices for Encoding and Retrieval. The undirected functional connectivity (uFC) was computed as the zero-lag correlation (r) between all unique pairs of nodes (i.e., r_AB ) for each subject, with the coefficient normalized (see Methods). Each correlation measure represents by convention a 2^nd order connectivity feature. (B) Across all participants in each of the HC and SCZ groups, higher order features were estimated from these 2^nd order matrices. As depicted for r _AB and r _CD, in each of the HC and SCZ groups, the resultant higher order feature (r _AB • r _CD) was estimated as the correlation in 2^nd order features across all 32 and 46 participants respectively. (C) The resultant higher order feature matrices represent the intra-group regularities across quadruples of network nodes. From these, intra-group differences in regularity were estimated (see Methods) and statistically thresholded (bottom).

Figure 3

The figure provides a comprehensive accounting of the behavioral results. (A) Averaged learning data at each time point are presented where the curves represent negatively accelerated learning functions, $y=1-e^{-bx}$ , fit to the averaged data (shaded areas represent 95% confidence intervals). As seen, patients display slower learning. This intuition was confirmed in (B). Here, learning functions were separately fit to performance data from each of the participants, and the average learning rates, b are depicted in the bar graph (± s.e.m.). The bar graph confirms that learning rates in patients was significantly lower. (C) Finally, individual performance data were rendered using heat maps. Here each row depicts data from an individual participant (each column represents time, i.e. epoch). Participants in each group are arranged in descending order of learning rate (Top: Fastest → Bottom: Slowest).

Figure 4

Differences in higher Order Network Features for each of Encoding (A) and Retrieval (B) are represented using chord diagrams. To denote the specific higher order feature, the 16 node identities are denoted on the outer ring. Then, the labels in the inner ring are organized to form a unique pair of nodes that form one of the elements of the higher order feature. Finally, each visible chord links pairs of network pairs denoted by the combination of the outer and inner label. Red chords connect pairs with greater higher order significance in in HC while blue chords connect pairs with greater higher order significance in SCZ. As seen, SCZ are characterized by a massive loss of higher order network features.

Figure 5

The stacked bar graphs represent the total number of instances a node is a member of a given chord in Figure 4 . These graphs are constructed separately for Encoding (left) and Retrieval (right) conditions. The graphs provide a measure of the relative contribution of any node to a loss or gain of higher order features in schizophrenia in each of the conditions. The color scheme (red/blue) is maintained from Figure 4 . In addition to the notable contributions of regions central to learning and memory, the contributions of nodes in the mesolimbic system (bottom), particularly during Retrieval, are highly salient.

Figure 6

The data from Figure 5 are re-visualized for each of the (A) Encoding and (B) Retrieval conditions as numerically scaled circles (only where there is a loss of higher order features) where each circle represents a node placed in its approximate anatomical location on a lateral or medial surface of the brain. Each node’s diameter is scaled to reflect its contribution to higher order feature loss: diameter $d=\frac{\sqrt{Chord} Frequency}{2 x 10}$ . The relative contributions of the nodes in the mesolimbic pathway (NAcc and VTA) are highly evident by their depicted sizes (highlighted for reference).