SCHRES-06993; No of Pages 9 Schizophrenia Research xxx (2016) xxx–xxx Contents lists available at ScienceDirect Schizophrenia Research journal homepage: www.elsevier.com/locate/schres Computational approaches to schizophrenia: A perspective on negative symptoms Lorenz Deserno a,b,c,⁎, Andreas Heinz a,b, Florian Schlagenhauf a,b a b c Max Planck Fellow Group ‘Cognitive and Affective Control of Behavioral Adaptation’, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany Department of Psychiatry and Psychotherapy, Campus Charité Mitte, Charité-Universitätsmedizin Berlin, Berlin, Germany Department of Child and Adolescent Psychiatry, Psychotherapy and Psychosomatics, University of Leipzig, Leipzig, Germany a r t i c l e i n f o Article history: Received 23 December 2015 Received in revised form 22 September 2016 Accepted 1 October 2016 Available online xxxx Keywords: Schizophrenia Decision-making Reinforcement learning Computational modelling Computational psychiatry Negative symptoms a b s t r a c t Schizophrenia is a heterogeneous spectrum disorder often associated with detrimental negative symptoms. In recent years, computational approaches to psychiatry have attracted growing attention. Negative symptoms have shown some overlap with general cognitive impairments and were also linked to impaired motivational processing in brain circuits implementing reward prediction. In this review, we outline how computational approaches may help to provide a better understanding of negative symptoms in terms of the potentially underlying behavioural and biological mechanisms. First, we describe the idea that negative symptoms could arise from a failure to represent reward expectations to enable flexible behavioural adaptation. It has been proposed that these impairments arise from a failure to use prediction errors to update expectations. Important previous studies focused on processing of so-called model-free prediction errors where learning is determined by past rewards only. However, learning and decision-making arise from multiple cognitive mechanisms functioning simultaneously, and dissecting them via well-designed tasks in conjunction with computational modelling is a promising avenue. Second, we move on to a proof-of-concept example on how generative models of functional imaging data from a cognitive task enable the identification of subgroups of patients mapping on different levels of negative symptoms. Combining the latter approach with behavioural studies regarding learning and decision-making may allow the identification of key behavioural and biological parameters distinctive for different dimensions of negative symptoms versus a general cognitive impairment. We conclude with an outlook on how this computational framework could, at some point, enrich future clinical studies. © 2016 Published by Elsevier B.V. 1. General introduction Schizophrenia patients often report inability to experience pleasure, withdrawal from social interactions, reduced ability to pursue meaningful goals, and they are characterized by a reduction in emotional and verbal expression. These negative symptoms have a high prevalence in schizophrenia patients with more than half of the patients displaying at least two of the PANSS items: blunted affect, emotional withdrawal, poor rapport, social withdrawal, or reduced verbal fluency (Bobes et al., 2010). Clinical records show that more than a third of patients are described as having poor motivation and blunted affect (Patel et al., 2015). Negative symptoms are important for functional outcome (Fervaha et al., 2013a, 2014; Galderisi et al., 2013, 2014) and considerably influence a patient's quality of life (Eack and Newhill, 2007). While the assessment of the domain of negative symptoms has ⁎ Corresponding author at: Max Planck Fellow Group ‘Cognitive and Affective Control of Behavioral Adaptation’, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany. E-mail address: deserno@cbs.mpg.de (L. Deserno). improved considerably, our understanding of the underlying pathophysiological mechanisms still remains limited. On a phenomenological level, negative symptoms such as anhedonia suggest a link with neurobiological systems underlying motivational and reward processes (Wise, 2008). Thus, the dopaminergic system, also referred to as the brain's reward system, is a likely candidate system on a biological level. Reward anticipation in the ventral striatum, a core dopaminoceptive region, during reward anticipation has been found to be associated with the degree of negative symptoms (Juckel et al., 2006b). However, there have also been reports of null findings (Esslinger et al., 2012), of an association between ventral striatal activation and positive symptoms (Nielsen et al., 2012b) or of correlations between negative symptoms and activation in dorsal rather than ventral subregions of the striatum (Mucci et al., 2015). Furthermore, evidence shows that antipsychotic treatment affects these ventral striatal signals (Juckel et al., 2006a; Nielsen et al., 2012a; Schlagenhauf et al., 2008). Interestingly, a recent meta-analysis of the available fMRI studies in schizophrenia patients confirmed a reduction of ventral striatal activation and supports an association with negative symptom severity (Radua et al., 2015). On the other hand, while elevated striatal http://dx.doi.org/10.1016/j.schres.2016.10.004 0920-9964/© 2016 Published by Elsevier B.V. Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 2 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx dopaminergic function is still a cornerstone in our pathophysiological understanding of schizophrenia, it has primarily been associated with positive symptoms (Heinz and Schlagenhauf, 2010; Howes and Murray, 2014). The aberrant salience hypothesis links dopaminergic hyperactivity with the development of positive symptoms via an inappropriate attribution of salience to otherwise neutral stimuli or internal representations (Heinz, 2002; Kapur, 2003). Besides the striking evidence for an involvement of hyperdopaminergic state with psychosis and especially positive symptoms (Howes and Murray, 2014), findings also suggest a link between dopaminergic dysfunction and negative symptoms. For example, negative symptoms were demonstrated to be negatively associated with D2 receptor availability (Heinz et al., 1998). Further, change in D2 receptor availability due to dopamine depletion, mirroring lower dopamine concentration, was inversely correlated with negative symptoms (Kegeles et al., 2010). Moreover, the inability to differentiate relevant compared to irrelevant stimuli, as a measure of aberrant salience, was not only associated with delusion severity but also significantly correlated with negative symptoms in patients with schizophrenia (Roiser et al., 2009). An involvement of the dopaminergic reward system in the development of both negative and positive symptoms seems to require a dysregulation in opposite directions (with an overactivation related to positive symptoms and an underactivation contributing to negative symptoms) or a contribution of distinct subsystems to either symptom dimension (Ziauddeen and Murray, 2010). Taken together, while evidence exists for dysfunctions in reward processing in schizophrenia, our understanding of how these relate to different aspects of psychopathology, in particular the different facets of negative symptoms, is still limited. It is noteworthy that there is shared variance of negative symptoms with well-known cognitive deficits of schizophrenia patients (Hartmann-Riemer et al., 2015; Strauss and Gold, 2012). A meta-analysis showed that dimensions of negative symptom dimensions are significantly but only modestly associated with cognitive deficits (effect sizes: −0.29 to −0.12, Dominguez et al., 2009). For example, a longitudinal study in first-episode schizophrenic patients revealed working memory performance to be a predictor of the degree of negative symptoms after a 5-year follow-up (Gonzalez-Ortega et al., 2013). As well as negative symptoms, cognitive deficits predict functional outcome (Bowie et al., 2008; Nuechterlein et al., 2011) and also the transition to schizophrenia (Fusar-Poli et al., 2012). Unfortunately still unrevised, conventional treatment strategies have largely failed to reduce working memory deficits and negative symptoms (Fusar-Poli et al., 2015; Hyman and Fenton, 2003; Lieberman et al., 2005). Accordingly, studies investigating associations between neurobiological measures and negative symptoms have not yet revealed a consistent pattern (Galderisi et al., 2015). Heterogeneity with regard to study design, sample characteristics and assessment of psychopathology may have contributed to this picture. For example, neuroimaging studies investigating reward anticipation and processing in schizophrenia often lack a detailed differentiation of negative symptom subdomains with some notable exceptions (Mucci et al., 2015). Psychometric research has defined different aspects of negative symptoms such as avolition, anhedonia, social withdrawal, reduced emotional and verbal expression (Kirkpatrick et al., 2006), which can be grouped into the factors avolition (including amotivation, anhedonia and asociality) and deficit of expression (including affective flattening and alogia) hypothetically arising from distinct pathophysiological mechanism. Nevertheless, a correlation between a certain symptom and a neurobiological measure in a particular sample does not allow any causal conclusions. For example, the presence (or absence) of an association might be due to other mediating factors or change over the course of the illness. This is further complicated by the fact that neuroimaging techniques like fMRI only provide limited mechanistic insights into the neurophysiological processes because they rely on proxy measures of neuronal activation, a problem we describe in more detail in Section 3 of this article. In this review, we strive to demonstrate that computational approaches may help to characterize behavioural and neurobiological processes contributing to negative symptoms more precisely. In the emerging field of computational psychiatry (Huys et al., 2016; Montague et al., 2012; Stephan and Mathys, 2014; Wang and Krystal, 2014) the term ‘computational’ is used broadly. In this article, we follow a nomenclature as in Stephan et al. (2015) by using the “umbrella term” computational models. Further, Stephan et al. (2015) summarized three categories of computational models: 1) biophysical network models; 2) so-called model-based neuroimaging analysis building upon computational models of behaviour; 3) generative models of neuroimaging data. In this article, we include the latter two of these three categories since these models can be inverted (see comment in Section 3). In Section 2, we first review evidence for altered reinforcement learning and decision-making in schizophrenia and their putative contribution to understand the origin of negative symptoms focusing on the behavioural level. We then discuss in Section 3 an example of how mechanistic models of functional imaging data measured during a cognitive task revealed distinct biological subgroups with regard to severity of negative symptoms. In the final Section 4, we outline how model parameters carrying mechanistic information (as discussed in Sections 2 and 3) could enable a biological characterization and stratification of cognitive and motivational mechanisms contributing to negative symptoms. 2. Understanding flexible goal-directed decision-making: a potential origin of the formation of negative symptoms It has been hypothesized that negative symptoms could arise from deficits in representing and using values for decision-making (Barch and Dowd, 2010; Gold et al., 2008). This notion has also received empirical evidence with particular noteworthy contribution from Gold, Waltz and colleagues (for reviews see: Gold et al., 2008; Strauss et al., 2014; Waltz and Gold, 2015). Schizophrenia patients are surprisingly unimpaired in hedonic (“in-the-moment”) experiences (Barch and Dowd, 2010). For example, it was demonstrated that stable-medicated, chronic patients do not differ in ratings of affective pictures with respect to either motor responses to repeat or to endure viewing affective material (Heerey and Gold, 2007). Similar results were reported in other studies using similar affective picture material (Dowd and Barch, 2010; Pankow et al., 2013; Ursu et al., 2011) and have been confirmed by meta-analysis (Cohen and Minor, 2010). Interestingly, and in contrast to these apparently intact in-the-moment experiences, schizophrenia patients are impaired in value-based decision-making. Using a probabilistic selection task in medicated chronic patients, Waltz et al. (2007) found an overall impairment in learning acquisition. Interestingly, however, in a novel post-acquisition testphase, the preference for previously rewarded stimuli was weakened in patients compared to controls, but patients were able to avoid stimuli associated with negative outcome (Waltz et al., 2007). Following up this finding in medicated, chronic schizophrenia patients showed an overall Go-bias together with a deficit in Go-learning during a Go–NoGo learning task. Although patients were impaired in rapid trial-by-trial adaptation to negative feedback, they gradually learned from negative feedback. These findings were predicted by a neurocomputational model of learning in fronto-striatal circuits (Frank et al., 2004) when increasing presynaptic (tonic) dopamine input: high levels of presynaptic dopamine may specifically impair learning in Go-pathways via D1 receptors due to drowning of phasic dopamine bursts facilitating reward-approach behaviour. However, spared punishment avoidance may result either from D2 hypersensitivity or antipsychotic medication. This deficit in Go-learning together with relatively intact NoGo-learning was confirmed in a subsequent study and was most pronounced in patients with high levels of negative symptoms (Strauss et al., 2011). Interestingly, the same study also showed that schizophrenia patients had reduced uncertainty-driven explorative behaviour (by applying computational modelling of behaviour) and that this was specifically correlated with inter-individual differences in anhedonia. This points towards the possibility of separating specific dimensions of negative symptoms via Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx decision-making tasks in combination with computational modelling of the behaviour. Gold et al. (2012) subsequently studied learning from rewards versus learning from punishments in stable-medicated, chronic patients (Fig. 1). Based on a median split on the avolition and anhedonia items from the Scale for the Assessment of Negative Symptoms, a deficit in learning from rewards was found in patients with high negative symptoms only, whereas learning from punishments was intact. Consistently, this association of learning from rewards with negative symptoms was also found in two independent studies in medicated patients (Somlai et al., 2011; Yilmaz et al., 2012). To better understand the underlying mechanism of learning, Gold et al. (2012) applied computational modelling to dissociate different types of learning signals to update values. Formally, one algorithm was model-free Q-learning, where each possible stimulus available for actions, both in reward and punishment conditions, becomes associated with a single value and these specific values are used to compute a prediction error. This model represents the specific outcome value and thus results in choosing frequently rewarded over frequently losing stimuli (Gold et al., 2012). The authors contrasted this to a more rigid model-free actor-critic algorithm (see Box 1 for description of different reinforcement learning models). This model cannot generate the pattern of choosing frequently rewarded over frequently losing stimuli but can explain the observed pattern of “choosing the frequent loss avoiders compared to infrequent winners”. These two patterns of behaviour took place in a post-acquisition test indicating that both mechanisms (Q-learning and actor-critic learning) were at work during acquisition, potentially represented by a rapid orbito/-prefrontal versus a slow basal-ganglia system. In this post-acquisition test, only patients with high-negative symptoms showed behaviour as predicted by actor-critic learning, while controls and low-negative symptom patients showed aspects of both learning mechanisms. Thus, when mixing both models in one algorithm and estimating the weight each individual attached to each mechanism, the influence of Q-learning was significantly reduced in patients with high negative symptoms compared to low negative symptom patients. This study presents a nice example of how computational modelling of choice behaviour can enable our understanding of potential mechanisms underlying negative symptoms in schizophrenia. 2.1. Goal-directed decision-making Given results of intact in-the-moment hedonic experiences but impaired decision-making, negative symptoms might relate to dysfunctions in the ability to learn values from such experiences and to guide flexible and goal-directed behaviour. Such values can be learned via reward 3 prediction errors representing the difference between an actually received outcome value and its expected value. Reward prediction errors were shown to scale with midbrain phasic dopamine releases (Montague et al., 1996; Schultz et al., 1997) and to be causally involved in learning cue-reward associations (Steinberg et al., 2013). Reward prediction errors are retrospective and strongly induce repetition of previously reinforced actions. The endpoint of such habitual learning, when action-outcome contingencies remain stable long enough over training, is a “habit”: a stimulus-response pattern leading to automatized responses even when outcome contingencies may have changed (Dickinson, 1985). This type of learning saves cognitive resources at the cost of flexibility. Based on a computational formulation of habitual learning, such learning is depicted in temporal-difference and Q-learning algorithms (i.e., retrospective learning to act via experiencing outcomes) and is “model-free” because it neglects the causal structure of the environment (e.g., without building maps of the action-outcome relationships). On the other hand, a goal-directed system of behavioural control is responsible for mapping the causal and most likely probabilistic relationship of actions and outcomes (Balleine and O'Doherty, 2010; Dolan and Dayan, 2013). Thus, a goal-directed system builds an internal representation (a “model”) of the environment (or the task) to track actions and their probabilistic outcome consequences. This enables forward planning of actions allowing for rapid behavioural adaptation after changes in action-outcome contingencies at the cost of considerable computational load (Doll et al., 2012). In computational modelling, this was described as “model-based”. Both modes of behavioural control can be considered adaptive per se. But based on the situational characteristics the dominance of one or the other control mechanism might be more appropriate and such arbitration is a demanding challenge likely taking place dynamically, for example, based on the uncertainties about values in each system (Daw et al., 2005). 2.1.1. Responding less towards devalued outcomes indicates goal-directed control The supposed “gold standard” in measuring goal-directed behaviour is outcome devaluation (Dolan and Dayan, 2013; Tricomi et al., 2009; Valentin et al., 2007). In such paradigms, after initial acquisition of action-outcome contingencies, an outcome is devalued, for example, due to feeding to satiety, and subsequently actions towards the devalued stimulus are tested in extinction, that is, with no outcome delivery. Morris et al. (2015) employed such a design in medicated schizophrenia patients and controls. Although devaluation of the reward value of a food snack worked similarly in patients and controls as reflected in Fig. 1. Exemplary study for behavioural subgroups taken from Gold et al. (2012) with permission. A) Reduced Q-learning (Q) and relative dominance of Actor-Critic (AC) learning in patients with high negative symptoms (HNS) versus healthy controls (HC) and patients with low negative symptoms (LNS). B) An illustration of the properties of Q-learning in generating preferences for frequent winners (FW) versus frequent loss avoiders (FLA) while AC-learning accounts for a preference for FLAs over infrequent winners (IW). Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 4 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx Box 1 Reinforcement Learning algorithms. A prediction error is the difference between an experienced reward R and an expected value Q. t, s and a denote indices that refer to time, state and chosen action. ½1 δs;t;a ¼ r t þ Q ðstþ1 ; atþ1 Þ—Q ðst ; at: Þ ½8 Q ðs; aÞ ¼ E½r t þ r tþ1 þrþ2 þ :::jst ¼ s; at ¼ a Essential for model-based learning is that it takes into account both T and R. This can be computationally laborious, in particular for large decision trees but feasible for relatively small problems, e.g., a car route. Despite the drawback of computational load, it has the advantage of rapid integration of new information and therefore enables flexible behavioural adaptation. In model-free learning, this error signal can be used to update action values: ½2 Q ðstþ1 ; atþ1 Þ ¼ Q ðst ; at Þ þ αδs;t;a Here, α represents a learning rate which weighs the influence of the prediction error [1] on the future expectation [2]. This can be repeated iteratively for a chosen action in each state s of each time point t and converges at “true” action value over time. Crucially, model-free learning has no knowledge about the structure of the world as it neglects the action-outcome contingencies. Thus, model-free learning bootstraps from experiences and runs towards the average expected reward. To do so, an experience of states and rewards is required as reflected in Eq. (2). This is comparably simple, thus, has relatively little computational demands. We can define prediction errors and corresponding update equations slightly differently than for example in model-free actor-critic models. The same error signal, generated by the critic, updates values of the critic and the actor: ½3 δCs;t ¼ r t þ cs;t ½4 cs;tþ1 ¼ cs;t þ α c δCs;t Notably, the critic [3,4] neglects the specific action and its expected outcome that was chosen in trial t. The actor learns specific action values via the same error signal: ½5 As;a;tþ1 ¼ As;a;t þ α A δCs;t This approach was applied in one clinical study (Gold et al., 2012). This variant of model-free learning thus neglects the specific expected value of outcomes. After learning, this results in a preference for an action that was rarely associated with a punishment in contrast to a rare winner. This preference would not emerge from model-free learning described in [1,2] because values for a rare winner would then always be preferred over a rare loser. A different way to deal with decision problems is model-based learning. This type of learning (sequentially) evaluates all possible actions and their hypothetical outcomes to work out the best action by considering the probabilistic structure (“model”) of the environment. The structure or model is described by the transition matrix T containing probabilities of the action-outcome contingencies. ½6 T ðs; a; s0 Þ ¼ P ðSt þ 1 ¼ s0 jst ¼ s; at ¼ aÞ As scholarly pointed out by Huys et al. (2014), knowing T refers to knowing the rules of a game. However, this also involves evaluating R, which is done separately. ½7 RðsÞ ¼ E½rt jst ¼ s Thus, R is the average reward as a function of state, which have been evaluated before via model-free learning. In the game analogy, R reflects the goal. Now, we aim to integrate R and T to obtain our action value Q. subjective ratings, the influence of devaluation on choice behaviour was reduced in patients confirming an impairment of goal-directed decision-making (Morris et al., 2015). Comparing valued versus devalued actions on the neural level revealed activation in medial prefrontal cortex and caudate nucleus in healthy controls but this effect was reduced in schizophrenia patients and was correlated with negative symptoms as measured with SANS, in particular avolition and alogia. Using a simpler paradigm not requiring initial learning, another recent study found that the effect of devaluation, via feeding to satiety, on pleasantness ratings was sensory specific in controls but not in patients who showed reduced rating for the sated and unsated food stimuli (Waltz et al., 2015). Interestingly, this unspecific effect of devaluation was correlated with negative but also positive symptoms. 2.1.2. Impairments in reversal learning As pointed out, a deficit in making accurate computations of values to guide goal-directed behaviour may emerge in situations requiring patients to flexibly adapt their behaviour. Early studies using the Wisconsin Card Sorting Task or probabilistic reversal learning found impairments in chronic, medicated schizophrenia patients (Elliott et al., 1995; McKirdy et al., 2009; Pantelis et al., 1999; Prentice et al., 2008; Waltz and Gold, 2007) as well as in medication-naïve, first-episode patients. Furthermore, these impairments are already present at the beginning of the disease (Murray et al., 2008a) and remain stable over time (for at least 6 years) independent of general effects of IQ (Leeson et al., 2009). In serial reversal learning, behavioural adaptation is improved by the individual's ability to gain insight into the anti-correlated task structure, for example, a drop in the value of one option implies an increase for the other option (Reiter et al., 2016a,b,c). Applying computational modelling of choice behaviour in unmedicated schizophrenia patients suggested that reduced reversal learning performance was due to a heightened belief about the occurrence of reversals revealed by a belief-based Hidden– Markov-Model (Fig. 2), which dynamically updates the probability of being in one of the two states of the tasks, thus representing the anticorrelated structure of the reversal task (Schlagenhauf et al., 2014). This increased tendency to switch between choice options was also found in independent medicated samples without applying computational modelling (Culbreth et al., 2015; Waltz et al., 2013). However, although reversal learning appears as an important clinical marker across the course of the illness (Leeson et al., 2009; Murray et al., 2008a), overall reversal learning performance was reported to be associated with higher negative symptom severity (Murray et al., 2008a) but this was not the case consistently across studies (Culbreth et al., 2015; Schlagenhauf et al., 2014; Waltz et al., 2013). Nevertheless, impaired reversal learning indicates a deficit in flexible value-based decision making and the mechanisms underlying reversal learning performance are likely to also contribute to the different domains of negative. 2.1.3. Using the task structure for forward planning of decision-making As mentioned above, one important feature of the model-based system is to use insight into the task structure to enable forward planning of choice behaviour. This can be tested via paradigms employing sequential decision-making challenging the individual with two (or Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx 5 B A C D Model parameters Symptoms 50 Symptoms 40 30 20 10 0 PS NS GP Fig. 2. Exemplary study for behavioural subgroups by Schlagenhauf et al. (2014) with permission. A) Shows the number of individuals not fitted better than chance by the best-fitting Hidden-Markov-Model (HMM). B) Reversal learning performance in controls (Ctrls, fit ok), patients with good fit (Sz, fit ok) and patients with poor fit (Poor Fit). C) Differences between groups in reward sensitivity were driven by poor fit patients alone, while enhanced switching behaviour was also observed in patients with good fit. D) Differences in symptom ratings between the two patient subgroups. They differed significantly in terms of positive symptoms (PS) but not negative symptoms (NS) nor general psychopathology (GP) as derived from the Positive and Negative Symptoms Scale (PANSS). more) consecutive decisions. The first decision can either be planned in a forward manner as predicted by a model-based account, or it can be driven retrospectively by rewards delivered after the second choice which would be characteristic of a model-free system (Daw et al., 2011). As replicated several times up to now, individuals show influences of both model-based and model-free behavioural control. A series of studies have demonstrated that disruption of the dorso-lateral prefrontal cortex reduces model-based control (Smittenaar et al., 2013) and that stress-induced inter-individual differences in cortisol increase, presumably affecting prefrontal functioning, correlate with the degree of model-based control (Otto et al., 2013b; Radenbach et al., 2015). Thus, in schizophrenia patients one would expect reduced modelbased control in this task based on evidence for prefrontal dysfunction and associated cognitive deficits. Indeed, schizophrenia patients demonstrate decreased model-based decision-making compared to controls during a sequential decision making task (Culbreth et al., 2016). This deficit in model-based learning was related to higher-order cognitive deficits such as impaired working memory in line with the tight relationship between model-based control in this task and cognitive function (Otto et al., 2013a, 2015), in particular working memory and cognitive processing speed (Schad et al., 2014). Evidence exists for a link between the dopaminergic system and model-based control. In healthy controls, pharmacological elevation of the brain's overall presynaptic dopamine levels enhanced model-based behaviour in this task (Wunderlich et al., 2012). Measuring dopamine in vivo in humans in combination with this task during fMRI (Deserno et al., 2015a), inter-individual differences in ventral striatal presynaptic dopamine positively correlated with the degree of model-based control. Further, inter-individual differences in ventral striatal presynaptic dopamine were also positively related to model-based signatures in the lateral prefrontal cortex, speaking for an effect of striatal dopamine on prefrontal model-based control. In the same study (Deserno et al., 2015a), ventral striatal presynaptic dopamine was negatively correlated with ventral striatal model-free reward prediction errors, replicating a previous finding during reversal learning (Schlagenhauf et al., 2013). This correlation is of interest in the context of schizophrenia because it was found twice with the neurochemical measure of presynaptic striatal dopamine as measured with FDOPA PET, which is well-known to be elevated in schizophrenia (Howes et al., 2009). This suggests that coding of model-free reward prediction errors in ventral striatum should be compromised in patients. Using reversal learning, this was indeed found in unmedicated patients (Schlagenhauf et al., 2014) and studies in medicated patients with different instrumental tasks found reduced reward prediction errors in other striatal subareas (Gradin et al., 2011; Koch et al., 2010; Murray et al., 2008b). In fact, as is known from fMRI studies applying the monetary incentive delay task, medication status could be crucial (Nielsen et al., 2012a; Schlagenhauf et al., 2008). Thus, whether a habitual or model-free reinforcement learning system itself is compromised in patients suffering from schizophrenia still remains to be answered. Two aspects appear crucial: First, medication status is a severe confound of model-free dopamine signals (a putative dopaminergic striatal dysfunction may depend on medication status; however, choice switching, as also seen in medicated patients, could result from a different putatively non-dopaminergic mechanism). Second, a series of recent studies in healthy individuals found that learning signatures in ventral striatum supposed to mainly reflect model-free learning signals can also carry model-based information (Daw et al., 2011), a finding that has now been replicated twice (Deserno et al., 2015a, 2015b) and has also been found with different tasks in other studies (Li and Daw, 2011; Simon and Daw, 2011; Wimmer et al., 2012). Thus, looking at process-pure reinforcement learning can be challenging given the apparent “ubiquity” of model-based evaluation (Doll et al., 2012). In line with this idea, even deterministic reinforcement learning tasks can, for example, require working memory but clever experimental designs in conjunction with computational modelling can help to untie the underlying mechanisms (Collins and Frank, 2012). When applying such a design in a sample of chronic medicated patients, the overall deficit in performing Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 6 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx deterministic reinforcement learning was entirely accounted for by parameters of working memory (Collins et al., 2014). 2.2. Effort-based decision making One promising avenue to further elucidate how mechanisms of goaldirected behaviour contribute to negative symptoms is to study effortbased decision-making driven by the hypothesis that patients overestimate the cost of effort, a putative mechanism underlying apathy and avolition. We keep this part relatively short—not reflecting the importance of the issue but acknowledging that there are excellent published discussions on this issue (for instance, Gold et al., 2015). Indeed there is evidence of a) group differences between patients with schizophrenia and controls in effort-based decision-making, in line with the idea that patients fail to make high-effort choices to maximize reward, and b) correlation between this impairment and the degree of negative symptoms (Barch et al., 2014; Fervaha et al., 2013b; Gold et al., 2013; Hartmann et al., 2015; Treadway et al., 2015; Wolf et al., 2014). A particularly strong correlation was found between apathy derived from the Brief Negative Symptom Scale and effort-discounting using a novel handgrip exertion task (Hartmann et al., 2015) although there have been negative findings, too (Docx et al., 2015). With regard to effortbased decision making and its relation to negative symptoms, Gold et al. (2015) suggest two possible interpretations to be disentangled experimentally: the (overestimation of) cost of effort and the (underestimation of) potential value of reward. Correspondingly, there is interesting work on how values invigorate behaviour in healthy individuals (Beierholm et al., 2013; Guitart-Masip et al., 2011). 3. Dissecting cognitive deficits and negative symptoms via generative models of connectivity: a proof-of-concept example So far, we have mostly discussed behavioural experiments with a focus on modelling approaches of such data and how this can be useful to understand negative symptoms. Turning to neurobiological underpinnings of symptoms, it remains an intriguing question why our understanding of neurobiological systems has not yet yielded a mechanistic understanding of mental illness (Kapur et al., 2012). Indeed, a certain position along the dimension of a behavioural mechanisms, even if expressed by a densely compressed computational parameter ideally being well-predictive for a certain clinical dimension, could emerge from a variety of different neuronal implementations. When arriving at this level, important biological information of clinical relevance could be revealed, as this could be the diversity underlying heterogeneity in clinical outcome and treatment response. One potentially fruitful approach to achieve this is to apply analytical methods to neural data that enable mechanistic insight. Functional Magnetic Resonance Imaging (fMRI) is one illustrative example of the limitation of mechanistic insight provided by most non-invasive neural measurement techniques in humans. FMRI does not observe neuronal activity directly but measures changes in blood oxygenation most likely reflecting changes in largescale neuronal populations due to neurovascular coupling. Besides the hope for future advent of technical innovation in non-invasive measurement techniques in humans, one promising avenue is to first formalize a potential mechanism (or even competing mechanisms) in a set of equations. Importantly, such hypotheses should reflect latent neuronal or psychological states, in other words mechanisms, which may have caused measured data. To see how these hidden neuronal states affect a neural measurement, a so-called observation or forward model is required to transform unknown neuronal states into measurable responses, for example, fMRI signal. This general two-step procedure has been in use for a long time in computational neuroscience and has reached psychiatric research in the larger framework of computational psychiatry (Montague et al., 2012; Stephan and Mathys, 2014; Wang and Krystal, 2014). In principle, some of these models can be inverted or fitted when measured data are available, which is of great importance because model inversion allows estimation of mechanistically informative parameters in single individuals. Broadly speaking, most complex models may have the appeal of rich biological realism but they cannot be inverted and, thus, are restrained to insightful simulations. In this third section, we focus on one generative model of neuroimaging data, namely dynamic causal modelling (DCM, Friston et al., 2003). The traditional version of deterministic DCM rests upon a set of differential bilinear equations to model interactions of large-scale neuronal ensembles. This comprises three types of mechanisms: 1) effects of inputs (e.g., sensory stimulation) eliciting neuronal activity in a region; 2) intrinsic connectivity transferring this formerly elicited activity between regions; and 3) context-dependent changes in connectivity due to experimentally induced systematic perturbations. DCM is generative because its implementation specifies a prior distribution and a likelihood function. First, this enables estimation of individual parameters via a probabilistic forward model as mentioned above, which represents the probability of observing a measurement based on the parameters of the neuronal states. Second, and of great importance, this enables approximation of the so-called model evidence by integrating out the dependency of the likelihood on the parameters and thus opens the possibility of performing Bayesian Model Selection. As DCM is a sophisticated tool, it is important to emphasize that it is available as open-source code, part of the widely used software package SPM. As an example of using such a generative modelling analysis, we will briefly summarize findings from a study by Deserno et al. (2012) where deterministic DCM was applied to fMRI data acquired in 42 controls and 41 medicated patients during a working memory task Based upon the dysconnectivity hypothesis of schizophrenia postulating a deficiency of NMDA-dependent cortico-cortical processing, a three-region DCM modelling working memory (WM)-dependent changes in prefrontalparietal connectivity, including the visual cortex as an input region for visual stimulation, was estimated. As hypothesized, this revealed a reduction in WM-dependent prefrontal to parietal effective connectivity in patients compared to controls. No correlation with positive or negative symptoms was observed with the modelling parameters. As introduced above, parameters of this DCM are mechanistically informative in the sense that they reflect inter-individual differences in how directed interactions between visual, parietal and dorsolateral prefrontal cortex change with respect to the experimental context working memory. In line with the idea that such mechanistically informative parameters could be helpful to dissect a spectrum disorder such as schizophrenia into different biological subgroups, a follow-up analysis of these data was performed. To do so, Brodersen et al. (2014) performed a proof-of-concept analysis based on the DCMs from Deserno et al. (2012). Before the final step of searching for connectivity-based subgroups, the mechanistically informative DCM parameters should meet two criteria: First, they should improve classification accuracy regarding controls versus patients in a supervised setting (support vector machine) and second, in an unsupervised setting (Gaussian Mixture Models) based on data of controls and patients, the two known groups should be recovered. Indeed, these two criteria were met. Importantly, when repeating the supervised analysis with measures of regional activity or functional connectivity within the 3-region network, classification accuracy dropped markedly (indistinguishable from chance based on regional activity) and the evidence for the known two subgroups was low in the unsupervised setting. This elegantly suggested the superiority of mechanistically informative parameters and motivates the search for biologically defined subgroups within patients based on these parameters. Thus, the unsupervised analysis was applied to the data of patients alone and interestingly revealed convincing evidence for the presence of three distinct subgroups (Fig. 3a). However, such subgroups may appear less useful if they had no clinical relevance. Thus, these subgroups were compared on their clinical ratings (using PANSS) and a difference between the three groups was observed in terms of their negative symptom severity only (Fig. 3c). Impressively, the three subgroups could not be recovered from estimates of regional activity nor functional connectivity. Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx 7 Fig. 3. Exemplary study for connectivity subgroups as taken from Brodersen et al. (2014) with permission. A) Model evidence for different numbers of clusters using unsupervised learning within patients alone. There is a clear preference for three subgroups that could not be obtained when using other measures such as functional connectivity or regional activity within the three-region network. B) The width of the arrows illustrates the connectivity profiles of the three subgroups. C) The three subgroups differed in terms of their negative symptom ratings as obtained with the Positive and Negative Symptom Scale (PANSS-NS). WM = working memory. Both of the latter findings revealed weak evidence for two subgroups, which did not, however, differ in symptom severity. Two main points can be concluded from these exemplary results: First, mechanistically informative parameters, from DCM, but potentially also from any other generative model, improve classification accuracy and have the potential to reveal distinct subgroups with clinical validity (also see Wiecki et al., 2015). Here it appears crucial that such clustering analyses depend on mechanistically informed parameters as provided by generative models. This approach was refined recently by including the identification of subgroups into the process of hierarchical model inversion based on an empirical Bayesian approach to DCM (Raman et al., 2016). Second, although the study by Brodersen et al. (2014) was a methodological proof-of-concept project, the identification of biologically distinct connectivity subgroups mapping on different levels of negative symptom severity has clinical plausibility in particular because both negative symptoms and cognitive deficits predict the clinical outcome of patients. These findings suggest that the application of such unsupervised classification techniques on parameters of generative models of behavioural and neural data can help improving a mechanistic understanding of the negative symptom domain. 4. Conclusion & outlook In the previous section we described how a generative model of neural data during working memory, in this case deterministic DCM for fMRI, can reveal mechanistically informed subgroups with clinical validity. Because negative symptoms also share variance with cognitive deficits, the application of such cognitive tasks during fMRI in combination with DCM appears a highly promising avenue to systematically separate the overlap of cognitive deficits and negative symptoms from a biological perspective. This could be done in large-scale multicentre studies, ideally of longitudinal nature, also with well-known tasks such as the n-back or MID tasks. This seems highly feasible because MRI is so widely available these days and such tasks have been used almost ubiquitously in cognitive and affective neuroscience studies in psychiatry. Data sharing based on already finished studies also seems desirable. This suggests that distinct neurobiological mechanisms, for example, reduced effective connectivity of dorsolateral PFC, as inferred by DCM, might be more informative than behavioural measures alone. In line with this and based on the computational approach pursued, one can further enrich the mechanistic information obtained from these models. One promising route seems to be the combination of DCM with learning trajectories from behavioural models of decision-making and learning. The latter models themselves provide a rich source of mechanistic information, and, to varying degrees, have biological plausibility regarding neuromodulatory systems, in particular dopamine (Deserno et al., 2015a; Montague et al., 1996; Schultz et al., 1997) but also serotonin (den Ouden et al., 2013) and acetylcholine (Iglesias et al., 2013). The mechanisms represented by such models could play a major role in the manifestation of negative symptoms. As reviewed in the second section, cognitive tasks studying reinforcement learning and decision-making have been shown to be associated with negative symptom severity with at least some consistency. This can be improved by mutually refining learning and decision-making tasks and computational models, as the analytic tool, based on the possibility of simulating data a-priori when designing a study and of inverting models a-posteriori after having collected data. Four lines of experiments should be followed systematically: the distinction of learning from rewards and punishments, effort-based decision-making, variants of reversal learning and sequential decision-making. Based on the currently available evidence, all four lines seem promising but the first two domains seemed to be more specific for negative symptoms or even particular subdomains. On the other hand, the investigation of reversal learning performance, or more broadly flexible behavioural adaptation, holds the opportunity to investigate updating beliefs about the stability or volatility of the environment by using a hierarchical learning approach, which might play a role in negative symptoms despite inconsistent findings regarding correlations with overall performance readouts. When investigating such processes, a possible overlap with general cognition such as working memory should be carefully taken into account and paradigms should be refined accordingly. This underlines that the application and further development of tasks suitable for patients are of crucial importance; otherwise group differences will be less informative. As pointed out, the approach illustrated here develops its potential with the mechanistic information carried by a model's parameters and this could be taken to a next level based on joint generative models of neural and behavioural data from dynamic tasks of learning and decision-making. Such a computational framework attempts to identify neurophysiological mechanisms contributing to different dimensions of negative symptoms but only if this takes place in conjunction with clinical guidance, in particular a further development and improvement of the clinical assessment taking into account the different domains of negative symptoms. Nonetheless, large samples of patients will have to be investigated, inevitably, in longitudinal design. Only longitudinal studies can give meaningful insight into the clinical usefulness of possibly identified subgroups based on mechanistically informed parameters. This includes relation to clinical symptomatology, stability over time as well as, most critically, their predictive value for clinical outcome and treatment response. Although the development of new treatment strategy is a goal, answers based on such studies regarding which of the available treatment strategies can be successful in which patients would already constitute a major success. Role of funding source This study was supported by the Max Planck Society and grants from the German Research Foundation awarded to FS (DFG SCHL1969/1-1, DFG SCHL 1969/2-1). Contributors All authors contributed to and have approved the final manuscript. Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 8 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx Financial disclosure All authors report no biomedical financial interests or potential conflicts of interest. Acknowledgements We would like to thank the European negative symptom research network (‘Eurones’, http://eurones.eu/) for fruitful discussions. In this regard, we are grateful to Ann Faerden, Stefan Kaiser, Andre Aleman and Silvana Galderisi for starting the initiative. We would like to thank Silvana Galderisi for her helpful comments on an earlier version of the manuscript. References Balleine, B.W., O'Doherty, J.P., 2010. Human and rodent homologies in action control: corticostriatal determinants of goal-directed and habitual action. Neuropsychopharmacology 35 (1), 48–69. Barch, D.M., Dowd, E.C., 2010. Goal representations and motivational drive in schizophrenia: the role of prefrontal–striatal interactions. Schizophr. Bull. 36 (5), 919–934. Barch, D.M., Treadway, M.T., Schoen, N., 2014. Effort, anhedonia, and function in schizophrenia: reduced effort allocation predicts amotivation and functional impairment. J. Abnorm. Psychol. 123 (2), 387–397. Beierholm, U., Guitart-Masip, M., Economides, M., Chowdhury, R., Duzel, E., Dolan, R., Dayan, P., 2013. Dopamine modulates reward-related vigor. Neuropsychopharmacology 38 (8), 1495–1503. Bobes, J., Arango, C., Garcia-Garcia, M., Rejas, J., Group, C.S.C., 2010. Prevalence of negative symptoms in outpatients with schizophrenia spectrum disorders treated with antipsychotics in routine clinical practice: findings from the CLAMORS study. J. Clin. Psychiatry 71 (3), 280–286. Bowie, C.R., Leung, W.W., Reichenberg, A., McClure, M.M., Patterson, T.L., Heaton, R.K., Harvey, P.D., 2008. Predicting schizophrenia patients' real-world behavior with specific neuropsychological and functional capacity measures. Biol. Psychiatry 63 (5), 505–511. Brodersen, K.H., Deserno, L., Schlagenhauf, F., Lin, Z., Penny, W.D., Buhmann, J.M., Stephan, K.E., 2014. Dissecting psychiatric spectrum disorders by generative embedding. NeuroImage Clinical 4, 98–111. Cohen, A.S., Minor, K.S., 2010. Emotional experience in patients with schizophrenia revisited: meta-analysis of laboratory studies. Schizophr. Bull. 36 (1), 143–150. Collins, A.G., Frank, M.J., 2012. How much of reinforcement learning is working memory, not reinforcement learning? A behavioral, computational, and neurogenetic analysis. Eur. J. Neurosci. 35 (7), 1024–1035. Collins, A.G., Brown, J.K., Gold, J.M., Waltz, J.A., Frank, M.J., 2014. Working memory contributions to reinforcement learning impairments in schizophrenia. J. Neurosci. 34 (41), 13747–13756. Culbreth, A.J., Gold, J.M., Cools, R., Barch, D.M., 2015. Impaired activation in cognitive control regions predicts reversal learning in schizophrenia. Schizophr. Bull. Culbreth, A.J., Westbrook, A., Daw, N.D., Botvinick, M., Barch, D.M., 2016. Reduced modelbased decision-making in schizophrenia. J. Abnorm. Psychol. Daw, N.D., Niv, Y., Dayan, P., 2005. Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nat. Neurosci. 8 (12), 1704–1711. Daw, N.D., Gershman, S.J., Seymour, B., Dayan, P., Dolan, R.J., 2011. Model-based influences on humans' choices and striatal prediction errors. Neuron 69 (6), 1204–1215. den Ouden, H.E., Daw, N.D., Fernandez, G., Elshout, J.A., Rijpkema, M., Hoogman, M., Franke, B., Cools, R., 2013. Dissociable effects of dopamine and serotonin on reversal learning. Neuron 80 (4), 1090–1100. Deserno, L., Sterzer, P., Wustenberg, T., Heinz, A., Schlagenhauf, F., 2012. Reduced prefrontal–parietal effective connectivity and working memory deficits in schizophrenia. J. Neurosci. 32 (1), 12–20. Deserno, L., Huys, Q.J., Boehme, R., Buchert, R., Heinze, H.J., Grace, A.A., Dolan, R.J., Heinz, A., Schlagenhauf, F., 2015a. Ventral striatal dopamine reflects behavioral and neural signatures of model-based control during sequential decision making. Proc. Natl. Acad. Sci. U. S. A. 112 (5), 1595–1600. Deserno, L., Wilbertz, T., Reiter, A.M.F., Horstmann, A., Neumann, J., Villringer, A., Heinze, H.J., Schlagenhauf, F., 2015b. Lateral prefrontal model-based signals are reduced in healthy individuals with high trait impulsivity. Transl. Psychiatry. Dickinson, A.D., 1985. Action and habits: the development of behavioural autonomy. Philos. Trans. R. Soc. Lond. Ser. B Biol. Sci. 308 (1135), 67–78. Docx, L., de la Asuncion, J., Sabbe, B., Hoste, L., Baeten, R., Warnaerts, N., Morrens, M., 2015. Effort discounting and its association with negative symptoms in schizophrenia. Cogn. Neuropsychiatry 20 (2), 172–185. Dolan, R.J., Dayan, P., 2013. Goals and habits in the brain. Neuron 80 (2), 312–325. Doll, B.B., Simon, D.A., Daw, N.D., 2012. The ubiquity of model-based reinforcement learning. Curr. Opin. Neurobiol. 22 (6), 1075–1081. Dominguez, M.d.G., Viechtbauer, W., Simons, C.J.P., van Os, J., Krabbendam, L., 2009. Are psychotic psychopathology and neurocognition orthogonal? A systematic review of their associations. Psychol. Bull. 135 (1), 157–171. Dowd, E.C., Barch, D.M., 2010. Anhedonia and emotional experience in schizophrenia: neural and behavioral indicators. Biol. Psychiatry 67 (10), 902–911. Eack, S.M., Newhill, C.E., 2007. Psychiatric symptoms and quality of life in schizophrenia: a meta-analysis. Schizophr. Bull. 33 (5), 1225–1237. Elliott, R., McKenna, P.J., Robbins, T.W., Sahakian, B.J., 1995. Neuropsychological evidence for frontostriatal dysfunction in schizophrenia. Psychol. Med. 25 (3), 619–630. Esslinger, C., Englisch, S., Inta, D., Rausch, F., Schirmbeck, F., Mier, D., Kirsch, P., MeyerLindenberg, A., Zink, M., 2012. Ventral striatal activation during attribution of stimulus saliency and reward anticipation is correlated in unmedicated first episode schizophrenia patients. Schizophr. Res. 140 (1–3), 114–121. Fervaha, G., Foussias, G., Agid, O., Remington, G., 2013a. Amotivation and functional outcomes in early schizophrenia. Psychiatry Res. 210 (2), 665–668. Fervaha, G., Graff-Guerrero, A., Zakzanis, K.K., Foussias, G., Agid, O., Remington, G., 2013b. Incentive motivation deficits in schizophrenia reflect effort computation impairments during cost-benefit decision-making. J. Psychiatr. Res. 47 (11), 1590–1596. Fervaha, G., Foussias, G., Agid, O., Remington, G., 2014. Impact of primary negative symptoms on functional outcomes in schizophrenia. Eur. Psychiatry 29 (7), 449–455. Frank, M.J., Seeberger, L.C., O'Reilly, R., C., 2004. By carrot or by stick: cognitive reinforcement learning in parkinsonism. Science 306 (5703), 1940–1943. Friston, K.J., Harrison, L., Penny, W., 2003. Dynamic causal modelling. NeuroImage 19 (4), 1273–1302. Fusar-Poli, P., Bonoldi, I., Yung, A.R., Borgwardt, S., Kempton, M.J., Valmaggia, L., Barale, F., Caverzasi, E., McGuire, P., 2012. Predicting psychosis: meta-analysis of transition outcomes in individuals at high clinical risk. Arch. Gen. Psychiatry 69 (3), 220–229. Fusar-Poli, P., Papanastasiou, E., Stahl, D., Rocchetti, M., Carpenter, W., Shergill, S., McGuire, P., 2015. Treatments of negative symptoms in schizophrenia: meta-analysis of 168 randomized placebo-controlled trials. Schizophr. Bull. 41 (4), 892–899. Galderisi, S., Bucci, P., Mucci, A., Kirkpatrick, B., Pini, S., Rossi, A., Vita, A., Maj, M., 2013. Categorical and dimensional approaches to negative symptoms of schizophrenia: focus on long-term stability and functional outcome. Schizophr. Res. 147 (1), 157–162. Galderisi, S., Rossi, A., Rocca, P., Bertolino, A., Mucci, A., Bucci, P., Rucci, P., Gibertoni, D., Aguglia, E., Amore, M., Bellomo, A., Biondi, M., Brugnoli, R., Dell'Osso, L., De Ronchi, D., Di Emidio, G., Di Giannantonio, M., Fagiolini, A., Marchesi, C., Monteleone, P., Oldani, L., Pinna, F., Roncone, R., Sacchetti, E., Santonastaso, P., Siracusano, A., Vita, A., Zeppegno, P., Maj, M., Italian Network For Research on, P., 2014. The influence of illness-related variables, personal resources and context-related factors on real-life functioning of people with schizophrenia. World Psychiatry 13 (3), 275–287. Galderisi, S., Merlotti, E., Mucci, A., 2015. Neurobiological background of negative symptoms. Eur. Arch. Psychiatry Clin. Neurosci. 265 (7), 543–558. Gold, J.M., Waltz, J.A., Prentice, K.J., Morris, S.E., Heerey, E.A., 2008. Reward processing in schizophrenia: a deficit in the representation of value. Schizophr. Bull. 34 (5), 835–847. Gold, J.M., Waltz, J.A., Matveeva, T.M., Kasanova, Z., Strauss, G.P., Herbener, E.S., Collins, A.G., Frank, M.J., 2012. Negative symptoms and the failure to represent the expected reward value of actions: behavioral and computational modeling evidence. Arch. Gen. Psychiatry 69 (2), 129–138. Gold, J.M., Strauss, G.P., Waltz, J.A., Robinson, B.M., Brown, J.K., Frank, M.J., 2013. Negative symptoms of schizophrenia are associated with abnormal effort-cost computations. Biol. Psychiatry 74 (2), 130–136. Gold, J.M., Waltz, J.A., Frank, M.J., 2015. Effort cost computation in schizophrenia: a commentary on the recent literature. Biol. Psychiatry. Gonzalez-Ortega, I., de Los Mozos, V., Echeburua, E., Mezo, M., Besga, A., Ruiz de Azua, S., Gonzalez-Pinto, A., Gutierrez, M., Zorrilla, I., Gonzalez-Pinto, A., 2013. Working memory as a predictor of negative symptoms and functional outcome in first episode psychosis. Psychiatry Res. 206 (1), 8–16. Gradin, V.B., Kumar, P., Waiter, G., Ahearn, T., Stickle, C., Milders, M., Reid, I., Hall, J., Steele, J.D., 2011. Expected value and prediction error abnormalities in depression and schizophrenia. Brain J. Neurol. 134 (Pt 6), 1751–1764. Guitart-Masip, M., Beierholm, U.R., Dolan, R., Duzel, E., Dayan, P., 2011. Vigor in the face of fluctuating rates of reward: an experimental examination. J. Cogn. Neurosci. 23 (12), 3933–3938. Hartmann, M.N., Hager, O.M., Reimann, A.V., Chumbley, J.R., Kirschner, M., Seifritz, E., Tobler, P.N., Kaiser, S., 2015. Apathy but not diminished expression in schizophrenia is associated with discounting of monetary rewards by physical effort. Schizophr. Bull. 41 (2), 503–512. Hartmann-Riemer, M.N., Hager, O.M., Kirschner, M., Bischof, M., Kluge, A., Seifritz, E., Kaiser, S., 2015. The association of neurocognitive impairment with diminished expression and apathy in schizophrenia. Schizophr. Res. 169 (1–3), 427–432. Heerey, E.A., Gold, J.M., 2007. Patients with schizophrenia demonstrate dissociation between affective experience and motivated behavior. J. Abnorm. Psychol. 116 (2), 268–278. Heinz, A., 2002. Dopaminergic dysfunction in alcoholism and schizophrenia–psychopathological and behavioral correlates. Eur. Psychiatry 17 (1), 9–16. Heinz, A., Schlagenhauf, F., 2010. Dopaminergic dysfunction in schizophrenia: salience attribution revisited. Schizophr. Bull. 36 (3), 472–485. Heinz, A., Knable, M.B., Coppola, R., Gorey, J.G., Jones, D.W., Lee, K.S., Weinberger, D.R., 1998. Psychomotor slowing, negative symptoms and dopamine receptor availability — an IBZM SPECT study in neuroleptic-treated and drug-free schizophrenic patients. Schizophr. Res. 31 (1), 19–26. Howes, O.D., Murray, R.M., 2014. Schizophrenia: an integrated sociodevelopmental–cognitive model. Lancet 383 (9929), 1677–1687. Howes, O.D., Montgomery, A.J., Asselin, M.C., Murray, R.M., Valli, I., Tabraham, P., BramonBosch, E., Valmaggia, L., Johns, L., Broome, M., McGuire, P.K., Grasby, P.M., 2009. Elevated striatal dopamine function linked to prodromal signs of schizophrenia. Arch. Gen. Psychiatry 66 (1), 13–20. Huys, Q.J., Tobler, P.N., Hasler, G., Flagel, S.B., 2014. The role of learning-related dopamine signals in addiction vulnerability. Prog. Brain Res. 211, 31–77. Huys, Q.J., Maia, T.V., Frank, M.J., 2016. Computational psychiatry as a bridge from neuroscience to clinical applications. Nat. Neurosci. 19 (3), 404–413. Hyman, S.E., Fenton, W.S., 2003. Medicine. What are the right targets for psychopharmacology? Science 299 (5605), 350–351. Iglesias, S., Mathys, C., Brodersen, K.H., Kasper, L., Piccirelli, M., den Ouden, H.E., Stephan, K.E., 2013. Hierarchical prediction errors in midbrain and basal forebrain during sensory learning. Neuron 80 (2), 519–530. Juckel, G., Schlagenhauf, F., Koslowski, M., Filonov, D., Wustenberg, T., Villringer, A., Knutson, B., Kienast, T., Gallinat, J., Wrase, J., Heinz, A., 2006a. Dysfunction of ventral striatal reward prediction in schizophrenic patients treated with typical, not atypical, neuroleptics. Psychopharmacology 187 (2), 222–228. Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004 L. Deserno et al. / Schizophrenia Research xxx (2016) xxx–xxx Juckel, G., Schlagenhauf, F., Koslowski, M., Wustenberg, T., Villringer, A., Knutson, B., Wrase, J., Heinz, A., 2006b. Dysfunction of ventral striatal reward prediction in schizophrenia. NeuroImage 29 (2), 409–416. Kapur, S., 2003. Psychosis as a state of aberrant salience: a framework linking biology, phenomenology, and pharmacology in schizophrenia. Am. J. Psychiatry 160 (1), 13–23. Kapur, S., Phillips, A.G., Insel, T.R., 2012. Why has it taken so long for biological psychiatry to develop clinical tests and what to do about it? Mol. Psychiatry 17 (12), 1174–1179. Kegeles, L.S., Abi-Dargham, A., Frankle, W.G., Gil, R., Cooper, T.B., Slifstein, M., Hwang, D.R., Huang, Y., Haber, S.N., Laruelle, M., 2010. Increased synaptic dopamine function in associative regions of the striatum in schizophrenia. Arch. Gen. Psychiatry 67 (3), 231–239. Kirkpatrick, B., Fenton, W.S., Carpenter Jr., W.T., Marder, S.R., 2006. The NIMH-MATRICS consensus statement on negative symptoms. Schizophr. Bull. 32 (2), 214–219. Koch, K., Schachtzabel, C., Wagner, G., Schikora, J., Schultz, C., Reichenbach, J.R., Sauer, H., Schlosser, R.G., 2010. Altered activation in association with reward-related trial-anderror learning in patients with schizophrenia. NeuroImage 50 (1), 223–232. Leeson, V.C., Robbins, T.W., Matheson, E., Hutton, S.B., Ron, M.A., Barnes, T.R., Joyce, E.M., 2009. Discrimination learning, reversal, and set-shifting in first-episode schizophrenia: stability over six years and specific associations with medication type and disorganization syndrome. Biol. Psychiatry 66 (6), 586–593. Li, J., Daw, N.D., 2011. Signals in human striatum are appropriate for policy update rather than value prediction. J. Neurosci. 31 (14), 5504–5511. Lieberman, J.A., Stroup, T.S., McEvoy, J.P., Swartz, M.S., Rosenheck, R.A., Perkins, D.O., Keefe, R.S., Davis, S.M., Davis, C.E., Lebowitz, B.D., Severe, J., Hsiao, J.K., 2005. Effectiveness of antipsychotic drugs in patients with chronic schizophrenia. N. Engl. J. Med. 353 (12), 1209–1223. McKirdy, J., Sussmann, J.E., Hall, J., Lawrie, S.M., Johnstone, E.C., McIntosh, A.M., 2009. Set shifting and reversal learning in patients with bipolar disorder or schizophrenia. Psychol. Med. 39 (8), 1289–1293. Montague, P.R., Dayan, P., Sejnowski, T.J., 1996. A framework for mesencephalic dopamine systems based on predictive Hebbian learning. J. Neurosci. 16 (5), 1936–1947. Montague, P.R., Dolan, R.J., Friston, K.J., Dayan, P., 2012. Computational psychiatry. Trends Cogn. Sci. 16 (1), 72–80. Morris, R.W., Quail, S., Griffiths, K.R., Green, M.J., Balleine, B.W., 2015. Corticostriatal control of goal-directed action is impaired in schizophrenia. Biol. Psychiatry 77 (2), 187–195. Mucci, A., Dima, D., Soricelli, A., Volpe, U., Bucci, P., Frangou, S., Prinster, A., Salvatore, M., Galderisi, S., Maj, M., 2015. Is avolition in schizophrenia associated with a deficit of dorsal caudate activity? A functional magnetic resonance imaging study during reward anticipation and feedback. Psychol. Med. 1–14. Murray, G.K., Cheng, F., Clark, L., Barnett, J.H., Blackwell, A.D., Fletcher, P.C., Robbins, T.W., Bullmore, E.T., Jones, P.B., 2008a. Reinforcement and reversal learning in first-episode psychosis. Schizophr. Bull. 34 (5), 848–855. Murray, G.K., Corlett, P.R., Clark, L., Pessiglione, M., Blackwell, A.D., Honey, G., Jones, P.B., Bullmore, E.T., Robbins, T.W., Fletcher, P.C., 2008b. Substantia nigra/ventral tegmental reward prediction error disruption in psychosis. Mol. Psychiatry 13(3) (239), 267–276. Nielsen, M.O., Rostrup, E., Wulff, S., Bak, N., Broberg, B.V., Lublin, H., Kapur, S., Glenthoj, B., 2012a. Improvement of brain reward abnormalities by antipsychotic monotherapy in schizophrenia. Arch. Gen. Psychiatry 69 (12), 1195–1204. Nielsen, M.O., Rostrup, E., Wulff, S., Bak, N., Lublin, H., Kapur, S., Glenthoj, B., 2012b. Alterations of the brain reward system in antipsychotic naive schizophrenia patients. Biol. Psychiatry 71 (10), 898–905. Nuechterlein, K.H., Subotnik, K.L., Green, M.F., Ventura, J., Asarnow, R.F., Gitlin, M.J., Yee, C.M., Gretchen-Doorly, D., Mintz, J., 2011. Neurocognitive predictors of work outcome in recent-onset schizophrenia. Schizophr. Bull. 37 (Suppl. 2), S33–S40. Otto, A.R., Gershman, S.J., Markman, A.B., Daw, N.D., 2013a. The curse of planning: dissecting multiple reinforcement-learning systems by taxing the central executive. Psychol. Sci. 24 (5), 751–761. Otto, A.R., Raio, C.M., Chiang, A., Phelps, E.A., Daw, N.D., 2013b. Working-memory capacity protects model-based learning from stress. Proc. Natl. Acad. Sci. U. S. A. 110 (52), 20941–20946. Otto, A.R., Skatova, A., Madlon-Kay, S., Daw, N.D., 2015. Cognitive control predicts use of model-based reinforcement learning. J. Cogn. Neurosci. 27 (2), 319–333. Pankow, A., Friedel, E., Sterzer, P., Seiferth, N., Walter, H., Heinz, A., Schlagenhauf, F., 2013. Altered amygdala activation in schizophrenia patients during emotion processing. Schizophr. Res. 150 (1), 101–106. Pantelis, C., Barber, F.Z., Barnes, T.R., Nelson, H.E., Owen, A.M., Robbins, T.W., 1999. Comparison of set-shifting ability in patients with chronic schizophrenia and frontal lobe damage. Schizophr. Res. 37 (3), 251–270. Patel, R., Jayatilleke, N., Broadbent, M., Chang, C.K., Foskett, N., Gorrell, G., Hayes, R.D., Jackson, R., Johnston, C., Shetty, H., Roberts, A., McGuire, P., Stewart, R., 2015. Negative symptoms in schizophrenia: a study in a large clinical sample of patients using a novel automated method. BMJ Open 5 (9), e007619. Prentice, K.J., Gold, J.M., Buchanan, R.W., 2008. The Wisconsin Card Sorting impairment in schizophrenia is evident in the first four trials. Schizophr. Res. 106 (1), 81–87. Radenbach, C., Reiter, A.M.F., Engert, V., Sjoerds, Z., Villringer, A., Heinze, H.J., Deserno, L., Schlagenhauf, F., 2015. The interaction of acute and chronic stress impairs modelbased behavioral control. Psychoneuroendocrinology. Radua, J., Schmidt, A., Borgwardt, S., et al., 2015. Ventral striatal activation during reward processing in psychosis: a neurofunctional meta-analysis. JAMA Psychiat. 1243–1251. Raman, S., Deserno, L., Schlagenhauf, F., Stephan, K.E., 2016. A hierarchical model for integrating unsupervised generative embedding and empirical Bayes. J. Neurosci. Methods. Reiter, A.M., Koch, S.P., Schroger, E., Hinrichs, H., Heinze, H.J., Deserno, L., Schlagenhauf, F., 2016a. The feedback-related negativity codes components of abstract inference during reward-based decision-making. J. Cogn. Neurosci. 28, 1127–1138. 9 Reiter, A.M.F., Deserno, L., Kallert, T., Heinze, H.J., Heinz, A., Schlagenhauf, F., 2016b. Behavioral and neural signatures of reduced updating of alternative options in alcohol-dependent patients during flexible decision-making. J. Neurosci. Reiter, A.M., Heinze, H.J., Schlagenhauf, F., Deserno, L., 2016c. Impaired flexible rewardbased decision-making in binge eating disorder: evidence from computational modeling and functional neuroimaging. Neuropsychopharmacology. Roiser, J.P., Stephan, K.E., den Ouden, H.E., Barnes, T.R., Friston, K.J., Joyce, E.M., 2009. Do patients with schizophrenia exhibit aberrant salience? Psychol. Med. 39 (2), 199–209. Schad, D.J., Junger, E., Sebold, M., Garbusow, M., Bernhardt, N., Javadi, A.H., Zimmermann, U.S., Smolka, M.N., Heinz, A., Rapp, M.A., Huys, Q.J., 2014. Processing speed enhances model-based over model-free reinforcement learning in the presence of high working memory functioning. Front. Psychol. 5, 1450. Schlagenhauf, F., Juckel, G., Koslowski, M., Kahnt, T., Knutson, B., Dembler, T., Kienast, T., Gallinat, J., Wrase, J., Heinz, A., 2008. Reward system activation in schizophrenic patients switched from typical neuroleptics to olanzapine. Psychopharmacology 196 (4), 673–684. Schlagenhauf, F., Rapp, M.A., Huys, Q.J., Beck, A., Wustenberg, T., Deserno, L., Buchholz, H.G., Kalbitzer, J., Buchert, R., Bauer, M., Kienast, T., Cumming, P., Plotkin, M., Kumakura, Y., Grace, A.A., Dolan, R.J., Heinz, A., 2013. Ventral striatal prediction error signaling is associated with dopamine synthesis capacity and fluid intelligence. Hum. Brain Mapp. 34 (6), 1490–1499. Schlagenhauf, F., Huys, Q.J., Deserno, L., Rapp, M.A., Beck, A., Heinze, H.J., Dolan, R., Heinz, A., 2014. Striatal dysfunction during reversal learning in unmedicated schizophrenia patients. NeuroImage 89, 171–180. Schultz, W., Dayan, P., Montague, P.R., 1997. A neural substrate of prediction and reward. Science 275 (5306), 1593–1599. Simon, D.A., Daw, N.D., 2011. Neural correlates of forward planning in a spatial decision task in humans. J. Neurosci. 31 (14), 5526–5539. Smittenaar, P., FitzGerald, T.H., Romei, V., Wright, N.D., Dolan, R.J., 2013. Disruption of dorsolateral prefrontal cortex decreases model-based in favor of model-free control in humans. Neuron 80 (4), 914–919. Somlai, Z., Moustafa, A.A., Keri, S., Myers, C.E., Gluck, M.A., 2011. General functioning predicts reward and punishment learning in schizophrenia. Schizophr. Res. 127 (1–3), 131–136. Steinberg, E.E., Keiflin, R., Boivin, J.R., Witten, I.B., Deisseroth, K., Janak, P.H., 2013. A causal link between prediction errors, dopamine neurons and learning. Nat. Neurosci. 16 (7), 966–973. Stephan, K.E., Mathys, C., 2014. Computational approaches to psychiatry. Curr. Opin. Neurobiol. 25, 85–92. Stephan, K.E., Iglesias, S., Heinzle, J., Diaconescu, A.O., 2015. Translational perspectives for computational neuroimaging. Neuron 87, 716–732. Strauss, G.P., Gold, J.M., 2012. A new perspective on anhedonia in schizophrenia. Am. J. Psychiatry 169 (4), 364–373. Strauss, G.P., Frank, M.J., Waltz, J.A., Kasanova, Z., Herbener, E.S., Gold, J.M., 2011. Deficits in positive reinforcement learning and uncertainty-driven exploration are associated with distinct aspects of negative symptoms in schizophrenia. Biol. Psychiatry 69 (5), 424–431. Strauss, G.P., Waltz, J.A., Gold, J.M., 2014. A review of reward processing and motivational impairment in schizophrenia. Schizophr. Bull. 40 (Suppl. 2), S107–S116. Treadway, M.T., Peterman, J.S., Zald, D.H., Park, S., 2015. Impaired effort allocation in patients with schizophrenia. Schizophr. Res. 161 (2–3), 382–385. Tricomi, E., Balleine, B.W., O'Doherty, J.P., 2009. A specific role for posterior dorsolateral striatum in human habit learning. Eur. J. Neurosci. 29 (11), 2225–2232. Ursu, S., Kring, A.M., Gard, M.G., Minzenberg, M.J., Yoon, J.H., Ragland, J.D., Solomon, M., Carter, C.S., 2011. Prefrontal cortical deficits and impaired cognition–emotion interactions in schizophrenia. Am. J. Psychiatry 168 (3), 276–285. Valentin, V.V., Dickinson, A., O'Doherty, J.P., 2007. Determining the neural substrates of goal-directed learning in the human brain. J. Neurosci. 27 (15), 4019–4026. Waltz, J.A., Gold, J.M., 2007. Probabilistic reversal learning impairments in schizophrenia: further evidence of orbitofrontal dysfunction. Schizophr. Res. 93 (1–3), 296–303. Waltz, J.A., Gold, J.M., 2015. Motivational deficits in schizophrenia and the representation of expected value. Curr. Top. Behav. Neurosci. Waltz, J.A., Frank, M.J., Robinson, B.M., Gold, J.M., 2007. Selective reinforcement learning deficits in schizophrenia support predictions from computational models of striatal–cortical dysfunction. Biol. Psychiatry 62 (7), 756–764. Waltz, J.A., Kasanova, Z., Ross, T.J., Salmeron, B.J., McMahon, R.P., Gold, J.M., Stein, E.A., 2013. The roles of reward, default, and executive control networks in set-shifting impairments in schizophrenia. PLoS One 8 (2), e57257. Waltz, J.A., Brown, J.K., Gold, J.M., Ross, T.J., Salmeron, B.J., Stein, E.A., 2015. Probing the dynamic updating of value in schizophrenia using a sensory-specific satiety paradigm. Schizophr. Bull. 41 (5), 1115–1122. Wang, X.J., Krystal, J.H., 2014. Computational psychiatry. Neuron 84 (3), 638–654. Wiecki, T.V., Poland, J., Frank, M.J., 2015. Model-based cognitive neuroscience approaches to computational psychiatry: clustering and classification. Clin. Psychol. Sci. 3 (3), 378–399. Wimmer, G.E., Daw, N.D., Shohamy, D., 2012. Generalization of value in reinforcement learning by humans. Eur. J. Neurosci. 35 (7), 1092–1104. Wise, R.A., 2008. Dopamine and reward: the anhedonia hypothesis 30 years on. Neurotox. Res. 14 (2–3), 169–183. Wolf, D.H., Satterthwaite, T.D., Kantrowitz, J.J., Katchmar, N., Vandekar, L., Elliott, M.A., Ruparel, K., 2014. Amotivation in schizophrenia: integrated assessment with behavioral, clinical, and imaging measures. Schizophr. Bull. 40 (6), 1328–1337. Wunderlich, K., Smittenaar, P., Dolan, R.J., 2012. Dopamine enhances model-based over model-free choice behavior. Neuron 75, 418–424. Yilmaz, A., Simsek, F., Gonul, A.S., 2012. Reduced reward-related probability learning in schizophrenia patients. Neuropsychiatr. Dis. Treat. 8, 27–34. Ziauddeen, H., Murray, G.K., 2010. The relevance of reward pathways for schizophrenia. Curr. Opin. Psychiatry 23 (2), 91–96. Please cite this article as: Deserno, L., et al., Computational approaches to schizophrenia: A perspective on negative symptoms, Schizophr. Res. (2016), http://dx.doi.org/10.1016/j.schres.2016.10.004