Maladaptive Avoidance Learning in the Orbitofrontal Cortex in Adolescents With Major Depression

  • David Willinger
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland

    Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland
    Search for articles by this author
  • Iliana I. Karipidis
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland

    Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland

    Center for Interdisciplinary Brain Sciences Research, Department of Psychiatry and Behavioral Sciences, School of Medicine, Stanford University, Stanford, California
    Search for articles by this author
  • Selina Neuer
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland
    Search for articles by this author
  • Sophie Emery
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland
    Search for articles by this author
  • Carolina Rauch
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland
    Search for articles by this author
  • Isabelle Häberling
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland
    Search for articles by this author
  • Gregor Berger
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland
    Search for articles by this author
  • Susanne Walitza
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland

    Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland
    Search for articles by this author
  • Silvia Brem
    Correspondence
    Address correspondence to Silvia Brem, Ph.D.
    Affiliations
    Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, Zurich, Switzerland

    Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland
    Search for articles by this author
Open AccessPublished:June 15, 2021DOI:https://doi.org/10.1016/j.bpsc.2021.06.005

      Abstract

      Background

      Understanding the mechanisms in the brain’s incentive network that give rise to symptoms of major depressive disorder (MDD) during adolescence provides new perspectives to address MDD in early stages of development. This functional magnetic resonance imaging study determines whether instrumental vigor and brain responses to appetitive and aversive monetary incentives are altered in adolescent MDD and associated with symptom severity.

      Methods

      Adolescents with moderate to severe MDD (n = 30, mean age [SD] = 16.1 [1.4] years) and healthy control subjects (n = 33, mean age = 16.2 [1.9] years) matched for age, sex, and IQ performed a monetary incentive delay task. During outcome presentation, prediction error signals were used to study the response and coupling of the incentive network during learning of cue-outcome associations. A computational reinforcement model was used to assess adaptation of response vigor. Brain responses and effective connectivity to model-derived prediction errors were assessed and related to depression severity and anhedonia levels.

      Results

      Participants with MDD behaved according to a more simplistic learning model and exhibited slower learning. Effective connectivity analysis of functional magnetic resonance imaging data revealed that impaired loss error processing in the orbitofrontal cortex was associated with aberrant gain control. Anhedonia scores correlated with loss-related error signals in the posterior insula and habenula.

      Conclusions

      Adolescent MDD is selectively related to impaired processing of error signals during loss, but not reward, in the orbitofrontal cortex. Aberrant evaluation of loss outcomes might reflect an early mechanism of how negative bias and helplessness manifest in the brain. This approach sheds light on pathomechanisms in MDD and may improve early diagnosis and treatment selection.

      Keywords

      Major depressive disorder (MDD) is among the most prevalent mental health problems in adolescents worldwide (
      ), with an estimated 12-month prevalence of 7.5% in mid to late adolescence (
      • Avenevoli S.
      • Swendsen J.
      • He J.P.
      • Burstein M.
      • Merikangas K.R.
      Major depression in the National Comorbidity Survey-Adolescent Supplement: Prevalence, correlates, and treatment.
      ). Adolescent MDD increases the risk for substance misuse; can severely impair success in school, social life, and cognitive functions (
      • Kieling C.
      • Adewuya A.
      • Fisher H.L.
      • Karmacharya R.
      • Kohrt B.A.
      • Swartz J.R.
      • Mondelli V.
      Identifying depression early in adolescence.
      ); and is a major risk factor for suicide, which is among the leading causes of death at this age (
      • Cha C.B.
      • Franz P.J.
      • M Guzmán E
      • Glenn C.R.
      • Kleiman E.M.
      • Nock M.K.
      Annual Research Review: Suicide among youth - epidemiology, (potential) etiology, and treatment.
      ). Despite these adverse outcomes, relatively little is known about brain mechanisms related to MDD with early onset. Recent evidence suggests that disrupted prediction error (PE) signaling constitutes a potential brain mechanism that promotes the persistence of negative beliefs and anhedonia (
      • Chen C.
      • Takahashi T.
      • Nakagawa S.
      • Inoue T.
      • Kusumi I.
      Reinforcement learning in depression: A review of computational research.
      ).
      It is widely established that the dopaminergic system is fundamental in encoding reward and loss PEs (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ), which are crucial in reinforcement learning and decision making. Influential computational models (
      • Rescorla R.A.
      • Wagner A.R.
      A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement.
      ,
      • Sutton R.S.
      Learning to predict by the methods of temporal differences.
      ) suggest that during anticipation of an incentive, an expected value (Q) signal is generated, which is the product of a learned probability and the magnitude of the incentive. During outcome receipt, the difference between the expected value signal and the actual outcome is signaled as PE to update predictions. While the ventral striatum primarily encodes PEs in reward contexts, the anterior insula does so in avoidance or loss contexts (
      • Palminteri S.
      • Justo D.
      • Jauffret C.
      • Pavlicek B.
      • Dauta A.
      • Delmaire C.
      • et al.
      Critical roles for anterior insula and dorsal striatum in punishment-based avoidance learning.
      ). Brain regions sensitive to errors in reward and loss contexts are the orbitofrontal cortex (OFC) and the dorsal anterior cingulate cortex (dACC) (
      • Knutson B.
      • Fong G.W.
      • Bennett S.M.
      • Adams C.M.
      • Hommer D.
      A region of mesial prefrontal cortex tracks monetarily rewarding outcomes: Characterization with rapid event-related fMRI.
      ,
      • Taylor S.F.
      • Martis B.
      • Fitzgerald K.D.
      • Welsh R.C.
      • Abelson J.L.
      • Liberzon I.
      • et al.
      Medial frontal cortex activity and loss-related responses to errors.
      ).
      Previous studies have demonstrated that striatal PE signaling is reduced in adult MDD in reward (
      • Gradin V.B.
      • Kumar P.
      • Waiter G.
      • Ahearn T.
      • Stickle C.
      • Milders M.
      • et al.
      Expected value and prediction error abnormalities in depression and schizophrenia.
      ,
      • Kumar P.
      • Waiter G.
      • Ahearn T.
      • Milders M.
      • Reid I.
      • Steele J.D.
      Abnormal temporal difference reward-learning signals in major depression.
      ) and loss (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ) contexts. Therefore, it is crucial to establish whether this aberrant signaling is already present in early-onset MDD or whether deviations in reward and loss processing are a downstream effect of chronicity and burden (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ). Studies in young cohorts suggest that blunted reward sensitivity in the ventral striatum predicts symptom deterioration (
      • Nelson B.D.
      • Perlman G.
      • Klein D.N.
      • Kotov R.
      • Hajcak G.
      Blunted neural response to rewards as a prospective predictor of the development of depression in adolescent girls.
      ,
      • Stringaris A.
      • Vidal-Ribas Belil P.
      • Artiges E.
      • Lemaitre H.
      • Gollier-Briant F.
      • Wolke S.
      • et al.
      The brain’s response to reward anticipation and depression in adolescence: Dimensionality, specificity, and longitudinal predictions in a community-based sample.
      ) and is present in individuals at high familial risk for depression (
      • Sharp C.
      • Kim S.
      • Herman L.
      • Pane H.
      • Reuter T.
      • Strathearn L.
      Major depression in mothers predicts reduced ventral striatum activation in adolescent female offspring with and without depression.
      ,
      • Luking K.R.
      • Pagliaccio D.
      • Luby J.L.
      • Barch D.M.
      Reward processing and risk for depression across development.
      ). In addition, there is emerging evidence of impaired loss sensitivity in the incentive network in high-risk groups (
      • Luking K.R.
      • Pagliaccio D.
      • Luby J.L.
      • Barch D.M.
      Depression risk predicts blunted neural responses to gains and enhanced responses to losses in healthy children.
      ) that predicts future depressive symptoms (
      • Jin J.
      • Narayanan A.
      • Perlman G.
      • Luking K.
      • DeLorenzo C.
      • Hajcak G.
      • et al.
      Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
      ). However, it remains unclear whether atypical learning signals are linked to deficient motivated behavior. Previous studies showed mixed results when applying computational models to behavioral data in adult MDD, with learning rates depending on the task used and the specific learning process probed (
      • Scholl J.
      • Klein-Flügge M.
      Understanding psychiatric disorder by capturing ecologically relevant features of learning and decision-making.
      ). This clearly indicates that more work is necessary to identify brain mechanisms that give rise to aberrant incentive processing in depression, particularly during development.
      In this functional magnetic resonance imaging (fMRI) study, we hypothesized that the encoding of reinforcement learning signals is impaired in adolescent MDD. We employed a monetary incentive delay task (
      • Knutson B.
      • Fong G.W.
      • Bennett S.M.
      • Adams C.M.
      • Hommer D.
      A region of mesial prefrontal cortex tracks monetarily rewarding outcomes: Characterization with rapid event-related fMRI.
      ,
      • Knutson B.
      • Westdorp A.
      • Kaiser E.
      • Hommer D.
      FMRI visualization of brain activity during a monetary incentive delay task.
      ) with varying magnitude (low, high) and valence (reward, loss) to probe the neural circuits supporting PE and expected value processing. On a neural level, we hypothesized 1) decreased reward PE signaling within the striatum in MDD (
      • Chen C.
      • Takahashi T.
      • Nakagawa S.
      • Inoue T.
      • Kusumi I.
      Reinforcement learning in depression: A review of computational research.
      ,
      • Stringaris A.
      • Vidal-Ribas Belil P.
      • Artiges E.
      • Lemaitre H.
      • Gollier-Briant F.
      • Wolke S.
      • et al.
      The brain’s response to reward anticipation and depression in adolescence: Dimensionality, specificity, and longitudinal predictions in a community-based sample.
      ), 2) a negative association between anhedonia scores and blunted responses to rewards in the striatum and OFC (
      • Der-Avakian A.
      • Markou A.
      The neurobiology of anhedonia and other reward-related deficits.
      ), and 3) reduced reactivity of the OFC during loss events (
      • Jin J.
      • Narayanan A.
      • Perlman G.
      • Luking K.
      • DeLorenzo C.
      • Hajcak G.
      • et al.
      Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
      ). Behaviorally, we tested whether response vigor, i.e., adaptive responses to reward and losses of varying magnitudes (
      • Niv Y.
      • Daw N.D.
      • Joel D.
      • Dayan P.
      Tonic dopamine: Opportunity costs and the control of response vigor.
      ), was differentially modulated in MDD and whether there are differences in the update of value representations in the instrumental learning task.

      Methods and Materials

       Participants

      A total of 30 adolescent patients and 33 healthy individuals matched for age, IQ, sex, and handedness participated in this study (Table 1). Participants with MDD were recruited through clinical services. All participants underwent a semistructured clinical interview [Schedule for Affective Disorders and Schizophrenia for School-Age Children--Present and Lifetime Version (
      • Kaufman J.
      • Birmaher B.
      • Brent D.
      • Rao U.
      • Flynn C.
      • Moreci P.
      • et al.
      Schedule for affective disorders and schizophrenia for school-age children-present and lifetime version (K-SADS-PL): Initial reliability and validity data.
      ) or Mini International Neuropsychiatric Interview for Children and Adolescents (
      • Sheehan D.V.
      • Sheehan K.H.
      • Shytle R.D.
      • Janavs J.
      • Bannon Y.
      • Rogers J.E.
      • et al.
      Reliability and validity of the mini international neuropsychiatric interview for children and adolescents (MINI-KID).
      )]. Participants with MDD fulfilled a diagnosis according to the DSM-IV (codes 296.20–296.23, 296.30–296.33). Past and present comorbid diagnoses in patients included anxiety disorders (n = 7), obsessive-compulsive disorder (n = 1), and attention-deficit/hyperactivity disorder (n = 2). Moreover, we assessed a battery of self-report questionnaires, IQ, and working memory of all participants (Table 1). We included total scores and scores from the anhedonia subscale from the German version of the Children's Depression Inventory (
      • Stiensmeier-Pelster J.
      • Braune-Krickau M.
      • Schürmann M.
      • Duda K.
      DIKJ: Depressionsinventar für Kinder und Jugendliche.
      ) in our neuroimaging analyses. Healthy control subjects were recruited through schools and volunteer websites. For control subjects, exclusion criteria included any current psychiatric disorder, other major medical illnesses, drug abuse, any MRI contraindication, pregnancy, and a history of brain injury. Three control participants had a past diagnostic work-up for attention-deficit/hyperactivity disorder, but they were currently symptom free and were not taking any medication during the study. All procedures contributing to this work comply with the ethical standards of the ethic committee of the Canton of Zurich, Switzerland (BASEC 2017-02179) and with the Helsinki Declaration of 1975, as revised in 2008. All participants gave their written informed consent; parents or legal guardians gave signed informed consent for children under the age of 14 years. They were reimbursed for participation and informed about the opportunity to additionally win up to CHF 20 during the task.
      Table 1Clinical and Demographic Characteristics of Study Participants
      CharacteristicsControl GroupMDD GroupTest Statisticp Value
      Uncorrected p values for between-group comparisons; significance threshold p < .05.
      Age, Years, Mean (SD), [Min–Max]16.2 (1.9), [11.2–18.8]16.1 (1.4), [12.8–18.7]U = 553.5.425
      Sex, Males, n (%)10 (30%)10 (33%)χ21 = 0.07.796
      Handedness, Right, n (%)32 (97%)28 (93%)χ21 = 0.46.500
      In-Scanner Movement, FD, mm0.18 (0.09)0.18 (0.08)t61 = 0.09.930
      CD-RISC72.9 (10.1)38.6 (15.6)t58 = 10.16< .001
      CDI8.4 (6.6)29.6 (9.3)U = 38.0< .001
       Anhedonia2.3 (2.2)10.5 (2.8)U = 13.5< .001
       Negative mood2.2 (2.0)6.4 (2.4)U = 88.0< .001
       Negative self-esteem1.0 (1.2)5.0 (1.7)U = 42.0< .001
       Ineffectiveness1.2 (1.2)5.0 (1.9)U = 54.5< .001
       Interpersonal problems1.1 (1.2)3.7 (1.5)U = 74.5< .001
       Stomach0.6 (0.6)1.1 (0.8)U = 301.5.018
      RIAS IQ104.5 (6.9)108.0 (8.7)t60 = −1.75.079
      PSS22.4 (6.6)28.8 (7.7)t57 = −3.44.001
      SDQ8.8 (5.3)16.3 (5.6)t56 = −5.26< .001
      WISC-IV Digit Span, Forward8.9 (2.1)8.8 (2.0)t60 = 0.32.747
      WISC-IV Digit Span, Backward8.6 (1.6)9.4 (2.0)t60 = −1.70.094
      WISC-IV Mosaic57.0 (5.7)59.0 (6.2)t56 = −1.27.208
      Current Medication, n (%)
       No medicationNA10 (33%)NANA
       SSRINA18 (60%)NANA
       Dual-action antidepressant
      Serotonin-noradrenaline reuptake inhibitor.
      NA2 (7%)NANA
       NERINA2 (7%)NANA
       Antipsychotic
      Used for behavioral control.
      NA2 (7%)NANA
       MethylphenidateNA2 (7%)NANA
      Data are presented as mean (SD) if not indicated otherwise.
      CDI, Children's Depression Inventory; CD-RISC, Connor-Davidson Resilience Scale; FD, framewise displacement; Max, maximum; MDD, major depressive disorder; Min, minimum; NA, not applicable; NERI, norepinephrine reuptake inhibitor; PSS, Perceived Stress Scale; RIAS, Reynolds Intellectual Assessment Scales; SDQ, Strengths and Difficulties Questionnaire; SSRI, selective serotonin reuptake inhibitor; WISC-IV, Wechsler Intelligence Scale for Children, fourth edition.
      a Uncorrected p values for between-group comparisons; significance threshold p < .05.
      b Serotonin-noradrenaline reuptake inhibitor.
      c Used for behavioral control.

       Experimental Task

      In this study, we used the monetary incentive delay task (Figure S1) to examine incentive processing in different valence and magnitude contexts (
      • Knutson B.
      • Westdorp A.
      • Kaiser E.
      • Hommer D.
      FMRI visualization of brain activity during a monetary incentive delay task.
      ). This fMRI task minimizes cognitive confounds (e.g., solving a task with different strategies) because of the simple decision processes (
      • Oldham S.
      • Murawski C.
      • Fornito A.
      • Youssef G.
      • Yücel M.
      • Lorenzetti V.
      The anticipation and outcome phases of reward and loss processing: A neuroimaging meta-analysis of the monetary incentive delay task.
      ). Importantly, it allowed us to assess individual trial-by-trial reward and loss PEs based on the associations participants built up relying on the outcome of the preceding trials, allowing us to distinguish it from outcome processing. Every trial began with the presentation of a cue indicating the level of magnitude (CHF 1, CHF 4) and the valence (reward, loss avoidance, null) for a button press on time (hit). Participants were instructed to use the index finger of their dominant hand to press a button on a 2-button fiberoptic response pad (Current Design, Inc., Philadelphia, PA) as soon as the go-signal target symbol, a star, appeared. In total, each cue was presented 24 times (i.e., 120 trials in total, mean stimulus onset asynchrony = 8500 ms) in two separate MRI runs. We used an adaptive algorithm that adjusted the presentation times of the target to the response time of the participant to ensure a hit rate of ~66%. The cue symbols indicating valence (square, triangle, and circle) were counterbalanced across subjects. A filled symbol indicated a trial with high magnitude (CHF 4), whereas an empty symbol indicated a trial with low magnitude (CHF 1). All participants had a short training session outside the scanner (~2 min) to familiarize themselves with the task, and we ensured that cue-outcome contingencies were understood. The task was implemented in python (pygame; https://www.pygame.org) and presented using video goggles (VisuaStim Digital; Resonance Technology, Northridge, CA) with a resolution of 800 × 600 pixels.
      Postscan ratings of subjective liking and arousal for rewards and losses were assessed at the end of the scanning session. Participants rated their subjective liking (How much did you like this outcome?) and arousal (How excited were you by the outcome?) on a continuous scale using a slider between 0 (strongly dislike, not aroused) and 100 (strongly like, highly aroused).

       Computational Modeling

      The task employed here allows for the assessment of mechanisms that determine behavior (i.e., instrumental response vigor). To assess these quantities, we constructed several competing generative behavioral models that predicted trial-by-trial reaction times for each participant. This allowed us to identify parameters with mechanistic meaning for observed response vigor, the latent representation of value, and the participant-specific learning rate. Even though no learning is necessary to perform this task well, previous studies have found that learning signals continue to be generated even if an association is well-learned (
      • Bayer H.M.
      • Glimcher P.W.
      Midbrain dopamine neurons encode a quantitative reward prediction error signal.
      ).
      The winning model included the parameters cue salience, i.e., the magnitude of possible outcomes × probability; novelty, i.e., the unsigned PE; a posterror term; a linear term; and the intercept. Group differences between controls and patients were assessed for parameters of the best-fitting model across both groups. A detailed description of the behavioral model, model simulations, and behavioral analysis is provided in the Supplement (Figures S2 and S3, Tables S1–S4).

       Functional MRI Analysis

       Model-Based fMRI—General Linear Model Analysis

      To investigate the trial-by-trial effect of the computational variables derived from the computational model, we used the variables of the best behavioral model across participants (Table S2) and entered expected value and PE as parametric modulators for cue onset and feedback, respectively (Supplemental Methods). Group effects were assessed with two-sample t tests, where we entered the contrast images for the expected values and the PEs for healthy control subjects and patients with MDD. The cluster-level significance threshold for the whole-brain group analyses (familywise error–corrected [FWEc]) was set to pFWEc < .05 with a cluster-defining threshold of pCDT < .001. All fMRI analyses were conducted in SPM12 (version 7487; Wellcome Trust Centre for Neuroimaging, University College London, London, UK); labels for brain regions are based on the Automated Anatomical Atlas (
      • Rolls E.T.
      • Huang C.C.
      • Lin C.P.
      • Feng J.
      • Joliot M.
      Automated anatomical labelling atlas 3.
      ).

       Effective Connectivity Analysis

      To reveal the functional coupling between regions of the incentive network in participants with and without MDD, we performed a dynamic causal modeling (DCM) analysis (
      • Friston K.J.
      • Harrison L.
      • Penny W.
      Dynamic causal modelling.
      ,
      • Zeidman P.
      • Jafarian A.
      • Corbin N.
      • Seghier M.L.
      • Razi A.
      • Price C.J.
      • Friston K.J.
      A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI.
      ). The regions for this analysis were selected based on 1) previously published findings of incentive processing (
      • Palminteri S.
      • Justo D.
      • Jauffret C.
      • Pavlicek B.
      • Dauta A.
      • Delmaire C.
      • et al.
      Critical roles for anterior insula and dorsal striatum in punishment-based avoidance learning.
      ,
      • Knutson B.
      • Fong G.W.
      • Bennett S.M.
      • Adams C.M.
      • Hommer D.
      A region of mesial prefrontal cortex tracks monetarily rewarding outcomes: Characterization with rapid event-related fMRI.
      ,
      • Hauser T.U.
      • Iannaccone R.
      • Walitza S.
      • Brandeis D.
      • Brem S.
      Cognitive flexibility in adolescence: Neural and behavioral mechanisms of reward prediction error processing in adaptive decision making during development.
      ), 2) findings of studies in participants with a history of and at risk for MDD (
      • Luking K.R.
      • Pagliaccio D.
      • Luby J.L.
      • Barch D.M.
      Depression risk predicts blunted neural responses to gains and enhanced responses to losses in healthy children.
      ,
      • Jin J.
      • Narayanan A.
      • Perlman G.
      • Luking K.
      • DeLorenzo C.
      • Hajcak G.
      • et al.
      Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
      ), and 3) our second-level general linear model analyses (Figure 1A; Tables S5–S8 and Figures S4–S6). The aforementioned studies have provided compelling evidence that the insula and dACC play a significant role in loss-avoidance learning, a finding we corroborate across groups during reward and loss processing. Furthermore, we found a significant group effect in the OFC that suggested aberrant network dynamics in MDD.
      Figure thumbnail gr1
      Figure 1(A) Effect of loss outcome error ϕtduring feedback presentation across all participants (n = 63). Patients showed significantly reduced responses related to errors during loss compared with control subjects in the orbitofrontal cortex (OFC, green). Consistent effects across groups (positive effect in yellow, negative effect in blue) of outcome error were found in the dorsal anterior cingulate cortex (dACC) and the anterior insula. Details are reported in and main effects in . pFWEc < .05, pCDT < .001. (B) Assessment of the effect of anhedonia within patients (n = 29). A negative relationship between magnitude-modulated loss prediction error δt-related actitivity and anhedonia scores was observed in a cluster in the medial thalamus/habenula and the posterior cingulate cortex (PCC). (C) A significant negative association between anhedonia scores and loss outcome error ϕt-related activity was found in the posterior insula. CDT, cluster-defining threshold; FWEc, familywise error–corrected.
      For the DCM analysis, we localized the effects of hits and misses across trials by performing whole-brain contrasts with the Children's Depression Inventory as covariate (Figure S5) and extracted the time series for each subject from activated voxels (p < .05, uncorrected) within a 12-mm spherical search volume around the group maxima from the miss-hit (insula [x = 35, y = 18, z = −18 mm Montreal Neurological Institute (MNI)]; dACC [x = 7, y = 36, z = 30 mm MNI]) (Table S8) and the all-events (inferior occipital gyrus [x = 29, y = −80, z = −16 mm MNI]) (Table S10) contrast. The search volume for the OFC comprised the voxels in the active OFC (x = −7, y = 46, z = −16 mm MNI) cluster of the second-level hit-miss contrast (Table S8) (cluster-extent threshold pFWEc < .05). If a participant’s maximum within the search volume differed from the group maximum, we centered the sphere around the participant’s maximum. The first eigenvariate of the time course of all active voxels (p < .05) was then extracted and adjusted for any motion effects. One patient and 2 control subjects had to be excluded from this analysis because they did not show any activation in the dACC for the defined threshold.
      The feedback regressor was the driving input for the visual region. The model was composed of direct forward connections from the visual area to all other fully interconnected regions. Because the general linear model analysis revealed loss-related group differences in OFC activity (see Results), our main interest was to study network effects during the loss condition; nevertheless, the reward condition was also included to improve the model fit. We included contextual modulation of PEs, magnitude (−1 for low and 1 for high), and their interaction term, i.e., magnitude-sensitive PEs on the self-connections of the regions (see Supplement). In this model, the self-connections embody the change in synaptic gain for a given task context (
      • Zeidman P.
      • Jafarian A.
      • Corbin N.
      • Seghier M.L.
      • Razi A.
      • Price C.J.
      • Friston K.J.
      A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI.
      ). Here, our goal was to identify the network dynamics that give rise to the lower error signal in the OFC in patients with MDD during loss processing. For this, we set up a DCM analysis within the parametric empirical Bayes (PEB) framework.
      On the first level, the full DCM model of each participant was estimated iteratively in an empirical Bayesian inversion scheme (
      • Zeidman P.
      • Jafarian A.
      • Corbin N.
      • Seghier M.L.
      • Razi A.
      • Price C.J.
      • Friston K.J.
      A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI.
      ). The individual DCM parameters from the first level were then entered in the second-level PEB model to determine 1) the differences between the MDD group and the control group and 2) the group mean. This analysis was carried out separately for intrinsic and modulatory connections. We performed Bayesian model reduction to iteratively discard those model parameters not contributing to the model evidence. Then, we averaged the parameters of the best PEB models weighted by the posterior probability of the respective model. A PEB model parameter was considered significant when exceeding a 95% posterior probability of being present versus absent based on the model evidence (
      • Zeidman P.
      • Jafarian A.
      • Corbin N.
      • Seghier M.L.
      • Razi A.
      • Price C.J.
      • Friston K.J.
      A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI.
      ,
      • Friston K.J.
      • Litvak V.
      • Oswal A.
      • Razi A.
      • Stephan K.E.
      • Van Wijk B.C.M.
      • et al.
      Bayesian model reduction and empirical Bayes for group (DCM) studies.
      ). Leave-one-out cross-validation was used to assess whether the predicted and actual group effect showed an independent out-of-sample correlation.

      Results

       Altered Learning of Cue-Outcome Associations in MDD

      Bayesian model comparison revealed that the response model including cue salience and novelty terms using a single learning rate fitted the response data best across groups. Nevertheless, we found that in patients, the simpler model with a single learning rate performed better (posterior probability = 57.7%) (Table S2), whereas a more complex model with separate learning rates for rewards and losses fitted data better in control subjects only (posterior probability = 51.6%). However, while we found this difference among groups, the difference between the simple and the more complex model was not very strong (Bayes factor = 2.00), and thus we decided to continue the analysis with the simpler model because this performed best for both groups. A between-group comparison of the learning rate of the best-fitting model across all participants showed that the learning rate was marginally lower in MDD (α: MDD, 0.050 [SD = 0.021]; control subjects, 0.052 [SD = 0.012]; U = 617; p = .095). This difference was significant after removing 1 outlier from to the MDD group (Grubb’s test: G = 6.107, U = 0.389, p < 10−11; α: MDD, 0.046 [SD = 0.009]; control subjects, 0.052 [SD = 0.012]; t60 = 2.04; p = .046) (Supplemental Methods). The response model parameters, in particular cue salience and novelty, did not differ between groups (all p > .10) (Table S4). These results demonstrate a nondiscriminable value update mechanism underlying adolescent MDD for both reward and loss conditions, while there was positive evidence that control subjects’ response vigor was best described by a more flexible dual update model for both valences. In addition, the comparison of the learning rate shows that participants with MDD changed their value expectations slower across the task.

       OFC Gain Control Explains Atypical Aversive Outcome Signaling in MDD

      When processing loss feedback, participants with MDD showed a significantly lower response to the outcome error signal ϕt in the OFC (Figure 1 and Table 2). While this region reflected a signal encoding the difference between subjective belief of the outcome and the actual outcome in reward and loss in control subjects (Table S5), patients only expressed this signal during rewarding and not loss-avoidance trials.
      Table 2Group Differences Between Patients With MDD and Healthy Control Subjects for ϕt Processing
      Brain RegionMNI Coordinates, mmSignificant ActivationPeak
      xyzpFWEckZ
      Controls > MDD
       Precuneus_L−3−4610.0221284.67
       Frontal_Sup_2_R172844.0461084.48
       Temporal_Mid_L−67−50−6.0002694.24
       Angular_L−55−6424.0061664.08
       Temporal_Mid_R65−10−16.0341164.05
       Frontal_Med_Orb_L−360−8.0191323.93
      MDD > Controls (NS)
      Significance level at whole-brain cluster-level pFWEc < .05, cluster-defining threshold pCDT < .001. Main effects are reported in Table S5.
      k, cluster size; FWEc, familywise error–corrected; L, left; MDD, major depressive disorder; Med, medial; Mid, middle; MNI, Montreal Neurological Institute; NS, not significant; Orb, orbital; R, right; Sup, superior.
      To further scrutinize the origin of this effect, we performed a DCM analysis (Figure 2 and Table 3). Bayesian model averaging showed that the effect of loss magnitude on the self-inhibition parameter of the OFC was significantly more negative in MDD, i.e., the region was more disinhibited during processing the outcome of high compared with low loss. These results indicate that MDD is related to aberrant gain control in the OFC, specifically in high-loss contexts. Moreover, self-inhibition of the dACC was significantly lower across task conditions in the MDD group. The posterior mean of the group effect on self-inhibition of the OFC was significantly related to the learning rate across all participants (Spearman’s ρ = 0.295, p = .023, n = 59), but not in the dACC (ρ = −0.135, p = .301, n = 59). A leave-one-out cross-validation using the loss magnitude–dependent difference in self-connection strength in the OFC explained a significant amount of the intersubject variability between MDD and controls, showing an independent out-of-sample correlation of r58 = 0.38, p = .001.
      Figure thumbnail gr2
      Figure 2Effective connectivity during reward and loss feedback processing. There was a significant group effect of the factor loss magnitude on the self-inhibition parameter in the orbitofrontal cortex (OFC), indicating aberrant input sensitivity during feedback in loss-avoidance contexts in adolescent major depressive disorder (MDD). Cross-validation showed that this effective connectivity parameter was able to predict the group variable indicated by a significant out-of-sample correlation. Importantly, this parameter was associated with the learning rate of participants, which was lower in adolescents with MDD. The arrows reflect the posterior estimates of the second-level parametric empirical Bayes model after Bayesian model reduction. Self-connections are depicted as a half-circle on each region. Solid lines indicate positive effective connectivity, whereas dashed lines represent negative effective connectivity. Details of the results are reported in . dACC, dorsal anterior cingulate cortex; INS, anterior insula; IOG, inferior occipital gyrus; ϕ, outcome error.
      Table 3Average Connectivity During Feedback Phase Obtained by Bayesian Model Averaging of PEB Model Parameters
      Connection TypeCommonalitiesPP CommonalitiesDepressionPP Depression
      Endogenous Parameters
       OFC→Insula0000
       OFC→dACC−0.19510.040.66
       dACC→Insula0.812100
       dACC→OFC−0.345100
       Insula→OFC0.284100
       Insula→dACC−0.0841−0.0490.85
       IOG→Insula0.144100
       IOG→dACC−0.22810.0170.53
       IOG→OFC−0.1081−0.0130.58
      Self-Inhibition Parameters
       OFC→OFC−0.4671
       Insula→Insula0.3111
       dACC→dACC−0.2881−0.1391
       IOG→IOG1.7401
      Modulatory Parameters
       Insula→Insula, ϕt+0000
       Insula→Insula, M+0000
       Insula→Insula, ϕt+× M+0000
       Insula→Insula, ϕt0000
       Insula→Insula, M0000
       Insula→Insula, ϕt× M−0.482100
       dACC→dACC, ϕt+1.506100
       dACC→dACC, M+−0.1030.5500
       dACC→dACC, ϕt+× M+000.1940.89
       dACC→dACC, ϕt1.986100
       dACC→dACC, M−0.628100
       dACC→dACC, ϕt× M−0.3080.9900
       OFC→OFC, ϕt+1.50610.3510.90
       OFC→OFC, M+0.1480.5400
       OFC→OFC, ϕt+× M+00−0.1400.51
       OFC→OFC, ϕt−0.748100
       OFC→OFC, M−0.4951−0.6631
       OFC→OFC, ϕt× M0.781100
      Input Parameter
       Feedback→IOG4.2871−0.2080.64
      Between-region connections are in units of Hz. Self-connections, where the source and target are the same, are the log of scaling parameters that multiply up or down the default value of −0.5 Hz. n = 60.
      dACC, dorsal anterior cingulate cortex; IOG, inferior occipital gyrus; M, magnitude; OFC, orbitofrontal cortex; PEB, parametric empirical Bayes; PP, posterior probability; ϕ, error signal.

       Neural Correlates of Depression Severity and Anhedonia

      Regression analysis of anhedonia scores within patients revealed significant associations in brain signaling of learning variables for loss. Particularly, we found that magnitude-modulated loss PE signaling δt was associated negatively with anhedonia scores in the medial thalamus/habenula, posterior cingulate cortex, postcentral gyrus, and fusiform gyrus (Table S6; Figure 1B). The loss-related outcome error signal ϕtin the posterior insula was associated negatively with anhedonia scores (peak Z = 4.73, pFWEc < .001) (Table S6; Figure 1C). No within-patient associations were observed for reward-related signals.

       No Effects of Selective Serotonin Reuptake Inhibitors on Brain Activity in Participants With MDD

      We conducted additional analyses comparing effects of the learning parameters, δ, ϕ, and Q to see if there were systematic differences of brain activity related to selective serotonin reuptake inhibitor (SSRI) medication. We compared groups of 18 patients medicated with SSRIs to 10 other patients (unmedicated or no SSRI). However, general linear model analyses of expected values and PE did not reveal any significant group differences. No clusters survived (pFWEc < .05) a threshold of pCDT < .001 or a more lenient cluster-defining threshold of pCDT < .005. Moreover, DCM parameter values of the OFC and dACC were assessed for SSRI effects. To this end, we extracted the individual posterior means for each patient and split them into two groups (SSRI, other or no medication). However, we did not find a significant difference between patients taking SSRIs and others in the OFC (t26 = −1.44, p = .162) or the dACC (t26 = −0.08, p = .936). Note that 1 patient was excluded from this analysis because their medication history was not disclosed.

      Discussion

      In this study, we used a combination of computational modeling, fMRI, and connectivity analysis to study reward and loss processing in adolescent MDD. We demonstrated that adolescent MDD is associated with 1) aberrant gain control in the OFC during loss feedback processing and 2) anhedonia-related reduction of representation of loss outcomes in the posterior insula and habenula, but we found no evidence for aberrant reward PE processing in the ventral striatum and the medial prefrontal cortex. Thus, this work provides novel insights into the neurobiological foundation of altered learning mechanisms in loss avoidance in early-onset MDD.
      Our computational modeling approach revealed differences in behavioral adaptation that were reflected in the learning rate to update one’s belief about prospective receipt of reward and loss. By testing a series of behavioral models, we showed that participants with MDD adapted their instrumental vigor according to a simpler learning model with a single learning rate for reward and loss and that they also updated the expected values slower than control subjects. While this does not indicate a different learning mechanism in MDD per se, it suggests that instrumental behavior does not rely on differential update rates to adapt behavior for approaching reward and avoiding loss. The speed of learning depends on the perceived volatility of the environment (
      • Behrens T.E.
      • Woolrich M.W.
      • Walton M.E.
      • Rushworth M.F.
      Learning the value of information in an uncertain world.
      ), which has been suggested to be affected in anhedonia and anxiety (
      • Bishop S.J.
      • Gagne C.
      Anxiety, depression, and decision making: A computational perspective.
      ), the latter being a highly prevalent comorbidity in depression (
      • Häberling I.
      • Baumgartner N.
      • Emery S.
      • Keller P.
      • Strumberger M.
      • Nalani K.
      • et al.
      Anxious depression as a clinically relevant subtype of pediatric major depressive disorder.
      ). Reduced learning and updating of one’s belief system and hence an inability to update and hold an appropriate structure of possible aversive outcomes might provide the basis for a biased evaluation of the environment.
      On the neural level, we linked these differences in value representation updates derived from the computational model to neural feedback processing. While in control subjects OFC activity was related to an outcome error signal across task conditions, this was absent in participants with MDD during loss processing. Effective connectivity analysis further showed that this effect was primarily driven by aberrant gain control in the OFC, specifically the sensitivity tuning in varying magnitude contexts. Gain in the OFC is modulated by various neurotransmitters (
      • Robbins T.W.
      • Arnsten A.F.
      The neuropsychopharmacology of fronto-executive function: Monoaminergic modulation.
      ), but there is evidence that dopamine plays a role in learning approach and avoidance behavior (
      • Palminteri S.
      • Justo D.
      • Jauffret C.
      • Pavlicek B.
      • Dauta A.
      • Delmaire C.
      • et al.
      Critical roles for anterior insula and dorsal striatum in punishment-based avoidance learning.
      ), supported by dense dopaminergic projections to the OFC (
      • Kahnt T.
      • Tobler P.N.
      Dopamine modulates the functional organization of the orbitofrontal cortex.
      ). This gain control might be critical to modulating the activity of the OFC during error processing in reward (
      • Ramnani N.
      • Elliott R.
      • Athwal B.S.
      • Passingham R.E.
      Prediction error for free monetary reward in the human prefrontal cortex.
      ) and loss (
      • Taylor S.F.
      • Martis B.
      • Fitzgerald K.D.
      • Welsh R.C.
      • Abelson J.L.
      • Liberzon I.
      • et al.
      Medial frontal cortex activity and loss-related responses to errors.
      ) contexts, updating neural representations of value (
      • Sul J.H.
      • Kim H.
      • Huh N.
      • Lee D.
      • Jung M.W.
      Distinct roles of rodent orbitofrontal and medial prefrontal cortex in decision making.
      ), and maintaining a representation of the task structure (
      • Wilson R.C.
      • Takahashi Y.K.
      • Schoenbaum G.
      • Niv Y.
      Orbitofrontal cortex as a cognitive map of task space.
      ). The latter entails updating an estimate of certainty for a specific outcome, which seems to fail in adolescent MDD when evaluating unexpected outcomes in avoidance learning. Strikingly, we found a significant negative association between learning rate and gain control of the OFC during loss processing.
      Although altered OFC activity and connectivity for aversive stimuli has been reported in various paradigms in adults with MDD (
      • Gao Q.
      • Zou K.
      • He Z.
      • Sun X.
      • Chen H.
      Causal connectivity alterations of cortical-subcortical circuit anchored on reduced hemodynamic response brain regions in first-episode drug-naïve major depressive disorder.
      ,
      • Schiller C.E.
      • Minkel J.
      • Smoski M.J.
      • Dichter G.S.
      Remitted major depression is characterized by reduced prefrontal cortex reactivity to reward loss.
      ), the few studies investigating adult MDD with the monetary incentive delay task did not report changes in loss-related processing (
      • Ubl B.
      • Kuehner C.
      • Kirsch P.
      • Ruttorf M.
      • Diener C.
      • Flor H.
      Altered neural reward and loss processing and prediction error signalling in depression.
      ,
      • Pizzagalli D.A.
      • Holmes A.J.
      • Dillon D.G.
      • Goetz E.L.
      • Birk J.L.
      • Bogdan R.
      • et al.
      Reduced caudate and nucleus accumbens response to rewards in unmedicated individuals with major depressive disorder.
      ). On the other hand, a longitudinal study in adolescents at risk for MDD (
      • Jin J.
      • Narayanan A.
      • Perlman G.
      • Luking K.
      • DeLorenzo C.
      • Hajcak G.
      • et al.
      Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
      ) showed that loss-related OFC-insula connectivity predicted depressive symptoms 9 months later. This suggests that altered OFC function during loss processing could be specific for early-onset depression. However, more research on OFC connectivity changes over the course of the disorder from adolescence into adulthood is needed. In accordance with Jin et al. (
      • Jin J.
      • Narayanan A.
      • Perlman G.
      • Luking K.
      • DeLorenzo C.
      • Hajcak G.
      • et al.
      Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
      ), we also showed that decreased loss error signaling in the posterior insula was related to anhedonia. While the encoding of magnitude-modulated loss PE signals did not significantly differ between groups, a within-patient analysis revealed that blood oxygen level–dependent responses related to loss PE in a cluster composed of the medial thalamus and habenula were significantly negatively associated with anhedonia. Previous work has implicated impaired habenula function and morphology in depression and anhedonia (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ,
      • Lawson R.P.
      • Nord C.L.
      • Seymour B.
      • Thomas D.L.
      • Dayan P.
      • Pilling S.
      • Roiser J.P.
      Disrupted habenula function in major depression.
      ). Thus, altered loss processing could reflect an important factor that contributes to increased susceptibility to adolescent MDD.
      Recent computational accounts on depression suggest that it is related to an aberrant cognitive prior that underlies negative bias in evaluation of the state of the environment (
      • Clark J.E.
      • Watson S.
      • Friston K.J.
      What is mood? A computational perspective.
      ). An overgeneralization of one’s own states might eventually lead to helplessness behavior, where negative outcomes are associated with poor performance and where failure of oneself and positive outcomes are regarded as mere random events. Our results could indicate that a negative prior about the outcome is not updated owing to dysfunction in the OFC, and this might contribute to maintaining a negative bias and a feeling of loss of control over outcomes. The latter is consistent with the differences of negative arousal derived from the postscan ratings, where participants with MDD expressed more relief (i.e., more deactivation) in rewarding outcomes and more fear (i.e., more activation) during loss outcomes. However, unlike in adult depression (
      • Steele J.D.
      • Kumar P.
      • Ebmeier K.P.
      Blunted response to feedback information in depressive illness.
      ), computational modeling did not indicate that this higher range of negative arousal significantly affected response vigor in MDD.
      Contrary to our hypothesis, we did not find any evidence of impact of depression on PE processing in rewarding contexts as previous studies in adult MDD (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ,
      • Gradin V.B.
      • Kumar P.
      • Waiter G.
      • Ahearn T.
      • Stickle C.
      • Milders M.
      • et al.
      Expected value and prediction error abnormalities in depression and schizophrenia.
      ,
      • Kumar P.
      • Waiter G.
      • Ahearn T.
      • Milders M.
      • Reid I.
      • Steele J.D.
      Abnormal temporal difference reward-learning signals in major depression.
      ). Here, we postulate that two factors could have led to this null finding. First, in our monetary incentive delay task, participants did not have to learn anything to perform well. This design was employed to minimize confounds of 1) brain maturation and development within participant groups (
      • Nussenbaum K.
      • Hartley C.A.
      Reinforcement learning across development: What insights can we draw from a decade of research?.
      ) and 2) diagnosis (
      • Snyder H.R.
      Major depressive disorder is associated with broad impairments on neuropsychological measures of executive function: A meta-analysis and review.
      ) on learning performance, which could be difficult to disentangle in more complex learning paradigms. Nevertheless, our results are in concordance with previous findings of intact reward PE signaling in a nonlearning task in adult MDD (
      • Rutledge R.B.
      • Moutoussis M.
      • Smittenaar P.
      • Zeidman P.
      • Taylor T.
      • Hrynkiewicz L.
      • et al.
      Association of neural and emotional impacts of reward prediction errors with major depression.
      ). Second, there is evidence that impairments of reward PE signaling are related to the number of depressive episodes across the lifetime (
      • Kumar P.
      • Goer F.
      • Murray L.
      • Dillon D.G.
      • Beltzer M.L.
      • Cohen A.L.
      • et al.
      Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
      ). This might explain the results in participants with an early onset as in our study and could indicate that previous reports of impaired reward PE signaling errors are related to the chronicity of the disorder.
      It has to be considered that the majority of participants with depression were receiving antidepressive medication (Table 1). It is possible that intake of SSRIs might have affected error signaling and learning of cue-outcome associations (
      • McCabe C.
      • Mishor Z.
      • Cowen P.J.
      • Harmer C.J.
      Diminished neural processing of aversive and rewarding stimuli during selective serotonin reuptake inhibitor treatment.
      ). Based on this assumption, we would expect blunting of reward responses due to the administration of SSRIs. However, additional control analyses of brain activity and connectivity comparing participants with MDD with (n = 18) and without (n = 10) SSRI intake revealed no significant effect. Although we cannot fully rule out that medication had an effect based on these rather small subsamples, we consider it unlikely that this was the case in this study. Nevertheless, the different medication status of patients should be considered when interpreting the results, and further investigation of effects of SSRIs in depression is warranted. Moreover, we note that it is not possible to make causal inferences using a cross-sectional design, and future work should assess larger, longitudinal samples to shed light on the neural mechanisms that could give rise to MDD in adolescence.
      In conclusion, this is the first study to our knowledge to show that adolescent MDD is associated with specific impairments of error processing in loss-avoidance contexts, whereas reward sensitivity is intact. Given the critical role of evaluating an action that led to an unexpected aversive outcome, this deficit could be directly related to severe difficulties in decision making and in social life and by contributing to the development and persistence of a negative bias in depression. Our study provides a first important step toward identifying computational mechanisms in adolescent MDD and paves the way for establishing computational assays (
      • Stephan K.E.
      • Mathys C.
      Computational approaches to psychiatry.
      ) that will facilitate the translation into clinical practice.

      Acknowledgments and Disclosures

      This research was funded by the Department of Child and Adolescent Psychiatry and Psychotherapy, University Hospital of Psychiatry Zurich, University of Zurich, and by the Olga Mayenfisch Foundation.
      We thank Plamina Dimanova, Nada Frei, Dr. Noemi Baumgartner, Mona Albermann, Kristin Nalani, Paola Keller, Desiree Thommen, Luana Signer, and Dr. Philipp Stämpfli for technical and clinical assistance.
      GB has received lecture honoraria from Lundbeck, Opopharma, and Antistress AG (Burgerstein) in the last 5 years. SW and SB report no conflicts of interest, but their outside professional activities and interests are declared under the link of the University of Zurich (www.uzh.ch/prof/ssl-dir/interessenbindungen/client/web/). All other authors report no biomedical financial interests or potential conflicts of interest.

      Supplementary Material

      References

      1. Lopez A.D. Mathers C.D. Ezzati M. Jamison D.T. Murray C.J. Global Burden of Disease and Risk Factors. The World Bank, Washington, D.C.2006
        • Avenevoli S.
        • Swendsen J.
        • He J.P.
        • Burstein M.
        • Merikangas K.R.
        Major depression in the National Comorbidity Survey-Adolescent Supplement: Prevalence, correlates, and treatment.
        J Am Acad Child Adolesc Psychiatry. 2015; 54: 37-44.e32
        • Kieling C.
        • Adewuya A.
        • Fisher H.L.
        • Karmacharya R.
        • Kohrt B.A.
        • Swartz J.R.
        • Mondelli V.
        Identifying depression early in adolescence.
        Lancet Child Adolesc Health. 2019; 3: 211-213
        • Cha C.B.
        • Franz P.J.
        • M Guzmán E
        • Glenn C.R.
        • Kleiman E.M.
        • Nock M.K.
        Annual Research Review: Suicide among youth - epidemiology, (potential) etiology, and treatment.
        J Child Psychol Psychiatry. 2018; 59: 460-482
        • Chen C.
        • Takahashi T.
        • Nakagawa S.
        • Inoue T.
        • Kusumi I.
        Reinforcement learning in depression: A review of computational research.
        Neurosci Biobehav Rev. 2015; 55: 247-267
        • Kumar P.
        • Goer F.
        • Murray L.
        • Dillon D.G.
        • Beltzer M.L.
        • Cohen A.L.
        • et al.
        Impaired reward prediction error encoding and striatal-midbrain connectivity in depression.
        Neuropsychopharmacology. 2018; 43: 1581-1588
        • Rescorla R.A.
        • Wagner A.R.
        A theory of Pavlovian conditioning: Variations in the effectiveness of reinforcement and nonreinforcement.
        in: Classical Conditioning: Current Research and Theory New York. Appleton-Century-Crofts, 1972: 64-99
        • Sutton R.S.
        Learning to predict by the methods of temporal differences.
        Mach Learn. 1988; 3: 9-44
        • Palminteri S.
        • Justo D.
        • Jauffret C.
        • Pavlicek B.
        • Dauta A.
        • Delmaire C.
        • et al.
        Critical roles for anterior insula and dorsal striatum in punishment-based avoidance learning.
        Neuron. 2012; 76: 998-1009
        • Knutson B.
        • Fong G.W.
        • Bennett S.M.
        • Adams C.M.
        • Hommer D.
        A region of mesial prefrontal cortex tracks monetarily rewarding outcomes: Characterization with rapid event-related fMRI.
        Neuroimage. 2003; 18: 263-272
        • Taylor S.F.
        • Martis B.
        • Fitzgerald K.D.
        • Welsh R.C.
        • Abelson J.L.
        • Liberzon I.
        • et al.
        Medial frontal cortex activity and loss-related responses to errors.
        J Neurosci. 2006; 26: 4063-4070
        • Gradin V.B.
        • Kumar P.
        • Waiter G.
        • Ahearn T.
        • Stickle C.
        • Milders M.
        • et al.
        Expected value and prediction error abnormalities in depression and schizophrenia.
        Brain. 2011; 134: 1751-1764
        • Kumar P.
        • Waiter G.
        • Ahearn T.
        • Milders M.
        • Reid I.
        • Steele J.D.
        Abnormal temporal difference reward-learning signals in major depression.
        Brain. 2008; 131: 2084-2093
        • Nelson B.D.
        • Perlman G.
        • Klein D.N.
        • Kotov R.
        • Hajcak G.
        Blunted neural response to rewards as a prospective predictor of the development of depression in adolescent girls.
        Am J Psychiatry. 2016; 173: 1223-1230
        • Stringaris A.
        • Vidal-Ribas Belil P.
        • Artiges E.
        • Lemaitre H.
        • Gollier-Briant F.
        • Wolke S.
        • et al.
        The brain’s response to reward anticipation and depression in adolescence: Dimensionality, specificity, and longitudinal predictions in a community-based sample.
        Am J Psychiatry. 2015; 172: 1215-1223
        • Sharp C.
        • Kim S.
        • Herman L.
        • Pane H.
        • Reuter T.
        • Strathearn L.
        Major depression in mothers predicts reduced ventral striatum activation in adolescent female offspring with and without depression.
        J Abnorm Psychol. 2014; 123: 298-309
        • Luking K.R.
        • Pagliaccio D.
        • Luby J.L.
        • Barch D.M.
        Reward processing and risk for depression across development.
        Trends Cogn Sci. 2016; 20: 456-468
        • Luking K.R.
        • Pagliaccio D.
        • Luby J.L.
        • Barch D.M.
        Depression risk predicts blunted neural responses to gains and enhanced responses to losses in healthy children.
        J Am Acad Child Adolesc Psychiatry. 2016; 55: 328-337
        • Jin J.
        • Narayanan A.
        • Perlman G.
        • Luking K.
        • DeLorenzo C.
        • Hajcak G.
        • et al.
        Orbitofrontal cortex activity and connectivity predict future depression symptoms in adolescence.
        Biol Psychiatry Cogn Neurosci Neuroimaging. 2017; 2: 610-618
        • Scholl J.
        • Klein-Flügge M.
        Understanding psychiatric disorder by capturing ecologically relevant features of learning and decision-making.
        Behav Brain Res. 2018; 355: 56-75
        • Knutson B.
        • Westdorp A.
        • Kaiser E.
        • Hommer D.
        FMRI visualization of brain activity during a monetary incentive delay task.
        Neuroimage. 2000; 12: 20-27
        • Der-Avakian A.
        • Markou A.
        The neurobiology of anhedonia and other reward-related deficits.
        Trends Neurosci. 2012; 35: 68-77
        • Niv Y.
        • Daw N.D.
        • Joel D.
        • Dayan P.
        Tonic dopamine: Opportunity costs and the control of response vigor.
        Psychopharmacology. 2007; 191: 507-520
        • Kaufman J.
        • Birmaher B.
        • Brent D.
        • Rao U.
        • Flynn C.
        • Moreci P.
        • et al.
        Schedule for affective disorders and schizophrenia for school-age children-present and lifetime version (K-SADS-PL): Initial reliability and validity data.
        J Am Acad Child Adolesc Psychiatry. 1997; 36: 980-988
        • Sheehan D.V.
        • Sheehan K.H.
        • Shytle R.D.
        • Janavs J.
        • Bannon Y.
        • Rogers J.E.
        • et al.
        Reliability and validity of the mini international neuropsychiatric interview for children and adolescents (MINI-KID).
        J Clin Psychiatry. 2010; 71: 313-326
        • Stiensmeier-Pelster J.
        • Braune-Krickau M.
        • Schürmann M.
        • Duda K.
        DIKJ: Depressionsinventar für Kinder und Jugendliche.
        3rd ed. Hogrefe Publishing, Göttingen2014
        • Oldham S.
        • Murawski C.
        • Fornito A.
        • Youssef G.
        • Yücel M.
        • Lorenzetti V.
        The anticipation and outcome phases of reward and loss processing: A neuroimaging meta-analysis of the monetary incentive delay task.
        Hum Brain Mapp. 2018; 39: 3398-3418
        • Bayer H.M.
        • Glimcher P.W.
        Midbrain dopamine neurons encode a quantitative reward prediction error signal.
        Neuron. 2005; 47: 129-141
        • Rolls E.T.
        • Huang C.C.
        • Lin C.P.
        • Feng J.
        • Joliot M.
        Automated anatomical labelling atlas 3.
        Neuroimage. 2020; 206: 116189
        • Friston K.J.
        • Harrison L.
        • Penny W.
        Dynamic causal modelling.
        Neuroimage. 2003; 19: 1273-1302
        • Zeidman P.
        • Jafarian A.
        • Corbin N.
        • Seghier M.L.
        • Razi A.
        • Price C.J.
        • Friston K.J.
        A guide to group effective connectivity analysis, part 1: First level analysis with DCM for fMRI.
        Neuroimage. 2019; 200: 174-190
        • Hauser T.U.
        • Iannaccone R.
        • Walitza S.
        • Brandeis D.
        • Brem S.
        Cognitive flexibility in adolescence: Neural and behavioral mechanisms of reward prediction error processing in adaptive decision making during development.
        Neuroimage. 2015; 104: 347-354
        • Friston K.J.
        • Litvak V.
        • Oswal A.
        • Razi A.
        • Stephan K.E.
        • Van Wijk B.C.M.
        • et al.
        Bayesian model reduction and empirical Bayes for group (DCM) studies.
        Neuroimage. 2016; 128: 413-431
        • Behrens T.E.
        • Woolrich M.W.
        • Walton M.E.
        • Rushworth M.F.
        Learning the value of information in an uncertain world.
        Nat Neurosci. 2007; 10: 1214-1221
        • Bishop S.J.
        • Gagne C.
        Anxiety, depression, and decision making: A computational perspective.
        Annu Rev Neurosci. 2018; 41: 371-388
        • Häberling I.
        • Baumgartner N.
        • Emery S.
        • Keller P.
        • Strumberger M.
        • Nalani K.
        • et al.
        Anxious depression as a clinically relevant subtype of pediatric major depressive disorder.
        J Neural Transm (Vienna). 2019; 126: 1217-1230
        • Robbins T.W.
        • Arnsten A.F.
        The neuropsychopharmacology of fronto-executive function: Monoaminergic modulation.
        Annu Rev Neurosci. 2009; 32: 267-287
        • Kahnt T.
        • Tobler P.N.
        Dopamine modulates the functional organization of the orbitofrontal cortex.
        J Neurosci. 2017; 37: 1493-1504
        • Ramnani N.
        • Elliott R.
        • Athwal B.S.
        • Passingham R.E.
        Prediction error for free monetary reward in the human prefrontal cortex.
        Neuroimage. 2004; 23: 777-786
        • Sul J.H.
        • Kim H.
        • Huh N.
        • Lee D.
        • Jung M.W.
        Distinct roles of rodent orbitofrontal and medial prefrontal cortex in decision making.
        Neuron. 2010; 66: 449-460
        • Wilson R.C.
        • Takahashi Y.K.
        • Schoenbaum G.
        • Niv Y.
        Orbitofrontal cortex as a cognitive map of task space.
        Neuron. 2014; 81: 267-279
        • Gao Q.
        • Zou K.
        • He Z.
        • Sun X.
        • Chen H.
        Causal connectivity alterations of cortical-subcortical circuit anchored on reduced hemodynamic response brain regions in first-episode drug-naïve major depressive disorder.
        Sci Rep. 2016; 6: 21861
        • Schiller C.E.
        • Minkel J.
        • Smoski M.J.
        • Dichter G.S.
        Remitted major depression is characterized by reduced prefrontal cortex reactivity to reward loss.
        J Affect Disord. 2013; 151: 756-762
        • Ubl B.
        • Kuehner C.
        • Kirsch P.
        • Ruttorf M.
        • Diener C.
        • Flor H.
        Altered neural reward and loss processing and prediction error signalling in depression.
        Soc Cogn Affect Neurosci. 2015; 10: 1102-1112
        • Pizzagalli D.A.
        • Holmes A.J.
        • Dillon D.G.
        • Goetz E.L.
        • Birk J.L.
        • Bogdan R.
        • et al.
        Reduced caudate and nucleus accumbens response to rewards in unmedicated individuals with major depressive disorder.
        Am J Psychiatry. 2009; 166: 702-710
        • Lawson R.P.
        • Nord C.L.
        • Seymour B.
        • Thomas D.L.
        • Dayan P.
        • Pilling S.
        • Roiser J.P.
        Disrupted habenula function in major depression.
        Mol Psychiatry. 2017; 22: 202-208
        • Clark J.E.
        • Watson S.
        • Friston K.J.
        What is mood? A computational perspective.
        Psychol Med. 2018; 48: 2277-2284
        • Steele J.D.
        • Kumar P.
        • Ebmeier K.P.
        Blunted response to feedback information in depressive illness.
        Brain. 2007; 130: 2367-2374
        • Nussenbaum K.
        • Hartley C.A.
        Reinforcement learning across development: What insights can we draw from a decade of research?.
        Dev Cogn Neurosci. 2019; 40: 100733
        • Snyder H.R.
        Major depressive disorder is associated with broad impairments on neuropsychological measures of executive function: A meta-analysis and review.
        Psychol Bull. 2013; 139: 81-132
        • Rutledge R.B.
        • Moutoussis M.
        • Smittenaar P.
        • Zeidman P.
        • Taylor T.
        • Hrynkiewicz L.
        • et al.
        Association of neural and emotional impacts of reward prediction errors with major depression.
        JAMA Psychiatry. 2017; 74: 790-797
        • McCabe C.
        • Mishor Z.
        • Cowen P.J.
        • Harmer C.J.
        Diminished neural processing of aversive and rewarding stimuli during selective serotonin reuptake inhibitor treatment.
        Biol Psychiatry. 2010; 67: 439-445
        • Stephan K.E.
        • Mathys C.
        Computational approaches to psychiatry.
        Curr Opin Neurobiol. 2014; 25: 85-92