Advertisement

Differential Impact of Anxious Misery Psychopathology on Multiple Representations of the Functional Connectome

  • Darsol Seok
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Joanne Beer
    Affiliations
    Penn Statistics in Imaging and Visualization Center, Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Marc Jaskir
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Nathan Smyk
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Adna Jaganjac
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Walid Makhoul
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Philip Cook
    Affiliations
    Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Mark Elliott
    Affiliations
    Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Russell Shinohara
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania

    Penn Statistics in Imaging and Visualization Center, Department of Biostatistics, Epidemiology, and Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania

    Center for Biomedical Image Computing and Analytics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
  • Yvette I. Sheline
    Correspondence
    Address correspondence to Yvette I. Sheline, M.D., M.S.
    Affiliations
    Center for Neuromodulation in Depression and Stress, Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania

    Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania

    Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania
    Search for articles by this author
Open AccessPublished:November 17, 2021DOI:https://doi.org/10.1016/j.bpsgos.2021.11.004

      Abstract

      Background

      One aim of characterizing dimensional psychopathology is associating different domains of affective dysfunction with brain circuitry. The functional connectome, as measured by functional magnetic resonance imaging, can be modeled and associated with psychopathology through multiple methods; some methods assess univariate relationships while others summarize broad patterns of activity. It remains unclear whether different dimensions of psychopathology require different representations of the connectome to generate reproducible associations.

      Methods

      Patients experiencing anxious misery symptomology (depression, anxiety, and trauma; n = 192) received resting-state functional magnetic resonance imaging scans. Three modeling approaches (seed-based correlation analysis, edgewise regression, and brain basis set modeling), each relying on increasingly broader representations of the functional connectome, were used to associate connectivity patterns with six data-driven dimensions of psychopathology: anxiety sensitivity, anxious arousal, rumination, anhedonia, insomnia, and negative affect. To protect against overfitting, 50 participants were held out in a testing dataset, leaving 142 participants as training data.

      Results

      Different modeling approaches varied in the extent to which they could model different symptom dimensions: seed-based correlation analysis failed to reproducibly model any symptoms, subsets of the connectome (edgewise regression) were sufficient to model insomnia and anxious arousal, and broad representations of the entire connectome (brain basis set modeling) were necessary to model negative affect and ruminative thought.

      Conclusions

      These results indicate that different methods of representing the functional connectome differ in the degree that they can model different symptom dimensions, highlighting the potential sufficiency of subsets of connections for some dimensions and the necessity of connectome-wide approaches in others.

      Keywords

      The DSM conceptualizes mental illnesses as discrete disorder classes. Notable issues with the DSM, such as heterogeneity within disorders and a lack of understanding of underlying mechanisms, have motivated the development of alternative frameworks such as the Research Domain Criteria (RDoC) (
      • Insel T.
      • Cuthbert B.
      • Garvey M.
      • Heinssen R.
      • Pine D.S.
      • Quinn K.
      • et al.
      Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders.
      ). Whereas the DSM groups symptoms together based on common co-occurrence, RDoC delineates these symptoms onto distinct dimensions of behavioral and cognitive (dys)function, with the goal of mapping these dimensions onto biological mechanisms. This study applies this dimensional approach to a class of disorders termed disorders of anxious misery (AM) (
      • Krueger R.F.
      The structure of common mental disorders.
      ,
      • Watson D.
      Rethinking the mood and anxiety disorders: A quantitative hierarchical model for DSM-V.
      ), which includes DSM diagnoses of generalized anxiety disorder, major depressive disorder, persistent depressive disorder, and posttraumatic stress disorder. Combined, these disorders affect more than 800 million people worldwide and constitute the leading cause of disability (
      • Murray C.J.L.
      • Atkinson C.
      • Bhalla K.
      • Birbeck G.
      • Burstein R.
      • Chou D.
      • et al.
      The state of US health, 1990–2010: Burden of diseases, injuries, and risk factors.
      ). Patients with these disorders exhibit dysfunction along various dimensions of neurocognitive functioning (constructs in the RDoC), such as anxiety, frustrative nonreward, and sleep-wakefulness.
      A key element for investigating these dimensions is the identification of their underlying neural circuits. Resting-state networks (
      • Biswal B.
      • Yetkin F.Z.
      • Haughton V.M.
      • Hyde J.S.
      Functional connectivity in the motor cortex of resting human brain using echo-planar MRI.
      ,
      • Fox M.D.
      • Raichle M.E.
      Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging.
      ) can be probed using noninvasive brain imaging to visualize the functional coactivation of specific brain regions, together forming the functional connectome (
      • Van Dijk K.R.A.
      • Hedden T.
      • Venkataraman A.
      • Evans K.C.
      • Lazar S.W.
      • Buckner R.L.
      Intrinsic functional connectivity as a tool for human connectomics: Theory, properties, and optimization.
      ). A key challenge in associating psychopathology with the functional connectome is the sheer size of connectomic data: a typical imaging sequence may measure signal from approximately 200,000 voxels, resulting in approximately 109 potential components of the connectome (
      • Glasser M.F.
      • Sotiropoulos S.N.
      • Wilson J.A.
      • Coalson T.S.
      • Fischl B.
      • Andersson J.L.
      • et al.
      The minimal preprocessing pipelines for the Human Connectome Project.
      ). Given the potentially massive size of these data, many modeling approaches only use a subset of the total information available within the connectome. At one extreme are univariate seed-based correlation analysis (SCA) approaches, which examine associations between behavioral variables and connectivity between just two regions, the selection of which is often determined a priori. In contrast, data-driven models such as edgewise support vector regression (
      • Davatzikos C.
      Machine learning in neuroimaging: Progress and challenges.
      ) typically rely on a small subset of functional connections within the connectome to generate multivariate symptom-brain associations. Finally, recent methods such as brain basis set modeling (BBS) (
      • Sripada C.
      • Angstadt M.
      • Rutherford S.
      • Kessler D.
      • Kim Y.
      • Yee M.
      • Levina E.
      Basic units of inter-individual variation in resting state connectomes.
      ) capitalize on connectivity patterns that can span the entire brain, using dimensionality reduction approaches such as principal component analysis (PCA) to efficiently summarize large-scale patterns. It remains an open question whether different modeling approaches are better suited for different dimensions of AM symptomology; dysfunction in a single circuit may be sufficient to model some dimensions, while broader, global patterns of connectivity may be necessary to model other dimensions.
      An important consideration in assessing the generalizability of symptom-brain associations is the ability to replicate performance in a held-out testing dataset (
      • Schnack H.G.
      • Kahn R.S.
      Detecting neuroimaging biomarkers for psychiatric disorders: Sample size matters.
      ). Systematic examinations of brain-behavior associations have revealed that samples far larger than those typically used in neuroimaging studies of psychopathology may be required to generate stable associations (
      • Marek S.
      • Tervo-Clemmens B.
      • Calabro F.J.
      • Montez D.F.
      • Kay B.P.
      • Hatoum A.S.
      • et al.
      Reproducible brain-wide association studies require thousands of individuals [published correction appears in Nature 2022; 605(7911):E11].
      ). Furthermore, notable examples of failed replications (
      • Müller V.I.
      • Cieslik E.C.
      • Serbanescu I.
      • Laird A.R.
      • Fox P.T.
      • Eickhoff S.B.
      Altered brain activity in unipolar depression revisited: Meta-analyses of neuroimaging studies.
      ,
      • Dinga R.
      • Schmaal L.
      • Penninx B.W.J.H.
      • van Tol M.J.
      • Veltman D.J.
      • van Velzen L.
      • et al.
      Evaluating the evidence for biotypes of depression: Methodological replication and extension of Drysdale et al..
      ) have revealed the perils of fitting multivariate models, particularly when models rely on a small subset of functional connections selected through data-driven methods. Many of these failures in replication stemmed from findings that incorrectly applied cross-validation, which requires complete encapsulation of information between folds to generate a reasonable estimate of out-of-sample performance (
      • Poldrack R.A.
      • Huckins G.
      • Varoquaux G.
      Establishment of best practices for evidence for prediction: A review.
      ). Direct validation of model performance in a held-out sample is therefore an important step to assessing the potential generalizability of any symptom-brain association.
      To this end, this work examined transdiagnostic associations between six data-driven dimensions of AM symptomology (anxiety sensitivity, anxious arousal, ruminative thought, anhedonia, insomnia, and negative affect) and functional connectivity using three different modeling approaches. The data were randomly divided into a training set and a held-out testing set to validate model performance.

      Methods and Materials

      Sample Characteristics

      For a complete description of sample characteristics, screening procedures, and participant assessment, readers are directed to the reference publication for this dataset (
      • Seok D.
      • Smyk N.
      • Jaskir M.
      • Cook P.
      • Elliott M.
      • Girelli T.
      • et al.
      Dimensional connectomics of anxious misery, a human connectome study related to human disease: Overview of protocol and data quality.
      ).
      Participants experiencing symptoms of AM (n = 194) were recruited through the community. Healthy comparators (HCs) (n = 48) were also recruited; only one modeling approach (BBS) used data from HC participants. Demographic information is summarized in Table 1, with additional demographic information in Table S1. Two AM participants were removed from analyses due to excessive in-scanner motion, so we present demographics for 192 AM and 48 HC participants.
      Table 1Participant Characteristics
      CharacteristicsTotal, N = 240Control, n = 48Anxious Misery, n = 192
      Sex
       Female168 (70%)34134
       Male72 (30%)1458
      Age, Years
      Median = 26 years.
       18–2368 (28.3%)1157
       24–2991 (37.9%)2368
       30–3542 (17.5%)933
       36–4130 (12.5%)426
       42–473 (1.3%)12
       48–532 (0.8%)02
       54–594 (1.7%)04
      Race
       Asian25 (10.4%)619
       Black45 (18.8%)936
       Multiracial13 (5.4%)211
       Other5 (2.1%)23
       Undisclosed8 (3.3%)08
       White144 (60.0%)29115
      Ethnicity
       Hispanic17 (7.1%)314
       Not Hispanic219 (91.2%)45174
       Undisclosed4 (1.7%)04
      Medication Status
       Medicated51 (26.5%)
       Unmedicated141 (73.4%)
      Primary DSM-5 Diagnosis
       Major depressive disorder91 (47.4%)
       Generalized anxiety disorder42 (21.9%)
       Posttraumatic stress disorder39 (20.3%)
       Persistent depressive disorder13 (6.77%)
       Social anxiety disorder6 (3.12%)
       Cyclothymic disorder1 (0.52%)
      Values are presented as n or n (%). Additional demographic information can be found in Table S1. Note that control participants were only used in the brain basis set modeling approach.
      a Median = 26 years.
      To ensure a transdiagnostic dataset within the AM category, participants were deemed eligible if their neuroticism score on the Neuroticism-Extraversion-Openness Five-Factor Inventory was greater than one standard deviation above the population mean (≥26.2 for males, ≥30.1 for females) (
      • McCrae R.R.
      • Costa Jr., P.T.
      Brief versions of the NEO-PI-3.
      ). Neuroticism was selected as an eligibility criterion because it captures general elements of psychopathology that are shared by participants diagnosed with depression, anxiety, and trauma-related disorders (
      • Andrews G.
      • Stewart G.
      • Morris-Yates A.
      • Holt P.
      • Henderson S.
      Evidence for a general neurotic syndrome.
      ,
      • Khan A.A.
      • Jacobson K.C.
      • Gardner C.O.
      • Prescott C.A.
      • Kendler K.S.
      Personality and comorbidity of common psychiatric disorders.
      ).
      We did not require participants to cease taking any psychotropic medications. The majority of our study participants (73.4%) were unmedicated. All participants provided informed consent, and all study procedures were conducted under the approval of the University of Pennsylvania Institutional Review Board.

      Participant Assessment

      To assess participants’ symptom profiles, seven clinician-administered and self-report scales measuring a broad range of depressive and anxious symptomology were administered to AM participants [see full list of assessments in Supplemental Methods and Materials and (
      • Seok D.
      • Smyk N.
      • Jaskir M.
      • Cook P.
      • Elliott M.
      • Girelli T.
      • et al.
      Dimensional connectomics of anxious misery, a human connectome study related to human disease: Overview of protocol and data quality.
      )].

      Hierarchical Clustering of Symptoms

      To derive data-driven symptom communities from our assessment data, we applied a hierarchical clustering algorithm (
      • Jeub L.G.S.
      • Sporns O.
      • Fortunato S.
      Multiresolution consensus clustering in networks.
      ) on the Pearson correlation matrix of the 113 constituent items of our seven instruments, using only data from the AM group. After selecting an appropriate resolution (see Supplemental Methods and Materials for details), we derived symptom dimension scores for each participant by computing the mean across items belonging to each symptom community after z-scoring.

      Imaging Acquisition

      Participants received T1-weighted imaging and four resting-state functional magnetic resonance imaging scans (total acquisition time = 23:01) after their symptom assessments. A summary of imaging parameters and image acquisition protocols are provided in Supplemental Methods and Materials, with full details provided elsewhere (
      • Seok D.
      • Smyk N.
      • Jaskir M.
      • Cook P.
      • Elliott M.
      • Girelli T.
      • et al.
      Dimensional connectomics of anxious misery, a human connectome study related to human disease: Overview of protocol and data quality.
      ).

      Image Preprocessing

      Preprocessing details are provided in Supplemental Methods and Materials and the reference publication for this dataset (
      • Seok D.
      • Smyk N.
      • Jaskir M.
      • Cook P.
      • Elliott M.
      • Girelli T.
      • et al.
      Dimensional connectomics of anxious misery, a human connectome study related to human disease: Overview of protocol and data quality.
      ). Two participants, both of whom were AM participants (0.83% of all participants), were removed from resting-state analyses due to excessive in-scanner movement (root-mean-squared movement > median (root-mean-squared movement) + 3 × interquartile range).

      Overview of Modeling Approaches

      We tested the ability of three different modeling approaches, each relying on increasingly dense representations of functional connectomics, to generate robust, generalizable associations between imaging data and symptom dimensions. An overview of each of these modeling approaches is provided (Figure 1), with complete details provided in Supplemental Methods and Materials. Implementations for modeling approaches are available at https://github.com/dseok/anxious_misery_connectomes.
      Figure thumbnail gr1
      Figure 1Overview of modeling approaches. All three modeling approaches (seed-based correlation analysis [SCA], edgewise regression, and brain basis set modeling [BBS]) undergo four common steps: 1) preprocessing, wherein the brain is parcellated or masked; 2) feature engineering and selection, wherein specific features are identified or generated; 3) extraction of individual values; and 4) modeling, wherein features are associated with symptom severity. PC, principal component.

      Modeling Approach 1: SCA

      SCA associates behavioral variables (in our case, symptom dimension scores) with connectivity between a seed region and another region. Each symptom dimension was assigned specific seeds and masks based on extant literature (Table 2). Exploratory whole-brain analyses (wherein significant clusters outside of a priori masks are permitted) were also conducted.
      Table 2Seed-Based Correlation Analysis Results
      SymptomReferencesSeedMaskxyzSignVoxels, nr, Trainingr, Testing
      Anxiety Sensitivity(
      • Laeger I.
      • Dobel C.
      • Dannlowski U.
      • Kugel H.
      • Grotegerd D.
      • Kissler J.
      • et al.
      Amygdala responsiveness to emotional words is modulated by subclinical anxiety and depression [published correction appears in Behav Brain Res 2014; 261:369–370].
      ,
      • Prater K.E.
      • Hosanagar A.
      • Klumpp H.
      • Angstadt M.
      • Phan K.L.
      Aberrant amygdala-frontal cortex connectivity during perception of fearful faces and at rest in generalized social anxiety disorder.
      ,
      • Liu W.J.
      • Yin D.Z.
      • Cheng W.H.
      • Fan M.X.
      • You M.N.
      • Men W.W.
      • et al.
      Abnormal functional connectivity of the amygdala-based network in resting-state FMRI in adolescents with generalized anxiety disorder.
      ,
      • Dong M.
      • Xia L.
      • Lu M.
      • Li C.
      • Xu K.
      • Zhang L.
      A failed top-down control from the prefrontal cortex to the amygdala in generalized anxiety disorder: Evidence from resting-state fMRI with Granger causality analysis.
      ,
      • Warren S.L.
      • Zhang Y.
      • Duberg K.
      • Mistry P.
      • Cai W.
      • Qin S.
      • et al.
      Anxiety and stress alter decision-making dynamics and causal amygdala-dorsolateral prefrontal cortex circuits during emotion regulation in children.
      )
      L amygdalaWhole brain
      L dlPFC−393331+740.3550.337
      Highest performing connections in the testing set for each symptom.
      R dlPFC423340+80.3350.069
      R amygdalaWhole brain
      L dlPFC−412935+310.3590.241
      R dlPFC
      Anxious Arousal(
      • Laeger I.
      • Dobel C.
      • Dannlowski U.
      • Kugel H.
      • Grotegerd D.
      • Kissler J.
      • et al.
      Amygdala responsiveness to emotional words is modulated by subclinical anxiety and depression [published correction appears in Behav Brain Res 2014; 261:369–370].
      ,
      • Prater K.E.
      • Hosanagar A.
      • Klumpp H.
      • Angstadt M.
      • Phan K.L.
      Aberrant amygdala-frontal cortex connectivity during perception of fearful faces and at rest in generalized social anxiety disorder.
      ,
      • Liu W.J.
      • Yin D.Z.
      • Cheng W.H.
      • Fan M.X.
      • You M.N.
      • Men W.W.
      • et al.
      Abnormal functional connectivity of the amygdala-based network in resting-state FMRI in adolescents with generalized anxiety disorder.
      ,
      • Dong M.
      • Xia L.
      • Lu M.
      • Li C.
      • Xu K.
      • Zhang L.
      A failed top-down control from the prefrontal cortex to the amygdala in generalized anxiety disorder: Evidence from resting-state fMRI with Granger causality analysis.
      ,
      • Warren S.L.
      • Zhang Y.
      • Duberg K.
      • Mistry P.
      • Cai W.
      • Qin S.
      • et al.
      Anxiety and stress alter decision-making dynamics and causal amygdala-dorsolateral prefrontal cortex circuits during emotion regulation in children.
      )
      L amygdalaWhole brain−50−244870.3760.007
      L dlPFC
      R dlPFC
      R amygdalaWhole brain
      L dlPFC
      R dlPFC302440180.3260.016
      Highest performing connections in the testing set for each symptom.
      Ruminative Thought(
      • Berman M.G.
      • Peltier S.
      • Nee D.E.
      • Kross E.
      • Deldin P.J.
      • Jonides J.
      Depression, rumination and the default network.
      ,
      • Hamilton J.P.
      • Furman D.J.
      • Chang C.
      • Thomason M.E.
      • Dennis E.
      • Gotlib I.H.
      Default-mode and task-positive network activity in major depressive disorder: Implications for adaptive and maladaptive rumination.
      ,
      • Berman M.G.
      • Misic B.
      • Buschkuehl M.
      • Kross E.
      • Deldin P.J.
      • Peltier S.
      • et al.
      Does resting-state connectivity reflect depressive rumination? A tale of two analyses.
      ,
      • Zhou H.X.
      • Chen X.
      • Shen Y.Q.
      • Li L.
      • Chen N.X.
      • Zhu Z.C.
      • et al.
      Rumination and the default mode network: Meta-analysis of brain imaging studies and implications for depression.
      )
      PCCWhole brain
      Precuneus10−7250+270.3450.135
      Highest performing connections in the testing set for each symptom.
      mPFC
      Anhedonia(
      • Keller J.
      • Young C.B.
      • Kelley E.
      • Prater K.
      • Levitin D.J.
      • Menon V.
      Trait anhedonia is associated with reduced reactivity and connectivity of mesolimbic and paralimbic reward pathways.
      ,
      • Satterthwaite T.D.
      • Kable J.W.
      • Vandekar L.
      • Katchmar N.
      • Bassett D.S.
      • Baldassano C.F.
      • et al.
      Common and dissociable dysfunction of the reward system in bipolar and unipolar depression.
      ,
      • Felger J.C.
      • Li Z.
      • Haroon E.
      • Woolwine B.J.
      • Jung M.Y.
      • Hu X.
      • Miller A.H.
      Inflammation is associated with decreased functional connectivity within corticostriatal reward circuitry in depression.
      ,
      • Ferenczi E.A.
      • Zalocusky K.A.
      • Liston C.
      • Grosenick L.
      • Warden M.R.
      • Amatya D.
      • et al.
      Prefrontal cortical regulation of brainwide circuit dynamics and reward-related behavior.
      ,
      • Rzepa E.
      • McCabe C.
      Anhedonia and depression severity dissociated by dmPFC resting-state functional connectivity in adolescents.
      )
      L iVSWhole brain
      L OFC
      R OFC
      R iVSWhole brain
      L OFC−1638−16190.3450.276
      Highest performing connections in the testing set for each symptom.
      R OFC
      Negative Affect(
      • Greicius M.D.
      • Flores B.H.
      • Menon V.
      • Glover G.H.
      • Solvason H.B.
      • Kenna H.
      • et al.
      Resting-state functional connectivity in major depression: Abnormally increased contributions from subgenual cingulate cortex and thalamus.
      ,
      • Sheline Y.I.
      • Price J.L.
      • Yan Z.
      • Mintun M.A.
      Resting-state functional MRI in depression unmasks increased connectivity between networks via the dorsal nexus.
      ,
      • Davey C.G.
      • Harrison B.J.
      • Yücel M.
      • Allen N.B.
      Regionally specific alterations in functional connectivity of the anterior cingulate cortex in major depressive disorder.
      ,
      • Koenigs M.
      • Grafman J.
      The functional neuroanatomy of depression: Distinct roles for ventromedial and dorsolateral prefrontal cortex.
      ,
      • Fox M.D.
      • Buckner R.L.
      • White M.P.
      • Greicius M.D.
      • Pascual-Leone A.
      Efficacy of transcranial magnetic stimulation targets for depression is related to intrinsic functional connectivity with the subgenual cingulate.
      ,
      • Baeken C.
      • Marinazzo D.
      • Wu G.R.
      • Van Schuerbeek P.
      • De Mey J.
      • Marchetti I.
      • et al.
      Accelerated HF-rTMS in treatment-resistant unipolar depression: Insights from subgenual anterior cingulate functional connectivity.
      ,
      • Avery J.A.
      • Drevets W.C.
      • Moseman S.E.
      • Bodurka J.
      • Barcalow J.C.
      • Simmons W.K.
      Major depressive disorder is associated with abnormal interoceptive activity and functional connectivity in the insula.
      ,
      • Shao R.
      • Lau W.K.W.
      • Leung M.K.
      • Lee T.M.C.
      Subgenual anterior cingulate-insula resting-state connectivity as a neural correlate to trait and state stress resilience.
      )
      sgACCWhole brain
      Precuneus
      mPFC
      PCC
      Insomnia(
      • Chuah Y.M.L.
      • Venkatraman V.
      • Dinges D.F.
      • Chee M.W.L.
      The neural basis of interindividual variability in inhibitory efficiency after sleep deprivation.
      ,
      • Koenigs M.
      • Holliday J.
      • Solomon J.
      • Grafman J.
      Left dorsomedial frontal brain damage is associated with insomnia.
      ,
      • Zhang Q.
      • Qin W.
      • He X.
      • Li Q.
      • Chen B.
      • Zhang Y.
      • Yu C.
      Functional disconnection of the right anterior insula in obstructive sleep apnea.
      ,
      • Park B.
      • Palomares J.A.
      • Woo M.A.
      • Kang D.W.
      • Macey P.M.
      • Yan-Go F.L.
      • et al.
      Aberrant insular functional network integrity in patients with obstructive sleep apnea.
      ,
      • Yu S.
      • Guo B.
      • Shen Z.
      • Wang Z.
      • Kui Y.
      • Hu Y.
      • Feng F.
      The imbalanced anterior and posterior default mode network in the primary insomnia.
      )
      L insulaWhole brain
      Precuneus
      mPFC11695+1110.3840.272
      Highest performing connections in the testing set for each symptom.
      R insulaWhole brain
      Precuneus
      mPFC11703+370.3530.245
      Unique seeds and masks for each symptom cluster (in addition to a whole-brain analysis) were assigned based on findings from extant literature (“Reference”). Whole-brain analyses indicate exploratory analyses wherein significant clusters were permitted anywhere in gray matter. Significant clusters were identified using a nonparametric permutation testing technique incorporating threshold-free cluster enhancement (implemented in FSL’s Randomise; FWER < 0.05). Coordinates for cluster center of mass are in MNI space. In the Sign column, the sign indicates the direction of association: “+” indicates that hyperconnectivity was associated with worse symptoms, while “−” indicates that hypoconnectivity was associated with worse symptoms. r (training) and r (testing) indicate Pearson correlations between observed and predicted symptom scores using a linear model using connectivity z scores.
      dlPFC, dorsolateral prefrontal cortex; FWER, familywise error rate; iVS, inferior ventral striatum; L, left; MNI, Montreal Neurological Institute; mPFC, medial prefrontal cortex; OFC, orbitofrontal cortex; PCC, posterior cingulate cortex; R, right; sgACC, subgenual anterior cingulate cortex.
      a Highest performing connections in the testing set for each symptom.
      Clusters that exhibited significant associations based on permutation testing (p < .05) are reported. Given that SCA was merely used to identify clusters for further validation in our testing set, this p-value threshold was not adjusted for multiple tests; Bonferroni correction of this p-value threshold resulted in no significant clusters being identified in the training set. After clusters were identified, linear models predicting symptom severity using connectivity were constructed using mean connectivity between the seed region and the identified cluster; significance of testing performances was corrected (Holm-Bonferroni) separately for each symptom cluster.

      Multivariate Methods

      The next two modeling approaches used the Power264 atlas, a well-validated parcellation that identifies 264 functional nodes of the brain (
      • Power J.D.
      • Cohen A.L.
      • Nelson S.M.
      • Wig G.S.
      • Barnes K.A.
      • Church J.A.
      • et al.
      Functional network organization of the human brain.
      ). In addition to these 264 nodes, we supplemented 14 additional nodes relevant to AM disorders, adopted from the additional nodes used in (
      • Drysdale A.T.
      • Grosenick L.
      • Downar J.
      • Dunlop K.
      • Mansouri F.
      • Meng Y.
      • et al.
      Resting-state connectivity biomarkers define neurophysiological subtypes of depression [published correction appears in Nat Med 2017; 23:264].
      ) (full table of additional seeds in Table S3).
      For both of the following methods, we extracted the mean time series of each node using spherical regions of interest (5-mm radius) and then computed the Pearson correlation between each time series to generate a 278 × 278 functional connectivity matrix for each scan (38,503 unique connections). Because participants completed four scans, the four resulting matrices were averaged to generate one matrix per participant.

      Modeling Approach 2: Edgewise Regression

      In this modeling approach, a subset of individual connection strengths (edges) was first selected using a cross-validated feature selection procedure and then associated with symptom severity using support vector regression (linear kernel). Our feature selection procedure indicated that edgewise regression models performed best with a small subset of connectivity features (approximately 5% of the full matrix).

      Modeling Approach 3: BBS

      BBS (
      • Sripada C.
      • Angstadt M.
      • Rutherford S.
      • Kessler D.
      • Kim Y.
      • Yee M.
      • Levina E.
      Basic units of inter-individual variation in resting state connectomes.
      ) uses PCA to generate a low-dimensional representation of interindividual differences in global connectivity patterns. After PCA, support vector regression (linear kernel) was used to associate each symptom dimension with connectivity component scores. To capture directions of connectivity variance that extend into healthy ranges, BBS uses data from HC participants; this was the only modeling approach that used our sample of HCs. PCA was only ever conducted within the training set to prevent data leakage from testing participant data.

      Modeling Assessment Framework

      All models underwent the following assessment framework (Figure 2) designed to test the performance of models on new, unseen data. First, age, sex, and motion (root-mean-squared) were regressed from connectivity matrices, and age and sex were regressed from symptom scales (see Figures S4 and S5 for correlations with motion and temporal signal-to-noise ratio). No significant associations between symptom scales and motion or temporal signal-to-noise ratio were detected. Second, while matching for age, sex, and symptom severity (see Supplemental Methods and Materials for details), a random subset of 50 AM participants were segregated to a testing set, leaving 142 AM participants for our training set. Next, within the training set, fivefold cross-validation was used to perform feature selection and hyperparameter tuning for the multivariate methods. Connections were selected for each symptom dimension by computing the Spearman correlation between each connection and the symptom dimension in question and then selecting the p connections with the highest absolute correlations, where p is a hyperparameter tuned through nested cross-validation.
      Figure thumbnail gr2
      Figure 2Flowchart of model assessment framework. Two participants were removed for excessive in-scanner motion during quality control (QC) procedures. After data are split into training and testing partitions, hyperparameter tuning in training data is accomplished using nested fivefold cross-validation (CV). The best-performing hyperparameters are used in a final model that is fit on the full training dataset. Finally, model performance is assessed using the testing data, and significance is determined using permutation testing. This process was repeated for each symptom dimension, each time using the same training/testing partition.
      Finally, model performance was assessed in the held-out testing set by computing the Pearson correlation between the predicted and actual symptom scores, using a model fit on the entire training set. Significance was determined using a permutation testing procedure wherein null distributions of correlations were generated by shuffling the rows of imaging data relative to symptom data in the training set, performing the entire model-fitting process using this permuted dataset and using this model to compute the correlation between predicted and actual symptom scores in the testing dataset, repeating this entire process 5000 times for each symptom and modeling approach. Given that each symptom dimension is being modeled using three different approaches, final model performances for each symptom were adjusted for multiple testing using separate Holm-Bonferroni corrections.

      Results

      Symptom Clustering

      Inspection of the symptom hierarchy revealed that items clustered according to clinically relevant dimensions (see Figure S1 for complete hierarchy). Symptoms first separated into two large communities, one with predominantly depressive symptoms and the other with largely anxiety-related symptoms. At finer resolutions, these two larger communities separated into three subcommunities each, resulting in six total communities (Figure 3). Anxiety-related symptoms divided into the following communities: anxiety sensitivity (representative item: “It scares me when my heart beats rapidly.”), anxious arousal (“Hands were shaky”), and ruminative thought (“I can’t stop thinking about some things.”). Depression-related symptoms divided into the following communities: anhedonia (“I would not find pleasure in my hobbies and pastimes.”), insomnia (“Difficulty falling asleep”), and negative affect (“Pessimistic thoughts”). See Table S4 for a complete list of symptom clusters and their constitutive items. Ultimately, this resolution was chosen for analysis because it 1) allowed for interrogation of symptom dimensions finer than those aligned with traditional diagnostic categories (i.e., anxiety and depressive disorders) and 2) resulted in communities that were broadly aligned with constructs of the RDoC, unlike finer resolutions that split groups of symptoms into clusters too narrow to be meaningfully interpreted. See Figure S3 for correlations between symptom clusters. As expected (
      • Krueger R.F.
      The structure of common mental disorders.
      ), all symptoms exhibited moderate, positive correlations with each other.
      Figure thumbnail gr3
      Figure 3Clustering revealed six communities of symptoms. Each circle represents an item on one of seven patient assessments of psychopathology. Indicated in each circle is the assessment to which the item belongs. Connections and proximity of circles indicate the strength of correlation between items. ASI, Anxiety Sensitivity Index; HAMD, Hamilton Depression Rating Scale; ISI, Insomnia Severity Index; MADRS, Montgomery–Åsberg Depression Rating Scale; MASQ, Anxiety Depression Distress Inventory-27; RTS, Ruminative Thought Style Questionnaire; SHAPS, Snaith-Hamilton Pleasure Scale.

      Training and Testing Partitions

      Training (n = 142) and testing (n = 50) partitions were well matched for age, sex, and z-scored symptom scores (Figure S2). Two-sample t tests and χ2 tests did not indicate significant differences between data partitions for all categories.

      SCA Results

      The complete list of significant clusters identified using SCA is detailed in Table 2. Only one significant cluster was identified using an exploratory whole-brain search, and association with this cluster did not replicate in the testing dataset (anxious arousal, left amygdala ↔ left primary motor cortex, rtesting = 0.007, puncorrected = .962).
      With the exception of negative affect, all symptom dimensions identified at least one significantly associated cluster within the training set. However, replication within the testing data was mixed. Only clusters for anxiety sensitivity (left amygdala ↔ left dorsolateral prefrontal cortex, rtesting = 0.337, puncorrected = .017) and anhedonia (right inferior ventral striatum ↔ left orbitofrontal cortex, rtesting = 0.276, puncorrected = .049) generated meaningful correlations in the testing set. Other symptoms, such as anxious arousal (right amygdala ↔ right dorsolateral prefrontal cortex, rtesting = 0.016, puncorrected = .909), completely failed to replicate in the testing set.

      Multivariate Models: Hyperparameter Tuning in the Training Dataset

      Hyperparameter tuning was accomplished using nested cross-validation within the training dataset (Figure S6). This procedure indicated that edgewise regression models generally favored smaller numbers of connections (approximately 500). In contrast, BBS models favored a high number of connections (approximately 38,000). Because BBS models generally exhibited increasing performance with more features, the full connectivity matrix was used for all BBS models.

      BBS: Distribution of Principal Components

      Alignment of BBS connectivity component weights with intrinsic connectivity networks was examined (Figure S7). While the first principal components exhibited greater mean loading on a small subset of networks (
      • Sripada C.
      • Angstadt M.
      • Rutherford S.
      • Kessler D.
      • Kim Y.
      • Yee M.
      • Levina E.
      Basic units of inter-individual variation in resting state connectomes.
      ), all components exhibited significant loading on multiple networks, suggesting that the interindividual variability captured by BBS components did not merely recapitulate the structure of intrinsic connectivity networks.

      BBS: Inclusion of HCs

      BBS is the only modeling approach that used HCs. Examination of testing performance with and without HCs revealed that the inclusion of HCs significantly improved performance for negative affect and rumination and had little effect for other symptom dimensions (Figure S8); as such, HCs were included in all BBS models.

      Final Model Performance in the Testing Dataset

      Final model performance for all three modeling approaches is detailed in Table 3. SCA models failed to replicate for all symptom dimensions. Insomnia (rtesting = 0.344, pcorrected = .018) and anxious arousal (rtesting = 0.336, pcorrected = .027) were only significantly modeled using edgewise regression models. Negative affect (rtesting = 0.323, pcorrected = .010) and ruminative thought (rtesting = 0.313, pcorrected = .045) were only significantly modeled by BBS models. Schematic representations of these symptoms and their best performing modeling approaches are displayed in Figure 4. Scatterplots of predicted and observed severities indicated normal distributions and absence of significant outliers (Figure S9).
      Table 3Pearson’s Correlations Between Actual and Predicted Symptom Constructs in the Held-Out Testing Set Using Three Different Modeling Approaches
      SymptomSCAEdgewise RegressionBBS
      rtestingpuncorrected
      Uncorrected p values for SCA have been corrected (Holm-Bonferroni) for multiple clusters for each symptom (see Methods and Materials for more details).
      pConnections, nrtestingpuncorrectedpConnections, nrtestingpuncorrectedpConnections, n
      Anxiety Sensitivity0.337
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      .048.14410.225.062.1445000.052.376.37638,503
      Anhedonia0.276
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      .049.14710.185.097.1945000.095.264.26438,503
      Insomnia0.272.117.11710.344
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      ,
      Correlation was significant at p < .05.
      .006.01846000.246.041.08238,503
      Anxious Arousal0.0161.0001.00010.336
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      ,
      Correlation was significant at p < .05.
      .009.02712000.143.164.32838,503
      Negative Affect0.016.449.4491400.367
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      ,
      Correlation was significant at p < .05.
      .005.01038,503
      Ruminative Thought0.135.360.36010.167.136.27230000.310
      Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      ,
      Correlation was significant at p < .05.
      .015.04538,503
      Displayed are uncorrected and corrected (Holm–Bonferroni) permutation-based p values where rows of imaging data were shuffled relative to symptom data in the training set, and the resulting trained models were used to predict testing set symptom data (repeated 5000 times for each symptom and modeling approach). Also listed are the number of functional connections used for each model.
      BBS, brain basis set modeling; SCA, seed-based correlation analysis.
      a Uncorrected p values for SCA have been corrected (Holm-Bonferroni) for multiple clusters for each symptom (see Methods and Materials for more details).
      b Best-performing model approaches for each symptom construct. For SCA, the best-performing association for each symptom was selected.
      c Correlation was significant at p < .05.
      Figure thumbnail gr4
      Figure 4Schematic representations of each symptom and its best-performing modeling approach. Symptoms in the first row were modeled using edgewise regression and in the second row using brain basis set modeling (BBS). A connection’s color represents either hyperconnectivity (red) or hypoconnectivity (blue), and a connection’s width represents the magnitude of its associated model weight. Green and yellow nodes denote the a priori connections tested in seed-based correlation analysis models; green nodes represent seeds while yellow nodes represent resulting clusters. Other nodes belonging to key connections are labeled and colored cyan for easier identification: anxious arousal was associated with hypoconnectivity between left (L) frontopolar cortex (FPC) and right (R) supramarginal gyrus (SMG) and hyperconnectivity between R insula (Ins) and L amygdala (Amy), ruminative thought was associated with hyperconnectivity between primary motor cortex (M1) and R angular gyrus (Ang), and negative affect was associated with hyperconnectivity between R Amy and anterior cingulate cortex (ACC). For edgewise regression and BBS models, connections in the strongest 1% of model weights are plotted for display purposes (approximately 30 connections for edgewise regression, approximately 300 for BBS models). mPFC, medial prefrontal cortex; PCC, posterior cingulate gyrus; PreC, precuneus; sgACC, subgenual anterior cingulate cortex.
      While modeling approaches were fit independently, connections weighted heavily in simpler models were also prominent in more complex modeling approaches. For example, a priori connections tested in SCA generally exhibited strong weights in multivariate models: for insomnia, the left insula ↔ medial prefrontal cortex connection was in the top 4.04% of connections in its BBS model; for ruminative thought, the precuneus ↔ posterior cingulate cortex connection was in the top 10.05% of connections in its BBS model (and this connection performed comparatively worse in SCA, with rtesting = 0.135). Only anxious arousal did not exhibit strong correspondence between its SCA model (which had poor replication in the testing set: rtesting = 0.016) and its multivariate models: edgewise regression did not select the right amygdala ↔ right dorsolateral prefrontal cortex connection, and this connection was only in the top 34.01% of connection weights in BBS. Finally, connection weights from edgewise regression models highly correlated with the subset of overlapping connection weights in the equivalent BBS model (range across six symptoms: r = 0.852–0.949).
      Network-level representations reveal common and distinct patterns of connectivity abnormalities associated with each symptom dimension (Figure 5). Of the four symptoms best represented by multivariate models, all exhibited abnormal connectivity patterns in the default mode network (DMN). Anxious arousal was characterized by hyperconnectivity between bilateral regions of the DMN, as well as hyperconnectivity between 1) sensory regions such as the visual and auditory cortices and 2) networks in association cortex, such as the cinguloopercular network (CIN), frontoparietal network (FPN), and DMN. Insomnia was associated with hyperconnectivity between the 1) DMN and 2) visual and auditory cortices, the right salience network, and limbic nodes. Ruminative thought was associated with hyperconnectivity between bilateral regions of the CIN, as well as hyperconnectivity between the CIN and visual cortex and hypoconnectivity between the right CIN and the DMN. Finally, negative affect was associated with decreased segregation (increased connectivity) between the left FPN and the DMN, as well as between the CIN and DMN. In addition, the left FPN exhibited decreased connectivity with the salience network and the visual cortex.
      Figure thumbnail gr5
      Figure 5Chord diagrams displaying network-level associations for symptoms best fit by multivariate models. Association strengths were collated by node community, according to the a priori communities delineated in the Power264 atlas. Communities were lateralized and plotted around each chord diagram, with left-lateralized nodes displayed on the left side of the plot and bar width proportional to the number of nodes in the community. Red ribbons indicate hyperconnectivity, while blue ribbons indicate hypoconnectivity in association with symptom scores. Ribbon opacity is proportional to the mean strength of association with the symptom dimension (more opaque = stronger association), and ribbon width is proportional to the number of connections participating in a plotted association (wider ribbon = more connections). Ribbon appearance was determined based on a formula that balanced 1) the mean strength of connections, 2) the variability of these connection strengths, and 3) the number of connections of each community-pair. Full details for how this visualization was generated can be found in (). Note that the left column of diagrams represents edgewise regression models, while the right column represents brain basis set modeling models. AUD, auditory; CER, cerebellar; CIN, cinguloopercular network; DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; LIM, limbic; MEM, memory retrieval; SAL, salience network; SMN, somatomotor network; SUB, subcortical; VAN, ventral attention network; VIS, visual.

      Discussion

      This work examines transdiagnostic associations between functional connectivity and different symptom dimensions, in line with the dimensional perspective espoused by the RDoC framework (
      • Insel T.
      • Cuthbert B.
      • Garvey M.
      • Heinssen R.
      • Pine D.S.
      • Quinn K.
      • et al.
      Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders.
      ). Our multimodeling approach suggests that different dimensions of psychopathology required different representations of the functional connectome. Univariate relationships failed to robustly model any symptom dimensions. Some dimensions, such as insomnia and anxious arousal, were sufficiently modeled by subsets of the connectome. Other dimensions, such as negative affect and rumination, required information from the entire connectome; for these dimensions, smaller subsets of connectome were not sufficient to generate a replicable association.
      Anhedonia and anxiety sensitivity were not reproducibly modeled by any of our modeling approaches, despite a broad body of work implicating specific connections for both anhedonia (
      • Keller J.
      • Young C.B.
      • Kelley E.
      • Prater K.
      • Levitin D.J.
      • Menon V.
      Trait anhedonia is associated with reduced reactivity and connectivity of mesolimbic and paralimbic reward pathways.
      ,
      • Satterthwaite T.D.
      • Kable J.W.
      • Vandekar L.
      • Katchmar N.
      • Bassett D.S.
      • Baldassano C.F.
      • et al.
      Common and dissociable dysfunction of the reward system in bipolar and unipolar depression.
      ,
      • Felger J.C.
      • Li Z.
      • Haroon E.
      • Woolwine B.J.
      • Jung M.Y.
      • Hu X.
      • Miller A.H.
      Inflammation is associated with decreased functional connectivity within corticostriatal reward circuitry in depression.
      ,
      • Ferenczi E.A.
      • Zalocusky K.A.
      • Liston C.
      • Grosenick L.
      • Warden M.R.
      • Amatya D.
      • et al.
      Prefrontal cortical regulation of brainwide circuit dynamics and reward-related behavior.
      ,
      • Rzepa E.
      • McCabe C.
      Anhedonia and depression severity dissociated by dmPFC resting-state functional connectivity in adolescents.
      ) and anxiety sensitivity (
      • Laeger I.
      • Dobel C.
      • Dannlowski U.
      • Kugel H.
      • Grotegerd D.
      • Kissler J.
      • et al.
      Amygdala responsiveness to emotional words is modulated by subclinical anxiety and depression [published correction appears in Behav Brain Res 2014; 261:369–370].
      ,
      • Prater K.E.
      • Hosanagar A.
      • Klumpp H.
      • Angstadt M.
      • Phan K.L.
      Aberrant amygdala-frontal cortex connectivity during perception of fearful faces and at rest in generalized social anxiety disorder.
      ,
      • Liu W.J.
      • Yin D.Z.
      • Cheng W.H.
      • Fan M.X.
      • You M.N.
      • Men W.W.
      • et al.
      Abnormal functional connectivity of the amygdala-based network in resting-state FMRI in adolescents with generalized anxiety disorder.
      ,
      • Dong M.
      • Xia L.
      • Lu M.
      • Li C.
      • Xu K.
      • Zhang L.
      A failed top-down control from the prefrontal cortex to the amygdala in generalized anxiety disorder: Evidence from resting-state fMRI with Granger causality analysis.
      ,
      • Warren S.L.
      • Zhang Y.
      • Duberg K.
      • Mistry P.
      • Cai W.
      • Qin S.
      • et al.
      Anxiety and stress alter decision-making dynamics and causal amygdala-dorsolateral prefrontal cortex circuits during emotion regulation in children.
      ). In fact, all associations from single connections failed to replicate in the testing set after correction for multiple comparisons. This finding recapitulates similar studies that have highlighted the challenges of replicating univariate associations in fully held-out samples (
      • Marek S.
      • Tervo-Clemmens B.
      • Calabro F.J.
      • Montez D.F.
      • Kay B.P.
      • Hatoum A.S.
      • et al.
      Reproducible brain-wide association studies require thousands of individuals [published correction appears in Nature 2022; 605(7911):E11].
      ).
      In contrast to anhedonia and anxiety sensitivity, anxious arousal and insomnia were sufficiently modeled using edgewise regression, which used a small subset of the connectome (∼5%). Similar to anxiety sensitivity, anxious arousal has been associated with abnormal amygdala connectivity (
      • Liu W.J.
      • Yin D.Z.
      • Cheng W.H.
      • Fan M.X.
      • You M.N.
      • Men W.W.
      • et al.
      Abnormal functional connectivity of the amygdala-based network in resting-state FMRI in adolescents with generalized anxiety disorder.
      ,
      • Davis M.
      Are different parts of the extended amygdala involved in fear versus anxiety?.
      ). However, cross-cultural studies have revealed that different ethnoracial groups experience different degrees of anxious arousal (somatization), suggesting that this symptom dimension likely involves cortical integration beyond the amygdala (
      • Davis M.
      Are different parts of the extended amygdala involved in fear versus anxiety?.
      ,
      • Ryder A.G.
      • Yang J.
      • Zhu X.
      • Yao S.
      • Yi J.
      • Heine S.J.
      • Bagby R.M.
      The cultural shaping of depression: Somatic symptoms in China, psychological symptoms in North America?.
      ). Indeed, anxious arousal was associated with hyperconnectivity between sensory regions and regions in the association cortex (including the CIN, FPN, and DMN), which may underlie the somatization of anxious states. Insomnia was associated with hyperconnectivity between the DMN and multiple cortical areas, including the primary sensory cortex, salience network, and limbic nodes. This exaggerated connectivity with the DMN may represent increased vigilance and awareness of emotional and bodily states, resulting in poor sleep initiation and continuity (
      • Khazaie H.
      • Veronese M.
      • Noori K.
      • Emamian F.
      • Zarei M.
      • Ashkan K.
      • et al.
      Functional reorganization in obstructive sleep apnoea and insomnia: A systematic review of the resting-state fMRI.
      ).
      Finally, two of our constructs, rumination and negative affect, were only robustly modeled by BBS, which used information from the entire connectome and drew from components with broad alignments with multiple intrinsic connectivity networks. Evolutionary theories of ruminative thought support the notion that this symptom dimension likely developed later in evolution because the emergence of complex social environments and the ability to engage in sustained processing are thought to be supported by the expansive primate neocortex (
      • Andrews P.W.
      • Thomson Jr., J.A.
      The bright side of being blue: Depression as an adaptation for analyzing complex problems.
      ,
      • Dunbar R.I.M.
      The social brain hypothesis and its implications for social evolution.
      ). Similar theories about negative affectivity suggest that stress drives the brain to reorganize to minimize surprise (prediction error) in the environment, a complex process that likely involves cortex-spanning, multinetwork interactions (
      • Badcock P.B.
      • Davey C.G.
      • Whittle S.
      • Allen N.B.
      • Friston K.J.
      The depressed brain: An evolutionary systems theory.
      ). Both ruminative thought and negative affect exhibited abnormalities with the DMN, which has been consistently implicated in AM disorders, particularly in depression (
      • Greicius M.D.
      • Flores B.H.
      • Menon V.
      • Glover G.H.
      • Solvason H.B.
      • Kenna H.
      • et al.
      Resting-state functional connectivity in major depression: Abnormally increased contributions from subgenual cingulate cortex and thalamus.
      ,
      • Sheline Y.I.
      • Barch D.M.
      • Price J.L.
      • Rundle M.M.
      • Vaishnavi S.N.
      • Snyder A.Z.
      • et al.
      The default mode network and self-referential processes in depression.
      ). Hyperactivity of the DMN during emotional perception and judgment (
      • Grimm S.
      • Boesiger P.
      • Beck J.
      • Schuepbach D.
      • Bermpohl F.
      • Walter M.
      • et al.
      Altered negative BOLD responses in the default-mode network during emotion processing in depressed subjects.
      ) and passive viewing and reappraisal of negative pictures (
      • Sheline Y.I.
      • Barch D.M.
      • Price J.L.
      • Rundle M.M.
      • Vaishnavi S.N.
      • Snyder A.Z.
      • et al.
      The default mode network and self-referential processes in depression.
      ) as well hyperconnectivity between the DMN and the FPN (
      • Sheline Y.I.
      • Price J.L.
      • Yan Z.
      • Mintun M.A.
      Resting-state functional MRI in depression unmasks increased connectivity between networks via the dorsal nexus.
      ) suggest that DMN abnormalities may underlie an inability to detach from internal emotional states [e.g., rumination and negative affect (
      • Sheline Y.I.
      • Barch D.M.
      • Price J.L.
      • Rundle M.M.
      • Vaishnavi S.N.
      • Snyder A.Z.
      • et al.
      The default mode network and self-referential processes in depression.
      )].
      This is not to suggest, however, that the nonoptimal modeling approaches are without merit. While only the highest performing models for each symptom survived testing for multiple comparisons (with the exception of SCA models; see Figure S10 for schematic diagrams of all six symptom dimensions and their best performing models), different models of the same symptom exhibited similar weights, suggesting that convergent signals informed these different models. Furthermore, differences in model performance for some symptom dimensions were relatively minor; for example, all modeling approaches for insomnia performed relatively well. While the use of permutation testing, validation in a held-out dataset, and correction for multiple comparisons helps to confirm the validity of each individual model, the comparison of performances between models is by no means conclusive without rigorous statistical tests that would require larger sample sizes. Future analyses in larger datasets using iterative cross-validation methods are necessary to confirm these hypotheses.
      Direct validation of our models in a held-out dataset was critical for verifying their performance. Indeed, edgewise regression models exhibited excellent performance in training set cross-validation (typically achieving correlations of 0.6) but did not generate significant associations in testing data for four of six symptom dimensions. Edgewise regression favored smaller subsets of features, typically on the order of 500 features (around approximately 1% of the full connectivity matrix), which is similar to the number of connections used in many other studies examining multivariate associations of psychopathology (
      • Drysdale A.T.
      • Grosenick L.
      • Downar J.
      • Dunlop K.
      • Mansouri F.
      • Meng Y.
      • et al.
      Resting-state connectivity biomarkers define neurophysiological subtypes of depression [published correction appears in Nat Med 2017; 23:264].
      ,
      • Xia C.H.
      • Ma Z.
      • Ciric R.
      • Gu S.
      • Betzel R.F.
      • Kaczkurkin A.N.
      • et al.
      Linked dimensions of psychopathology and connectivity in functional brain networks.
      ,
      • Grosenick L.
      • Shi T.C.
      • Gunning F.M.
      • Dubin M.J.
      • Downar J.
      • Liston C.
      Functional and optogenetic approaches to discovering stable subtype-specific circuit mechanisms in depression.
      ,
      • Yamashita A.
      • Sakai Y.
      • Yamada T.
      • Yahata N.
      • Kunimatsu A.
      • Okada N.
      • et al.
      Generalizable brain network markers of major depressive disorder across multiple imaging sites.
      ). Notably, many of these studies did not perform a direct validation of their models on held-out data which, as the results from this study suggest, is a crucial step in assessing the generalizability of multivariate models.
      While we strove to implement a rigorous analytic design and findings are reinforced by and expand on prior literature, these results should also be considered in the context of the following limitations. To minimize the number of hypotheses tested for SCA, we limited seeds and masks for SCA to a circumscribed set based on extant literature. It is possible that univariate connections outside of those tested could replicate in a testing set. However, we took care to focus on those connections that were best supported by the literature. Furthermore, exploratory whole-brain analyses only generated one cluster (anxious arousal; left amygdala ↔ left primary motor cortex), which failed to replicate in the testing dataset. In addition, most seed regions used for SCA were defined using anatomical atlases; it has been demonstrated that functional regions may be better captured through different means (
      • Dadi K.
      • Rahim M.
      • Abraham A.
      • Chyzhyk D.
      • Milham M.
      • Thirion B.
      • et al.
      Benchmarking functional connectome-based predictive models for resting-state fMRI.
      ,
      • Zhi D.
      • King M.
      • Hernandez-Castillo C.R.
      • Diedrichsen J.
      Evaluating brain parcellations using the distance-controlled boundary coefficient.
      ), and the use of alternative seeds in future studies may yield stronger results for SCA. Additional limitations include the theory-driven selection of symptom structure resolution; other resolutions in the symptom hierarchy could reveal additional insights. Entirely data-driven multiview analyses that simultaneously model symptom structures and associated connectivity patterns are possible (
      • Kaczkurkin A.N.
      • Moore T.M.
      • Sotiras A.
      • Xia C.H.
      • Shinohara R.T.
      • Satterthwaite T.D.
      Approaches to defining common and dissociable neurobiological deficits associated with psychopathology in youth.
      ). However, recent simulation studies suggest that very high sample sizes (approaching thousands of individuals for an r = 0.3) may be necessary to generate reproducible associations using these methods (
      • Helmer M.
      • Warrington S.
      • Mohammadi-Nejad A.R.
      • Ji J.L.
      • Howell A.
      • Rosand B.
      • et al.
      On stability of Canonical Correlation Analysis and Partial Least Squares with application to brain-behavior associations.
      ). Given our sample size (n = 192), we chose instead to fix our symptom structure based on the emergence of psychologically plausible, coherent symptom communities. Finally, the testing dataset was of moderate size (n = 50); future analyses pooling data from multiple sites will prove invaluable for testing the robustness and generalizability of symptom-brain associations (
      • Marek S.
      • Tervo-Clemmens B.
      • Calabro F.J.
      • Montez D.F.
      • Kay B.P.
      • Hatoum A.S.
      • et al.
      Reproducible brain-wide association studies require thousands of individuals [published correction appears in Nature 2022; 605(7911):E11].
      ).
      Despite these limitations, this work represents one of the only studies to examine transdiagnostic AM psychopathology using multiple representations of brain connectivity. Our rigorous performance assessment framework revealed that different dimensions of AM psychopathology required different representations of functional connectivity, highlighting the potential sufficiency of single-circuit, univariate approaches for some dimensions and the necessity of connectome-wide, multivariate approaches in others.

      Acknowledgments and Disclosures

      This work was supported by the National Institute of Mental Health Grant No. U01 MH109991 (to YIS).
      We acknowledge Maria Prociuk for her help in preparing the manuscript for publication. We also thank all participants for their participation.
      A previous version of this article was published as a preprint on bioRxiv: https://www.biorxiv.org/content/10.1101/2021.03.05.434151v1.
      The authors report no biomedical financial interests or potential conflicts of interest.

      References

        • Insel T.
        • Cuthbert B.
        • Garvey M.
        • Heinssen R.
        • Pine D.S.
        • Quinn K.
        • et al.
        Research domain criteria (RDoC): Toward a new classification framework for research on mental disorders.
        Am J Psychiatry. 2010; 167: 748-751
        • Krueger R.F.
        The structure of common mental disorders.
        Arch Gen Psychiatry. 1999; 56: 921-926
        • Watson D.
        Rethinking the mood and anxiety disorders: A quantitative hierarchical model for DSM-V.
        J Abnorm Psychol. 2005; 114: 522-536
        • Murray C.J.L.
        • Atkinson C.
        • Bhalla K.
        • Birbeck G.
        • Burstein R.
        • Chou D.
        • et al.
        The state of US health, 1990–2010: Burden of diseases, injuries, and risk factors.
        JAMA. 2013; 310: 591-608
        • Biswal B.
        • Yetkin F.Z.
        • Haughton V.M.
        • Hyde J.S.
        Functional connectivity in the motor cortex of resting human brain using echo-planar MRI.
        Magn Reson Med. 1995; 34: 537-541
        • Fox M.D.
        • Raichle M.E.
        Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging.
        Nat Rev Neurosci. 2007; 8: 700-711
        • Van Dijk K.R.A.
        • Hedden T.
        • Venkataraman A.
        • Evans K.C.
        • Lazar S.W.
        • Buckner R.L.
        Intrinsic functional connectivity as a tool for human connectomics: Theory, properties, and optimization.
        J Neurophysiol. 2010; 103: 297-321
        • Glasser M.F.
        • Sotiropoulos S.N.
        • Wilson J.A.
        • Coalson T.S.
        • Fischl B.
        • Andersson J.L.
        • et al.
        The minimal preprocessing pipelines for the Human Connectome Project.
        Neuroimage. 2013; 80: 105-124
        • Davatzikos C.
        Machine learning in neuroimaging: Progress and challenges.
        Neuroimage. 2019; 197: 652-656
        • Sripada C.
        • Angstadt M.
        • Rutherford S.
        • Kessler D.
        • Kim Y.
        • Yee M.
        • Levina E.
        Basic units of inter-individual variation in resting state connectomes.
        Sci Rep. 2019; 9: 1900
        • Schnack H.G.
        • Kahn R.S.
        Detecting neuroimaging biomarkers for psychiatric disorders: Sample size matters.
        Front Psychiatry. 2016; 7: 50
        • Marek S.
        • Tervo-Clemmens B.
        • Calabro F.J.
        • Montez D.F.
        • Kay B.P.
        • Hatoum A.S.
        • et al.
        Reproducible brain-wide association studies require thousands of individuals [published correction appears in Nature 2022; 605(7911):E11].
        Nature. 2022; 603: 654-660
        • Müller V.I.
        • Cieslik E.C.
        • Serbanescu I.
        • Laird A.R.
        • Fox P.T.
        • Eickhoff S.B.
        Altered brain activity in unipolar depression revisited: Meta-analyses of neuroimaging studies.
        JAMA Psychiatry. 2017; 74: 47-55
        • Dinga R.
        • Schmaal L.
        • Penninx B.W.J.H.
        • van Tol M.J.
        • Veltman D.J.
        • van Velzen L.
        • et al.
        Evaluating the evidence for biotypes of depression: Methodological replication and extension of Drysdale et al..
        Neuroimage Clin. 2019; 22: 101796
        • Poldrack R.A.
        • Huckins G.
        • Varoquaux G.
        Establishment of best practices for evidence for prediction: A review.
        JAMA Psychiatry. 2020; 77: 534-540
        • Seok D.
        • Smyk N.
        • Jaskir M.
        • Cook P.
        • Elliott M.
        • Girelli T.
        • et al.
        Dimensional connectomics of anxious misery, a human connectome study related to human disease: Overview of protocol and data quality.
        Neuroimage Clin. 2020; 28: 102489
        • McCrae R.R.
        • Costa Jr., P.T.
        Brief versions of the NEO-PI-3.
        J Individ Differ. 2007; 28: 116-128
        • Andrews G.
        • Stewart G.
        • Morris-Yates A.
        • Holt P.
        • Henderson S.
        Evidence for a general neurotic syndrome.
        Br J Psychiatry. 1990; 157: 6-12
        • Khan A.A.
        • Jacobson K.C.
        • Gardner C.O.
        • Prescott C.A.
        • Kendler K.S.
        Personality and comorbidity of common psychiatric disorders.
        Br J Psychiatry. 2005; 186: 190-196
        • Jeub L.G.S.
        • Sporns O.
        • Fortunato S.
        Multiresolution consensus clustering in networks.
        Sci Rep. 2018; 8: 3259
        • Power J.D.
        • Cohen A.L.
        • Nelson S.M.
        • Wig G.S.
        • Barnes K.A.
        • Church J.A.
        • et al.
        Functional network organization of the human brain.
        Neuron. 2011; 72: 665-678
        • Drysdale A.T.
        • Grosenick L.
        • Downar J.
        • Dunlop K.
        • Mansouri F.
        • Meng Y.
        • et al.
        Resting-state connectivity biomarkers define neurophysiological subtypes of depression [published correction appears in Nat Med 2017; 23:264].
        Nat Med. 2017; 23: 28-38
        • Keller J.
        • Young C.B.
        • Kelley E.
        • Prater K.
        • Levitin D.J.
        • Menon V.
        Trait anhedonia is associated with reduced reactivity and connectivity of mesolimbic and paralimbic reward pathways.
        J Psychiatr Res. 2013; 47: 1319-1328
        • Satterthwaite T.D.
        • Kable J.W.
        • Vandekar L.
        • Katchmar N.
        • Bassett D.S.
        • Baldassano C.F.
        • et al.
        Common and dissociable dysfunction of the reward system in bipolar and unipolar depression.
        Neuropsychopharmacology. 2015; 40: 2258-2268
        • Felger J.C.
        • Li Z.
        • Haroon E.
        • Woolwine B.J.
        • Jung M.Y.
        • Hu X.
        • Miller A.H.
        Inflammation is associated with decreased functional connectivity within corticostriatal reward circuitry in depression.
        Mol Psychiatry. 2016; 21: 1358-1365
        • Ferenczi E.A.
        • Zalocusky K.A.
        • Liston C.
        • Grosenick L.
        • Warden M.R.
        • Amatya D.
        • et al.
        Prefrontal cortical regulation of brainwide circuit dynamics and reward-related behavior.
        Science. 2016; 351: aac9698
        • Rzepa E.
        • McCabe C.
        Anhedonia and depression severity dissociated by dmPFC resting-state functional connectivity in adolescents.
        J Psychopharmacol. 2018; 32: 1067-1074
        • Laeger I.
        • Dobel C.
        • Dannlowski U.
        • Kugel H.
        • Grotegerd D.
        • Kissler J.
        • et al.
        Amygdala responsiveness to emotional words is modulated by subclinical anxiety and depression [published correction appears in Behav Brain Res 2014; 261:369–370].
        Behav Brain Res. 2012; 233: 508-516
        • Prater K.E.
        • Hosanagar A.
        • Klumpp H.
        • Angstadt M.
        • Phan K.L.
        Aberrant amygdala-frontal cortex connectivity during perception of fearful faces and at rest in generalized social anxiety disorder.
        Depress Anxiety. 2013; 30: 234-241
        • Liu W.J.
        • Yin D.Z.
        • Cheng W.H.
        • Fan M.X.
        • You M.N.
        • Men W.W.
        • et al.
        Abnormal functional connectivity of the amygdala-based network in resting-state FMRI in adolescents with generalized anxiety disorder.
        Med Sci Monit. 2015; 21: 459-467
        • Dong M.
        • Xia L.
        • Lu M.
        • Li C.
        • Xu K.
        • Zhang L.
        A failed top-down control from the prefrontal cortex to the amygdala in generalized anxiety disorder: Evidence from resting-state fMRI with Granger causality analysis.
        Neurosci Lett. 2019; 707: 134314
        • Warren S.L.
        • Zhang Y.
        • Duberg K.
        • Mistry P.
        • Cai W.
        • Qin S.
        • et al.
        Anxiety and stress alter decision-making dynamics and causal amygdala-dorsolateral prefrontal cortex circuits during emotion regulation in children.
        Biol Psychiatry. 2020; 88: 576-586
        • Berman M.G.
        • Peltier S.
        • Nee D.E.
        • Kross E.
        • Deldin P.J.
        • Jonides J.
        Depression, rumination and the default network.
        Soc Cogn Affect Neurosci. 2011; 6: 548-555
        • Hamilton J.P.
        • Furman D.J.
        • Chang C.
        • Thomason M.E.
        • Dennis E.
        • Gotlib I.H.
        Default-mode and task-positive network activity in major depressive disorder: Implications for adaptive and maladaptive rumination.
        Biol Psychiatry. 2011; 70: 327-333
        • Berman M.G.
        • Misic B.
        • Buschkuehl M.
        • Kross E.
        • Deldin P.J.
        • Peltier S.
        • et al.
        Does resting-state connectivity reflect depressive rumination? A tale of two analyses.
        Neuroimage. 2014; 103: 267-279
        • Zhou H.X.
        • Chen X.
        • Shen Y.Q.
        • Li L.
        • Chen N.X.
        • Zhu Z.C.
        • et al.
        Rumination and the default mode network: Meta-analysis of brain imaging studies and implications for depression.
        Neuroimage. 2020; 206: 116287
        • Greicius M.D.
        • Flores B.H.
        • Menon V.
        • Glover G.H.
        • Solvason H.B.
        • Kenna H.
        • et al.
        Resting-state functional connectivity in major depression: Abnormally increased contributions from subgenual cingulate cortex and thalamus.
        Biol Psychiatry. 2007; 62: 429-437
        • Sheline Y.I.
        • Price J.L.
        • Yan Z.
        • Mintun M.A.
        Resting-state functional MRI in depression unmasks increased connectivity between networks via the dorsal nexus.
        Proc Natl Acad Sci U S A. 2010; 107: 11020-11025
        • Davey C.G.
        • Harrison B.J.
        • Yücel M.
        • Allen N.B.
        Regionally specific alterations in functional connectivity of the anterior cingulate cortex in major depressive disorder.
        Psychol Med. 2012; 42: 2071-2081
        • Koenigs M.
        • Grafman J.
        The functional neuroanatomy of depression: Distinct roles for ventromedial and dorsolateral prefrontal cortex.
        Behav Brain Res. 2009; 201: 239-243
        • Fox M.D.
        • Buckner R.L.
        • White M.P.
        • Greicius M.D.
        • Pascual-Leone A.
        Efficacy of transcranial magnetic stimulation targets for depression is related to intrinsic functional connectivity with the subgenual cingulate.
        Biol Psychiatry. 2012; 72: 595-603
        • Baeken C.
        • Marinazzo D.
        • Wu G.R.
        • Van Schuerbeek P.
        • De Mey J.
        • Marchetti I.
        • et al.
        Accelerated HF-rTMS in treatment-resistant unipolar depression: Insights from subgenual anterior cingulate functional connectivity.
        World J Biol Psychiatry. 2014; 15: 286-297
        • Avery J.A.
        • Drevets W.C.
        • Moseman S.E.
        • Bodurka J.
        • Barcalow J.C.
        • Simmons W.K.
        Major depressive disorder is associated with abnormal interoceptive activity and functional connectivity in the insula.
        Biol Psychiatry. 2014; 76: 258-266
        • Shao R.
        • Lau W.K.W.
        • Leung M.K.
        • Lee T.M.C.
        Subgenual anterior cingulate-insula resting-state connectivity as a neural correlate to trait and state stress resilience.
        Brain Cogn. 2018; 124: 73-81
        • Chuah Y.M.L.
        • Venkatraman V.
        • Dinges D.F.
        • Chee M.W.L.
        The neural basis of interindividual variability in inhibitory efficiency after sleep deprivation.
        J Neurosci. 2006; 26: 7156-7162
        • Koenigs M.
        • Holliday J.
        • Solomon J.
        • Grafman J.
        Left dorsomedial frontal brain damage is associated with insomnia.
        J Neurosci. 2010; 30: 16041-16043
        • Zhang Q.
        • Qin W.
        • He X.
        • Li Q.
        • Chen B.
        • Zhang Y.
        • Yu C.
        Functional disconnection of the right anterior insula in obstructive sleep apnea.
        Sleep Med. 2015; 16: 1062-1070
        • Park B.
        • Palomares J.A.
        • Woo M.A.
        • Kang D.W.
        • Macey P.M.
        • Yan-Go F.L.
        • et al.
        Aberrant insular functional network integrity in patients with obstructive sleep apnea.
        Sleep. 2016; 39: 989-1000
        • Yu S.
        • Guo B.
        • Shen Z.
        • Wang Z.
        • Kui Y.
        • Hu Y.
        • Feng F.
        The imbalanced anterior and posterior default mode network in the primary insomnia.
        J Psychiatr Res. 2018; 103: 97-103
        • Davis M.
        Are different parts of the extended amygdala involved in fear versus anxiety?.
        Biol Psychiatry. 1998; 44: 1239-1247
        • Ryder A.G.
        • Yang J.
        • Zhu X.
        • Yao S.
        • Yi J.
        • Heine S.J.
        • Bagby R.M.
        The cultural shaping of depression: Somatic symptoms in China, psychological symptoms in North America?.
        J Abnorm Psychol. 2008; 117: 300-313
        • Khazaie H.
        • Veronese M.
        • Noori K.
        • Emamian F.
        • Zarei M.
        • Ashkan K.
        • et al.
        Functional reorganization in obstructive sleep apnoea and insomnia: A systematic review of the resting-state fMRI.
        Neurosci Biobehav Rev. 2017; 77: 219-231
        • Andrews P.W.
        • Thomson Jr., J.A.
        The bright side of being blue: Depression as an adaptation for analyzing complex problems.
        Psychol Rev. 2009; 116: 620-654
        • Dunbar R.I.M.
        The social brain hypothesis and its implications for social evolution.
        Ann Hum Biol. 2009; 36: 562-572
        • Badcock P.B.
        • Davey C.G.
        • Whittle S.
        • Allen N.B.
        • Friston K.J.
        The depressed brain: An evolutionary systems theory.
        Trends Cogn Sci. 2017; 21: 182-194
        • Sheline Y.I.
        • Barch D.M.
        • Price J.L.
        • Rundle M.M.
        • Vaishnavi S.N.
        • Snyder A.Z.
        • et al.
        The default mode network and self-referential processes in depression.
        Proc Natl Acad Sci U S A. 2009; 106: 1942-1947
        • Grimm S.
        • Boesiger P.
        • Beck J.
        • Schuepbach D.
        • Bermpohl F.
        • Walter M.
        • et al.
        Altered negative BOLD responses in the default-mode network during emotion processing in depressed subjects.
        Neuropsychopharmacology. 2009; 34: 932-943
        • Xia C.H.
        • Ma Z.
        • Ciric R.
        • Gu S.
        • Betzel R.F.
        • Kaczkurkin A.N.
        • et al.
        Linked dimensions of psychopathology and connectivity in functional brain networks.
        Nat Commun. 2018; 9: 3003
        • Grosenick L.
        • Shi T.C.
        • Gunning F.M.
        • Dubin M.J.
        • Downar J.
        • Liston C.
        Functional and optogenetic approaches to discovering stable subtype-specific circuit mechanisms in depression.
        Biol Psychiatry Cogn Neurosci Neuroimaging. 2019; 4: 554-566
        • Yamashita A.
        • Sakai Y.
        • Yamada T.
        • Yahata N.
        • Kunimatsu A.
        • Okada N.
        • et al.
        Generalizable brain network markers of major depressive disorder across multiple imaging sites.
        PLoS Biol. 2020; 18e3000966
        • Dadi K.
        • Rahim M.
        • Abraham A.
        • Chyzhyk D.
        • Milham M.
        • Thirion B.
        • et al.
        Benchmarking functional connectome-based predictive models for resting-state fMRI.
        Neuroimage. 2019; 192: 115-134
        • Zhi D.
        • King M.
        • Hernandez-Castillo C.R.
        • Diedrichsen J.
        Evaluating brain parcellations using the distance-controlled boundary coefficient.
        Hum Brain Mapp. 2022; 43: 3706-3720
        • Kaczkurkin A.N.
        • Moore T.M.
        • Sotiras A.
        • Xia C.H.
        • Shinohara R.T.
        • Satterthwaite T.D.
        Approaches to defining common and dissociable neurobiological deficits associated with psychopathology in youth.
        Biol Psychiatry. 2020; 88: 51-62
        • Helmer M.
        • Warrington S.
        • Mohammadi-Nejad A.R.
        • Ji J.L.
        • Howell A.
        • Rosand B.
        • et al.
        On stability of Canonical Correlation Analysis and Partial Least Squares with application to brain-behavior associations.
        bioRxiv. 2021; https://doi.org/10.1101/2020.08.25.265546

      Linked Article

      • The Role of Analytical Flexibility in Determining Mental Health Biomarkers
        Biological Psychiatry: Global Open ScienceVol. 2Issue 4
        • Preview
          Many studies have reported alterations in functional brain organization and connectivity associated with mental health disorders such as anxiety and depression, also known as disorders of anxious misery. However, clinical translation of this work into robust magnetic resonance imaging–based mental health biomarkers has remained elusive. There are several challenges that contribute to this lack of clinical translation, including limited reproducibility of findings (1) and analytical flexibility (2) in terms of both brain representations (3) and brain-symptom modeling approaches (4).
        • Full-Text
        • PDF
        Open Access