Discriminating Heterogeneous Trajectories of Resilience and Depression After Major Life Stressors Using Polygenic Scores | Depressive Disorders | JAMA Psychiatry | JAMA Network
[Skip to Navigation]
Sign In
Figure 1.  Trajectories of Depression Using Combined Health and Retirement Study Data
Trajectories of Depression Using Combined Health and Retirement Study Data

The points represent the mean (SE) Center for Epidemiologic Studies-Depression (CES-D) score per trajectory class at each time point (a 10-item scale, with scores greater than or equal to 4 indicating depression). Depressive symptoms were measured every 2 years in the Health and Retirement Study. The index major life stressor occurred between the Prior to the index event and First follow-up points.

Figure 2.  Receiver Operating Characteristic (ROC) Curves
Receiver Operating Characteristic (ROC) Curves

Multinomial classification performance presenting the discriminatory accuracy over all trajectories (A) and the discriminatory accuracy for each individual trajectory (B). AUC indicates area under the curve.

Figure 3.  Variable Importance Based on Local Interpretable Model-Agnostic Explanations for Discriminating Each of the 4 Trajectories
Variable Importance Based on Local Interpretable Model-Agnostic Explanations for Discriminating Each of the 4 Trajectories

Trajectories shown for estimating resilience trajectory (A), preexisting/chronic trajectory (B), recovery trajectory (C), and emerging depression trajectory (D). Values of a given feature are indexed between 0 and 1, with lower feature values closer to 0 and higher values closer to 1. The vertical line at the 0 mark of the x-axis represents no association, and positive values on the x-axis (ie, blue bars) represent a positive association of feature values and outcome. Negative values on the x-axis (ie, gray bars) represent a negative association of feature values and outcome. The y-axis shows the top 15 features. Because the features are used as continuous variables in the model, different value ranges can be associated both positively and negatively. For instance, in panel D, the most important feature positively associated with membership in the resilient latent growth mixture modeling class are values of the schizophrenia polygenic score (PGS) of less than or equal to 0.39, while higher schizophrenia PGSs (ie, >0.57) are negatively associated with resilience. ADHD indicates attention-deficit/hyperactivity disorder; BMI, body mass index; LDL, low-density lipoprotein; MDD, major depressive disorder; and PTSD, posttraumatic stress disorder.

Table 1.  Overview About the PGS Candidate Features in the Data Seta
Overview About the PGS Candidate Features in the Data Seta
Table 2.  Participant Characteristics
Participant Characteristics
1.
Maccallum  F, Galatzer-Levy  IR, Bonanno  GA.  Trajectories of depression following spousal and child bereavement: a comparison of the heterogeneity in outcomes.   J Psychiatr Res. 2015;69:72-79. doi:10.1016/j.jpsychires.2015.07.017 PubMedGoogle ScholarCrossref
2.
Malgaroli  M, Galatzer-Levy  IR, Bonanno  GA.  Heterogeneity in trajectories of depression in response to divorce is associated with differential risk for mortality.   Clin Psychol Sci. 2017;5(5):843-850. doi:10.1177/2167702617705951 PubMedGoogle ScholarCrossref
3.
Stolove  CA, Galatzer-Levy  IR, Bonanno  GA.  Emergence of depression following job loss prospectively predicts lower rates of reemployment.   Psychiatry Res. 2017;253:79-83. doi:10.1016/j.psychres.2017.03.036 PubMedGoogle ScholarCrossref
4.
Breslau  N.  Epidemiologic studies of trauma, posttraumatic stress disorder, and other psychiatric disorders.   Can J Psychiatry. 2002;47(10):923-929. doi:10.1177/070674370204701003 PubMedGoogle ScholarCrossref
5.
Galatzer-Levy  IR, Bonanno  GA.  Optimism and death: predicting the course and consequences of depression trajectories in response to heart attack.   Psychol Sci. 2014;25(12):2177-2188. doi:10.1177/0956797614551750 PubMedGoogle ScholarCrossref
6.
Burton  CL, Galatzer-Levy  IR, Bonanno  GA.  Treatment type and demographic characteristics as predictors for cancer adjustment: prospective trajectories of depressive symptoms in a population sample.   Health Psychol. 2015;34(6):602-609. doi:10.1037/hea0000145 PubMedGoogle ScholarCrossref
7.
Galatzer-Levy  IR, Huang  SH, Bonanno  GA.  Trajectories of resilience and dysfunction following potential trauma: a review and statistical evaluation.   Clin Psychol Rev. 2018;63:41-55. doi:10.1016/j.cpr.2018.05.008 PubMedGoogle ScholarCrossref
8.
Bonanno  GA.  Loss, trauma, and human resilience: have we underestimated the human capacity to thrive after extremely aversive events?   Am Psychol. 2004;59(1):20-28. doi:10.1037/0003-066X.59.1.20 PubMedGoogle ScholarCrossref
9.
Schultebraucks  K, Shalev  AY, Michopoulos  V,  et al.  A validated predictive algorithm of post-traumatic stress course following emergency department admission after a traumatic stressor.   Nat Med. 2020;26(7):1084-1088. doi:10.1038/s41591-020-0951-z PubMedGoogle ScholarCrossref
10.
Schultebraucks  K, Qian  M, Abu-Amara  D,  et al.  Pre-deployment risk factors for PTSD in active-duty personnel deployed to Afghanistan: a machine-learning approach for analyzing multivariate predictors.   Mol Psychiatry. 2020;1-12.PubMedGoogle Scholar
11.
Hinrichs  R, van Rooij  SJ, Michopoulos  V,  et al.  Increased skin conductance response in the immediate aftermath of trauma predicts PTSD risk.   Chronic Stress (Thousand Oaks). 2019;3:2470547019844441. doi:10.1177/2470547019844441 PubMedGoogle Scholar
12.
Michopoulos  V, Beurel  E, Gould  F,  et al.  Association of prospective risk for chronic PTSD symptoms with low TNFα and IFNγ concentrations in the immediate aftermath of trauma exposure.   Am J Psychiatry. 2020;177(1):58-65. doi:10.1176/appi.ajp.2019.19010039 PubMedGoogle ScholarCrossref
13.
Bonanno  GA, Romero  SA, Klein  SI.  The temporal elements of psychological resilience: an integrative framework for the study of individuals, families, and communities.   Psychol Inquir. 2015;26(2):139-169. doi:10.1080/1047840X.2015.992677Google ScholarCrossref
14.
Bonanno  GA, Westphal  M, Mancini  AD.  Resilience to loss and potential trauma.   Annu Rev Clin Psychol. 2011;7:511-535. doi:10.1146/annurev-clinpsy-032210-104526 PubMedGoogle ScholarCrossref
15.
Boardman  JD, Blalock  CL, Button  TM.  Sex differences in the heritability of resilience.   Twin Res Hum Genet. 2008;11(1):12-27. doi:10.1375/twin.11.1.12 PubMedGoogle ScholarCrossref
16.
Amstadter  AB, Myers  JM, Kendler  KS.  Psychiatric resilience: longitudinal twin study.   Br J Psychiatry. 2014;205(4):275-280. doi:10.1192/bjp.bp.113.130906 PubMedGoogle ScholarCrossref
17.
Waaktaar  T, Torgersen  S.  Genetic and environmental causes of variation in trait resilience in young people.   Behav Genet. 2012;42(3):366-377. doi:10.1007/s10519-011-9519-5 PubMedGoogle ScholarCrossref
18.
Schizophrenia Working Group of the Psychiatric Genomics Consortium.  Biological insights from 108 schizophrenia-associated genetic loci.   Nature. 2014;511(7510):421-427. doi:10.1038/nature13595 PubMedGoogle ScholarCrossref
19.
Hyde  CL, Nagle  MW, Tian  C,  et al.  Identification of 15 genetic loci associated with risk of major depression in individuals of European descent.   Nat Genet. 2016;48(9):1031-1036. doi:10.1038/ng.3623 PubMedGoogle ScholarCrossref
20.
Psychiatric GWAS Consortium Bipolar Disorder Working Group.  Large-scale genome-wide association analysis of bipolar disorder identifies a new susceptibility locus near ODZ4.   Nat Genet. 2011;43(10):977-983. doi:10.1038/ng.943 PubMedGoogle ScholarCrossref
21.
Robinson  EB, St Pourcain  B, Anttila  V,  et al; iPSYCH-SSI-Broad Autism Group.  Genetic risk for autism spectrum disorders and neuropsychiatric variation in the general population.   Nat Genet. 2016;48(5):552-555. doi:10.1038/ng.3529 PubMedGoogle ScholarCrossref
22.
Nagel  M, Jansen  PR, Stringer  S,  et al; 23andMe Research Team.  Meta-analysis of genome-wide association studies for neuroticism in 449,484 individuals identifies novel genetic loci and pathways.   Nat Genet. 2018;50(7):920-927. doi:10.1038/s41588-018-0151-7 PubMedGoogle ScholarCrossref
23.
Watanabe  K, Taskesen  E, van Bochoven  A, Posthuma  D.  Functional mapping and annotation of genetic associations with FUMA.   Nat Commun. 2017;8(1):1826. doi:10.1038/s41467-017-01261-5 PubMedGoogle ScholarCrossref
24.
Wray  NR, Goddard  ME, Visscher  PM.  Prediction of individual genetic risk to disease from genome-wide association studies.   Genome Res. 2007;17(10):1520-1528. doi:10.1101/gr.6665407 PubMedGoogle ScholarCrossref
25.
Sugrue  LP, Desikan  RS.  What are polygenic scores and why are they important?   JAMA. 2019;321(18):1820-1821. doi:10.1001/jama.2019.3893 PubMedGoogle ScholarCrossref
26.
Dudbridge  F.  Power and predictive accuracy of polygenic risk scores.   PLoS Genet. 2013;9(3):e1003348. doi:10.1371/journal.pgen.1003348 PubMedGoogle Scholar
27.
Khera  AV, Chaffin  M, Aragam  KG,  et al.  Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations.   Nat Genet. 2018;50(9):1219-1224. doi:10.1038/s41588-018-0183-z PubMedGoogle ScholarCrossref
28.
Wray  NR, Lee  SH, Mehta  D, Vinkhuyzen  AA, Dudbridge  F, Middeldorp  CM.  Research review: polygenic methods and their application to psychiatric traits.   J Child Psychol Psychiatry. 2014;55(10):1068-1087. doi:10.1111/jcpp.12295 PubMedGoogle ScholarCrossref
29.
Schoeler  T, Choi  SW, Dudbridge  F,  et al.  Multi–polygenic score approach to identifying individual vulnerabilities associated with the risk of exposure to bullying.   JAMA Psychiatry. 2019;76(7):730-738. doi:10.1001/jamapsychiatry.2019.0310 PubMedGoogle ScholarCrossref
30.
Krapohl  E, Patel  H, Newhouse  S,  et al.  Multi-polygenic score approach to trait prediction.   Mol Psychiatry. 2018;23(5):1368-1374. doi:10.1038/mp.2017.163 PubMedGoogle ScholarCrossref
31.
Stein  MB, Choi  KW, Jain  S,  et al.  Genome-wide analyses of psychological resilience in U.S. Army soldiers.   Am J Med Genet B Neuropsychiatr Genet. 2019;180(5):310-319. doi:10.1002/ajmg.b.32730 PubMedGoogle ScholarCrossref
32.
Sonnega  A, Faul  JD, Ofstedal  MB, Langa  KM, Phillips  JW, Weir  DR.  Cohort Profile: the Health and Retirement Study (HRS).   Int J Epidemiol. 2014;43(2):576-585. doi:10.1093/ije/dyu067 PubMedGoogle ScholarCrossref
33.
Juster  FT, Suzman  R.  An overview of the Health and Retirement Study.   J Hum Resour. 1995;30:S7-S56. doi:10.2307/146277Google ScholarCrossref
34.
Steffick  DE. Documentation of affective functioning measures in the Health and Retirement Study: HRS/AHEAD Documentation Report 2000. Accessed February 16, 2021. https://hrs.isr.umich.edu/publications/biblio/5411
35.
Kohout  FJ, Berkman  LF, Evans  DA, Cornoni-Huntley  J.  Two shorter forms of the CES-D (Center for Epidemiological Studies Depression) depression symptoms index.   J Aging Health. 1993;5(2):179-193. doi:10.1177/089826439300500202 PubMedGoogle ScholarCrossref
36.
Radloff  LS.  The use of the Center for Epidemiologic Studies Depression Scale in adolescents and young adults.   J Youth Adolesc. 1991;20(2):149-166. doi:10.1007/BF01537606 PubMedGoogle ScholarCrossref
37.
Irwin  M, Artin  KH, Oxman  MN.  Screening for depression in the older adult: criterion validity of the 10-item Center for Epidemiological Studies Depression Scale (CES-D).   Arch Intern Med. 1999;159(15):1701-1704. doi:10.1001/archinte.159.15.1701 PubMedGoogle ScholarCrossref
38.
Muthén  L, Muthén  B. Mplus. Statistical analysis with latent variables: user’s guide. Accessed February 16, 2021. http://www.statmodel.com/download/usersguide/Mplus%20user%20guide%20Ver_7_r3_web.pdf
39.
Galatzer-Levy  IR, Bonanno  GA, Mancini  AD.  From marianthal to latent growth mixture modeling: a return to the exploration of individual differences in response to unemployment.   J Neuroscience Psychology Econ. 2010;3(2):116. doi:10.1037/a0020077Google ScholarCrossref
40.
Ware  E, Schmitz  L, Faul  JD. HRS polygenic scores 2006-2010: genetic data. HRS documentation report. Accessed February 16, 2021. https://hrs.isr.umich.edu/news/hrs-polygenic-scores-2006-2010-genetic-data
41.
Wray  NR, Ripke  S, Mattheisen  M,  et al; eQTLGen; 23andMe; Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium.  Genome-wide association analyses identify 44 risk variants and refine the genetic architecture of major depression.   Nat Genet. 2018;50(5):668-681. doi:10.1038/s41588-018-0090-3 PubMedGoogle ScholarCrossref
42.
Duncan  LE, Ratanatharathorn  A, Aiello  AE,  et al.  Largest GWAS of PTSD (N=20 070) yields genetic overlap with schizophrenia and sex differences in heritability.   Mol Psychiatry. 2018;23(3):666-673. doi:10.1038/mp.2017.77 PubMedGoogle ScholarCrossref
43.
Demontis  D, Walters  RK, Martin  J,  et al; ADHD Working Group of the Psychiatric Genomics Consortium (PGC); Early Lifecourse & Genetic Epidemiology (EAGLE) Consortium; 23andMe Research Team.  Discovery of the first genome-wide significant risk loci for attention deficit/hyperactivity disorder.   Nat Genet. 2019;51(1):63-75. doi:10.1038/s41588-018-0269-7 PubMedGoogle ScholarCrossref
44.
Arnold  PD, Askland  KD, Barlassina  C,  et al; International Obsessive Compulsive Disorder Foundation Genetics Collaborative (IOCDF-GC) and OCD Collaborative Genetics Association Studies (OCGAS).  Revealing the complex genetic architecture of obsessive-compulsive disorder using meta-analysis.   Mol Psychiatry. 2018;23(5):1181-1188. doi:10.1038/mp.2017.154 PubMedGoogle ScholarCrossref
45.
de Moor  MH, van den Berg  SM, Verweij  KJ,  et al; Genetics of Personality Consortium.  Meta-analysis of genome-wide association studies for neuroticism, and the polygenic association with major depressive disorder.   JAMA Psychiatry. 2015;72(7):642-650. doi:10.1001/jamapsychiatry.2015.0554 PubMedGoogle ScholarCrossref
46.
Ripke  S, Wray  NR, Lewis  CM,  et al; Major Depressive Disorder Working Group of the Psychiatric GWAS Consortium.  A mega-analysis of genome-wide association studies for major depressive disorder.   Mol Psychiatry. 2013;18(4):497-511. doi:10.1038/mp.2012.21 PubMedGoogle ScholarCrossref
47.
Otowa  T, Hek  K, Lee  M,  et al.  Meta-analysis of genome-wide association studies of anxiety disorders.   Mol Psychiatry. 2016;21(10):1391-1399. doi:10.1038/mp.2015.197 PubMedGoogle ScholarCrossref
48.
Okbay  A, Baselmans  BM, De Neve  J-E,  et al; LifeLines Cohort Study.  Genetic variants associated with subjective well-being, depressive symptoms, and neuroticism identified through genome-wide analyses.   Nat Genet. 2016;48(6):624-633. doi:10.1038/ng.3552 PubMedGoogle ScholarCrossref
49.
Davies  G, Armstrong  N, Bis  JC,  et al; Generation Scotland.  Genetic contributions to variation in general cognitive function: a meta-analysis of genome-wide association studies in the CHARGE consortium (N=53949).   Mol Psychiatry. 2015;20(2):183-192. doi:10.1038/mp.2014.188 PubMedGoogle ScholarCrossref
50.
Lee  JJ, Wedow  R, Okbay  A,  et al; 23andMe Research Team; COGENT (Cognitive Genomics Consortium); Social Science Genetic Association Consortium.  Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals.   Nat Genet. 2018;50(8):1112-1121. doi:10.1038/s41588-018-0147-3 PubMedGoogle ScholarCrossref
51.
Bolton  JL, Hayward  C, Direk  N,  et al; CORtisol NETwork (CORNET) Consortium.  Genome wide association identifies common variants at the SERPINA6/SERPINA1 locus influencing plasma cortisol and corticosteroid binding globulin.   PLoS Genet. 2014;10(7):e1004474. doi:10.1371/journal.pgen.1004474 PubMedGoogle Scholar
52.
Furberg  H, Kim  Y, Dackor  J,  et al; Tobacco and Genetics Consortium.  Genome-wide meta-analyses identify multiple loci associated with smoking behavior.   Nat Genet. 2010;42(5):441-447. doi:10.1038/ng.571 PubMedGoogle ScholarCrossref
53.
Shungin  D, Winkler  TW, Croteau-Chonka  DC,  et al; ADIPOGen Consortium; CARDIOGRAMplusC4D Consortium; CKDGen Consortium; GEFOS Consortium; GENIE Consortium; GLGC; ICBP; International Endogene Consortium; LifeLines Cohort Study; MAGIC Investigators; MuTHER Consortium; PAGE Consortium; ReproGen Consortium.  New genetic loci link adipose and insulin biology to body fat distribution.   Nature. 2015;518(7538):187-196. doi:10.1038/nature14132 PubMedGoogle ScholarCrossref
54.
Locke  AE, Kahali  B, Berndt  SI,  et al; LifeLines Cohort Study; ADIPOGen Consortium; AGEN-BMI Working Group; CARDIOGRAMplusC4D Consortium; CKDGen Consortium; GLGC; ICBP; MAGIC Investigators; MuTHER Consortium; MIGen Consortium; PAGE Consortium; ReproGen Consortium; GENIE Consortium; International Endogene Consortium.  Genetic studies of body mass index yield new insights for obesity biology.   Nature. 2015;518(7538):197-206. doi:10.1038/nature14177 PubMedGoogle ScholarCrossref
55.
Willer  CJ, Schmidt  EM, Sengupta  S,  et al; Global Lipids Genetics Consortium.  Discovery and refinement of loci associated with lipid levels.   Nat Genet. 2013;45(11):1274-1283. doi:10.1038/ng.2797 PubMedGoogle ScholarCrossref
56.
Ware  E, Schmitz  L, Gard  A, Faul  JD. HRS polygenic scores—release 3, 2006-2012 genetic data. HRS documentation report. Published October 2018. Accessed January 7, 2020. https://hrsonline.isr.umich.edu/modules/meta/xyear/pgs/desc/PGENSCORES3DD.pdf
57.
Ribeiro  MT, Singh  S, Guestrin  C. ”Why should I trust you?”: explaining the predictions of any classifier. Presented at: 22nd Association for Computing Machinery Special Interest Group on Knowledge Discovery and Data Mining International Conference on Knowledge Discovery and Data Mining; August 2016.
58.
Muthén  LK, Muthén  BO.  Mplus (version 7.3). Muthén & Muthén; 2014.
59.
Keras  CF.  The Python Deep Learning Library. Astrophysics Source Code Library; 2018.
60.
Pedregosa  F, Varoquaux  G, Gramfort  A,  et al.  Scikit-learn: machine learning in Python.   J Mach Learn Res. 2011;12:2825-2830.Google Scholar
61.
Kwong  ASF, López-López  JA, Hammerton  G,  et al.  Genetic and environmental risk factors associated with trajectories of depression symptoms from adolescence to young adulthood.   JAMA Netw Open. 2019;2(6):e196587. doi:10.1001/jamanetworkopen.2019.6587 PubMedGoogle Scholar
62.
Choi  KW, Chen  C-Y, Ursano  RJ,  et al  Prospective study of polygenic risk, protective factors, and incident depression following combat deployment in US Army soldiers.   bioRxiv. 2018:361725. doi:10.1101/361725Google Scholar
63.
Choi  KW, Stein  MB, Dunn  EC, Koenen  KC, Smoller  JW.  Genomics and psychological resilience: a research agenda.   Mol Psychiatry. 2019;24(12):1770-1778. doi:10.1038/s41380-019-0457-6 PubMedGoogle ScholarCrossref
64.
Wainberg  M, Merico  D, Delong  A, Frey  BJ.  Deep learning in biomedicine.   Nat Biotechnol. 2018;36(9):829-838. doi:10.1038/nbt.4233 PubMedGoogle ScholarCrossref
65.
Torkamani  A, Wineinger  NE, Topol  EJ.  The personal and clinical utility of polygenic risk scores.   Nat Rev Genet. 2018;19(9):581-590. doi:10.1038/s41576-018-0018-x PubMedGoogle ScholarCrossref
66.
Lipton  ZC.  The mythos of model interpretability.   Queue. 2018;16(3):31-57. doi:10.1145/3236386.3241340Google ScholarCrossref
67.
Howard  DM, Adams  MJ, Clarke  T-K,  et al; 23andMe Research Team; Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium.  Genome-wide meta-analysis of depression identifies 102 independent variants and highlights the importance of the prefrontal brain regions.   Nat Neurosci. 2019;22(3):343-352. doi:10.1038/s41593-018-0326-7 PubMedGoogle ScholarCrossref
Limit 200 characters
Limit 25 characters
Conflicts of Interest Disclosure

Identify all potential conflicts of interest that might be relevant to your comment.

Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.

Err on the side of full disclosure.

If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.

Not all submitted comments are published. Please see our commenting policy for details.

Limit 140 characters
Limit 3600 characters or approximately 600 words
    Original Investigation
    March 31, 2021

    Discriminating Heterogeneous Trajectories of Resilience and Depression After Major Life Stressors Using Polygenic Scores

    Author Affiliations
    • 1Department of Emergency Medicine, Columbia University Irving Medical Center, New York, New York
    • 2Data Science Institute, Columbia University, New York, New York
    • 3Department of Psychiatry, NYU Grossman School of Medicine, New York, New York
    • 4Department of Psychiatry, Massachusetts General Hospital, Boston
    • 5Department of Psychiatry, NYU Grossman School of Medicine, New York, New York
    • 6Teachers College, Columbia University, New York, New York
    JAMA Psychiatry. 2021;78(7):744-752. doi:10.1001/jamapsychiatry.2021.0228
    Key Points

    Question  Is it possible to accurately discriminate longitudinal trajectories of depression and resilience by using multiple polygenic scores as psychiatric risk and health indicators?

    Findings  In this longitudinal cohort study including 2071 participants, resilience and symptomatic trajectories were accurately discriminated using 21 polygenic scores using deep neural nets. The resilience trajectory was associated with lower polygenic scores for several psychiatry disorders as well as metabolic risk.

    Meaning  The results of this study suggest that polygenic scores can be used to determine long-term risk for depression and resilience.

    Abstract

    Importance  Major life stressors, such as loss and trauma, increase the risk of depression. It is known that individuals show heterogeneous trajectories of depressive symptoms following major life stressors, including chronic depression, recovery, and resilience. Although common genetic variation has been associated with depression risk, genomic factors that could help discriminate trajectories of risk vs resilience following adversity have not been identified.

    Objective  To assess the discriminatory accuracy of a deep neural net combining joint information from 21 psychiatric and health-related multiple polygenic scores (PGSs) for discriminating resilience vs other longitudinal symptom trajectories with use of longitudinal, genetically informed data on adults exposed to major life stressors.

    Design, Setting, and Participants  The Health and Retirement Study is a longitudinal panel cohort study in US citizens older than 50 years, with data being collected once every 2 years between 1992 and 2010. A total of 2071 participants who were of European ancestry with available depressive symptom trajectory information after experiencing an index depressogenic major life stressor were included. Latent growth mixture modeling identified heterogeneous trajectories of depressive symptoms before and after major life stressors, including stable low symptoms (ie, resilience), as well as improving, emergent, and preexisting/chronic symptom patterns. Twenty-one PGSs were examined as factors distinctively associated with these heterogeneous trajectories. Local interpretable model-agnostic explanations were applied to examine PGSs associated with each trajectory. Data were analyzed using the DNN model from June to July 2020.

    Exposures  Development of depression and resilience were examined in older adults after a major life stressor, such as bereavement, divorce, and job loss, or major health events, such as myocardial infarction and cancer.

    Main Outcomes and Measures  Discriminatory accuracy of a deep neural net model trained for the multinomial classification of 4 distinct trajectories of depressive symptoms (Center for Epidemiologic Studies–Depression scale) based on 21 PGSs using supervised machine learning.

    Results  Of the 2071 participants, 1329 were women (64.2%); mean (SD) age was 55.96 (8.52) years. Of these, 1638 (79.1%) were classified as resilient, 160 (7.75) in recovery (improving), 159 (7.7%) with emerging depression, and 114 (5.5%) with preexisting/chronic depression symptoms. Deep neural nets distinguished these 4 trajectories with high discriminatory accuracy (multiclass micro-average area under the curve, 0.88; 95% CI, 0.87-0.89; multiclass macro-average area under the curve, 0.86; 95% CI, 0.85-0.87). Discriminatory accuracy was highest for preexisting/chronic depression (AUC 0.93), followed by emerging depression (AUC 0.88), recovery (AUC 0.87), resilience (AUC 0.75).

    Conclusions and Relevance  The results of the longitudinal cohort study suggest that multivariate PGS profiles provide information to accurately distinguish between heterogeneous stress-related risk and resilience phenotypes.

    Introduction

    Exposure to major life stressors, such as bereavement,1 divorce,2 and job loss,3 can increase the risk of major depression.4 Similarly, major health events, such as myocardial infarction5 and cancer,6 can increase the risk for psychiatric disorders. Nonetheless, it is well established that psychological responses to such events tend to follow heterogeneous symptom trajectories.7,8 Some individuals exposed to major life stressors will exhibit persistent symptom elevations (chronic), while others will show initially high symptoms that decrease with time (recovery) or low-to-moderate symptoms that worsen with time (emergent). However, the most common trajectory is one of stable good mental health (resilience).7

    Although a number of factors associated with these trajectory patterns have been identified—ranging from personality and behavioral factors to neurobiological markers9-12—their individual effects have been modest, suggesting that other key factors may be needed to increase explanatory power.6,13,14 One potential yet unexplored strategy for discriminating resilience and risk trajectories is using genetic factors that may underlie these potential adaptive mechanisms. The regularity with which these trajectories have been observed across a wide range of adverse life events7,13,14 is consistent with possible underlying genetic factors. Moreover, evidence suggests that resilience, although multifactorial, has a notable genetic component. Twin studies indicate that 31% to 52% of observed variance in resilience phenotypes (eg, positive psychological functioning despite life stressors) can be explained by genetic differences.15-17 However, the extent to which genetic data can be used to accurately discriminate more precisely defined longitudinal trajectories of resilience to adversity is not well understood.

    Genome-wide association studies (GWAS) have been used to estimate associations between millions of genetic variants and diverse phenotypes, ranging from health conditions to psychiatric disorders such as schizophrenia and major depression.18-23 Polygenic scores (PGSs)24 aggregate genome-wide contributions into an overall score reflecting an individual’s overall genetic propensity for a given trait or disorder25 and have been shown to explain the risk of various psychiatric disorders.26-28 In addition, PGS have been used as genetic proxies of health-related, cognitive, and mental health traits to enable the identification of underlying factors that might contribute to psychosocial outcomes.29 Although the explanatory power of individual PGSs is modest, combining multiple PGSs has been applied to increase the predictive power for psychosocial outcomes.29,30 Despite the utility of combining multiple PGSs, to our knowledge, this approach has not yet been used to distinguish multinomial trajectory outcomes following adversity. Doing so can lead to several benefits, including a better understanding of the utility of genetic factors associated with resilience and enhancing the programmatic identification and allocation of clinical resources for those in need. However, to date, genomic studies of psychological resilience have relied on outcomes defined by cross-sectional designs or, at most, prospective studies with data from baseline and a single follow-up.31

    To address this gap, powerful computational methods are required that combine multiple PGSs and can accurately discriminate between longitudinal trajectories of resilience and depression after major life stressors. Data-driven deep learning is particularly useful for handling the joint probabilistic information of multiple PGSs. This method allows for potential nonlinear and higher-order dependencies, handles multicollinearity, requires no assumptions about the association between different PGSs, and is well equipped to maximize the discriminatory accuracy required for the multinomial classification of heterogeneous trajectories.

    This study used large-scale GWAS results for psychiatric and health phenotypes and a genetically informed, longitudinal cohort of adults exposed to major life stressors. Specifically, trajectories of depression symptoms were based on the results of 5 previous studies in a nationally representative cohort of adults1-3,5,6 in which depression has been measured before and following major life stressors. Using these data, we tested the discriminatory accuracy of a deep learning model combining joint information from 21 psychiatric and health-related PGSs for discriminating resilience vs other longitudinal trajectory patterns.

    Methods

    We included data from 2071 participants from the US nationally representative cohort of older adults in the Health and Retirement Study (HRS).32,33 The HRS is a longitudinal prospective study of US citizens born between 1931 and 1947 with data collected once every 2 years between 1992 and 2010.32 Data were analyzed using the DNN model from June to July 2020. Participants were surveyed on mental and physical health-related aspects as described in detail elsewhere.34 The HRS measures depressive symptom severity using the abbreviated 10-item version35 of the Center for Epidemiologic Studies–Depression (CES-D) scale.36 The optimal cutoff score of the 10-item CES-D is 4, with scores greater than or equal to 4 indicating depression with a sensitivity of 97%, a specificity of 84%, and a positive predictive value of 85%.37 For the present study, we included all participants of European ancestry with available depressive symptom trajectory information who experienced 1 of the following index depressogenic major life stressors: bereavement,1 myocardial infarction,5 divorce,2 cancer,6 or job loss.3 Participants who experienced more than 1 of these major life stressors during this period were excluded from the analysis to define a cohort of patients with a single index event as a common reference point for the longitudinal assessment of pre- and post-event follow-up. eFigure 1 in the Supplement presents a flowchart describing the sample selection. The HRS data are deidentified and publicly available, and all participants provided written informed consent in the HRS study; participants received financial compensation in the HRS study. Our study was a secondary data analysis for which no reimbursement was paid.32,33 This secondary data analysis was exempt from institutional review board approval in accordance with the policies of the New York University Institutional Review Board and the Teachers College Columbia University Institutional Review Board.

    The outcome of interest was defined as the longitudinal trajectory course of depressive symptoms measured before and in the years following an index major life stressor. We combined data from 5 different samples from the HRS1-3,5,6 that all previously identified the same 4 prospective trajectories of depressive symptoms: resilience, characterized as a stable trajectory of low symptom severity before and following the index event; recovery, characterized by initial clinically elevated symptoms that steadily decrease following the index event; emerging depression, characterized by low to moderate symptoms that increased above the clinically significant threshold following the index event; and preexisting and chronic depression marked by clinically elevated symptoms before and subsequent to the index event. All trajectories were identified using latent growth mixture modeling38 with a floating baseline method39 in which participants’ data were centered on the year of the index event.39 Figure 1 shows depression symptom trajectories at 1 time point before and 2 time points following the index stressor. In the present study, we examined all participants previously assigned to 1 of these 4 latent growth mixture modeling trajectories with available PGSs. Further details are reported in the eMethods and eResults in the Supplement.

    We included 21 different PGSs as candidate features (Table 1). These PGSs were selected for their potential relevance for discriminating heterogeneous stress responses based on previous analyses and publications1-3,5,6 of the HRS data set.40 Details on genomic data processing and PGS construction in the HRS are provided elsewhere.56 A brief description is presented in the eMethods in the Supplement.

    Statistical Analysis

    We applied supervised learning for multinomial classification using a multilayer feedforward neural network to train a deep neural net (DNN). The primary outcome of this study was the multinomial classification task to differentiate the 4 symptom trajectories based on the 21 PGSs (Table 1), age, and the types of major life stressors. Benchmark models as well as further technical details about feature preprocessing, model development, and model validation are presented in the eMethods and eResults in the Supplement.

    To examine why the model assigned each participant to a given latent trajectory of depression symptoms, we applied methods for explainable machine learning. Local interpretable model-agnostic explanations (LIME) via submodular optimization57 were used for the explanations of classifications in human-interpretable form by approximating the output of the DNN model locally with penalized general linear models. The results were averaged per latent growth mixture modeling class to estimate which set of features, on average, influenced the classification of depressive symptom trajectories the most. The trajectories were built using MPlus, version 7.3,58 the DNN was built using Keras Tensorflow, version 2.1.0 in Python 3.7,59 and nested cross-validation was performed using Scikit-learn, version 0.23.60

    Results

    We examined a set of 21 PGSs as candidate features (Table 1) in a sample of 2071 participants from the HRS cohort. The cohort included 1329 women (64.2%) and 739 men (35.7%); mean (SD) age was 55.96 (8.52) years. Table 2 provides other descriptive statistics, including years of education, for each trajectory.

    Neural networks achieved good discriminatory power to distinguish all 4 trajectories with a multiclass macro-average AUC of 0.86 (95% CI, 0.85-0.87) and a micro-average AUC of 0.88 (95% CI, 0.87-0.89) (Figure 2A; average precision, 0.79; average recall, 0.60; average specificity, 0.82; average F1, 0.64; average geometric mean, 0.70; and average index balanced accuracy, 0.48). All 4 trajectories were classified with high discriminatory accuracy (Figure 2B). A comparison of the nested vs nonnested cross-validation performance is shown in eFigure 2 on the Supplement.

    The 15 most important features identified using LIME for each trajectory are shown in Figure 3. LIME feature importance should not be considered a true explanation but as a heuristic approach that can lead to novel hypotheses about the input-output association.

    The LIME results (Figure 3) indicated that each trajectory was characterized by a unique profile of PGSs. For example, membership in the resilience trajectory was positively associated (LIME >0) with lower PGSs for schizophrenia (≤0.39) and attention-deficit/hyperactivity disorder (≤0.46), lower-to-medium PGSs for posttraumatic stress disorder (>0.29 to ≤0.65), lower PGSs for neuroticism (>0.22 to ≤0.62), lower PGSs for waist circumference (≤0.36) and body mass index (>0.17 to ≤0.64), and high PGSs for educational attainment (>0.60). The resilience trajectory was negatively associated (LIME <0) with lower PGSs for well-being (≤0.66) and higher PGSs for total cholesterol level (>0.56) and extraversion (>0.50). By contrast, membership in the emerging depression trajectory was positively associated (LIME >0) with higher PGSs for schizophrenia (>0.57 to ≤0.80) and attention-deficit/hyperactivity disorder (>0.47 to ≤0.68) as well as body mass index (>0.64 to ≤0.89), and the recovery trajectory was associated with higher PGSs for schizophrenia (>0.57 to ≤0.80), and attention-deficit/hyperactivity disorder (>0.47 to ≤0.68), and lower PGSs for depressive symptoms (>0.30 to ≤0.57), and negatively associated (LIME <0) with higher major depressive disorder (>0.38 to ≤0.61) PGSs, and the preexisting/chronic depression trajectory was associated with higher PGSs for depressive symptoms (>0.61), and anxiety symptoms (>0.26 to ≤0.65), as well as lower educational level attainment PGSs (≤0.25).

    Discussion

    To our knowledge, this study is the first investigation of polygenic contributions to longitudinal trajectories of risk and resilience following major life stressors. Previous work has identified a set of heterogeneous patterns of response to stressful life events, ranging from chronically elevated depressive symptoms to the stable absence of such symptoms (ie, resilience).1-3,5,6 The consistency with which these trajectory patterns have been identified across diverse stressor events suggests a plausible genetic basis. Using DNNs to combine PGSs for a range of health and psychiatric traits, we were able to accurately classify longitudinal trajectories of depression-related risk and resilience following a major life stressor. The prognosis of depressive symptoms in the aftermath of major life stressors is clinically important and the accurate discrimination between distinct trajectories such as resilience and emerging depressive symptoms potentially opens new windows for targeted interventions across time. Our results show that individual PGSs alone were not able to discriminate the 4 trajectories in this sample (eResults in the Supplement). By combining different PGSs for a range of psychiatric and health-related characteristics, the computational power of a DNN model increased discriminatory accuracy and allowed us to distinguish these heterogeneous outcome patterns. Our results further show that the classification of the preexisting chronic depression was the most accurate while the classification of resilience was most difficult. This finding suggests that resilience is a more complex construct that is influenced by many risk and protective factors and it seems that the genetic component can explain only part of it, whereas in chronic depression the contribution of genetic factors for the model’s performance seems to be more pronounced.

    Further research to elucidate the association between PGSs and resilience is necessary to probe the mechanistic underpinnings that underlie these findings. Although such explanation is beyond the methodologic approach of this study, our results nonetheless provide a demonstration of value of multiple PGSs in accurately discriminating depression and resilience. Moreover, although PGSs do not provide specific mechanistic information, they provide accessible proxy information that may help researchers better target systems for further investigation. For example, a PGS of educational level reflects several single-nucleotide variants that represent systems involved in cognitive function and susceptibility to environmental stress. Thus, the PGS approach offers 2 advantages: the ability to compare genetic neurobiological systems in a multivariate manner and the ability to reduce the dimensionality of genetic information so that it can be used for a workable clinical risk profile.

    Previous work has shown that polygenic risk for depression is associated with depression symptom trajectories across time61 and also separately with the risk of depression after stressful life events.62 However, to our knowledge, no published research has examined polygenic influences on depression symptom trajectories following exposure to stressful life events. Resilience to adversity is best understood as a longitudinal process rather than discrete points in time,8,13,14 but few genetic studies have been able to examine resilience in this way.63 By evaluating polygenic contributions to symptom trajectories in the aftermath of stressful life events, we can assess the relevance of genetic factors for their ability to discriminate longitudinal patterns of psychological response to stress, including resilience. We found that multinomial logistic regression models based only on PGSs for major depression and depressive symptoms were unable to discriminate between the 4 identified risk and resilience trajectories. Rather, as our results suggest, a broader range of PGSs and a computationally more complex model, such as deep learning, are better suited to account for such nonlinear dependencies.

    The main advantage of our DNN approach is the ability to identify and use information in the data that is a priori unknown. Deep neural nets produce nonlinear mappings between the values of the candidate features and the outcome of interest.64 Currently, genotype-phenotype association studies often use linear additive models to assess polygenic influences without accounting for potentially more complex interactions among variables, as is the case in DNNs.64 Using flexible DNNs, we explored the ability for multiple PGSs to discriminate different trajectories of depressive symptoms and resilience. Our results suggest that, although individual PGSs show limited utility for discriminating longitudinal trajectories of risk vs resilience, combining multiple PGSs may yield informative probabilistic information for this task.

    As a trade-off, a limitation of the computationally more powerful DNN approach is that results cannot be easily interpreted as linear associations between an outcome and a limited set of PGSs. Potential higher-order interactions and nonlinear association complicate the interpretations, and we did not test such higher-order interactions and nonlinear associations, nor do we see a feasible and theoretically sound possibility of modeling such associations in a cogent framework. Although deep learning results tend to be computationally demanding and difficult to interpret, promising methods have emerged to enhance interpretability of DNNs, including the strategy we used (LIME via submodular optimization) to approximate the deep learning model and estimate by rank order which variable—at a specific range of values—is most important for the classification task.

    Although the identified features should not be interpreted as etiologic factors, several associations are noteworthy. For instance, lower polygenic risk for schizophrenia and attention-deficit/hyperactivity disorder were relevant for discriminating the resilience trajectory from other trajectories, and higher polygenic risk for depression was associated with long-lasting depressive symptoms (ie, the chronic symptom trajectory). Other relevant factors associated with a resilience trajectory included higher PGSs for well-being and educational attainment, lower neuroticism PGSs, lower body mass index and waist circumference PGSs, and lower total cholesterol level PGSs. Taken together, the associations suggest that lower vulnerability for psychiatric risk and greater likelihood for well-being as indicated by PGSs increase the chances of enduring psychological health after major life stressors. Overall, these findings align with previous research on psychiatric, cognitive, and biological factors that may be relevant for stress responses. However, owing to the exploratory nature of our DNN analysis, further studies using a null-hypothesis significance testing design are needed to establish such associations with scientific confidence. The goal of this study was to assess for what we believe to be the first time whether a combination of PGSs could show utility for discriminating longitudinal trajectories of risk vs resilience. The findings suggest that polygenic information for a range of psychiatric and health traits can reveal probabilistic information to help identify subgroups of individuals who could benefit from the knowledge of their most likely response to major life stressors, including resilience, that may guide the targeting of preventive strategies.65

    Strengths and Limitations

    There are several limitations of this study. First, DNNs are less transparent compared with linear statistical models.66 As mentioned, we used the most up-to-date methods to provide interpretable estimates. Second, we used the most recent PGSs derived and made available for research by the HRS study team.56 For some traits, there may be even more updated GWAS based on larger samples. However, we selected the most recent available PGSs in the database where possible and provide these citations in Table 1. Although the HRS study yields comprehensive information about PGSs that have been used to build the DNN model, there are additional potentially relevant PGSs, such as more recently published major depressive disorder and other psychiatric GWAS67 that could further increase the discriminatory accuracy of this approach. Despite these limitations, this study was conducted using a relatively large sample of adults from a well-reputed longitudinal cohort with extensive genomic and phenotypic data, allowing us to test polygenic influences on outcome trajectories following major life stressors.

    Conclusions

    Prospective longitudinal investigations that capture changes over time are ideal for examining the genetic influences on resilience. Drawing on a prospective population-based sample of older adults exposed to major life stressors, our results suggest how genetic information may be used to identify protective genomic factors of resilience. The algorithm can be used to discriminate distinct trajectories of depressive symptoms in response to major life stressors as diverse as bereavement, job loss, divorce, myocardial infarction, or cancer. A focus on resilience is important as it helps to identify individuals who have a lower propensity to experience stress-related psychiatric morbidity across time. This information is useful because it might lead to retargeting individuals who may benefit more from intervention and may help to prevent overtreatment or less-efficient enrollment in clinical research. To our knowledge, this represents the first investigation of the discriminatory ability of PGSs for heterogeneous trajectories following major life stressors and would benefit from replication efforts by future studies that combine genomic and longitudinal outcome data following major life stressors. Because this data-driven study is inherently exploratory, external validation of the findings is an important next step and a prerequisite before the clinical use of the model is justified. Owing to the importance of accurately distinguishing between resilience and risk for emergent depressive symptoms following major life stressors, the presented approach to combine multiple PGSs using computational methods may be a useful approach for developing prognostic models that have potential to provide new areas for targeted interventions over time.

    Back to top
    Article Information

    Accepted for Publication: February 4, 2021.

    Published Online: March 31, 2021. doi:10.1001/jamapsychiatry.2021.0228

    Corresponding Author: Katharina Schultebraucks, PhD, Department of Emergency Medicine, Columbia University Medical Center, 622 W 168th St, New York, NY 10032 (ks3796@cumc.columbia.edu).

    Author Contributions: Drs Schultebraucks and Bonanno had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

    Concept and design: All authors.

    Acquisition, analysis, or interpretation of data: All authors.

    Drafting of the manuscript: All authors.

    Critical revision of the manuscript for important intellectual content: All authors.

    Statistical analysis: Schultebraucks, Galatzer-Levy.

    Obtained funding: Bonanno.

    Administrative, technical, or material support: Galatzer-Levy, Bonanno.

    Supervision: Bonanno.

    Conflict of Interest Disclosures: Dr Galatzer-Levy has received a grant from the National Institute of Mental Health (K01MH102415). No other disclosures were reported.

    References
    1.
    Maccallum  F, Galatzer-Levy  IR, Bonanno  GA.  Trajectories of depression following spousal and child bereavement: a comparison of the heterogeneity in outcomes.   J Psychiatr Res. 2015;69:72-79. doi:10.1016/j.jpsychires.2015.07.017 PubMedGoogle ScholarCrossref
    2.
    Malgaroli  M, Galatzer-Levy  IR, Bonanno  GA.  Heterogeneity in trajectories of depression in response to divorce is associated with differential risk for mortality.   Clin Psychol Sci. 2017;5(5):843-850. doi:10.1177/2167702617705951 PubMedGoogle ScholarCrossref
    3.
    Stolove  CA, Galatzer-Levy  IR, Bonanno  GA.  Emergence of depression following job loss prospectively predicts lower rates of reemployment.   Psychiatry Res. 2017;253:79-83. doi:10.1016/j.psychres.2017.03.036 PubMedGoogle ScholarCrossref
    4.
    Breslau  N.  Epidemiologic studies of trauma, posttraumatic stress disorder, and other psychiatric disorders.   Can J Psychiatry. 2002;47(10):923-929. doi:10.1177/070674370204701003 PubMedGoogle ScholarCrossref
    5.
    Galatzer-Levy  IR, Bonanno  GA.  Optimism and death: predicting the course and consequences of depression trajectories in response to heart attack.   Psychol Sci. 2014;25(12):2177-2188. doi:10.1177/0956797614551750 PubMedGoogle ScholarCrossref
    6.
    Burton  CL, Galatzer-Levy  IR, Bonanno  GA.  Treatment type and demographic characteristics as predictors for cancer adjustment: prospective trajectories of depressive symptoms in a population sample.   Health Psychol. 2015;34(6):602-609. doi:10.1037/hea0000145 PubMedGoogle ScholarCrossref
    7.
    Galatzer-Levy  IR, Huang  SH, Bonanno  GA.  Trajectories of resilience and dysfunction following potential trauma: a review and statistical evaluation.   Clin Psychol Rev. 2018;63:41-55. doi:10.1016/j.cpr.2018.05.008 PubMedGoogle ScholarCrossref
    8.
    Bonanno  GA.  Loss, trauma, and human resilience: have we underestimated the human capacity to thrive after extremely aversive events?   Am Psychol. 2004;59(1):20-28. doi:10.1037/0003-066X.59.1.20 PubMedGoogle ScholarCrossref
    9.
    Schultebraucks  K, Shalev  AY, Michopoulos  V,  et al.  A validated predictive algorithm of post-traumatic stress course following emergency department admission after a traumatic stressor.   Nat Med. 2020;26(7):1084-1088. doi:10.1038/s41591-020-0951-z PubMedGoogle ScholarCrossref
    10.
    Schultebraucks  K, Qian  M, Abu-Amara  D,  et al.  Pre-deployment risk factors for PTSD in active-duty personnel deployed to Afghanistan: a machine-learning approach for analyzing multivariate predictors.   Mol Psychiatry. 2020;1-12.PubMedGoogle Scholar
    11.
    Hinrichs  R, van Rooij  SJ, Michopoulos  V,  et al.  Increased skin conductance response in the immediate aftermath of trauma predicts PTSD risk.   Chronic Stress (Thousand Oaks). 2019;3:2470547019844441. doi:10.1177/2470547019844441 PubMedGoogle Scholar
    12.
    Michopoulos  V, Beurel  E, Gould  F,  et al.  Association of prospective risk for chronic PTSD symptoms with low TNFα and IFNγ concentrations in the immediate aftermath of trauma exposure.   Am J Psychiatry. 2020;177(1):58-65. doi:10.1176/appi.ajp.2019.19010039 PubMedGoogle ScholarCrossref
    13.
    Bonanno  GA, Romero  SA, Klein  SI.  The temporal elements of psychological resilience: an integrative framework for the study of individuals, families, and communities.   Psychol Inquir. 2015;26(2):139-169. doi:10.1080/1047840X.2015.992677Google ScholarCrossref
    14.
    Bonanno  GA, Westphal  M, Mancini  AD.  Resilience to loss and potential trauma.   Annu Rev Clin Psychol. 2011;7:511-535. doi:10.1146/annurev-clinpsy-032210-104526 PubMedGoogle ScholarCrossref
    15.
    Boardman  JD, Blalock  CL, Button  TM.  Sex differences in the heritability of resilience.   Twin Res Hum Genet. 2008;11(1):12-27. doi:10.1375/twin.11.1.12 PubMedGoogle ScholarCrossref
    16.
    Amstadter  AB, Myers  JM, Kendler  KS.  Psychiatric resilience: longitudinal twin study.   Br J Psychiatry. 2014;205(4):275-280. doi:10.1192/bjp.bp.113.130906 PubMedGoogle ScholarCrossref
    17.
    Waaktaar  T, Torgersen  S.  Genetic and environmental causes of variation in trait resilience in young people.   Behav Genet. 2012;42(3):366-377. doi:10.1007/s10519-011-9519-5 PubMedGoogle ScholarCrossref
    18.
    Schizophrenia Working Group of the Psychiatric Genomics Consortium.  Biological insights from 108 schizophrenia-associated genetic loci.   Nature. 2014;511(7510):421-427. doi:10.1038/nature13595 PubMedGoogle ScholarCrossref
    19.
    Hyde  CL, Nagle  MW, Tian  C,  et al.  Identification of 15 genetic loci associated with risk of major depression in individuals of European descent.   Nat Genet. 2016;48(9):1031-1036. doi:10.1038/ng.3623 PubMedGoogle ScholarCrossref
    20.
    Psychiatric GWAS Consortium Bipolar Disorder Working Group.  Large-scale genome-wide association analysis of bipolar disorder identifies a new susceptibility locus near ODZ4.   Nat Genet. 2011;43(10):977-983. doi:10.1038/ng.943 PubMedGoogle ScholarCrossref
    21.
    Robinson  EB, St Pourcain  B, Anttila  V,  et al; iPSYCH-SSI-Broad Autism Group.  Genetic risk for autism spectrum disorders and neuropsychiatric variation in the general population.   Nat Genet. 2016;48(5):552-555. doi:10.1038/ng.3529 PubMedGoogle ScholarCrossref
    22.
    Nagel  M, Jansen  PR, Stringer  S,  et al; 23andMe Research Team.  Meta-analysis of genome-wide association studies for neuroticism in 449,484 individuals identifies novel genetic loci and pathways.   Nat Genet. 2018;50(7):920-927. doi:10.1038/s41588-018-0151-7 PubMedGoogle ScholarCrossref
    23.
    Watanabe  K, Taskesen  E, van Bochoven  A, Posthuma  D.  Functional mapping and annotation of genetic associations with FUMA.   Nat Commun. 2017;8(1):1826. doi:10.1038/s41467-017-01261-5 PubMedGoogle ScholarCrossref
    24.
    Wray  NR, Goddard  ME, Visscher  PM.  Prediction of individual genetic risk to disease from genome-wide association studies.   Genome Res. 2007;17(10):1520-1528. doi:10.1101/gr.6665407 PubMedGoogle ScholarCrossref
    25.
    Sugrue  LP, Desikan  RS.  What are polygenic scores and why are they important?   JAMA. 2019;321(18):1820-1821. doi:10.1001/jama.2019.3893 PubMedGoogle ScholarCrossref
    26.
    Dudbridge  F.  Power and predictive accuracy of polygenic risk scores.   PLoS Genet. 2013;9(3):e1003348. doi:10.1371/journal.pgen.1003348 PubMedGoogle Scholar
    27.
    Khera  AV, Chaffin  M, Aragam  KG,  et al.  Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations.   Nat Genet. 2018;50(9):1219-1224. doi:10.1038/s41588-018-0183-z PubMedGoogle ScholarCrossref
    28.
    Wray  NR, Lee  SH, Mehta  D, Vinkhuyzen  AA, Dudbridge  F, Middeldorp  CM.  Research review: polygenic methods and their application to psychiatric traits.   J Child Psychol Psychiatry. 2014;55(10):1068-1087. doi:10.1111/jcpp.12295 PubMedGoogle ScholarCrossref
    29.
    Schoeler  T, Choi  SW, Dudbridge  F,  et al.  Multi–polygenic score approach to identifying individual vulnerabilities associated with the risk of exposure to bullying.   JAMA Psychiatry. 2019;76(7):730-738. doi:10.1001/jamapsychiatry.2019.0310 PubMedGoogle ScholarCrossref
    30.
    Krapohl  E, Patel  H, Newhouse  S,  et al.  Multi-polygenic score approach to trait prediction.   Mol Psychiatry. 2018;23(5):1368-1374. doi:10.1038/mp.2017.163 PubMedGoogle ScholarCrossref
    31.
    Stein  MB, Choi  KW, Jain  S,  et al.  Genome-wide analyses of psychological resilience in U.S. Army soldiers.   Am J Med Genet B Neuropsychiatr Genet. 2019;180(5):310-319. doi:10.1002/ajmg.b.32730 PubMedGoogle ScholarCrossref
    32.
    Sonnega  A, Faul  JD, Ofstedal  MB, Langa  KM, Phillips  JW, Weir  DR.  Cohort Profile: the Health and Retirement Study (HRS).   Int J Epidemiol. 2014;43(2):576-585. doi:10.1093/ije/dyu067 PubMedGoogle ScholarCrossref
    33.
    Juster  FT, Suzman  R.  An overview of the Health and Retirement Study.   J Hum Resour. 1995;30:S7-S56. doi:10.2307/146277Google ScholarCrossref
    34.
    Steffick  DE. Documentation of affective functioning measures in the Health and Retirement Study: HRS/AHEAD Documentation Report 2000. Accessed February 16, 2021. https://hrs.isr.umich.edu/publications/biblio/5411
    35.
    Kohout  FJ, Berkman  LF, Evans  DA, Cornoni-Huntley  J.  Two shorter forms of the CES-D (Center for Epidemiological Studies Depression) depression symptoms index.   J Aging Health. 1993;5(2):179-193. doi:10.1177/089826439300500202 PubMedGoogle ScholarCrossref
    36.
    Radloff  LS.  The use of the Center for Epidemiologic Studies Depression Scale in adolescents and young adults.   J Youth Adolesc. 1991;20(2):149-166. doi:10.1007/BF01537606 PubMedGoogle ScholarCrossref
    37.
    Irwin  M, Artin  KH, Oxman  MN.  Screening for depression in the older adult: criterion validity of the 10-item Center for Epidemiological Studies Depression Scale (CES-D).   Arch Intern Med. 1999;159(15):1701-1704. doi:10.1001/archinte.159.15.1701 PubMedGoogle ScholarCrossref
    38.
    Muthén  L, Muthén  B. Mplus. Statistical analysis with latent variables: user’s guide. Accessed February 16, 2021. http://www.statmodel.com/download/usersguide/Mplus%20user%20guide%20Ver_7_r3_web.pdf
    39.
    Galatzer-Levy  IR, Bonanno  GA, Mancini  AD.  From marianthal to latent growth mixture modeling: a return to the exploration of individual differences in response to unemployment.   J Neuroscience Psychology Econ. 2010;3(2):116. doi:10.1037/a0020077Google ScholarCrossref
    40.
    Ware  E, Schmitz  L, Faul  JD. HRS polygenic scores 2006-2010: genetic data. HRS documentation report. Accessed February 16, 2021. https://hrs.isr.umich.edu/news/hrs-polygenic-scores-2006-2010-genetic-data
    41.
    Wray  NR, Ripke  S, Mattheisen  M,  et al; eQTLGen; 23andMe; Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium.  Genome-wide association analyses identify 44 risk variants and refine the genetic architecture of major depression.   Nat Genet. 2018;50(5):668-681. doi:10.1038/s41588-018-0090-3 PubMedGoogle ScholarCrossref
    42.
    Duncan  LE, Ratanatharathorn  A, Aiello  AE,  et al.  Largest GWAS of PTSD (N=20 070) yields genetic overlap with schizophrenia and sex differences in heritability.   Mol Psychiatry. 2018;23(3):666-673. doi:10.1038/mp.2017.77 PubMedGoogle ScholarCrossref
    43.
    Demontis  D, Walters  RK, Martin  J,  et al; ADHD Working Group of the Psychiatric Genomics Consortium (PGC); Early Lifecourse & Genetic Epidemiology (EAGLE) Consortium; 23andMe Research Team.  Discovery of the first genome-wide significant risk loci for attention deficit/hyperactivity disorder.   Nat Genet. 2019;51(1):63-75. doi:10.1038/s41588-018-0269-7 PubMedGoogle ScholarCrossref
    44.
    Arnold  PD, Askland  KD, Barlassina  C,  et al; International Obsessive Compulsive Disorder Foundation Genetics Collaborative (IOCDF-GC) and OCD Collaborative Genetics Association Studies (OCGAS).  Revealing the complex genetic architecture of obsessive-compulsive disorder using meta-analysis.   Mol Psychiatry. 2018;23(5):1181-1188. doi:10.1038/mp.2017.154 PubMedGoogle ScholarCrossref
    45.
    de Moor  MH, van den Berg  SM, Verweij  KJ,  et al; Genetics of Personality Consortium.  Meta-analysis of genome-wide association studies for neuroticism, and the polygenic association with major depressive disorder.   JAMA Psychiatry. 2015;72(7):642-650. doi:10.1001/jamapsychiatry.2015.0554 PubMedGoogle ScholarCrossref
    46.
    Ripke  S, Wray  NR, Lewis  CM,  et al; Major Depressive Disorder Working Group of the Psychiatric GWAS Consortium.  A mega-analysis of genome-wide association studies for major depressive disorder.   Mol Psychiatry. 2013;18(4):497-511. doi:10.1038/mp.2012.21 PubMedGoogle ScholarCrossref
    47.
    Otowa  T, Hek  K, Lee  M,  et al.  Meta-analysis of genome-wide association studies of anxiety disorders.   Mol Psychiatry. 2016;21(10):1391-1399. doi:10.1038/mp.2015.197 PubMedGoogle ScholarCrossref
    48.
    Okbay  A, Baselmans  BM, De Neve  J-E,  et al; LifeLines Cohort Study.  Genetic variants associated with subjective well-being, depressive symptoms, and neuroticism identified through genome-wide analyses.   Nat Genet. 2016;48(6):624-633. doi:10.1038/ng.3552 PubMedGoogle ScholarCrossref
    49.
    Davies  G, Armstrong  N, Bis  JC,  et al; Generation Scotland.  Genetic contributions to variation in general cognitive function: a meta-analysis of genome-wide association studies in the CHARGE consortium (N=53949).   Mol Psychiatry. 2015;20(2):183-192. doi:10.1038/mp.2014.188 PubMedGoogle ScholarCrossref
    50.
    Lee  JJ, Wedow  R, Okbay  A,  et al; 23andMe Research Team; COGENT (Cognitive Genomics Consortium); Social Science Genetic Association Consortium.  Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals.   Nat Genet. 2018;50(8):1112-1121. doi:10.1038/s41588-018-0147-3 PubMedGoogle ScholarCrossref
    51.
    Bolton  JL, Hayward  C, Direk  N,  et al; CORtisol NETwork (CORNET) Consortium.  Genome wide association identifies common variants at the SERPINA6/SERPINA1 locus influencing plasma cortisol and corticosteroid binding globulin.   PLoS Genet. 2014;10(7):e1004474. doi:10.1371/journal.pgen.1004474 PubMedGoogle Scholar
    52.
    Furberg  H, Kim  Y, Dackor  J,  et al; Tobacco and Genetics Consortium.  Genome-wide meta-analyses identify multiple loci associated with smoking behavior.   Nat Genet. 2010;42(5):441-447. doi:10.1038/ng.571 PubMedGoogle ScholarCrossref
    53.
    Shungin  D, Winkler  TW, Croteau-Chonka  DC,  et al; ADIPOGen Consortium; CARDIOGRAMplusC4D Consortium; CKDGen Consortium; GEFOS Consortium; GENIE Consortium; GLGC; ICBP; International Endogene Consortium; LifeLines Cohort Study; MAGIC Investigators; MuTHER Consortium; PAGE Consortium; ReproGen Consortium.  New genetic loci link adipose and insulin biology to body fat distribution.   Nature. 2015;518(7538):187-196. doi:10.1038/nature14132 PubMedGoogle ScholarCrossref
    54.
    Locke  AE, Kahali  B, Berndt  SI,  et al; LifeLines Cohort Study; ADIPOGen Consortium; AGEN-BMI Working Group; CARDIOGRAMplusC4D Consortium; CKDGen Consortium; GLGC; ICBP; MAGIC Investigators; MuTHER Consortium; MIGen Consortium; PAGE Consortium; ReproGen Consortium; GENIE Consortium; International Endogene Consortium.  Genetic studies of body mass index yield new insights for obesity biology.   Nature. 2015;518(7538):197-206. doi:10.1038/nature14177 PubMedGoogle ScholarCrossref
    55.
    Willer  CJ, Schmidt  EM, Sengupta  S,  et al; Global Lipids Genetics Consortium.  Discovery and refinement of loci associated with lipid levels.   Nat Genet. 2013;45(11):1274-1283. doi:10.1038/ng.2797 PubMedGoogle ScholarCrossref
    56.
    Ware  E, Schmitz  L, Gard  A, Faul  JD. HRS polygenic scores—release 3, 2006-2012 genetic data. HRS documentation report. Published October 2018. Accessed January 7, 2020. https://hrsonline.isr.umich.edu/modules/meta/xyear/pgs/desc/PGENSCORES3DD.pdf
    57.
    Ribeiro  MT, Singh  S, Guestrin  C. ”Why should I trust you?”: explaining the predictions of any classifier. Presented at: 22nd Association for Computing Machinery Special Interest Group on Knowledge Discovery and Data Mining International Conference on Knowledge Discovery and Data Mining; August 2016.
    58.
    Muthén  LK, Muthén  BO.  Mplus (version 7.3). Muthén & Muthén; 2014.
    59.
    Keras  CF.  The Python Deep Learning Library. Astrophysics Source Code Library; 2018.
    60.
    Pedregosa  F, Varoquaux  G, Gramfort  A,  et al.  Scikit-learn: machine learning in Python.   J Mach Learn Res. 2011;12:2825-2830.Google Scholar
    61.
    Kwong  ASF, López-López  JA, Hammerton  G,  et al.  Genetic and environmental risk factors associated with trajectories of depression symptoms from adolescence to young adulthood.   JAMA Netw Open. 2019;2(6):e196587. doi:10.1001/jamanetworkopen.2019.6587 PubMedGoogle Scholar
    62.
    Choi  KW, Chen  C-Y, Ursano  RJ,  et al  Prospective study of polygenic risk, protective factors, and incident depression following combat deployment in US Army soldiers.   bioRxiv. 2018:361725. doi:10.1101/361725Google Scholar
    63.
    Choi  KW, Stein  MB, Dunn  EC, Koenen  KC, Smoller  JW.  Genomics and psychological resilience: a research agenda.   Mol Psychiatry. 2019;24(12):1770-1778. doi:10.1038/s41380-019-0457-6 PubMedGoogle ScholarCrossref
    64.
    Wainberg  M, Merico  D, Delong  A, Frey  BJ.  Deep learning in biomedicine.   Nat Biotechnol. 2018;36(9):829-838. doi:10.1038/nbt.4233 PubMedGoogle ScholarCrossref
    65.
    Torkamani  A, Wineinger  NE, Topol  EJ.  The personal and clinical utility of polygenic risk scores.   Nat Rev Genet. 2018;19(9):581-590. doi:10.1038/s41576-018-0018-x PubMedGoogle ScholarCrossref
    66.
    Lipton  ZC.  The mythos of model interpretability.   Queue. 2018;16(3):31-57. doi:10.1145/3236386.3241340Google ScholarCrossref
    67.
    Howard  DM, Adams  MJ, Clarke  T-K,  et al; 23andMe Research Team; Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium.  Genome-wide meta-analysis of depression identifies 102 independent variants and highlights the importance of the prefrontal brain regions.   Nat Neurosci. 2019;22(3):343-352. doi:10.1038/s41593-018-0326-7 PubMedGoogle ScholarCrossref
    ×