Arcos-Burgos et al. Translational Psychiatry (2019)9:42 https://doi.org/10.1038/s41398-019-0396-7 Translational Psychiatry ARTICLE Open Access ADGRL3 (LPHN3) variants predict substance use disorder Mauricio Arcos-Burgos1,2,3, Jorge I. Vélez 1,4, Ariel F. Martinez 1, Marta Ribasés 5,6,7, Josep A. Ramos-Quiroga5,6,7,8, Cristina Sánchez-Mora5,6,7, Vanesa Richarte6,7,8, Carlos Roncero 6,7,8,9, Bru Cormand 10,11,12,13, Noelia Fernández-Castillo10,11,12,13, Miguel Casas5,6,7,8, Francisco Lopera14, David A. Pineda14, Juan D. Palacio14, Johan E. Acosta-López15, Martha L. Cervantes-Henriquez4,15, Manuel G. Sánchez-Rojas15, Pedro J. Puentes-Rozo15,16, Brooke S. G. Molina17, MTA Cooperative Group, Margaret T. Boden18, Deeann Wallis19, Brett Lidbury20, Saul Newman 20, Simon Easteal 20, James Swanson21,22, Hardip Patel23, Nora Volkow 24, Maria T. Acosta1, Francisco X. Castellanos 25,26, Jose de Leon 18, Claudio A. Mastronardi2,27 and Maximilian Muenke1 Abstract Genetic factors are strongly implicated in the susceptibility to develop externalizing syndromes such as attention- deficit/hyperactivity disorder (ADHD), oppositional defiant disorder, conduct disorder, and substance use disorder (SUD). Variants in the ADGRL3 (LPHN3) gene predispose to ADHD and predict ADHD severity, disruptive behaviors comorbidity, long-term outcome, and response to treatment. In this study, we investigated whether variants within ADGRL3 are associated with SUD, a disorder that is frequently co-morbid with ADHD. Using family-based, case-control, and longitudinal samples from disparate regions of the world (n = 2698), recruited either for clinical, genetic 1234567890():,; 1234567890():,; 1234567890():,; 1234567890():,; epidemiological or pharmacogenomic studies of ADHD, we assembled recursive-partitioning frameworks (classification tree analyses) with clinical, demographic, and ADGRL3 genetic information to predict SUD susceptibility. Our results indicate that SUD can be efficiently and robustly predicted in ADHD participants. The genetic models used remained highly efficient in predicting SUD in a large sample of individuals with severe SUD from a psychiatric institution that were not ascertained on the basis of ADHD diagnosis, thus identifying ADGRL3 as a risk gene for SUD. Recursive-partitioning analyses revealed that rs4860437 was the predominant predictive variant. This new methodological approach offers novel insights into higher order predictive interactions and offers a unique opportunity for translational application in the clinical assessment of patients at high risk for SUD. Introduction alcohol and drug abuse were reported (Whiteford et al., Substance use disorders (SUD) and addiction represent 2015), and SUD prevalence is expected to increase over a global public health problem of substantial socio- time. Genetic factors have been implicated in SUD etiol- economic implications1,2. In 2010, 147.5 million cases of ogy, with genes involved in the regulation of several neurobiological systems (including dopaminergic and glutamatergic) found to be important (for a review see Correspondence: Mauricio Arcos-Burgos (mauricio.arcos@udea.edu.co) or Prom-Wormley et al., 20173). However, limitations Maximilian Muenke (mamuenke@mail.nih.gov) 1 intrinsic to most genetic epidemiological studies support Medical Genetics Branch, National Human Genome Research Institute, National Institutes of Health, Bethesda, MD, USA the search for additional risk genes. 2 INPAC Research Group, Fundación Universitaria Sanitas, Bogotá, Colombia Attention-deficit/hyperactivity disorder (ADHD), the Full list of author information is available at the end of the article. most common neurodevelopmental behavioral dis- These authors contributed equally: Mauricio Arcos-Burgos, Jorge I. Vélez, Ariel F. Martinez, Maria T. Acosta, Francisco X. Castellanos, Jose de Leon and Claudio order4,5, is frequently co-morbid with disruptive behaviors A. Mastronardi © The Author(s) 2019 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 2 of 15 such as oppositional defiant disorder (ODD), conduct Subjects and methods disorder (CD), and SUD6,7. The close association between Subjects ADHD and disruptive behaviors is summarized by long- We used independent populations from disparate itudinal observations in ADHD cohorts6,8,9. Children regions of the world (n = 2698) ascertained through diagnosed with ADHD monitored during the transition patients affected with ADHD co-morbid with disruptive into adolescence exhibit higher rates of alcohol, tobacco, behaviors (Paisa, Spanish and MTA samples) or SUD and psychoactive drug use than control groups of children (Spanish and Kentucky samples). without ADHD10,11. It has been estimated that the life- time risk for SUD is ~50% in subjects with childhood Paisa sample ADHD persisting into adulthood12,13. Reciprocally, the This population isolate is unique in that it was used to prevalence of ADHD is high in adolescents with identify ADHD susceptibility genes by linkage and asso- SUD9,14,15 and the presence of an ADHD diagnosis affects ciation strategies. Detailed clinical and demographic SUD prognosis, with ADHD being associated with both information on this sample has been published else- earlier and more frequent alcohol-related relapses16 and where23,25,29. The sample consists of 1176 people (adults, lower likelihood of cannabis-dependence treatment adolescents, and children), mean age 28 ± 17 years, completion17. ascertained from 18 extended multigenerational and 136 Strong evidence from family, twin, and genome-wide nuclear Paisa families inhabiting the Medellin metropo- linkage and association studies suggests that genetic fac- litan area in the State of Antioquia, Colombia. Initial tors play a crucial role in shaping the susceptibility to both coded pedigrees were obtained through a fixed sampling ADHD and SUD18–21. During the last 15 years, we have scheme from a parent or grandparent of an index proband collected families clustering individuals affected with after having collected written informed consent from all ADHD and disruptive behaviors from disparate regions subjects or their parent/guardian, as approved by the around the world6,18,22,23. Although the prevalence of University of Antioquia and the NIH Ethics Committees, ADHD co-morbid with disruptive behaviors is variable and in accordance with the Helsinki Declaration. Patients across populations, we found a higher frequency of CD, were recruited under NHGRI protocol 00-HG-0058 ODD, and SUD (mainly nicotine dependence and alcohol (NCT00046059). abuse) in ADHD individuals than in unaffected rela- Exclusion criteria for ADHD participants were IQ < 80, tives6,22,24. Using genome-wide data from extended mul- or any autistic or psychotic disorders. Parents underwent tigenerational families, we found evidence of linkage of a full psychiatric structured interview regarding their ADHD to markers in chromosomes 4q13.2, 5q33.3, offspring (Diagnostic Interview for Children and Adoles- 8q11.23, 11q22, and 17p1125, and co-segregation of cents—Revised—Parents version (DICA-IV-P, Spanish ADHD and disruptive behaviors with loci at 2p21-22.3, version translated with permission from Dr. Wendy Reich 4q13.2, 5p13.1-p13.3, 8q24, 8q15, 11q22, 12p11.23-13.3, (Washington University, St. Louis). All adult participants and 14q21.1-22.28. Fine mapping of the 4q13.2 region were assessed using the Composite International Diag- identified variants in the adhesion G-protein-coupled nostic Interview (CIDI), as well as the Disruptive Behavior receptor L3 gene (ADGRL3, also known as latrophilin 3 or Disorders module from the DICA-IV-P modified for ret- LPHN3) that predispose to ADHD22,24,26–30. rospective use. The interview was conducted by a “blind” Characterization of the association between ADHD and rater (either a psychologist, a neuropsychologist, or a ADGRL3 has provided key information to better predict psychiatrist) at the Neurosciences Clinic of the University the severity of ADHD, the long-term outcome, the pat- of Antioquia, or during home visits. ADHD status was terns of brain metabolism, and the response to stimulant defined by the best estimate method. Specific information medication24,27,29,31–34. To the best of our knowledge, regarding clinical diagnoses and co-morbid disruptive ADGRL3 linkage and association results represent some disorders, affective disorders, anxiety, and substance use of the most robustly replicated genetic and pharmaco- has been published elsewhere3. genetic findings in ADHD genetic research. While From the 1176 individuals in this cohort, only founder ADGRL3 has also shown association with disruptive members were included in analyses (n = 472). This was behaviors in the context of ADHD18,24,35, a direct link to done to avoid kinship relatedness bias and to exclude SUD has not been systematically investigated. In this children and adolescents, as they may have not been manuscript we tested the hypothesis that ADHD risk exposed to substances of abuse yet. Of these 472 indivi- variants harbored at the ADGRL3 locus interact with duals, 17% (n = 79) fulfilled criteria for ADHD, 17% (n = clinical, demographic, and environmental variables 78) for ODD, 18% (n = 84) for CD, 22% nicotine depen- associated with SUD. dence (n = 102), 27% alcohol dependence (n = 124), 3% Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 3 of 15 Table 1 A concise description of the cohorts’ principal Table 1 continued demographic and clinical data. Paisa Spanish MTA Kentucky Paisa Spanish MTA Kentucky sample sample sample sample sample sample sample sample n % n % n % n % n % n % n % n % Phobias Sex a a a a Affected 156 37% 50 3% Males 231 49% 1193 72% 287 76% 285 53% a a a a Unaffected 172 33% 584 36% Females 241 51% 454 28% 89 24% 248 47% a a a a Unknown 144 30% 1013 61% Total 472 100% 1647 100% 376 100% 533 Anxiety ADHD a a a a Affected 58 13% 107 6% a a Affected 79 17% 670 41% 140 37% a a a a Unaffected 270 57% 818 50% a a Unaffected 249 53% 486 29% 236 63% a a a a Unknown 144 30% 722 44% a a Unknown 144 30% 491 30% 376 100% Depression ODD a a a a Affected 117 25% 143 9% a a a a Affected 78 17% 81 5% a a a a Unaffected 211 45% 490 30% a a a a Unaffected 250 53% 391 24% a a a a Unknown 144 30% 1014 61% a a a a Unknown 144 30% 1175 71% Mood CD a a a a a a Affected 157 29% a a a a Affected 84 18% 102 6% a a a a a a Unaffected 376 71% a a a a Unaffected 244 52% 357 22% a a a a a a Unknown 0 0% a a a a Unknown 144 30% 1188 72% Schizophrenia Nicotine a a a a a a Affected 253 47% Affected 102 22% 646 39% 97 26% 372 70% a a a a a a Unaffected 280 53% Unaffected 226 48% 613 37% 40 11% 161 30% a a a a a a Unknown 0 0% Unknown 144 30% 388 24% 239 63% 0 0% For the Paisa cohort, only information for founder members (adults) used in the Alcohol ARPA-based predictive model for SUD is shown. See Methods section for more details Affected 124 27% 396 24% 120 32% 342 64% ADHD attention-deficit/hyperactivity disorder, CD conduct disorder, ODD oppositional defiant disorder, SUD substance use disorders Unaffected 204 43% 637 39% 106 28% 191 36% a Data not available Unknown 144 30% 614 37% 150 40% 0 0% Cannabis drug dependence (n = 12), 37% social/simple phobia (n = Affected a a a a 94 25% a a 156), 13% any other anxiety disorder (n = 58), and 25% a a a a a a major depressive disorder (n = 117) (Table 1). Unaffected 71 19% a a a a a a Unknown 211 56% Spanish sample Other drugs The ADHD sample consisted of 670 adult ADHD Affected 12 3% a a a a 147 28% patients, mean age 33 ± 10 years, 69% males (n = 461), a a a a recruited and evaluated at the Psychiatry Department of Unaffected 197 56% 386 72% the Hospital Universitari Vall d’Hebron (Barcelona, Spain) a a a a Unknown 263 41% 0 0% according to DSM-IV TR criteria. ADHD diagnosis was SUD based on the Spanish version of the Conners Adult Affected a a 768 47% a a 452 85% ADHD Diagnostic Interview for DSM-IV (CAADID)36. a a a a Comorbidity was assessed by Structured Clinical Inter- Unaffected 879 53% 81 15% view for DSM-IV Axis I and Axis II Disorders (SCID-I a a a a a a Unknown 0 0% and SCID-II). ODD during childhood and adolescence was retrospectively evaluated with the Schedule for Affective Disorders and Schizophrenia (SADS) for Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 4 of 15 School-Age Children, Present and Lifetime Version (K- classmates group-matched for grade and sex was added SADS). Thirty-nine percent of ADHD patients (n = 263) when the ADHD participants were between 9–12 years of fulfilled diagnostic criteria for SUD, 21% for disruptive age. The outcomes in childhood (14, 24, and 36 months behavior disorders (CD and/or ODD; n = 142), 21% for after baseline), and adolescence (6 and 8 years after depression (n = 143), 13% for anxiety (n = 89), and 8% for baseline) and into adulthood (12, 14, and 16 years after phobias (n = 50). The level of impairment was measured baseline) have been reported10,11,37–43. Substance use was with the Clinical Global Impression (CGI) included in the assessed with a child/adolescent-reported questionnaire CAADID Part II and the Sheehan Disability Inventory. adapted for the MTA11,43. The measure included items Exclusion criteria for ADHD patients were IQ < 80; per- for lifetime and current (previous 6 months) use of vasive developmental disorders; schizophrenia or other alcohol, cigarettes, tobacco, cannabis, and other recrea- psychotic disorders; presence of mood, anxiety or per- tional drugs. Also included were items for non-prescribed sonality disorders that might explain ADHD symptoms; use or misuse of psychoactive medications, including birth weight ≤ 1.5 kg; and other neurological or systemic stimulants. The measure was modeled after similar sub- disorders that might explain ADHD symptoms. stance use measures in longitudinal or national survey The SUD sample consisted of 494 adults (mean age studies of alcohol and other drug use that also rely on 37 ± 9 years and 76% males, n = 376) recruited and eval- confidential youth self-report as the best source of uated at the Addiction and Dual Diagnosis Unit of the data44,45. A National Institutes of Health (NIH) Certificate Psychiatry Department at the Hospital Universitari Vall of Confidentiality further strengthened the assurance of d’Hebron with the Structured Clinical Interview for DSM- privacy. Substance use was coded positive if any of the IV Axis I Disorders (SCID-I). All patients fulfilled DSM- following behaviors, selected after examining distribu- IV criteria for drug dependence beyond nicotine depen- tions, were endorsed as occurring in the participant’s dence. None were evaluated for ADHD. lifetime up to 8 years post-baseline: (1) alcohol con- The control sample consisted of 483 blood donors sumption (more than just a sip) more than five times or (mean age 42 ± 20 years, 74% males) in which DSM-IV drunk at least once; (2) cigarette smoking or tobacco lifetime ADHD symptomatology was excluded under the chewing more than a few times; (3) cannabis use more following criteria: (1) not having been diagnosed with than once; or (4) use of inhalants, hallucinogens, cocaine, ADHD and (2) answering negatively to the lifetime pre- or any of amphetamines/stimulants, barbiturates/seda- sence of the following DSM-IV ADHD symptoms: (a) tives, and opioids/narcotics without a prescription or often has trouble keeping attention on tasks, (b) often misused a prescription (used in greater quantity or more loses things needed for tasks, (c) often fidgets with hands often than prescribed). Each of the four types of sub- or feet or squirms in seat, and (d) often gets up from seat stances, as well as daily use of tobacco and the number of when remaining in seat is expected. Individuals affected substance use classes endorsed (0, 1, 2, or more), were with SUD were excluded from this sample. None of them explored in secondary analyses. had self-administered drugs intravenously. It is important DSM-IV abuse or dependence was based on a positive to mention that the exposure criterion was not applied; parent or child report with the Diagnostic Interview therefore, this set cannot be classified as “pure” controls. Schedule for Children version 2.3/3.0 (DISC)46 at the 6- All patients and controls were Spanish of Caucasian and 8-year follow-up assessments. The DISC includes descent. This study was approved by the ethics committee both lifetime and past year diagnoses. The Diagnostic of the Hospital Universitari Vall d’Hebron and informed Interview Schedule-IV47 was used at the 8-year follow-up consent was obtained from all subjects in accordance with for 18 + year-olds (n = 111). SUD was defined as the the Helsinki Declaration. lifetime presence of any abuse or dependence (excluding tobacco dependence, due to differences in the meaning of MTA sample abuse/dependence for tobacco versus other substances). The Multimodal Treatment Study of Children with Additional analyses explored SUD for alcohol, tobacco, ADHD (MTA) was designed to evaluate the relative effi- and cannabis/other drugs (recreational or misused pre- cacy of treatments for childhood ADHD, combined sub- scription medications) separately10. All patients in this type, in a 14-month randomized controlled trial of 579 study provided informed written consent as approved by children assigned to four treatment groups: medication the NIH Ethics Committee. management, behavior modification, their combination, and treatment as usual in community care. After the 14- Kentucky sample month treatment-by-protocol phase, the MTA continued A sample of 560 inpatients and outpatients with severe as a naturalistic follow-up in which self-selected use of SUD from Central Kentucky psychiatric facilities was psychoactive medication was monitored. A local norma- collected during a pharmacogenetics investigation48. tive comparison group of 289 randomly selected Patient interviews and medical record information Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 5 of 15 (including urine drug screens and substance abuse analysis. Markers with a minor allele frequency (MAF) < counselor notes) were used by the research nurse to assess 0.01 (rare variants), significant deviation from the Clinician Rating of Alcohol (CRAUD) and Drug Use Hardy–Weinberg equilibrium (P-values < 0.0001), and a Disorder (CRDUD)49,50 that provides a score from 1 = genotyping success rate < 90%, were excluded. For the abstinence (not used in the assessed period) to 5 = severe Paisa and Spanish samples, a subset of variants in the dependence. Scores of 3 and higher are pathological and ADGRL3 minimal critical region (MCR), 5′UTR and 3′ were considered positive in our analyses. All drugs were UTR were selected based on a previous ADHD associa- combined into one rating48. Descriptions of the training tion study30. Because the Paisa sample is a family-based provided to research nurses to assess the CRAUD and cohort and recursive-partitioning analysis does not cor- CRDUD were published elsewhere48,51. rect for kinship relatedness, only founder members from DNA was available from 533 of 560 study subjects. Of the pedigrees were included in the analyses. For the MTA the 533 subjects with available DNA, 53% (n = 285) were sample, a total of 8568 markers with a MAF ≥ 1.0 % from male, 82% (n = 436) were Caucasian, 16% (n = 87) were the 244,414 markers genotyped with the exome chip were African American, and 2% (n = 10) were from other filtered out using linkage disequilibrium (LD) pruning, ethnicities. Additional clinical information for this sample and variants within ADGRL3 were selected for analyses. has been described elsewhere48,51 and included: (1) clin- For the Kentucky sample, only four ADGRL3 variants ical diagnosis obtained from medical records, (2) prior were selected for analyses after LD pruning of a list of psychiatric history, (3) history of daily smoking, (4) markers located within the ADGRL3 5′UTR and MCR reviews of current and psychiatric medication use, and (5) regions that was available to us. Variants rs7659636 and body mass index (Supplemental Table 2). All participants rs5010235 had been imputed from ADHD genome-wide in the Kentucky study provided informed written consent association data funded through the Genetics Analysis as approved by the University of Kentucky IRB. Information Network (GAIN) initiative, a public-private partnership between the NIH and the private sector Genotyping (https://www.genome.gov/19518664/genetic-association- DNA was extracted from whole blood (Paisa, Spanish information-network-gain/#al-4). ADGRL3 variants used and MTA sample) or buccal swabs (Kentucky sample) in this study for each cohort are presented in Supple- using standard protocols. The Paisa sample was geno- mental Table 3. typed using the service provided by Illumina (San Diego, CA). The Spanish, MTA, and Kentucky samples were Advanced recursive-partitioning (tree-based) approach genotyped for select variants using pre-designed TaqMan® (ARPA) SNP genotyping assays (Thermo Fisher Scientific, Wal- Association studies of ADGRL3 variants with ADHD, tham, MA). Allelic discrimination real-time PCR reac- ODD, CD, response to stimulant treatment and severity tions were performed in a 384-well plate format for each outcome have been published elsewhere for the Paisa and individual sample according to the manufacturer’s Spanish populations24,29,32,52. We used ARPA to build a instructions. Briefly, 20 ng of genomic DNA were mixed predictive framework to forecast the behavioral outcome with 2.5 μL of 2X TaqMan Universal PCR Master Mix and of children with ADHD, suitable for translational appli- 0.25 μL of 20X SNP Genotyping Assay in a total volume of cations. Our goal was to test the hypothesis that ADGRL3 5 μL per reaction. Assays were run in an ABI 7900HT Fast variants predisposing to ADHD also increase the risk of Real-Time PCR System (Thermo Fisher Scientific). Allele co-morbid disruptive symptoms, including SUD. calling was made by end-point fluorescent signal analysis ARPA is a tree-based method widely used in predictive using the ABI’s SDS2.3 software. In addition, we had analyses because it accounts for non-linear and interac- previously collected exome genotype data from the MTA tion effects, offers fast solutions to reveal hidden complex sample26 using the Infinium® HumanExome-12 v1.2 substructures and provides truly non-biased statistically BeadChip kit (Illumina), which covers putative functional significant analyses of high-dimension, seemingly unre- exonic variants selected from over 12,000 individual lated data53. In a visionary manuscript, D.C. Rao sug- exome and whole-genome sequences. Processed and raw gested that recursive-partitioning techniques could be intensity signals for the array data can be accessed at GEO useful for genetic dissection of complex traits54. ARPA (GSE112652). SNP markers harbored at the ADGRL3 accounts for the effect of hidden interactions better than gene were filtered in from this dataset and added to those alternative methods, and is independent of the type of genotyped using TaqMan® assays. data (i.e., categorical, continuous, ordinal, etc.) and of the type of data distribution (i.e., fitting or not fitting nor- Dataset quality control and preparation for analysis mality)54. Furthermore, results supplied by tree-based Genotype data were imported into Golden Helix® SVS analytics are easy to interpret visually and logically53. 8.3.1 (Golden Helix, Bozeman, MT) for quality control Therefore, to generate the most comprehensive and Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 6 of 15 parsimonious classificatory model to predict the sus- such as CART or by ensembles such as bagging (i.e., ceptibility to disruptive behaviors, we applied ARPA using RF)57. The TreeNet algorithm generates thousands of a set of different modules implemented in the Salford small decision trees built in a sequential error-correcting Predictive Modeler® (SPM) software, namely, Classifica- process converging on an accurate model57. The number tion and Regression Trees (CART), Random Forest, and of variables considered to derive each node with RF was pffiffiffi TreeNet (http://www.salford-systems.com). One impor- n, where n is the number of independent variables tant advantage of SPM when compared to other available (either 3 or 4). data mining software is its ability to use raw data with To derive honest assessments of the derived models and sparse or empty cells, a problem frequently encountered have a better view of their performance on future unseen in genetic data. data, we applied a cross-validation strategy where both Briefly, CART is a non-parametric approach whereby a training with all the data and then indirectly testing with series of recursive subdivisions separate the data by all the data were performed. To do so, we randomly dichotomization55. The aim is to identify, at each partition divided the data into separate partitions (folds) of differ- step, the best predictive variable and its best corre- ent sizes. This strategy allowed us to review the stability of sponding splitting value while optimizing a splitting sta- results across multiple replications55. We used a 10-fold tistical criterion, so that the dataset can be successfully cross-validation as implemented in the SPM software. split into increasingly homogeneous subgroups55. We A fixed-effects meta-analysis of the overall fraction of used a battery of different statistical criteria as splitting correctly classified individuals (accuracy) using the rules (e.g., GINI Index, Entropy, and Twoing) to deter- derived models from each of the four samples was applied mine the splitting rule, maximally decreasing the relative to derive a general perspective of the SUD predictive cost of the tree while increasing the prediction accuracy of capacity of this demographic-clinical-genetic framework. target variable categories55. The best split at each dichotomous node was chosen by either a measure of Results between-node dissimilarity or iterative hypothesis testing A series of predictive models were built on our data of all possible splits to find the most homogeneous split using combinations of the following criteria: (i) the rules (lowest impurity). Similarly, we used a wide range of of splitting (GINI index, twoing, order twoing, and empirical probabilities (priors) to model numerous sce- entropy); (ii) the priors; (iii) the size of the terminal nodes; narios recreating the distribution of the targeted variable (iv) the costs; (v) the depth of branching; and (vi) the size categories in the population55. Following this iterative of the folds for cross-validation, to maximize the accuracy process, each terminal node was assigned to a class out- of the derived classification tree while considering class come. To avoid finishing with an over-fitted CART pre- assignment, tree pruning, testing and cross-validation. dictive model (a common problem in CART analyses), A parsimonious and informative reconstructed pre- and to ensure that the final splits were well substantiated, dictive tree derived from CART for the Paisa sample we applied tree pruning. During the procedure, predictor revealed demographic (age), clinical (CD), and genetic variables that were close competitors (surrogate pre- variables (rs5010235 and rs4860437) (Fig. 1a). The dictors with comparable overall classification error to the importance of these variables was corroborated, and their optimal predictors) were pruned to eliminate redundant potential over fitting discarded by the TreeNet analyses commonalities among variables, so the most parsimo- that revealed a set of predictors for SUD containing those nious tree would have the lowest misclassification rate for derived by CART (Fig. 1b). This predictive model displays an individual not included in the original data55. good sensitivity and specificity as shown by areas under Additionally, we applied the Random Forest (RF) the receiver-operating characteristic (ROC) curve (0.954 methodology using a bagging strategy to exactly identify and 0.87 for the learning and the test data, respectively) the most important set of variables predicting disruptive during TreeNet cross-validation using folding (Fig. 1c). behaviors56. The RF strategy differs from CART in the use The proportions of misclassification for SUD cases in the of a limited number of variables to derive each node while cross-validation experiment for the learning and testing creating hundreds to thousands of trees. This strategy has data were 0.124 and 0.177, respectively (Fig. 1d). proved to be immune to the over fitting generated by In the case of the Spanish sample, a parsimonious and CART56. In RF, variables that appeared repeatedly as informative tree was reconstructed with CART revealing predictors in the trees were identified. The misclassifica- demographic (sex), clinical (CD, ODD, depression, and tion rate was recorded for each approach. ADHD), and genetic variables (rs4860437 and rs1868790) The TreeNet strategy was used as a complement to the (Fig. 2a). The TreeNet analysis revealed a set of predictors CART and RF strategies because it reaches a level of for SUD containing those derived by CART (Fig. 2b). This accuracy that is usually not attainable by single models predictive model displayed good sensitivity and specificity Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 7 of 15 Fig. 1 Advanced Recursive Partitioning Analysis (ARPA) for the Paisa sample. a Derived Classification and Regression Tree (CART) for SUD status as categorical target variable (disjunctive affection status, i.e., substance use of either alcohol, or nicotine, or other drugs). Only founder individuals were included in the analysis to avoid kinship relatedness bias. Class 0 (unaffected) is indicated in red and class 1 (affected) in blue. This derived tree for the Paisa sample included demographic (age), clinical (conduct disorder (CD)), and genetic variables (markers rs5010235 and rs4860437). The T allele of the rs4860437 variant (node 4) generates a highly discriminant split in combination with age (45.5 years) to terminal node 3 of ADHD individuals without CD (see root node 1). b Variable importance scores derived by Random Forest and TreeNet analysis were compatible with the variables included in the tree derived by CART. c, d TreeNet analysis to maximize the ROC area and minimize the classification error using 200 trees. The areas under the ROC curve (AUC) were 0.954 and 0.87 for learning and testing samples (blue and red curves and values, respectively), while the proportions of misclassification for SUD cases in the cross-validation experiment were 0.124 and 0.177 for learning and testing data sets, respectively as shown by areas under the ROC curve (AUC) of 0.911 derived by CART (Fig. 3b). This predictive model displays and 0.897 for learning and testing samples, respectively, good sensitivity and specificity as showed by AUC of 0.808 during TreeNet cross-validation using folding (Fig. 2c). and 0.643 for learning and testing samples, respectively, The proportions of misclassification for SUD cases during TreeNet cross-validation using folding (Fig. 3c). obtained by TreeNet analysis for learning and testing data The proportions of misclassification for SUD cases were 0.151 and 0.175, respectively (Fig. 2d). obtained by TreeNet analysis for learning and testing data As in the previous cohorts, for the MTA sample we were 0.314 and 0.358, respectively (Fig. 3d). derived a parsimonious and informative predictive tree Finally, for the Kentucky sample, we derived a parsi- with CART depicting demographic (site of ascertain- monious and informative predictive tree with CART ment), and genetic variables (rs2172802, rs61747658, involving demographic (sex), clinical (high body mass rs12509110, and rs6856328) (Fig. 3a). The TreeNet ana- index (HBMI) and schizophrenia diagnosis), and genetic lyses revealed a set of predictors for SUD containing those variables (rs4860437 and rs7659636) (Fig. 4a). The Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 8 of 15 Fig. 2 ARPA for the Spanish sample. a Derived tree by CART for SUD status as categorical target variable (disjunctive affection status, i.e., substance use of either alcohol, or nicotine, or other drugs). This derived tree for the Spanish sample included demographic (sex), clinical (CD, ODD, depression, and ADHD), and genetic variables (markers rs677642, rs4860437, rs1868790). b Variable importance scores derived by Random Forest and TreeNet analysis were compatible with the variables included in the tree derived by CART. c, d TreeNet analysis to maximize the AUC and minimize the classification error using 200 trees. The AUC were 0.911 and 0.897 for learning and testing samples while the proportions of misclassification for SUD cases in the cross-validation experiment were 0.151 and 0.175 for learning and testing data sets, respectively. Conventions as in Fig. 1 TreeNet analyses revealed a set of predictors for SUD values. Overall, ADGRL3 marker rs4860437 was the most containing those derived by CART (Fig. 4b). This pre- important variant predicting susceptibility to SUD, a dictive model displays good sensitivity and specificity as commonality suggesting that these networks may be showed by AUC of 0.811 and 0.744 for learning and accurate in predicting the development of SUD based on testing samples, respectively, during TreeNet cross- ADGRL3 genotypes. validation using folding (Fig. 4c). The proportions of We conducted independent analyses for alcohol or misclassification for SUD cases obtained by TreeNet nicotine dependence and compared these results with analysis for learning and testing data were 0.285 and those of our composite SUD phenotype, as defined by the 0.252, respectively (Fig. 4d). The results from the RF disjunctive presence of substance use phenotypes and analysis were consistent with those produced by TreeNet explained by likely common neuropathophysiological cross-validation using folding. mechanisms. In general, across cohorts, we found sig- A fixed-effects meta-analysis for overall accuracy nificant alcohol and nicotine risk variants, some of which returned a value of 0.727 (95% CI = 0.710–0.744) (Fig. 5), have reasonably high odd ratios (OR). For instance, in the suggesting potential eventual clinical utility of predictive Spain sample, marker rs2271339 conferred significant risk Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 9 of 15 Fig. 3 ARPA for the MTA sample. a Derived tree by CART for the SUD status as categorical target variable (disjunctive affection status, i.e., substance use of either alcohol, or nicotine, cannabis, or other drugs). As the MTA is a longitudinal study, we used SUD status at 96 and 120 month follow-ups and applied a lag analysis of SUD emergence. The derived tree included demographic (site of ascertainment), and genetic variables (markers rs2172802, rs61747658, rs12509110, and rs6856328). The combination of variants rs61747658 and rs2172802 generated an important discriminant splitting of SUD affected and unaffected classes. b Variable importance scores derived by Random Forest and TreeNet analysis were compatible with the variables included in the tree derived by CART. c, d TreeNet analysis to maximize ROC area and minimize classification error using 200 trees. The AUC were 0.808 and 0.643 for learning and testing samples, while the proportions of misclassification for SUD cases in the cross-validation experiment, for learning and testing data were 0.314 and 0.358, respectively. Conventions as in Fig. 1 to nicotine use: the heterozygote genotype A/G confers the AUC is greater than 0.7 in both cases, which suggests 43% increased risk of being diagnosed with nicotine use a straight performance of markers rs1456862 and (OR = 1.43, 95% CI = 1.12–1.82). In the same vein, we rs2159140 in predicting nicotine and alcohol use, found in the Paisa sample that the heterozygote A/T respectively. genotype for rs1456862 confers 83% increased risk to To determine the significance of improvement of pre- nicotine use (corrected OR = 1.84, 95 CI% = 1.03–3.38) diction when genetic markers are introduced in the than the A/A genotype. Regarding alcohol use, we found ARPA-based predictive model for SUD, we compared the in the Paisas that the heterozygote C/T genotype for performance measures (i.e., sensitivity, specificity, classi- rs2159140 confers susceptibility, whereas the C/C geno- fication rate, and lift) across all cohorts under two dis- type does not (corrected OR = 1.64, 95 CI% = 1.01–2.72). junctive scenarios: inclusion of genetic markers or not. Supplemental Fig. 1 shows the ROC curves of nicotine We found that including genetic markers improved the and alcohol use prediction in the Paisa sample. Note that performance measures of the resulting ARPA-based Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 10 of 15 Fig. 4 ARPA for the Kentucky sample. a Derived CART tree for SUD status as categorical target variable (disjunctive affection status, i.e., substance use of either alcohol, or nicotine, or other drugs). This derived tree for the Kentucky sample included demographic (sex), clinical (high Body Mass index [HBMI], and schizophrenia diagnosis), and genetic variables (markers rs4860437 and rs7659636). Notably, the T allele of the rs4860437 variant generated a split in the same direction as occurred for the derived tree in the Paisa and in the Spain samples. b Variable importance scores derived by Random Forest and TreeNet analysis were compatible with the variables included in the tree derived by CART. c, d TreeNet analysis to maximize ROC area and minimize classification error using 200 trees. The AUC were 0.811 and 0.744 for learning and testing samples, respectively, while the proportions of misclassification for SUD cases in the cross-validation experiment, for learning and testing data were 0.285 and 0.252, respectively. Conventions as in Fig. 1 Fig. 5 Fixed-effects meta-analysis for the prediction accuracy of the ARPA-based predictive model for SUD derived in each cohort. The overall SUD correct classification rate is ~73%. CI Confidence Interval, SE Standard Error, Z test statistic, P P-value predictive model of SUD, regardless of cohort (Supple- 77.5 (95% CI = 75.9–79.1) when it was excluded. A mental Fig. 2 and Supplemental Table 1). For instance, the bootstrap-based test with 10,000 replicates revealed that AUC for the Spain sample was 81.6% (95% CI = the former AUC was statistically greater than the latter (P 79.8–83.4) when genetic information was included, and < 0.0001, Supplemental Table 1). Similar results were Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 11 of 15 obtained for the Paisa sample: the AUC was 90% (95% CI included in the analyses for this sample. Given the limited = 86.6–93.0) when genetic information was included overlap of markers across datasets and possible stratifi- versus 78.8% (95% CI = 75.8–81.7) when it was not (P < cation differences among study populations, a gene- 0.0001, Supplemental Table 1). Improvements were also rather than a marker-level approach has been observed in the correct classification rate for the Spanish advocated64. and Paisa samples, the sensitivity values in all samples, the Adopting such a perspective, our results suggest that specificity in the Spanish and Paisa samples, and the lift in genetic variants harbored in the ADGRL3 locus confer the Paisa sample (Supplemental Table 1). Similar results susceptibility to SUD in populations from disparate were observed for the MTA and Kentucky samples, where regions of the world. These populations are from three including genetic information in the predictive model for different countries and involve different investigators, SUD drastically improved these performance measures diverse inclusion criteria, and different clinical assess- (Supplemental Table 1). ments, which suggests that our results may replicate in other settings and are likely to be clinically relevant. Of Discussion particular interest is the generalization of our findings to a SUD genetic epidemiological studies across multiple longitudinal study (the MTA sample), where adding substances have been plagued with inconsistency in the genetic information to baseline data predicted the devel- replication of genetic association results. This may be due opment of SUD at later ages, as determined from infor- to reasons such as: (i) small effect size of variants expected mation gathered over a period of more than 10 years. to influence the SUD phenotype, as with any complex Additionally, our results generalized to a sample of disease;58 (ii) insufficient power to detect significant patients with severe SUD from Kentucky (U.S.) that were associations due to small sample size;59 (iii) phenotypic not ascertained on the basis of ADHD diagnosis. heterogeneity of SUD across samples that may reflect The first genome-wide significant ADHD risk loci were different disease stages or multiple subtypes (i.e., single- published recently65. Marker rs4860437 is not represented drug versus poly-drug dependence/use); (iv) genetic het- in this dataset; however, this study was not aimed at erogeneity arising from distinct risk genes sets; (v) eth- identifying loci shared between ADHD and SUD. In any nicity inconsistencies between discovery and replication case, while genome-wide association studies are a useful samples;60,61 and (vi) comorbidity with other psychiatric tool for discovering novel risk variants—as it involves a conditions (e.g., ADHD) with shared genetic and envir- hypothesis-free interrogation of the entire genome—the onmental architecture62,63. Consequently, additional stu- lack of genetic association may be a reflection of the dies are required to identify new SUD candidate genes polygenic, multifactorial nature of ADHD, with both and to help dissect genetic contributions in the context of common and rare variants likely contributing small effects complex interactions with co-morbid conditions. to its etiology66–68. In addition, an important factor may In this study, we present a demographic, clinical and be the genetic heterogeneity of ADHD subtypes, which genetic framework generated using ARPA that is able to may have different underlying genetic mechanisms. predict the risk of developing SUD. Interestingly, marker Therefore, genome-wide significance may identify loci rs4860437 showed a differential splitting pattern in the with larger genetic effects, while others with smaller Paisa, Spain, and Kentucky cohorts. For instance, in Fig. effects remain undetected for a given population size. 1a, rs4860437 splits into (G/G, G/T) and T/T; in Fig. 2a, Variation in ADGRL3 has been implicated in ADHD in the same variable splits into (G/G, T/T) and G/T; and in diverse populations24,27–29,31–34,69. ADGRL3 is a member Fig. 4a, it splits into (G/T, T/T) and G/G. The most of the latrophilin subfamily of G-protein-coupled recep- parsimonious and plausible explanation of this splitting tors (GPCR)70 and is most strongly expressed in brain pattern is the presence of genomic variability surrounding regions implicated in the neurophysiological basis of this proxy marker, reflecting ancestral composition. ADHD32,52,71. Mouse and zebrafish knockout models also Future studies of genomic regions surrounding rs4860437 support ADGRL3 implication in ADHD pathophysiol- might reveal a cryptic mechanism. It is particularly ogy72,73. More recently, Martinez et al35. identified a compelling that ADGRL3 marker rs4860437, which is a brain-specific transcriptional enhancer within ADGRL3 major predictor variable component in the trees for SUD, that contains an ADHD risk haplotype associated with is in complete LD with ADHD susceptibility markers reduced ADGRL3 mRNA expression in the thalamus. rs6551665 and rs1947274 in Caucasians28,30,52, suggesting This haplotype was associated not only with ADHD, but that the phenotype underpinning SUD is under the also with disruptive behaviors, including SUD35. A pleiotropic effect of ADGRL3 variants. Unfortunately, member of the family of leucine-rich repeat transmem- rs4860437 was not included in the exome chip used to brane (FLRT) proteins has been identified as an endo- genotype the MTA sample and, therefore, could not be genous postsynaptic ligand for latrophilins74. Interference Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 12 of 15 with this interaction reduces excitatory synapse density in differential genetic susceptibility not only to SUD, but also cultured neurons and decreases afferent input strength to the long-term protective effects of medication treat- and dendritic spine number in dentate granule cells, ment. Confirmation of such hypothesis would have sub- which implicates ADGRL3 and FLRT3 in glutamatergic stantial public health implications. Inasmuch as ADGRL3 synapse development74. Similarly, convergent evidence participates in synaptic formation and function, its from a network analysis of a gene set significantly asso- involvement in SUD could be mediated by either influ- ciated and/or linked to ADHD and SUD revealed path- encing brain development or moderating drug-induced ways involved in axon guidance, regulation of synaptic changes in synaptic strength. Molecular studies are transmission, and regulation of transmission of nerve required to elucidate the pathogenic mechanism(s) asso- impulse18. These data altogether suggest that ADGRL3 ciated with ADGRL3 dysfunction in SUD. may be an important SUD susceptibility gene. Strong evidence from clinical and genetic association Acknowledgements First and foremost, we thank all the patients who participated in this study. studies suggests that genetic factors play a crucial role in This study was supported by the National Human Genome Research Institute shaping the susceptibility to both ADHD and SUD75–81. intramural funds (to M.M.) for the ascertainment of patients from the Paisa More strikingly, ADHD treatment has been shown to population. The Kentucky sample was ascertained using funds from the Eli Lilly Research Foundation (to J.d.L.), a National Alliance for Research on reduce the risk of SUD82–84. Though the neurobiological Schizophrenia and Depression (NARSAD) Independent Award (to J.d.L.), basis for this association remains unclear, a variety of University of Kentucky internal funding (to J.d.L.), and Roche Molecular causal pathways from ADHD to SUD have been proposed Systems, Inc., who provided free genotyping and laboratory supplies (to J.d.L.). J.I.V. is partially supported by research grant FOFICO 32101-511035-PE0031 that involve conduct problems77,79. Clinical studies have from Universidad del Norte, Barranquilla, Colombia. The Multimodal Treatment suggested that the link between SUD and ADHD dis- Study of Children with ADHD (MTA) was a National Institute of Mental Health appears after controlling for co-morbid CD84–86. In (NIMH) cooperative agreement randomized clinical trial, continued under an NIMH contract as a follow-up study and finally under a National Institute on agreement with these studies, the presence of CD was a Drug Abuse (NIDA) contract followed by a data analysis grant (DA039881). The major predictor of SUD in the ARPA-based predictive Spanish sample was recruited, assessed and genotyped using funds from the models for SUD in the Paisa and Spanish cohorts (both Instituto de Salud Carlos III, Spain (PI12/01139, PI14/01700, PI15/01789, PI16/ 01505), and co-financed by the European Regional Development Fund (ERDF); assessed first for ADHD) (Figs. 1 and 2). Agència de Gestió d’Ajuts Universitaris i de Recerca-AGAUR, Generalitat de Some researchers implicate genetically mediated per- Catalunya (2014SGR1357, 2014SGR0932); Departament de Salut, Generalitat de sonality traits, such as impulsivity and lack of inhibitory Catalunya; Ministerio de Economía, Industria y Competitividad, Spain (SAF2015-68341-R); the European College of Neuropsychopharmacology control (common to ADHD and disruptive behaviors) as a (ECNP network: ‘ADHD across the lifespan’) and a NARSAD Young Investigator link between ADHD and SUD resulting from common Grant from the Brain & Behavior Research Foundation. This project has also neurological substrates87. Some investigators have pro- received funding from the European Union’s Horizon 2020 research and innovation programme (CoCA under grant agreement 667302). C.S.-M. is posed that patients with ADHD use addictive substances recipient of a Sara Borrell contract from the Instituto de Salud Carlos III, to self-medicate88 and that the differential response to Ministerio de Economía, Industria y Competitividad, Spain (CD15/00199). M.R. is drugs of abuse and atypical behavioral regulation in a recipient of a Miguel de Servet contract from the Instituto de Salud Carlos III, Ministerio de Economía, Industria y Competitividad, Spain (CP09/00119 and response to social cues (e.g., social modeling and peer CPII15/00023). N.F.-C. was supported by a contract of the Centro de selection) may fuel substance use77,89. Others suggest that Investigación Biomédica en Red de Enfermedades Raras (CIBERER). the poor judgment and impulsivity associated with ADHD contribute to the development of substance depen- Collaborators from NIMH dence79. Clinical variables from childhood have also been Benedetto Vitiello, M.D. (Child & Adolescent Treatment and Preventive Interventions Research Branch), Joanne B. Severe, M.S. (Clinical Trials associated with SUD in patients with ADHD, such ADHD Operations and Biostatistics Unit, Division of Services and Intervention subtype, temper characteristics (fear, accident propensity), Research), Peter S. Jensen, M.D. (currently at The REACH Institute), L. Eugene sexual abuse, suspension from school, and a family history Arnold, M.D., M.Ed. (currently at Ohio State University), and Kimberly Hoagwood, Ph.D. (currently at NYU School of Medicine); previous contributors of ADHD13. from NIMH to the early phases: John Richters, Ph.D. (currently at National In summary, our results support a possible functional Institute of Nursing Research) and Donald Vereen, M.D. (currently at NIDA). role for ADGRL3 in modulating drug seeking behavior. Principal investigators and co-investigators from the sites are: University of California, Berkeley/San Francisco: Stephen P. Hinshaw, Ph.D. (Berkeley), Glen R. Regardless of the type of abused substance, longitudinal Elliott, Ph.D., M.D. (San Francisco); Duke University: Karen C. Wells, Ph.D., Jeffery studies generally find that the onset of ADHD precedes N. Epstein, Ph.D. (currently at Cincinnati Children's Hospital Medical Center), that of SUD, suggesting that the psychopathology of and Desiree W. Murray, Ph.D.; previous Duke contributors to early phases: C. Keith Conners, Ph.D. (former PI; deceased) and John March, M.D., M.P.H.; ADHD is not secondary to SUD in most patients79. University of California, Irvine: James Swanson, Ph.D. and Timothy Wigal, Ph.D.; Accordingly, it is reasonable to consider that timely previous contributor from UCLA to the early phases: Dennis P. Cantwell, M.D. diagnosis and treatment of ADHD with stimulant medi- (deceased); New York University School of Medicine: Howard B. Abikoff, Ph.D.; Montreal Children's Hospital/ McGill University: Lily Hechtman, M.D.; and New cation may reduce the occurrence and/or severity of SUD. York State Psychiatric Institute/Columbia University/Mount Sinai Medical Based on the relationship with medication response30, we Center: Laurence L. Greenhill, M.D. (Columbia) and Jeffrey H. Newcorn, M.D. speculate that ADGRL3 variants may underlie a (Mount Sinai School of Medicine). University of Pittsburgh: Brooke Molina, Ph. D., Betsy Hoza, Ph.D. (currently at University of Vermont), and William E. Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 13 of 15 Pelham, Ph.D. (PI for early phases, currently at Florida International University). Received: 15 February 2017 Revised: 18 December 2018 Accepted: 2 Follow-up phase statistical collaborators: Robert D. Gibbons, Ph.D. (University January 2019 of Illinois, Chicago); Sue Marcus, Ph.D. (Icahn School of Medicine at Mt. Sinai) and Kwan Hur, Ph.D. (University of Illinois, Chicago). Original study statistical and design consultant: Helena C. Kraemer, Ph.D. (Stanford University). Collaborator from the Office of Special Education Programs/US Department of Education: Thomas Hanley, Ed.D. Collaborator from Office of Juvenile Justice References and Delinquency Prevention/Department of Justice: Karen Stern, Ph.D. 1. Whiteford, H. A., Ferrari, A. J., Degenhardt, L., Feigin, V. & Vos, T. The global burden of mental, neurological and substance use disorders: an analysis from the Global Burden of Disease Study 2010. PLoS ONE 10, e0116820 (2015). 2. Heslin K. C., Elixhauser A., Steiner C. A. Hospitalizations Involving Mental and Substance Use Disorders Among Adults, 2012: Statistical Brief #191 Healthcare Author details Cost and Utilization Project (HCUP) Statistical Briefs. (Agency for Healthcare 1 Medical Genetics Branch, National Human Genome Research Institute, Research and Quality, Rockville, MD, 2010). National Institutes of Health, Bethesda, MD, USA. 2INPAC Research Group, 3. Prom-Wormley, E. C., Ebejer, J., Dick, D. M. & Bowers, M. S. The genetic epi- Fundación Universitaria Sanitas, Bogotá, Colombia. 3Instituto de demiology of substance use disorder: a review. Drug Alcohol. Depend. 180, Investigaciones Médicas (IIM), Facultad de Medicina, Universidad de Antioquia, 241–259 (2017). Medellín, Colombia. 4Universidad del Norte, Barranquilla, Colombia. 5Psychiatric 4. Centers for Disease C, Prevention. Increasing prevalence of parent-reported Genetics Unit, Group of Psychiatry, Mental Health and Addiction, Vall d’Hebron attention-deficit/hyperactivity disorder among children—United States, 2003 Research Institute (VHIR), Universitat Autònoma de Barcelona, Barcelona, Spain. and 2007. Mmwr. Morb. Mortal. Wkly. Rep. 59, 1439–1443 (2010). 6 Department of Psychiatry, Hospital Universitari Vall d’Hebron, Barcelona, Spain. 5. Visser, S. N. et al. Trends in the parent-report of health care provider-diagnosed 7 Biomedical Network Research Centre on Mental Health (CIBERSAM), and medicated attention-deficit/hyperactivity disorder: United States, Barcelona, Spain. 8Department of Psychiatry and Legal Medicine, Universitat 2003–2011. J. Am. Acad. Child Adolesc. Psychiatry 53, 34–46 e2 (2014). Autònoma de Barcelona, Barcelona, Spain. 9Addiction and Dual Diagnosis Unit, 6. Palacio, J. D. et al. Attention-deficit/hyperactivity disorder and comorbidities in Departament of Psychiatry, Hospital Universitari Vall d’Hebron-Public Health 18 Paisa Colombian multigenerational families. J. Am. Acad. Child Adolesc. Agency, Barcelona, Spain. 10Department of Genetics, Microbiology and Psychiatry 43, 1506–1515 (2004). Statistics, University of Barcelona, Barcelona, CAT, Spain. 11Centro de 7. Sibley, M. H. et al. The delinquency outcomes of boys with ADHD with and Investigación Biomédica en Red de Enfermedades Raras (CIBERER), Instituto de without comorbidity. J. Abnorm. Child Psychol. 39, 21–32 (2011). Salud Carlos III, Madrid, Spain. 12Institut de Biomedicina de la Universitat de 8. Jain, M. et al. Attention-deficit/hyperactivity disorder and comorbid disruptive Barcelona (IBUB), Barcelona, CAT, Spain. 13Institut de Recerca Sant Joan de Déu behavior disorders: evidence of pleiotropy and new susceptibility loci. Biol. (IRSJD), Esplugues, CAT, Spain. 14Neuroscience Research Group, Universidad de Psychiatry 61, 1329–1339 (2007). Antioquia, Medellín, Colombia. 15Grupo de Neurociencias del Caribe, Unidad 9. Kuperman, S. et al. Developmental sequence from disruptive behavior diag- de Neurociencias Cognitivas, Universidad Simón Bolívar, Barranquilla, nosis to adolescent alcohol dependence. Am. J. Psychiatry 158, 2022–2026 Colombia. 16Grupo de Neurociencias del Caribe, Universidad del Atlántico, (2001). Barranquilla, Colombia. 17Departments of Psychiatry and Psychology, University 10. Molina, B. S. et al. Adolescent substance use in the multimodal treatment of Pittsburg, Pittsburg, PA, USA. 18University of Kentucky Mental Health study of attention-deficit/hyperactivity disorder (ADHD) (MTA) as a function of Research Center at Eastern State Hospital, Lexington, KY, USA. 19Department of childhood ADHD, random assignment to childhood treatments, and sub- Biochemistry and Biophysics, Texas A&M University, College Station, TX, USA. sequent medication. J. Am. Acad. Child Adolesc. Psychiatry 52, 250–263 (2013). 20 National Center for Indigenous Genomics, Genome Biology Department, 11. Molina, B. S. G. & Pelham, W. E. Childhood predictors of adolescent substance John Curtin School of Medical Research, ANU College of Medicine, Biology and use in a longitudinal study of children with ADHD. J. Abnorm. Psychol. 112, Environment, The Australian National University, Canberra, ACT, Australia. 497–507 (2003). 21 Department of Psychiatry, Florida International University, Miami, FL, USA. 12. Biederman, J. et al. Psychoactive substance use disorders in adults with 22 Child Development Center, University of California at Irvine, Irvine, CA, USA. attention-deficit hyperactivity Disorder (ADHD)—effects of ADHD and psy- 23 Genome Discovery Unit, Genome Biology Department, John Curtin School of chiatric comorbidity. Am. J. Psychiatry 152, 1652–1658 (1995). Medical Research, ANU College of Medicine, Biology and Environment, The 13. Nogueira, M. et al. Early-age clinical and developmental features associated to Australian National University, Canberra, ACT, Australia. 24Office of the Director, substance use disorders in attention-deficit/hyperactivity disorder in adults. National Institute on Drug Abuse, National Institutes of Health, Rockville, MD, Compr. Psychiatry 55, 639–649 (2014). USA. 25Department of Child and Adolescent Psychiatry, Hassenfeld Children’s 14. DeMilio, L. Psychiatric syndromes in adolescent substance abusers. Am. J. Hospital at NYU Langone, New York, NY, USA. 26Nathan Kline Institute for Psychiatry 146, 1212–1214 (1989). Psychiatric Research, Orangeburg, NY, USA. 27Center for Research in Genetics 15. Horner, B. R. & Scheibe, K. E. Prevalence and implications of attention-deficit and Genomics, Institute of Translational Medicine, School of Medicine and hyperactivity disorder among adolescents in treatment for substance abuse. J. Health Sciences, Universidad del Rosario, Bogotá, Colombia Am. Acad. Child Adolesc. Psychiatry 36, 30–36 (1997). 16. Ercan, E. S., Coskunol, H., Varan, A. & Toksoz, K. Childhood attention deficit/ hyperactivity disorder and alcohol dependence: a 1-year follow-up. Alcohol. Alcohol. 38, 352–356 (2003). Conflict of interest 17. White, A. M. et al. Predictors of relapse during treatment and treatment J.A.R.-Q has served on the speakers’ bureau and acted as consultant for Eli Lilly completion among marijuana-dependent adolescents in an intensive out- and Co., Janssen-Cilag, Novartis, Lundbeck, Shire, Ferrer, and Laboratorios patient substance abuse program. Subst. Abus. 25, 53–59 (2004). Rubió. He has received travel awards from Eli Lilly and Co., Janssen-Cilag, and 18. Arcos-Burgos, M., Velez, J. I., Solomon, B. D. & Muenke, M. A common genetic Shire for participating in psychiatric meetings. The ADHD Program chaired by network underlies substance use disorders and disruptive or externalizing J.A.R.-Q. has received unrestricted educational and research support from Eli disorders. Hum. Genet. 131, 917–929 (2012). Lilly and Co., Janssen-Cilag, Shire, Rovi, Ferrer and Laboratorios Rubió in the 19. Gorwood, P. et al. Genetics of dopamine receptors and drug addiction. Hum. past two years. The remaining authors declare that they have no conflict of Genet. 131, 803–822 (2012). interest. 20. Hart, A. B. et al. Genome-wide association study of d-amphetamine response in healthy volunteers identifies putative associations, including cadherin 13 (CDH13). PLoS ONE 7, e42646 (2012). Publisher’s note 21. Li, M. D. & Burmeister, M. New insights into the genetics of addiction. Nat. Rev. Springer Nature remains neutral with regard to jurisdictional claims in Genet. 10, 225–231 (2009). published maps and institutional affiliations. 22. Acosta, M. T. et al. Latent class subtyping of attention-deficit/hyperactivity disorder and comorbid conditions. J. Am. Acad. Child Adolesc. Psychiatry 47, Supplementary information accompanies this paper at (https://doi.org/ 797–807 (2008). 10.1038/s41398-019-0396-7). Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 14 of 15 23. Arcos-Burgos, M. et al. Attention-deficit/hyperactivity disorder (ADHD): feasi- 46. Shaffer, D., Fisher, P., Lucas, C. P., Dulcan, M. K. & Schwab-Stone, M. E. NIMH bility of linkage analysis in a genetic isolate using extended and multi- Diagnostic Interview Schedule for Children Version IV (NIMH DISC-IV): generational pedigrees. Clin. Genet. 61, 335–343 (2002). description, differences from previous versions, and reliability of some com- 24. Acosta, M. T. et al. A two-locus genetic interaction between LPHN3 and 11q mon diagnoses. J. Am. Acad. Child Adolesc. Psychiatry 39, 28–38 (2000). predicts ADHD severity and long-term outcome. Transl. Psychiatry 1, e17 47. Robins L. N., Cottler L., Bucholz K., Compton W. The Diagnostic Interview (2011). Schedule for DSM-IV (DIS-IV). (Washington University School of Medicine, St. 25. Arcos-Burgos, M. et al. Attention-deficit/hyperactivity disorder in a population Louis, MO, 2000). isolate: linkage to loci at 4q13.2, 5q33.3, 11q22, and 17p11. Am. J. Hum. Genet. 48. de Leon, J., Susce, M. T., Diaz, F. J., Rendon, D. M. & Velasquez, D. M. Variables 75, 998–1014 (2004). associated with alcohol, drug, and daily smoking cessation in patients with 26. Acosta, M. T. et al. ADGRL3 (LPHN3) variants are associated with a refined severe mental illnesses. J. Clin. Psychiatry 66, 1447–1455 (2005). phenotype of ADHD in the MTA study. Mol. Genet. Genom. Med. 4, 540–547 49. Drake, R. E., Mchugo, G. J. & Biesanz, J. C. The test-retest reliability of stan- (2016). dardized instruments among homeless persons with substance use disorders. 27. Bruxel, E. M. et al. LPHN3 and attention-deficit/hyperactivity disorder: a sus- J. Stud. Alcohol. 56, 161–167 (1995). ceptibility and pharmacogenetic study. Genes. Brain. Behav. 14, 419–427 50. Drake, R. E., Rosenberg, S. D. & Mueser, K. T. Assessing substance use disorder (2015). in persons with severe mental illness. New Dir. Ment. Health Serv. 1996, 3–17 28. Hwang, I. W., Lim, M. H., Kwon, H. J. & Jin, H. J. Association of LPHN3rs6551665 (1996). A/G polymorphism with attention deficit and hyperactivity disorder in Korean 51. de Leon, J., Armstrong, S. C. & Cozza, K. L. Clinical guidelines for psychiatrists for children. Gene 566, 68–73 (2015). the use of pharmacogenetic testing for CYP450 2D6 and CYP450 2C19. Psy- 29. Jain, M. et al. A cooperative interaction between LPHN3 and 11q doubles the chosomatics 47, 75–85 (2006). risk for ADHD. Mol. Psychiatry 17, 741–747 (2012). 52. Arcos-Burgos, M. & Muenke, M. Toward a better understanding of ADHD: 30. Arcos-Burgos, M. et al. A common variant of the latrophilin 3 gene, LPHN3, LPHN3 gene variants and the susceptibility to develop ADHD. Atten. Defic. confers susceptibility to ADHD and predicts effectiveness of stimulant medi- Hyperact Disord. 2, 139–147 (2010). cation. Mol. Psychiatry 15, 1053–1066 (2010). 53. Wong, M. L., Dong, C., Andreev, V., Arcos-Burgos, M. & Licinio, J. Prediction of 31. Labbe, A. et al. Refining psychiatric phenotypes for response to treatment: susceptibility to major depression by a model of interactions of multiple contribution of LPHN3 in ADHD. Am. J. Med. Genet. B. Neuropsychiatr. Genet. functional genetic variants and environmental factors. Mol. Psychiatry 17, 159B, 776–785 (2012). 624–633 (2012). 32. Ribases, M. et al. Contribution of LPHN3 to the genetic susceptibility to ADHD 54. Rao, D. C. CAT scans, PET scans, and genomic scans. Genet. Epidemiol. 15, 1–18 in adulthood: a replication study. Genes. Brain. Behav. 10, 149–157 (2011). (1998). 33. Song, J. et al. Association of SNAP-25, SLC6A2, and LPHN3 with OROS 55. Ryzin, J. V., Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, C. J. Classification methylphenidate treatment response in attention-deficit/hyperactivity dis- and regression trees. J. Am. Stat. Assoc. 81, 253 (1986). order. Clin. Neuropharmacol. 37, 136–141 (2014). 56. Breiman L. Random Forests. In Machine Learning, Vol. 45.(ed. Schapire, R. E.) 34. Gomez-Sanchez, C. I. et al. Attention deficit hyperactivity disorder: genetic (Kluwer Academic Publishers, The Netherlands, 2001). association study in a cohort of Spanish children. Behav. Brain. Funct. 12, 2 57. Friedman J. H. Greedy Function Approximation: a Gradient Boosting Machine. (2016). (University of Stanford, Stanford, CA, 1999). 35. Martinez, A. F. et al. An ultraconserved brain-specific enhancer within ADGRL3 58. Marjoram, P., Zubair, A. & Nuzhdin, S. V. Post-GWAS: where next? More sam- (LPHN3) underpins attention-deficit/hyperactivity disorder susceptibility. Biol. ples, more SNPs or more biology? Hered. (Edinb.) 112, 79–88 (2014). Psychiatry 80, 943–954 (2016). 59. Visscher, P. M., Brown, M. A., McCarthy, M. I. & Yang, J. Five years of GWAS 36. Ramos-Quiroga, J. A. et al. Criterion and concurrent validity of Conners Adult discovery. Am. J. Hum. Genet. 90, 7–24 (2012). ADHD Diagnostic Interview for DSM-IV (CAADID) Spanish version. Rev. Psi- 60. Melroy-Greif, W. E. et al. Examination of the involvement of cholinergic- quiatr Salud Ment. 5, 229–235 (2012). associated genes in nicotine behaviors in European and African Americans. 37. Group, T. M. C. A 14-month randomized clinical trial of treatment strategies for Nicotine. Tob. Res. 19, 417–425 (2017). attention-deficit/hyperactivity disorder. The MTA Cooperative Group. Multi- 61. Polimanti, R., Yang, C., Zhao, H. & Gelernter, J. Dissecting ancestry genomic modal treatment study of children with ADHD. Arch. Gen. Psychiatry 56, background in substance dependence genome-wide association studies. 1073–1086 (1999). Pharmacogenomics 16, 1487–1498 (2015). 38. Jensen, P. S. et al. 3-year follow-up of the NIMH MTA study. J. Am. Acad. Child 62. Bi, J., Gelernter, J., Sun, J. & Kranzler, H. R. Comparing the utility of homo- Adolesc. Psychiatry 46, 989–1002 (2007). geneous subtypes of cocaine use and related behaviors with DSM-IV cocaine 39. Molina, B. S. et al. The MTA at 8 years: prospective follow-up of children treated dependence as traits for genetic association analysis. Am. J. Med. Genet. B. for combined-type ADHD in a multisite study. J. Am. Acad. Child Adolesc. Neuropsychiatr. Genet. 165B, 148–156 (2014). Psychiatry 48, 484–500 (2009). 63. Palmer, R. H. et al. Examining the role of common genetic variants on alcohol, 40. Hechtman, L. et al. Functional adult outcomes 16 years after childhood tobacco, cannabis and illicit drug dependence: genetics of vulnerability to diagnosis of attention-deficit/hyperactivity disorder: MTA results. J. Am. Acad. drug dependence. Addiction 110, 530–537 (2015). Child Adolesc. Psychiatry 55, 945–52 e2 (2016). 64. Neale, B. M. & Sham, P. C. The future of association studies: gene-based 41. Swanson, J. M. et al. Young adult outcomes in the follow-up of the multi- analysis and replication. Am. J. Hum. Genet. 75, 353–362 (2004). modal treatment study of attention-deficit/hyperactivity disorder: symptom 65. Demontis, D. et al. Discovery of the first genome-wide significant risk loci for persistence, source discrepancy, and height suppression. J. Child Psychol. attention deficit/hyperactivity disorder. Nat. Genet. 51, 63–75 (2019). Psychiatry 58, 663–678 (2017). 66. Stergiakouli, E. et al. Investigating the contribution of common genetic var- 42. Roy, A. et al. Childhood predictors of adult functional outcomes in the mul- iants to the risk and pathogenesis of ADHD. Am. J. Psychiatry 169, 186–194 timodal treatment study of attention-deficit/hyperactivity disorder (MTA). J. (2012). Am. Acad. Child Adolesc. Psychiatry 56, 687–95 e7 (2017). 67. Poelmans, G., Pauls, D. L., Buitelaar, J. K. & Franke, B. Integrated genome-wide 43. Molina, B. S. et al. Delinquent behavior and emerging substance use in the association study findings: identification of a neurodevelopmental network for MTA at 36 months: prevalence, course, and treatment effects. J. Am. Acad. attention deficit hyperactivity disorder. Am. J. Psychiatry 168, 365–377 (2011). Child Adolesc. Psychiatry 46, 1028–1040 (2007). 68. Williams, N. M. et al. Genome-wide analysis of copy number variants in 44. Johnson, C. et al. Pooled association genome scanning for alcohol depen- attention deficit hyperactivity disorder: the role of rare variants and duplica- dence using 104,268 SNPs: validation and use to identify alcoholism vulner- tions at 15q13.3. Am. J. Psychiatry 169, 195–204 (2012). ability loci in unrelated individuals from the collaborative study on the 69. Choudhry, Z. et al. LPHN3 and attention-deficit/hyperactivity disorder: inter- genetics of alcoholism. Am. J. Med. Genet. B. Neuropsychiatr. Genet. 141B, action with maternal stress during pregnancy. J. Child Psychol. Psychiatry 53, 844–853 (2006). 892–902 (2012). 45. Uhl GR. Molecular genetics of substance abuse vulnerability: remarkable 70. Martinez, A. F., Muenke, M. & Arcos-Burgos, M. From the black widow spider to recent convergence of genome scan results. Ann. N. Y. Acad. Sci. 1025, 1–13 human behavior: Latrophilins, a relatively unknown class of G protein-coupled (2004). receptors, are implicated in psychiatric disorders. Am. J. Med. Genet. B. Neu- ropsychiatr. Genet. 156B, 1–10 (2011). Arcos-Burgos et al. Translational Psychiatry (2019)9:42 Page 15 of 15 71. Koob, G. F. & Volkow, N. D. Neurocircuitry of addiction. Neuropsychopharma- 81. Chen, Q. et al. Common psychiatric and metabolic comorbidity of adult cology 35, 217–238 (2010). attention-deficit/hyperactivity disorder: a population-based cross-sectional 72. Lange, M. et al. The ADHD-susceptibility genelphn3.1 modulates dopami- study. PLoS ONE 13, e0204516 (2018). nergic neuron formation and locomotor activity during zebrafish develop- 82. Quinn, P. D. et al. ADHD medication and substance-related problems. Am. J. ment. Mol. Psychiatry 17, 946–954 (2012). Psychiatry 174, 877–885 (2017). 73. Wallis, D. et al. Initial characterization of mice null for Lphn3, a gene implicated 83. Chang, Z. et al. Stimulant ADHD medication and risk for substance abuse. J. in ADHD and addiction. Brain Res. 1463, 85–92 (2012). Child Psychol. Psychiatry 55, 878–885 (2014). 74. O'Sullivan, M. L. et al. FLRT proteins are endogenous latrophilin ligands and 84. Groenman, A. P. et al. Substance use disorders in adolescents with attention regulate excitatory synapse development. Neuron 73, 903–910 (2012). deficit hyperactivity disorder: a 4-year follow-up study. Addiction 108, 75. Chassin L., Colder C. R., Hussong A., Sher K. J. Substance Use and Substance 1503–1511 (2013). Use Disorders. (ed. Cicchetti, D.) Developmental Psychopathology, 2nd edn, Vol. 85. Flory, K. & Lynam, D. R. The relation between attention deficit hyperactivity 3. (John Wiley & Sons, Inc., Hoboken, NJ, 2016) pp 833-897. disorder and substance abuse: what role does conduct disorder play? Clin. 76. Brook, J. S., Balka, E. B., Zhang, C. & Brook, D. W. ADHD, conduct disorder, Child Fam. Psychol. Rev. 6, 1–16 (2003). substance use disorder, and nonprescription stimulant use. J. Atten. Disord. 21, 86. Lee, S. S., Humphreys, K. L., Flory, K., Liu, R. & Glass, K. Prospective association of 776–782 (2017). childhood attention-deficit/hyperactivity disorder (ADHD) and substance use 77. Molina, B. S. & Pelham, W. E. Jr. Attention-deficit/hyperactivity disorder and risk and abuse/dependence: a meta-analytic review. Clin. Psychol. Rev. 31, 328–341 of substance use disorder: developmental considerations, potential pathways, (2011). and opportunities for research. Annu. Rev. Clin. Psychol. 10, 607–639 (2014). 87. Valero, S. et al. Personality profile of adult ADHD: the alternative five factor 78. Zulauf, C. A., Sprich, S. E., Safren, S. A. & Wilens, T. E. The complicated rela- model. Psychiatry Res. 198, 130–134 (2012). tionship between attention deficit/hyperactivity disorder and substance use 88. Khantzian, E. J. The self-medication hypothesis of substance use disorders: a disorders. Curr. Psychiatry Rep. 16, 436 (2014). reconsideration and recent applications. Harv. Rev. Psychiatry 4, 231–244 79. Wilens, T. E. & Biederman, J. Alcohol, drugs, and attention-deficit/ hyperactivity (1997). disorder: a model for the study of addictions in youth. J. Psychopharmacol. 20, 89. Pedersen, S. L., Harty, S. C., Pelham, W. E., Gnagy, E. M. & Molina, B. S. G. 580–588 (2006). Differential associations between alcohol expectancies and adolescent alcohol 80. Estevez-Lamorte, N. et al. Adult attention-deficit/hyperactivity disorder, risky use as a function of childhood ADHD. J. Stud. Alcohol. Drugs 75, 145–152 substance use and substance use disorders: a follow-up study among young (2014). men. Eur. Arch. Psychiatry Clin. Neurosci. (2018). https://doi.org/10.1007/s00406- 018-0958-3.