Untitled

Genetic markers anticipate response to citalopramin a majority of patientsFarrokh Alemia, Manaf Zargoushf, Harold Erdmana, Jee Vanga,Steve Epsteinb and Fanous Aymanb,c,d,e Objective Scientists have concluded that genetic profiles prescription of antidepressants. Psychiatr Genet cannot predict a large percentage of variation in response to citalopram, a common antidepressant. Using the same data, we examined if a different conclusion can be arrived at when the results are personalized to fit specific patients.
Keywords: antidepressant, citalopram, Classification and Regression Trees, Methods We used data available through the Sequenced interaction among genetic markers, personalized medicine Treatment Alternatives to Relieve Depression database. We aDepartment of Health Systems Administration, bDepartment of Psychiatry, created three boosted Classification and Regression Trees Georgetown University, cMental Health Service Line, Washington VA Medical to identify 16 subgroups of patients, among whom Center, Northwest, Washington, District of Columbia, dDepartment of Psychiatry,Virginia Commonwealth University School of Medicine, Richmond, Virginia, anticipation of positive or negative response to citalopram eDepartment of Psychiatry, Keck School of Medicine of the University of Southern was significantly different from 0.5 (P r 0.1).
California, Los Angeles, California, USA and fESSEC Business School, AvenueBernard Hirsch, Cergy-Pontoise Cedex, France Results In a 10-fold cross-validation, this ensemble of Correspondence to Farrokh Alemi, PhD, Department of Health Systems trees made no predictions in 33% of cases. In the Administration, Georgetown University, 3700 Reservoir Road, Washington, DC20007, USA remaining 67% of cases, it accurately classified response Tel: + 1 703 283 3100; fax: + 1 202 784 3127; Received 22 September 2010 Revised 7 December 2010 Conclusion For the majority of the patients, genetic markers can be used to guide selection of citalopram.
The rules identified in this study can help personalize It is generally assumed that the discovery of genetic processes and markers should precede translation to Despite progress, the effect size is small and the practice. This is not always the case. It is theoretically percentage of outcomes correctly predicted from any possible that scientists may find no markers or genetic single genetic marker is near random chance processes that are relevant in the general population but . This has led the investigators to conclude when the findings are restricted to specific patients, as that there is ‘limited clinical utility in matching when the data are personalized, the findings may be antidepressants to patient’s genetic profile’ radically different. In short, recommendations to one The studies to date have not examined the patient may differ from advice to the average patient.
combination of various genetic markers or examined This possibility changes how scientific findings should be subgroups of patients among whom response might be approached in practice setting. Translational research predicted more accurately. In this study, we will examine requires us to pay less attention to conclusions of the the ability to predict response to citalopram from a scientist and more attention to the data collected by combination of genetic markers within subgroups of the scientist. In particular, it requires us to reanalyze the subset of scientist’s data that is relevant to the patient athand. To demonstrate this possibility, we set out to Citalopram (brand names Celexa, Cipramil, etc.) is a reexamine scientific data on response to antidepressants common antidepressant prescribed in United States and and to see if personalized recommendations could differ from the general scientific findings.
more than 18 million patients have taken this medication The search for genetic markers for response to anti- but the majority of patients (60%) do not achieve depressants has identified a number of markers. Variants in HTR2A, GRIK4, KCNK2, FKBP5, PDE11A and Untreated or poorly treated depression leads BDNF, and SLC6A4 have been identified in some but to significant functional impairment and may even lead to not all studies to be predictive of response to anti- self-medication through alcohol and illicit drugs and c 2011 Wolters Kluwer Health | Lippincott Williams & Wilkins Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
sometimes suicide. Poorly responsive or untreated response to treatment is statistically significant. In both patients are twice as likely to be hospitalized and incur approaches the advice to the patient is based on the 19 times greater cost than patients with effectively outcomes of similar cases within the database. Both approaches are part of algorithms called by statisticians as statistics is to anticipate response to citalopram and prescribe it for patients who are likely to benefit from it.
As it may take weeks to determine if an antidepressant ‘patients-like-me algorithms’. These approaches allow will be helpful, it would be useful to know in advance if a one to select a small subset of data and anticipate a particular medication is likely to be successful for a given patient’s response from the experience of cases within that subset. Contrary to usual statistical procedures, theobjective is not to rely on a large data set but on the Data suggest that response to antidepressants depends smallest most relevant subset of data.
on a host of factors including age ),comorbidities such as diabetes sex, subtype of depression, for example This study examines the response to citalopram through a reanalysis of the data available from National Institute of factors. The wide variation in phenotypic predictors of Mental Health: the Sequenced Treatment Alternatives to response to antidepressants suggests genetic factors Relieve Depression database. National Institute of might be affecting the outcomes of different subgroups.
Mental Health provided a public release of the STAR*D Of particular interest are studies of differences among sex database in August 2008. The design of the STAR*D and race in response to antidepressants ( study has been described elsewhere .
The STAR*D project enrolled more than 4200 out- variants of polymorphisms in the CYP2D6, CYP2C19, patients (aged 18–75 years) diagnosed with nonpsychotic CYP3A4, CYP3A5, and ABCB1 genes were associated major depressive disorder. Data were collected from 41 with citalopram metabolism. Six of the seven variants primary care and mental health clinics. All of these were found in African–Americans and not in Caucasians.
patients were prescribed citalopram. If patients did not These data support the notion that there are different achieve remission or could not tolerate the medication, markers for response to treatment in different subgroups they were encouraged to proceed to the next random assignment, in which they received other medications orcognitive therapy. Those who achieved remission or Subgroup analysis is not new and many studies examine reduction in symptoms and tolerated acute treatment response to antidepressants among different groups.
Looking at racial subgroups is one way to do so. Onecan also group patients by age, by comorbidity, depression Genetic profiles were available on 1933 cases within the subtype or clinical features, and by a host of different STAR*D database. The DNA sampling and the definition factors including genetic markers. In fact, any one of the of remission have been described in a previous publica- genetic markers in a patient profile can be used to define a new subgroup. If there are ‘n’ predictors of response to considered responsive to treatment if at follow-up they therapy, there are 2n possible subgroups. This raises the were in remission; defined as patients scoring 5 or less on possibility of a very large number of subgroups, too many the Quick Inventory of Depressive Symptomatology to be practical. Not surprisingly, none of the studies (Clinician Rating) at follow-up. Genotyping was con- reported to date have implemented a complete subgroup ducted on two platforms: Affymetrix Human Mapping analysis. One way to limit the number of subgroups 500 K Array Set and the Affymetrix Genome-Wide Human examined is to examine only those groups within which SNP Array 5.0 (SeqWright Inc., Houston, Texas, USA).
an accurate prediction can be made. To accomplish this, This resulted in 430 198 validated single nuclide poly- we used Classification and Regression Trees (CART) as implemented in the SPSS 17 Tree add-on package (IBM identified the top 25 SNPs associated with response to North America, New York, New York, USA).
citalopram in a genome-wide association study. Wefocused on these 25 markers because they were the most CART is used to group cases. When a new patient likely SNPs that might affect response.
presents, the group corresponding to the patient isidentified and used to guide the patient. This approach The classification of cases within the STAR*D database into different subgroups was done using a variation of the which cases in the database are ordered in the sequence CART procedure. A traditional CART classifies cases of their similarity to the patient and the sequential within the data set by progressively splitting the data into probability ratio test is used on the next case until two additional classes (two child nodes) so that the new Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Anticipating response to citalopram Alemi et al.
classes can explain the most variance of the response to subgroups in which the probability of positive or negative citalopram. In particular, at every step, the CART response to citalopram is significantly different from 0.5 procedure used the Gini index (quadratic entropy), to (P r 0.10). This approach does not classify all cases, but select one of the 25 SNPs to classify the cases within the the cases it does classify fall into categories in which the database. The procedure continued until all cases were probability of response is significantly different from 0.5.
classified into subgroups of at least 10 cases. The traditional CART procedure leads to a classification tree.
(1) Fit a CART (10-fold cross-validation and pruning Each node along a branch in the tree describes the factor of 1) to all unclassified cases and identify the presence of a specific genetic marker. The entire branch subgroups classified by each branch in the tree.
describes a particular genetic profile. The final node in (2) Stop if none of the subgroups identified have a the branch shows the subgroup of patients with the same probability of positive or negative response signifi- genetic profile and, as far as possible, homogenous cantly different from 0.5 (P r 0.10).
response to citalopram. For example, the right hand (3) Rename cases within subgroups with probability of branch of the tree in describes the situation in positive or negative response not significantly which rs7239368 is AG or AA. A total of 482 patients fall different from 0.5 (P > 0.10) to ‘Unclassified cases’.
into this branch, and 84.4% of these patients respond Specify exclusion rules by using the branches that negatively to citalopram. Therefore, if the patient-at- defined the subgroups that were renamed.
hand has the genetic profile that matches this group, (4) Apply CART (10-fold cross-validation and pruning then there is a good chance that this patient will also factor of 1) to cases not excluded in the previous step. This results in a new tree. The branches withinthis tree specify the inclusion rules.
Since 1990, several investigators have shown methods for This procedure leads to an ensemble of trees. Each tree . In boosted trees, the analyst selects misclassified classifies the response of a particular set of cases. The last cases and fits a separate tree to these cases. The tree contains cases from subgroups, in which none meet procedure is repeated hundreds of times and the the significance criterion. This procedure improves the weighted average of the ensemble of trees is used to performance of the initial tree similar to other boosting make the final predictions. Several studies have shown methods. Unlike other boosting procedures, it has the that boosted trees are more accurate than other advantage that it stops after a few iterations and is easily classification systems such as logistic regression or trees interpretable as every branch within each of the without boosting; but boosted trees are difficult to ensemble of trees provides a rule for defining a specific subgroup of cases. Additional detail on our method of of boosted trees, in which the branches of the tree lead to boosting the performance of CART is made available bythe first two researchers.
The CART procedure was applied with 10-fold cross- validation and pruning, therefore, the percentage of casescorrectly classified was unlikely to be due to over-fitting.
The results we present are data on the accuracy of thecross-validated trees.
The initial CART (without boosting) yielded a tree that correctly classified 56% of cases (standard error in risk 0.01). A lasso-penalized logistic regression with 10-fold cross-validation and 50 indicator variables was assessed.
Two indicator variables represented the three possible genotypes of each one of 25 SNPs. At the optimal tuning parameter l of 0.97, there were 47 nonzero indicator Classification of response to citalopram for 762 cases.
Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
In the above regression, if the patient’s SNP has the The corresponding author of this paper can provide 91 indicated allele, then it has a score of 1, otherwise 0. The rules, combinations of SNPs, that identified which one of regression correctly classified 66% of cases the trees were appropriate. These rules were derivedfrom branches we excluded from the analysis and we refer One could accurately classify 61% of cases by merely to them as exclusion rules. In cases in which none of the predicting no one would benefit from citalopram. There- three trees are appropriate, we cannot make a prediction.
fore, the initial performance of CART or logistic regression Once a tree has been selected, then each branch within was not adequate; we used our boosting procedure to the tree identifies a unique subgroup of patients. We identify subsets of patients for whom we could more refer to these rules as the inclusion rules. There were 16 accurately anticipate response to citalopram. The initial inclusion rules (see ). These rules can be used to tree provided 71 subgroups of patients (branches in the further classify the patient into relevant subgroups. Once tree), each with a different probability of positive response a patient has been classified into a relevant subgroup, to citalopram. We excluded subgroups of patients, in which then the average response of the group can be used to the probability of positive or negative response was not anticipate the outcome for the patient.
significantly different from random tossup. We thenrepeated our CART analysis for these more homogenous identifies 16 inclusion rules identifying subgroups subgroups, as per algorithm provided earlier.
with different probability of positive response to citalo- The boosted CART procedure created an ensemble of pram, ranging from 84 to 6% (94% probability of negative three related trees. There were 1933 cases (see response). The number of cases that fall within each The first boosted tree was organized for 762 cases and subgroup is different, ranging from 13 to 482 cases. A accurately classified 81% of these cases in cross-valida- patient may have more confidence in the advice of the tion. In step two, a tree was organized for 351 cases and system when they fall in the larger groups or if they fall in accurately classified 74% of these cases in cross-valida- a subgroup with more extreme probability of positive/ tion. In step three, a tree was organized that accurately negative response. A statistical test can be done for the classified 76% of 176 cases. This algorithm left 644 (33%) individual patient to see if the probability of positive of cases as unclassified; on these cases the algorithm response within their subgroup is significantly different from a particular value, say 0.5. Such a test providesguidance only to the patient that falls within thesubgroup.
The overall accuracy of the ensemble of the three trees is the sum of the accuracy of each one, weighted by the percentage of cases classified by the trees. Therefore, the ensemble as a whole makes no predictions on 644 (33%) of cases and correctly classifies 78% of the remaining 1289 Anticipating response to citalopram using three trees no predictions were made in 644 (33%) cases Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Anticipating response to citalopram Alemi et al.
Inclusion rules for defining subgroups within the three trees Tree Node rs2744692 rs2697992 rs4017724 rs6817919 rs10499638 rs809736 rs7239368 rs6046805 rs6127921 rs11701162 Our analysis showed that for two thirds of patients, the The purpose of this study was to identify groups of patients, majority of cases, we can accurately classify their response to on the basis of combinations of marker alleles, explaining citalopram. The unaided clinician’s prescription of citalo- the variance in citalopram response. The study provided 91 pram was effective in only 39% of cases. If a clinician would exclusion and 16 inclusion rules for anticipating response to have followed our algorithm, the accuracy of his prescription citalopram. It may be helpful to provide an example of how would have nearly doubled to 78% of cases. Similar to the these 107 rules can be used in a clinical setting. First, one obtains the patient’s genetic profile, using the 25 SNPs significant relationship between the genetic profiles and identified in this study. Second, the exclusion rules are observed response to citalopram. Unlike their study, the applied to determine which one of the three trees is effect size was large and therefore, clinically more relevant.
appropriate. Third, the inclusion rules are applied to seewhich branch of the selected tree describes the patient. If a Our findings were based on retrospective data analysis branch is identified, then the number of cases within the and the results that can be obtained prospectively in a branch and the statistical significance of the finding are new clinic could be very different. Before these findings reported to the clinician/patient. For example, if a patient can be used, we encourage prospective, double-blind has allele AG for rs11128623, TC for rs6817919, and GG for randomized studies to test whether the 107 rules we have rs7239368, then this patient is excluded from the first tree.
identified can improve anticipating response to anti- If this patient has allele GG for rs2697992 and allele TT for rs6817919, then this patient is classified by the second tree.
It is difficult to relate the 107 rules identified in this According to this subgroup, there are 62 cases that have the study to any specific biological process, and this was not same pattern of alleles; citalopram reduced depression the intent of our analysis. Nevertheless, a few specula- symptoms in 82% of these cases. Therefore, the evidence tions can be made. The existence of so many rules from the cases in STAR*D suggests that the patient-at- suggests that there may be multiple pathophysiological hand should try citalopram. Clearly, no patient or clinician processes that lead to depression. Depression is not only can be expected to go through the rules by themselves. A clinically, but also likely genetically heterogeneous. It computer program can facilitate the interpretation of the may be possible that the rules identified in the study patient’s genetic profile. These reports can improve in correspond to etiological subtypes of depression.
accuracy by including not only the patient’s genetic profilebut also the evidence-based implication of the profile for Among the combinations of alleles at different genes selection of appropriate medication.
comprising the subgroups in we observed two Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
notable ones. These were characterized by the prediction They do not ask why the medication works but whether it of either success or failure in relatively large groups of works. If it helps their patient, they are happy to patients, as well as a possibility of biological interaction, prescribe it. The paradox is that we can help these in addition to their observed statistical interaction. Of the clinicians improve their prescriptions, even though we are three markers, identifying 110 cases with an 89% chance of nonresponse listed above, rs7238368 and rs809736occur in the genes nucleolar protein 4 (NOL4), and RAR- related orphan receptor A isoform A (RORA), respec- The authors thank Panagiota Kitsantas PhD, who tively, whereas rs10499638 is intergenic. NOL4 is highly provided early advice on the methods discussed in this expressed in both brain and testis. Its function is poorly study. Dr Guoqing Diao conducted the analysis using understood. Nucleoli are the sites of ribosome-subunit logistic regression. The authors have benefited from production, although they are increasingly thought to discussions with Dr Gonzalo Laje. The boosted CART have a variety of other functions. Furthermore, nucleolar procedure described in this study is protected by a patent dysfunction may be related to disease etiology, including application 20090132460 from George Mason University in neuropsychiatric conditions such as Alzheimer’s and a provisional application 61/351,749 from Georgetown University. Scientists and students are allowed to use the to the NR1 subfamily of nuclear hormone receptors procedure with citation and without written permission.
Commercial use requires written permission. During the tection and cerebellar development, and is a recent preparation of this manuscript Dr Alemi was supported by the National Institute of Heart Lung and Blood Grant normalities in neuroendocrine systems (of which hor- mone receptors are a component), particularly in thehypothalamic–pituitary–Adrenal Axis, have long been Alemi F, Erdman H, Griva I, Evans CH (2009). Improved statistical methods are An interaction between a nuclear hormone receptor and needed to advance personalized medicine. Open Transl Med J 1:16–20.
proteins involved in nucleolar processes, although spec- Bryan C, Songer T, Brooks MM, Thase ME, Gaynes B, Klinkman M, et al. (2009).
Do depressed patients with diabetes experience more side effects when ulative at this time, would not be an unreasonable treated with citalopram than their counterparts without diabetes? A STAR*D study. Prim Care Companion J Clin Psychiatry 11:186–196.
Buyse IM, Shao G, Huang S (1995). The retinoblastoma protein binds to RIZ, a Another subgroup was defined by markers in rs2697992, zinc-finger protein that shares an epitope with the adenovirus E1A protein.
which occurs in retinoblastoma protein-binding zinc Proc Natl Acad Sci U S A 92:4467–4471.
Cabanero M, Laje G, Detera-Wadleigh S, McMahon FJ (2009). Association study finger (PRDM2), rs6127921, which is approximately of phosphodiesterase genes in the sequenced treatment alternatives to 106 kb downstream of bone morphogenetic protein 7 relieve depression sample. Pharmacogenet Genomics 19:235–238.
Cleveland WS (1979). Robust locally weighted regression and smoothing precursor (BMP7), along with rs6046805, which occurs in scatterplots. J Am Stat Assoc 74:829–836. MR0556476.
the open reading frame C20orf26. This subgroup Cleveland WS, Devlin SJ (1988). Locally-weighted regression: an approach to consisted of 128 individuals and had a 69% probability regression analysis by local fitting. J Am Stat Assoc 83:596–610.
Crown WH, Finkelstein S, Berndt ER, Ling D, Poret AW, Rush AJ, Russell JM of response. PRDM2 binds to RIZ, a zinc-finger protein (2002). The impact of treatment-resistant depression on health care which regulates transcription during neuronal differentia- utilization and costs. J Clin Psychiatry 63:963–971.
Do¨nmez-Altunta ¸s H, Akalin H, Karaman Y, Demirta ¸s H, Imamog˘lu N, Ozkul Y (2005). Evaluation of the nucleolar organizer regions in Alzheimer’s disease.
Fava M (2009). Partial responders to antidepressant treatment: switching strategies. J Clin Psychiatry 70:e24.
Fava M, Rush AJ, Trivedi MH, Nierenberg AA, Thase ME, Sackeim HA, et al.
These two proteins are therefore involved in the (2003). Background and rationale for the sequenced treatment alternatives to development of cell types in neural tissue whose relieve depression (STAR*D) study. Psychiatr Clin North Am 26:457–494.
differentiation is mutually antagonistic, lending face Fava M, Rush AJ, Alpert JE, Balasubramani GK, Wisniewski SR, Carmin CN, et al.
(2008). Difference in treatment outcome in outpatients with anxious versus validity to a biological interaction at either the protein nonanxious depression: a STAR*D report. Am J Psychiatry 165:342–351.
or systems level. We caution against overinterpreting the Freund Y, Schapire RE (1996). Experiments with a new boosting algorithm. In: meaning of rules discovered through our statistical Saitta L, editor. Proceedings of the Thirteenth International Conference onMachine Learning. San Francisco: Morgan Kauffman; pp. 148–156.
process, until they are further confirmed in future Friedman J (2001). Greedy function approximation: a gradient boosting machine.
Friedman J, Hastie T, Tibshirani R (2000). Additive logistic regression: a statistical This study has not clarified the underlying genetic view of boosting. Ann Statist 38:337–374.
underpinnings of response to antidepressants. For scien- Garriock HA, Kraft JB, Shyn SI, Peters EJ, Yokoyama JS, Jenkins GD, et al.
(2010). A genomewide association study of citalopram response in major tists, this study provides some new hypotheses to test.
depressive disorder. Biol Psychiatry 67:133–138.
This study sheds light on the possible complex biological Gigue`re V, Beatty B, Squire J, Copeland NG, Jenkins NA (1995). The orphan nuclear receptor ROR alpha (RORA) maps to a conserved region of processes mediating antidepressant response. In contrast homology on human chromosome 15q21-q22 and mouse chromosome 9.
to scientists, clinicians focus on one patient at a time.
Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.
Anticipating response to citalopram Alemi et al.
Hastie T, Tibshirani R (1996). Discriminant adaptive nearest neighbor classifica- Peters EJ, Slager SL, McGrath PJ, Knowles JA, Hamilton SP (2004). Investigation of tion. IEEE Trans Pattern Anal Mach Intell Arch 18:607–616.
serotonin-related genes in antidepressant response. Mol Psychiatry 9:879–889.
Heath S (2007). Reactive depression and remission rates in the STAR*D trials.
Peters EJ, Slager SL, Kraft JB, Jenkins GD, Reinalda MS, McGrath PJ, Hamilton SP (2008). Pharmacokinetic genes do not influence response or tolerance to Laje G, Perlis RH, Rush AJ, McMahon FJ (2009). Pharmacogenetics studies in citalopram in the STAR*D sample. PLoS One 3:e1872.
STAR*D: strengths, limitations, and results. Psychiatr Serv 60:1446–1457.
Schapire RE (1990). The strength of weak learnability. Mach Learn 5:197–227.
Lesser IM, Castro DB, Gaynes BN, Gonzalez J, Rush AJ, Alpert JE, et al. (2007).
Shen H, Luo Y, Kuo CC, Wang Y (2008). BMP7 reduces synergistic injury Ethnicity/race and outcome in the treatment of depression: results from induced by methamphetamine and ischemia in mouse brain. Neurosci Lett Nemeroff CB, Vale WW (2005). The neurobiology of depression: inroads to Tsai MJ, Weng CF, Shyue SK, Liou DY, Chen CH, Chiu CW, et al. (2007). Dual treatment and new drug discovery. J Clin Psychiatry 66 (Suppl 7):5–13.
effect of adenovirus-mediated transfer of BMP7 in mixed neuron-glial cultures: Neumann A, Holstein J, Le Gall JR, Lepage E (2004). Measuring performance in neuroprotection and cellular differentiation. J Neurosci Res 85:2950–2959.
health care: case-mix adjustment by boosted decision trees. Artif Intell Med Weissman MM, Pilowsky DJ, Wickramaratne PJ, Talati A, Wisniewski SR, Fava M, et al.; STAR*D-Child Team (2006). Remissions in maternal depression and Nguyen A, Rauch TA, Pfeifer GP, Hu VW (2010). Global methylation profiling of child psychopathology: a STAR*D-child report. JAMA 295:1389–1398.
lymphoblastoid cell lines reveals epigenetic contributions to autism spectrum disorders and a novel autism candidate gene, RORA, whose protein product Willetts J, Lippa A, Beer B (1999). Clinical development of citalopram. J Clin is reduced in autistic brain. FASEB J 24:3036–3051.
Psychopharmacol 19 (5 Suppl 1):36S–46S.
Ortega JA, Alca´ntara S (2010). BDNF/MAPK/ERK-induced BMP7 expression in Young EA, Kornstein SG, Marcus SM, Harvey AT, Warden D, Wisniewski SR, the developing cerebral cortex induces premature radial glia differentiation et al. (2009). Sex differences in response to citalopram: a STAR*D report.
and impairs neuronal migration. Cereb Cortex 20:2132–2144.
Paddock S, Laje G, Charney D, Rush AJ, Wilson AF, Sorant AJ, et al. (2007).
Zisook S, Lesser I, Stewart JW, Wisniewski SR, Balasubramani GK, Fava M, et al.
Association of GRIK4 with outcome of antidepressant treatment in the (2007). Effect of age at onset on the course of major depressive disorder. Am STAR*D cohort. Am J Psychiatry 164:1181–1188.
Copyright Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited.

Source: http://openonlinecourses.com/ehr/genetic%20markers%20anticipate%20response%20to%20citalopram%20antidepressant.pdf