Genome-wide approaches to antidepressant treatment: working towards understanding and predicting response

Antidepressants are among the most commonly prescribed drugs, and a range of medications are available. However, treatment response to a particular drug varies greatly between patients, with only 30% of patients responding well to the first treatment administered. Given evidence that antidepressant treatment response is a heritable trait, together with technological advances in genetic research, three recently published genome-wide investigations into antidepressant responses have examined the determinants of variability in treatment outcomes between depressed patients. Here, we review these studies within the context of wider research efforts to identify treatment response predictors. Some interesting genes have been implicated, but no variants have yet been robustly and reliably linked to response. This may suggest that genetic effect sizes are smaller than originally anticipated. Candidate gene approaches in these samples have lent support to the involvement of serotonergic, glutamatergic and stress-response systems in treatment response, although corroborative evidence from genome-wide analyses indicates these results should be interpreted cautiously. Closer examination of antidepressant response, considering it as a complex trait, has indicated that multiple genes of small effect are likely to be involved. Furthermore, there is some evidence that genetic influence on response to treatment may vary between patients with different symptom profiles or environmental exposures. This has implications for the translation of pharmacogenetic findings into clinical practice: genotypic information from multiple loci and data on non-genetic factors are likely to be needed to tailor antidepressant treatment to the individual patient.

concentration, low selfesteem and suicidal ideations, as well as somatic symptoms such as early morning waken ing, and loss of appetite and libido. There is a relatively high degree of symptomatic heterogeneity between depressed patients, with some showing 'atypical' features such as increased sleep and appetite.
Both the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSMIV) [3], and the Inter national Classification of Diseases, 10th revision (ICD10) [4], give classification criteria for depression ( Table 1). The disorder is considered a single diagnostic entity, and the separation of depression into 'neurotic' and 'endoge nous' subtypes has fallen out of favor. However, other additional specifiers can be used within both classifi ca tion systems to more precisely describe patients. To establish if a patient fulfils the criteria for depression as defined in DSMIV or ICD10, the majority of research studies use structured or semistructured diagnostic inter views such as the Schedules for Clinical Assessment in Neuropsychiatry [5] or the Composite International Diagnostic Interview [6]. These methods attempt to achieve both diagnostic validity and reliability.
Nevertheless, definitions of depression should not be considered absolute or immutable; there is continued debate over the best way to understand and define the illness. Indeed there is an ongoing research effort to confirm whether depression is best considered as a homogeneous clinical entity, given the variability seen between patients in terms of symptoms, course of illness and treatment response.
In order to measure symptoms and establish treatment response over time, numerous assessment tools are avail able, including clinicianrated scales such as the Hamil ton Depression Rating Scale (HAMD) [7], as well as self report alternatives such as the Beck Depression Inventory (BDI) [8]. However, there are differences in the items included in each of these scale, as well as variability in the weightings given to certain symptoms, reflecting the lack of consensus regarding the core features of depression.

Antidepressant treatment options
There are various treatment options for depressed patients, including pharmacotherapy, psychological treatments and, in more severe cases, electroconvulsive therapy. To date, however, the majority of work looking at genetic predictors of response has focused on pharmacological treat ments. The main classes of commonly used anti de pres sant medications are tricyclic antidepressants (TCAs), selective serotonin reuptake inhibitors (SSRIs), and serotoninnorepinephrine reuptake inhibitors (SNRIs) ( Table 2).
TCAs act by blocking both norepinephrine and sero tonin transporters. This inhibits reuptake of these two neurotransmitters into the presynaptic neuron, so increasing available levels in the synaptic cleft to bind to postsynaptic receptors. TCAs also act as antagonists at  (1) Depressed mood for at least 2 weeks Five or more symptoms, including (1) or (2) Mild: four or more symptoms, including two of (1), (2) or (3) Moderate: six or more symptoms, including two of (1), (2) or (3) Severe: eight or more symptoms, including (1), (2) and (3) (2) Loss of interest and enjoyment (3) Increased fatigability (4) Loss of confidence/self-esteem a (5) Self-reproach/guilt (6)  serotonergic, histaminergic, cholinergic and adrenergic receptors. However, these actions are associated with commonly reported side effects (such as dry mouth and constipation). Furthermore, in some cases TCAs can cause cardiovascular issues and be fatal in overdose. Given these adverse effects, SSRIs were developed to target the serotonin transporter much more selectively. SSRIs are widely used in depression, and have replaced TCAs as the first line of treatment for depression due to their more limited sideeffect profile and improved safety. Nevertheless, TCAs remain useful for patients who do not respond to newer antidepressants. The dualaction SNRIs have similar selectivity and sideeffect profiles to SSRIs, but inhibit the reuptake of both serotonin and norepinephrine.
The antidepressants described above have comparable average efficacy [9]. However, individuals vary widely in their response to treatments, with only an estimated 30% of patients responding sufficiently to their first treatment [10,11]. Antidepressants have a delayed onset of action, and it can take up to 8 weeks to assess treatment response [11] and, prior to commencing treatment, it is unclear which drug will produce an optimal response for a particular patient. Given the time and costs required to identify the most effective medication, together with the negative impact of multiple treatment steps (which are associated with increased risk of relapse [11]), there is considerable clinical value in identifying predictors of treatment response.

The role of genetics in determining antidepressant response
While it is logistically difficult to collect the samples required to investigate whether antidepressant treatment response is a heritable trait, family studies looking at response to antidepressants indicate that variability in response to these medications is likely to be at least partially genetic in nature [12,13]. Given this evidence, a growing number of studies are investigating which genes might determine antidepressant treatment response. This research has the dual aims of better understanding how these drugs exert their therapeutic action, and enabling individualized treatment recommendations.

Pharmacogenomic approaches
One key recent development within the rapidly evolving area of genetic research is the ability to perform genome wide tests of association, where variation across the entire genome is explored in relation to a particular phenotype (in this case, antidepressant treatment res ponse). Genotyping is performed at 500,000 to upwards of 1 million single nucleotide polymorphisms (SNPs), as made possible by commercially available microarray tech nology (for example: Affymetrix, South San Fran cisco, CA, USA; Illumina Inc., San Diego, CA, USA). Genomic approaches allow researchers to explore genetic variation in a hypothesisfree fashion, with no require ment to preselect genes of interest. If previously un sus pected genes are found to be associated with variability in treatment response through these syste matic analyses, 'pharmacogenomics' has the potential to push forward scientific understanding of how antidepressants work. To date there have been three key genomewide association studies (GWASs) of antidepressant treatment response [1416], as summarized in Figure 1.

Sequenced Treatment Alternatives to Relieve Depression
The USbased Sequenced Treatment Alternatives to Relieve Depression (STAR*D) project [17] enrolled a total of 4,041 treatmentseeking adults. All patients included in the study had a diagnosis of nonpsychotic major depressive disorder (as confirmed by DSMIV criteria), scoring greater than 14 on the HAMD (that is, suffered from moderate to severe depression). Among the sample, there were high rates of chronic or recurrent depression. STAR*D had a multistep design; all patients received citalopram (an SSRI) in the first treatment level, then those without sufficient symptomatic benefit moved onto further treatment levels where they received either alternative or augmented treatment options. Common side effects similar to SSRIs; also associated with increased blood pressure STAR*D is the largest and longest study to ever be conducted evaluating antidepressant treatments. It was found that with firstlevel treatment with citalopram, approximately 30% of patients achieved remission. While the study was not expressly designed for pharmacogenetic purposes, DNA was collected from 1,953 patients, allowing a GWAS of treatment outcomes [15].
Looking at those patients who were enrolled on the trial for at least 6 weeks, the investigators elected to explore two outcome phenotypes: response and remis sion to citalopram treatment. Treatment response was defined as a greater than 50% reduction in scores on the Quick Inventory of Depressive Symptomatology Self Report version (QIDSSR) [18] at the patient's final visit, giving 883 responders and 608 nonresponders. Remis sion was defined as scoring 5 or less on QIDSSR at followup, and patients achieving remission (n = 743) were compared against those classified as non responders. Genotyping was performed on either the Affymetrix 500K or 5.0 Human SNP arrays, and 430,198 SNPs were used in the genomewide analysis. The study was designed to be representative of the ancestral diversity within the US population and so stringent statistical controls for population stratification were required.
Three SNPs were associated with treatment response at a suggested level of significance (P <1 × 10 5 ), and these same SNPs were also among the top 25 hits associated with the remission phenotype (however, it should be noted that remitters represent a subgroup of responding patients). Rs6966038, which is located near the ubiquitin protein ligase E3C gene (UBE3C; involved in protein ubiquitination to promote degradation [19]), was asso ciated with response at P = 4.65 × 10 7 and remission at P = 3.63 × 10 7 . Rs6127921 is closest in location to the bone morphogenic protein 7 gene (BMP7; linked to bone development), and was associated with response at P = 3.45 × 10 6 and remission at P = 1.07 × 10 6 . Finally, rs809736 is an intronic SNP in the RARrelated orphan receptor alpha gene (RORA; a nuclear receptor [20]) and was associated with response at P = 8.19 × 10 6 and remission at P = 7.64 × 10 5 . There was no evidence that treatment side effects might confound the above findings.

Munich Antidepressant Response Signatures
The Munich Antidepressant Response Signatures (MARS) study is a German pharmacogenomic project, with genomewide data available for 339 inpatients with either major depression or bipolar disorder (11.2%) [14]. The majority of patients were German (85.1%), and all  [5,9,13]. The differences between studies may play some role in the lack of overlap in results between studies, although a more major factor is likely to be limitations in statistical power to detect smaller genetic effects. GWAS, genome-wide association study; HAM-D, Hamilton Depression Rating Scale; MADRS, Montgomery-Asberg Depression Rating scale; QID-SR, Quick Inventory of Depressive Symptomatology Self-Report version. reported white European ancestry. The study was naturalistic in design; patients were treated with a range of antidepressants. Depression symptoms were measured using the HAMD [7] and three outcome phenotypes were considered: early partial response (HAMD score improvement greater than 25% after 2 weeks), response (greater than 50% improvement in scores after 5 weeks) and remission (a score of <10 on the HAMD after 5 weeks or before hospital discharge). Genotyping was performed using the Sentrix Human1 and HumanHap300 arrays. While MARS was a smaller study than STAR*D, the genomewide analysis not only used a replication sample of 361 German inpatients, but also a subsample of 832 white outpatients from the STAR*D study to further validate findings. The most significant association (P = 7.6 × 10 7 ) was seen between early partial response and rs6989467, a SNP in the 5' region of the cadherin 17 gene (CDH17; encoding a calciumdependent membraneassociated glycoprotein). Looking across all three phenotypes, the SNP most consistently associated with treatment out come was rs1502174 (P = 8.5 × 10 5 ), located downstream of EPHB1 (encoding ephrin typeB receptor 1 protein, a transmembrane protein involved in developmental processes in the nervous system).

Genome Based Therapeutic Drugs for Depression
The Genome Based Therapeutic Drugs for Depression (GENDEP) study is a European, multicenter, 12week study with two active pharmacological treatment arms (for further details see [21]). A total of 811 outpatients with moderate to severe depression received one of two antidepressants, nortriptyline (a TCA) or escitalopram (an SSRI), in a partially randomized design. The primary outcome measure was the clinicianrated Montgomery Asberg Depression Rating scale (MADRS) [22], on which each patient was measured weekly. For the pharmaco genomic analysis [16], treatment response was defined using percentage change in the MADRS from baseline to week 12 of the study and genomewide data were available (using the Illumnia Human610 Quad chip) for 706 patients, 394 of whom were taking escitalopram and 312 nortriptyline. In a wholesample analysis, associa tions at a suggestive level of significance were seen in intergenic regions on chromosome 1 (rs2136093, rs6701608, rs2136094) and chromosome 10 (rs16920624, rs11598854, rs7081156). However, when drugspecific analyses were performed, a genomewide significant association with nortriptyline response was seen with an intronic SNP (rs 2500535) contained within the UST gene. UST encodes uronyl2sulphotransferase, an enzyme with an important link to neurotrophic mechanisms [23]. This SNP also reached suggested levels of significance when testing genedrug interactions (that is, the SNP differentially predicted treatment outcome with escitalo pram and nortriptyline). The best performing SNP within the escitalopramspecific analysis was rs1126757 (P = 2.83 × 10 6 ), a synonymous SNP within the coding region of the IL11 (interleukin 11) gene. Interleukins are neuropoietic cytokines involved in inflammatory res ponses, and inflammation processes have been associated with depression [24].
Across these three pharmacogenomic studies, only one genomewide significant finding was observed (the relationship between UST and response to nortriptyline in GENDEP), and, as can been seen from Table 3, there is no overlap between the three samples in the observed suggestive genetic signals. This may be due to differences in sample characteristics (inpatients versus outpatients, variation in cohort ancestry), the antidepressant treat ment received (medication taken, length of time receiving drug), or analysis approach (treatment outcome defini tions used). Alternatively, this could simply reflect the role of chance; even with true genetic associations, when looking at Pvalues for over 500,000 markers, the 'top hits' of a genomewide study can be highly unstable. An additional consideration when assessing pharmaco genomic research is that sample sizes are still very small in comparison with cohorts collected in the field of disease genetics. Studies looking at drugresponse phenotypes are limited by the greatly increased cost of sample collection (for example, due to the requirement of following patients over a period of time to assess treat ment response) and so the cohorts collected to date may not be sufficiently large to detect the genetic effects that influence treatment response at a genomewide level of significance.

Candidate gene approaches
By testing for genetic association at over 500,000 SNPs, genomewide analyses must compensate for the very large number of hypotheses tested with an appropriate adjustment of the significance thresholds used (generally set at P < 5 × 10 8 ) [25]. Aiming to avoid this multiple hypothesis testing burden, candidate gene studies pre select appropriate genes of interest. Although this hypothesisdriven method cannot reveal novel treatment pathways, this approach is better powered to detect genes with smaller effect sizes. There is a lot of published literature on candidate genes in antidepressant response. However, candidate gene studies are vulnerable to publication bias, where the same few 'favorite' genes have been repeatedly tested in many small cohorts, with only the more interesting (generally significant) results being published. Therefore, when interpreting results, key considera tions are replication and sample size (as emphasized by Horstmann and Binder [26]). Given this, the focus is placed on replicated candidate gene findings emerging from large cohort studies (Table 4). We describe findings for two genes involved in serotonergic pathways (SLC6A4 and HTR2A), one gene linked to the hypothalamicpituitaryadrenal axis (FKBP5) and one glutamatergic receptor gene (GRIK4).

SLC6A4
The serotonin transporter is a key target for many antidepressants (including SSRIs, SNRIs and TCAs), and so the gene encoding this protein (SLC6A4) represents a prominent candidate gene for pharmacogenetics. In the promoter region of this gene there is a repeat polymorphism comprising a 44 bp deletion [27], known as 5HTTLPR (serotonin transportlinked polymorphic region). This variant was first associated with treatment response by Smeraldi et al. [28], a finding supported by a metaanalysis in 2007 [29]. However, when 5HTTLPR was examined within large cohort samples, the results were mixed. Within the GENDEP sample, the poly morphism did indeed predict treatment outcome, but only for those patients on escitalopram [30]. Sexspecific analysis suggested that this effect was concentrated in males. Furthermore, an additional variant (rs2020933) within the SLC6A4 gene also demonstrated a significant effect on treatment outcome.However, other large studies report conflicting results. The Genetic and Clinical Pre dictors of Treatment Response in Depression (GenPod) project is a pharmacogenetic study based in the UK, where genotypic and outcome data are available for 520 patients, randomized to receive citalopram or reboxetine [31]. Reports from GenPod as well as those from STAR*D [32] found no association between 5HTTLPR and treatment outcome. Inconsistent findings could result from ethnic variability [33] or differences in outcome response definitions, but in the recent metaanalysis by Taylor et al. [34], including both GENDEP and STAR*D data (although not GenPod), a weak effect of 5HTTLPR genotype on remission did not reach significance once potential publication bias was accounted for.

HTR2A
Another serotonergic gene that has been implicated in antidepressant response is HTR2A. This encodes the serotonin receptor 2A, which is downregulated by anti depressant treatment in parallel with clinical improve ment [35]. A significant association between a marker in the HTR2A gene and treatment outcomes was observed when looking at 68 candidate genes (768 SNPs) in the STAR*D cohort (n = 1,297) [36]. A significant association was seen between rs7997012 and treatment response (the association with remission was not significant), and other markers within this gene also reached suggested levels of significance. Given differences in allele frequencies, the authors suggested that HTR2A may play a role in the racial differences in treatment response found in their mixedancestry sample. The top two SNPs implicated in STAR*D were also followed up in the MARS cohort [37]. In a combined sample of 637 inpatients (as in the genomewide analysis, a second independent German sample was used in addition to the 451 patients recruited as part of the MARS project) a significant association was seen between rs7997012 and both remission and response phenotypes. In that report, however, the allele association was the inverse of that observed in STAR*D; in MARS the AA genotype was associated with impaired treatment response whereas in STAR*D the G allele was associated with impaired treatment response. When HTR2A was examined in the GENDEP cohort, as part of an analysis looking at 10 different candidate genes (n = 760), once again an association was seen with treatment outcomes, but there was no replication of the specific allelic relationship seen in STAR*D [38]. In an escitalopramspecific analysis, no association was obs erved with the rs7997012 SNP. Instead, two intronic markers in the HTR2A gene (rs931623 and rs2224721, which are in linkage disequilibrium) were both signifi cantly associated with response to the SSRI. Therefore, while several studies indicated a relationship between genetic variation in the HTR2A gene and treatment response, there has been no replicated evidence that specific variants confer increased likelihood of favorable outcomes. Further attempts at replication in large cohorts are needed to clarify this issue, and it may also be valuable to perform sequencing in this region in an attempt to identify the causal variant that might underlie these findings. Interestingly, Murphy et al. [39] have implicated genetic variation in HTR2A in side effects to the SSRI paroxetine. Thus, the potentially confounding effects of treatment discontinuation should also be considered.

FKBP5
In addition to serotonergic pathways, the hypothalamic pituitaryadrenal (HPA) axis has also been implicated in antidepressant treatment response [40]. When looking across eight genes in the HPA axis in the MARS sample [41], significant associations were observed between treat ment outcomes and three SNPs within the gluco corticoidreceptorassociated cochaperone gene FKBP5 (rs1360780, rs3800373 and rs4713916). Followup work demonstrated that these SNPs were linked to increased intracellular expression of FKBP5, which leads to regu lation of the HPA axis via adaptive changes in the gluco corticoid receptor. This evidence supports the proposal that effective antidepressant treatment may involve modulation of the HPA axis. Examination of these three SNPs in STAR*D replicated the association between rs4713916 and remission [42], and it was noted that the linkage disequilibrium patterns observed in the mixed ancestry sample of STAR*D could suggest that the causal variant underlying this relationship may lie in the pro moter region of the gene. However, the association with FKBP5 was not replicated in GENDEP [38]. As part of a tencandidate gene analysis, the HPA axis genes FKBP5 and NR3C1 (which encodes the glucocorticoid receptor) were both examined. In a wholesample analysis, no asso ciations between treatment response and any of the markers tested in the FKBP5 gene (including the rs1360780 SNP implicated by Binder et al. [41]) were identified, although two SNPs within NR3C1 were associated at suggested levels of significance (rs852977 and rs10482633).

GRIK4
The final gene implicated in these large cohort studies of antidepressant treatment response is GRIK4, a gene encoding glutamatergic receptor, kainate 4. Paddock et al. [43] examined the 68 candidate genes previously considered in the STAR*D sample [36], but were able to incorporate genetic data from a second wave of geno typing, taking the sample size from 1,297 to 1,816. The authors not only replicated the HTR2A signal outlined above [36], but also reported evidence implicating rs1954787 within GRIK4 in this enlarged sample. The glutamatergic system has previously been considered within treatment response pathways [44] and this asso cia tion could lend weight to this hypothesis, but, as highlighted, further replication is crucial. It was also observed that the combined effect of GRIK4 and HTR2A on treatment outcomes was more robust than when each gene was considered separately. Despite the evidence above linking candidate genes and treatment outcome, if these findings were true, it would be expected that there would be a consistent pattern of associations seen across GWASs. Nevertheless, in both the STAR*D and GENDEP genomewide analyses [15,16], a priori candidate genes did not show associa tions with treatment outcomes above the levels expected by chance (although the STAR*D analysis does draw attention to 'nominally significant' associations with several SNPs of interest in the HTR2A gene). Although metaanalyses are needed, the results so far indicate that the current genes of interest do not exert a sizable direct effect on antidepressant treatment response.

Gene-gene interactions
Alternatively, the observation of combined effects of GRIK4 and HTR2A in the STAR*D sample may indicate that genetic variation has a more complex relationship with treatment outcomes. Horstmann et al. [45] used data from MARS to consider whether three of the genes highlighted in these large cohort studies (HTR2A, GRIK4 and FKBP5) may contribute to treatment response varia bility, in an additive or interactive fashion ( Figure 2). While the specific SNPs in HTR2A and GRIK4 implicated in the STAR*D sample did not replicate in MARS, when looking across all the variants analyzed in each gene, significant associations with treatment response were found. Exploring the most predictive SNPs, the bestfitting model of genetic moderation of treatment res ponse included a main effect of GRIK4 genotype, and two interaction terms, between GRIK4 and FKBP5, and GRIK4 and HTR2A. This threeSNP interaction model accounted for 13.1% of the variation in treatment res ponse, a substantial improvement over simpler, main effect models. This supports the proposal that anti depressant treatment response phenotypes are influenced by multiple genes, and that these genes are likely to have both independent and interactive (or epistatic) effects. This idea was also investigated in the MARS sample where a multilocus analysis was used to score patients as possessing high or low numbers of response alleles, based on the top 310 most informative SNPs in their GWAS analysis [14].
It has been calculated that STAR*D and GENDEP are powered to detect single SNP main effects on general treatment response that are clinically significant in size [46]. However, it now seems likely that treatment res ponse is a complex trait that is not determined by a single genetic variant, but results from multiple interacting genetic effects. In which case, the failure of pharmaco genetic and pharmacogenomic approaches to identify any replicated genetic variants associated with clinically meaningful effects on treatment response may be because studies are underpowered to detect these effects.

Gene-environment interactions in predicting treatment response
Effects at one gene locus may not only be moderated by other genetic variants, but may also be dependent on environmental exposures. Environments such as stressful life events (SLEs) have been associated with depression [47], and investigations into this relationship have indi cated that 5HTTLPR moderates the effect of SLEs on the emergence of depressive symptoms [48,49]. Mandelli et al. [50] reported initial evidence that this same inter action may also predict antidepressant treatment res ponse, in a sample of 159 mooddisorder patients treated with an SSRI. This interaction between SLEs and 5HTTLPR was also replicated in GENDEP, although only among patients taking escitalopram, not among those taking nortriptyline [51] (Figure 2). This could suggest that the geneenvironment interaction is drugspecific, and may be why no statistically significant interactions were found with 5HTTLPR in a sample of 290 patients who were receiving a range of different antidepressants [52]. Thus, it is likely that the accurate prediction of anti depressant treatment response will require the considera tion of multiple factors, both genetic and environmental, and the interaction between them.

Symptomatic heterogeneity in depression
Beyond SLEs, other nongenetic factors that may influence antidepressant response, such as symptomatic differences between patients, have also been explored as potential indicators of antidepressant response. This reflects the ongoing debate as to whether depression is a single, homogeneous clinical entity. Three commonly considered subgroups, defined on their symptomatic presentation, are: melancholic depression (akin to 'endogenous' depression, with pervasive anhedonia, sleep and appetite disturbance plus diurnal variation of mood); atypical depression (with features including mood reactivity, hyperphagia and hypersomnia); and anxious depression (reflecting the high rates of anxious symptoms and comorbid anxiety disorders seen in depression [53,54]). In the STAR*D sample, differences in response to citalopram seen between melancholic and non melancholic and between atypical and nonatypical depression were accounted for by demographic and clinical factors at baseline [55,56], but anxious depression was linked to poorer treatment response [57]. However, this pattern with anxious depression is not consistently replicated [58,59]. Studies comparing the efficacy of different antidepressant drugs in these subtypes have also been inconsistent in their findings [6063].
As noted by Uher et al. [46], it is important to consider not only whether associations between depression sub type and treatment response can be replicated, but also whether the observed effect sizes are sufficiently large to be of clinical value.
Interestingly, more robust predictors of treatment response were identified when the symptomatic variation in depression was explored in terms of dimensional traits (rather than the above categorical subtypes). Using detailed symptomatic data available in GENDEP, six dimensional facets of depressive symptoms were described: mood, anxiety, pessimism, interest activity (describ ing symptoms of indecisiveness and reduced activity), sleep and appetite [64]. The interestactivity facet was the most robust predictor of response in GENDEP; symptoms of low interest and decreased activity predicted significantly worse treatment outcomes [65]. Importantly, the association between interest activity symptoms and treatment outcome was also replicated in the STAR*D sample.

Differential response prediction
The predictive value of the interestactivity facet is sub stantial enough to be of clinical importance (one additional remission could be accurately predicted for every three or four patients where interest activity was measured), and obtaining the required information is relatively inexpensive. Thus, interestactivity scores appear to be valuable in translational terms. Nonetheless, low levels of interest and activity predicted poorer outcomes for patients, regardless of which antidepressant was used. As Simon and Perlis [66] highlight, the greatest clinical utility is obtained from factors that differentially predict response to alternative treatments, as these may be used to guide treatment recommendations. Conse quently, assessment of whether other treatment options might be effective in patients displaying loss of interest and decreased activity would be of translational value. Therefore, in addition to betweendrug comparisons, work in 'therapygenetics' [67] (where predictors of response to talking therapies are considered) points to a potentially fruitful direction for exploring differential predictors of treatment response. If studies allowing direction comparisons between psychological and pharmacological treatments could be undertaken, response predictors with high translational value might be revealed. For example, there is preliminary evidence that the 5HTTLPR variant linked with poorer antidepressant response is associated with better response to cognitive behavioral therapy [67].
The observed heterogeneity in treatment response associated with symptomatic differences also points to interesting new avenues for pharmacogenetic research, as it may indicate that genetic moderation of response to antidepressants could be symptomspecific. For example, among patients taking nortriptyline in GENDEP, genetic variation in the beta3 subunit of the Gprotein complex (GNB3) was associated specifically with differences in improvement of neurovegetative symptoms (that is, symp toms relating to sleep and appetite), but was not asso ciated with improvement in other symptom dimensions (such as observed mood or cognitive symptoms) [68]. Therefore, consideration of the symptomatic diversity encom passed by the diagnostic category of depression (possibly through dimensional rather than categorical models) may allow the identification of genetic effects on specific components of treatment response that are not otherwise seen.

Towards an understanding of mechanisms underlying antidepressant treatments
Genetic research into antidepressant response not only aims to identify outcome predictors, but also provides valuable insight into the physiological processes under lying successful treatment of depression. The lack of Figure 2. Examples of reported interactions between genetic and non-genetic factors in determining antidepressant treatment response. Work looking at the relationship between potential response predictors indicates that antidepressant treatment response is determined by multiple interacting factors. For example, genetic variation in GRIK4 (glutamate receptor, ionotropic, kainate 4) has been reported to have both a main effect and interactive effects with FKBP5 (FK506 binding protein 5) and HTR2A (5-hydroxytryptamine (serotonin) receptor 2A, G-protein-coupled) in its influence on response outcomes [45]. Similarly, an association between the occurrence of stressful life events and response may be moderated by genetic variation in SLC6A4 (solute carrier family 6 (neurotransmitter transporter, serotonin), member 4) [50,51].  [45] strong genetic signals in either genomewide or candidate gene studies indicates that antidepressant treatment response is not a phenotype determined by a small number of highly influential genes. Instead, the genetic evidence suggests that antidepressant treatment response is best considered a complex phenotype, determined by multiple genes of small effect. While candidate gene findings require further replication and confirmation, the evidence suggests that these genes lie in multiple, interacting biological systems.
The serotonergic system, as the canonical target for antidepressants, is a key candidate for investigators, and has been implicated with associations observed between treatment response and variation in SLC6A4 [30] and HTR2A [3638], although effect sizes are small when considered alone. The effects of successful antidepressant treatment on the HPA axis suggest that this system also plays an important role in the physiological processes that lead to remission [69]. This proposal is supported by the observed association between FKBP5 (which acts as a regulator within this system) and treatment response [41,42]. Indeed, there has been a large research effort in attempting to develop novel antidepressants targeting the HPA system [70], although this is yet to achieve success [71]. Similarly, evidence of glutamatergic involvement in both depression and treatment response [72] has led to investigations of ketamine as another potential new treatment for depression [73], and the association between variability in the glutamatergic gene GRIK4 and treatment response strengthens the hypothesis that this system also plays an important role in antidepressant action.
Arguably, what is most interesting about the genetic findings is that they provide preliminary indications that, rather than purely additive effects of each of the different pathways contributing to treatment response, the serotonergic, glutamatergic and HPA systems may have interactive effects. This might indicate a physiological interaction, as experimental observations in humans seem to suggest (for example, [74]).
The nongenetic predictors of response outlined here also may give an insight into the mechanisms underlying antidepressant response. For example, the finding that interestactivity symptoms (of low interest, reduced activity, indecisiveness and lack of enjoyment) predict poorer treatment response [65] could be interpreted to suggest that the pathophysiological processes producing these symptoms are, to some extent, distinct, and less well targeted by antidepressants.
Nevertheless, the identification of treatment response predictors is only the first step in understanding the actions of antidepressants; further exploration of the biological impact of these genetic variants and the mechanism by which suggested predictors may be associated with treatment response is needed, as well as a greater understanding of how these predictors may interact.

Conclusions and future directions
Genomewide analyses of antidepressant treatment response have proved disappointing to date, with no strong signals emerging from the three studies conducted thus far. Given the small genetic effect sizes observed in candidate gene studies, it seems that existing studies are likely to be underpowered on a genomewide scale, and strong conclusions cannot be drawn without increasing sample sizes and undertaking formal metaanalyses.
Nonetheless, there is some evidence from candidate gene approaches within these large samples indicating that genetic variability in the serotonergic, glutamatergic and stressresponse pathways may influence anti depressant treatment response. These associations remain small in effect size; no genes of interest have yet been reliably linked with clinically significant effects. This suggests that, rather than single gene effects, there may instead be multiple genetic influences that determine treatment outcomes in combination with environmental factors. The interactions between each of these various elements must be considered if we wish to better provide treatment recommendations that are optimized for each individual.
Indeed, the bestperforming predictor identified in the three large GWAS samples has been the symptom dimension of interest activity, and there is some evidence for symptomspecific improvements. This indicates that the future for pharmacogenomic studies lies not only in larger samples, but also in the collection of detailed and accurate patient data, to allow the identification of any subgroupspecific relationships.
The translational value of pharmacogenetic work with antidepressants has yet to be fully realized; indeed, for real clinical utility, studies must focus on comparing different treatments for depression in order to identify differential predictors of response. Furthermore, it is likely that translation into clinics will only be achieved by combining information from multiple response pre dictors. Nonetheless, the candidate genes identified to date provide an insight into the mechanisms of anti depressant action, and may enable the development of novel treatments for depression. reuptake inhibitor; STAR*D, Sequenced Treatment Alternatives to Relieve Depression; TCA, tricyclic antidepressant; 5HTTLPR, serotonin transport-linked polymorphic region.