Understanding genetic risk factors for common side effects of antidepressant medications.

Background: Major depression is one of the most disabling health conditions internationally. In recent years, new generation antidepressant medicines have become very widely prescribed. While these medicines are efficacious, side effects are common and frequently result in discontinuation of treatment. Compared with specific pharmacological properties of the different medications, the relevance of individual vulnerability is understudied.
Methods: We used data from the Australian Genetics of Depression Study to gain insights into the aetiology and genetic risk factors to antidepressant side effects. To this end, we employed structural equation modelling, polygenic risk scoring and regressions.
Results: Here we show that participants reporting a specific side effect for one antidepressant are more likely to report the same side effect for other antidepressants, suggesting the presence of shared individual or pharmacological factors. Polygenic risk scores (PRS) for depression associated with side effects that overlapped with depressive symptoms, including suicidality and anxiety. Body Mass Index PRS are strongly associated with weight gain from all medications. PRS for headaches are associated with headaches from sertraline. Insomnia PRS show some evidence of predicting insomnia from amitriptyline and escitalopram. Conclusions: Our results suggest a set of common factors underlying the risk for antidepressant side effects. These factors seem to be partly explained by genetic liability related to depression severity and the nature of the side effect. Future studies on the genetic aetiology of side effects will enable insights into their underlying mechanisms and the possibility of risk stratification and prophylaxis strategies.

Introduction

The World Health Organisation predicts that depression will become the leading cause of disability globally by 2030[1]. The symptomatology, longitudinal course, response to treatment and functional impact of depressive disorders are highly variable. Antidepressant medicines are widely prescribed across the spectrum of depression severity and subtypes, alone or in combination with psychological therapies.

Effective pharmacotherapies for depression were first developed in the 1960s, following the identification of antipsychotic therapies. A clear focus was on the regulation of brain monoamine systems (dopamine, serotonin and noradrenaline). These agents, including tricyclic antidepressants (TCAs) and Monoamine Oxidase Inhibitors (MAOIs), were limited in the extent of their use by considerable side effect burdens and potential toxicity. From the 1980s onwards, further pharmacological developments have been dominated by the establishment of second-generation antidepressant classes, including selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs)[2]. Although second-generation antidepressants have been shown to alleviate depression[3], treatment response is heterogeneous, and new side effect profiles have emerged (gastrointestinal, weight gain, sexual dysfunction). The degree of individual variation in the incidence and severity of these difficulties is high.

Treatment failure is commonly caused by the discontinuation of antidepressants from adverse side effects. Over half of individuals have been recorded to cease medication within the first six months of initial prescription[4]. Previously reported antidepressant adverse effects include sexual dysfunction[5-7], weight changes[8-11], insomnia[12-15], and suicidality[16-18]. However, these ‘side effects’ may also reflect ongoing symptoms of the depressive illness. For example, anhedonia is a cardinal symptom of major depressive disorder (MDD) which could explain lower levels of sexual interest and arousal leading to sexual function impairments[19,20]. Weight changes, sleep disturbances, and suicidality are also symptoms of depression[21] and its various phenotypic subtypes. Finally, other comorbid mental health disorders may amplify or trigger suicidal behaviours[22]. Whether these side effects stem from adverse reactions to antidepressants or whether they are extensions or exacerbations of characteristics of an individual’s depression or a consequence of comorbidity with another disorder remains unclear.

Variability in medication response and tolerability may be inherited. For instance, genetic variation leading to changes in the function of antidepressant metabolising enzymes are believed to underlie side effects due to drug overexposure[23]. Five to seven percent of European ancestry individuals are estimated to be poor CYP2D6 metabolisers[24], one of the major metabolising enzymes of fluoxetine, paroxetine and fluvoxamine. Furthermore, variants in genes such as CYP2C19 and CYP3A4 have been linked to citalopram[25] and sertraline[26] differential metabolism and clearance. Metabolising enzymes are relevant hypotheses for understanding adverse side effects. Nonetheless, genetic variants within these enzymes have failed to reach significance in recent genome-wide association studies on treatment resistance[27] and response[28], suggesting that treatment outcomes might be more complex than previously thought. It is likely that genetic factors underlying antidepressant side effects are a product of drug-specific factors such as variation within drug-metabolising enzymes, as well as common (or non-drug-specific) factors, the nature of which remains elusive.

In general, the aetiology of antidepressant adverse side effects remains largely understudied. Thus, we aim to bridge this research gap by leveraging data from the Australian Genetics of Depression Study (AGDS) to gain insights into the prevalence, aetiology and genetic underpinnings of adverse side effects associated with antidepressant use. We investigate the prevalence and demographic risk factors for 23 side effects across ten commonly prescribed antidepressants. We test for SSRI or SNRI specificity and provide evidence for a co-occurring relationship between adverse side effects across different antidepressant medications. That is, participants who took two or more antidepressants were more likely to report the same side effects regardless of the antidepressant used. This co-occurrence would suggest a set of common risk factors underlie these side effects. Here, we use polygenic risk scores (PRS) to study the genetic aetiology of specific antidepressant adverse side effects to understand the nature of these common risk factors. PRS are an estimate of an individual’s genetic risk for a given trait. They are calculated based on genome-wide association study (GWAS) results whereby genetic variants are linked to a trait of interest through an effect size (i.e., the increased risk per copy of the genetic variant). PRS are calculated in an independent sample by performing a sum of risk variants weighted by their effect size. PRS are gaining popularity due to their potential to enable many applications such as testing genetic overlap between traits, enabling risk stratification, and aiding diagnosis and personalised treatment[29]. We use PRS for MDD, BMI, insomnia and headaches to test for evidence of non-specific or shared genetic factors underpinning specific side effects. Overall, our results suggest drug exposure alone does not explain the occurrence of side effects, and a combination of specific and non-specific factors underlie their aetiology.

Methods

Sample recruitment and genotyping

We use the Australian Genetics of Depression Study (AGDS), in which participants provide self-report responses on psychosocial factors of depression heterogeneity and antidepressant treatment outcomes (N = 20,941 with reported depression diagnosis) as well as DNA samples for genetic analysis. Sample recruitment has been described in detail elsewhere[30]. Briefly, 14.3% of volunteers were recruited by mail invitations distributed by the Australian Department of Human Services (DHS) and encouraged individuals who had previously used prescription antidepressants to participate in the last 4.5 years. Secondly, a nationwide media publicity campaign was broadcast. This campaign, targeted individuals who have sought medical attention from a psychiatrist or a psychologist for clinical depression. Recruited participants were directed to the study website to complete consent forms before answering the instruments. Once the instruments had been completed and informed consent for donation of a DNA sample was given, a GeneFix GFX-02 DNA extraction kit (Isohelix plc) was sent to participants to collect 2 mL of saliva for DNA extraction. Genotyping was performed using the Illumina Global Screening Array (GSA V.2.0.). Genotype data were cleaned by removing unknown or ambiguous map position, strand alignment, high missingness (>5%), deviation from Hardy–Weinberg equilibrium, low minor allele frequency (<1%) and GenTrain score <0.6 variants. Imputation was performed through the Michigan imputation server web service using the HRCr1.1 reference panel. Genotyped individuals were excluded from PRS analyses based on high genotype missingness, inconsistent and unresolvable sex or if deemed ancestry outliers from the European population, based on principal components derived from the 1000Genomes reference panel. The protocol for approaching participants through the DHS, enroling them in the study and consenting for all phases of the study (including invitation to future related studies) and accessing MBS and PBS records was approved by the Ethics Department of the Department of Human Services. The QIMR Human Research Ethics Committee also approved all protocols for the ADGS data collection and scope for downstream studies under project number 2118. The study presented here falls within the scope of the analyses reviewed and approved under project 2118.

Phenotype ascertainment

This study focuses on participant-reported antidepressant adverse side effects. Participants first confirmed they had taken any of the ten most commonly prescribed antidepressants in Australia (sertraline, escitalopram, venlafaxine, fluoxetine, citalopram, desvenlafaxine, duloxetine, mirtazapine, amitriptyline and paroxetine). For each antidepressant taken, participants were asked whether they had experienced side effects and, if they did, to select which from a checklist with the twenty-three most commonly reported antidepressant side effects, including reduced sexual drive or desire, weight gain, dry mouth, nausea, drowsiness, insomnia, dizziness, fatigue, sweating, headache, suicidal thoughts, anxiety, agitation, shaking, constipation, diarrhoea, suicide attempt, blurred vision, muscle pain, vomiting, weight loss, runny nose and rash.

Side effect correlations and structural equation modelling

We used tetrachoric correlations as implemented in the psych library in R v3.6.1 to estimate the correlation (i.e. co-occurrence within the same set of people) of side effects across medications. Pairwise complete observations were used for these analyses. The correlation matrix was transformed into a distance matrix subjected to a minimum variance hierarchical clustering analysis using the scipy library in Python 3. The results are visualised with a clustergram generated using the seaborn and matplotlib libraries in Python 3.6. We further used structural equation modelling (OpenMx Rv3.6.2) to assess whether, for each side effect, there was evidence for drug-class-specific factors over and above a common factor. For each side effect, we fit a bifactor model consisting of a general factor loading onto the ten binary side effects, and two drug-class factors, “SSRI” and “SNRI” loading onto side effects from their respective drug class. The general factor is orthogonal to the drug-class factors. We refer to this model as the full model. Reduced models are also fit by removing the drug-class factors one at a time; these are the SSRI and SNRI models (Supplementary Fig. 3). Finally, a model consisting of a single general factor is also used for completeness. The four models are fit to the data using a full information maximum likelihood estimation assuming a liability threshold model for the binary manifest variables. After fitting, the simpler models are compared to the full model by the Akaike information criterion (AIC) and likelihood ratio test (LRT) with the mxCompare function implemented in OpenMx. Under this approach, the p-value represents whether a nested reduced model loses a significant amount of information compared to its full counterpart. Thus, a statistically significant p-value indicates that removing that drug-class factor results in a poorer fit.

Genetic instruments and polygenic risk scoring

To avoid biases due to population stratification and cryptic relatedness, only unrelated individuals of European ancestry were included in the genetic part of this study. PRS were calculated as a proxy for an individual’s genetic liability to a trait. This study used publicly available GWAS results for depression[31], insomnia[32], chronic headaches[33], and BMI[34]. Genetic variant effect sizes were acquired from the GWAS data and used to calculate the predictive genetic risk for the traits investigated. Before estimating PRS, we excluded low (r2 < 0.6) imputation quality and strand-ambiguous variants. We used two approaches to deal with correlation among genetic variants emerging through linkage disequilibrium (LD). First, we employed a recently developed powerful method named SBayesR[35]. SBayesR estimates a conditional GWAS (i.e., one including all of the genetic variants as predictors simultaneously) using marginal GWAS summary statistics and LD measures between genetic variants (LD matrix) under a Bayesian multiple regression framework. This method has been shown to improve the polygenic prediction of complex traits. We also employed a more traditional clumping and thresholding procedure as sensitivity analyses. Briefly, PLINK (1.9)[36] was used to detect independent SNPs through a conservative clumping (p1 = 1, p2 = 1, r2 = 0.1, kb=10,000) adjustment of linkage disequilibrium. Various p-value thresholds (p < 5 × 10−8, p < 1 × 10−5, p < 0.001, p < 0.01, p < 0.05, p < 0.1, p < 0.5, p < 1) were used to determine which variants to include for PRS calculation. Imputed genotype dosage data were used to calculate PRS by multiplying the variant effect size times the dosage of the effect allele. Finally, the total sum was calculated across all variants.

PRS side effect association

Logistic regressions were used to examine the association between participant-reported side effects and PRS. The regressions were adjusted for sex, age at study enrolment and the first 20 genetic principal components to further adjust for potential population stratification. Variance explained was calculated as Nagelkerke’s pseudo R2:where LLfull and LLnull are the log-likelihoods for the model with and without the PRS, respectively. Nominally significant results are defined as those with p < 0.05, and statistical significance was defined after Bonferroni correction for multiple testing. For the MDD analysis, we adjusted for the association of MDD PRS with the 25 side effects across medications (p < 0.002). For the other PRS, we adjusted for the testing of ten drugs (p < 0.005), and for the sensitivity analyses (using clumping and thresholding), we adjusted for eight thresholds times ten medications (p < 0.000625). This method is relatively conservative, as it does not account for the moderate to high co-linearity within the eight PRS and within the side effects across medications. Results are visualised as heat maps of variance explained using seaborn and matplotlib in Python 3.6. These analyses were performed using complete case data. Scripts and data for this study, including PRS (i.e. SBayesR effect sizes) are available online at doi: 10.5281/zenodo.5533372.

Table 1

AGDS demographics and side effect prevalence across medications.

	Males N	Females N	Sex p-valuea	Endorsed side effect age (s.d.)	Not endorsed side effect age (s.d.)	Age p-valueb
N	5111	15830	—	—	—	—
Age mean (s.d.)	47.99 (15)	41.41 (14)	6.6e−160	—	—	—
Reduced sexual desire	2251 (44%)	6264 (39%)	1.5e−08	41.0 (13.98)	44.4 (16.01)	1.1e−58
Weight gain	1402 (27%)	5695 (35%)	3.2e−29	41.9 (13.89)	43.6 (15.96)	1.1e−13
Dry mouth	1236 (24%)	4544 (28%)	3.2e−10	42.7 (14.20)	43.1 (15.71)	0.071
Nausea	867 (16%)	4352 (27%)	1e−51	37.7 (13.34)	44.8 (15.52)	1.6e−187
Headaches	704 (13%)	3282 (20%)	3.1e−28	38.3 (13.74)	44.1 (15.45)	2e−106
Dizziness	959 (18%)	3930 (24%)	5.2e−19	38.4 (13.43)	44.4 (15.57)	5.3e−130
Shakes	571 (11%)	2466 (15%)	7.4e−15	38.8 (14.22)	43.7 (15.37)	3.8e−62
Muscle pain	234 (4%)	837 (5%)	0.045	42.0 (14.98)	43.1 (15.33)	0.029
Sweating	997 (19%)	3291 (20%)	0.048	40.3 (13.75)	43.7 (15.61)	6.8e−38
Vomit	147 (2%)	826 (5%)	4.7e−12	35.7 (12.65)	43.4 (15.34)	6e−53
Constipation	395 (7%)	1489 (9%)	2.70E−04	43.3 (15.01)	43.0 (15.34)	0.440
Diarrhoea	368 (7%)	1176 (7%)	0.590	39.9 (13.87)	43.3 (15.39)	8.8e−17
Drowsiness	1173 (22%)	3709 (23%)	0.480	39.7 (14.30)	44.0 (15.46)	6.8e−67
Trouble sleeping	1052 (20%)	3672 (23%)	1.00e−04	39.6 (14.35)	44.0 (15.43)	2.3e−70
Anxiety	794 (15%)	2973 (18%)	1.5e−07	39.1 (14.24)	43.9 (15.40)	1.6e−68
Agitation	786 (15%)	2816 (17%)	7.2e−05	39.6 (14.12)	43.7 (15.45)	1.3e−50
Fatigue	912 (17%)	3181 (20%)	4.20e−04	39.4 (14.49)	43.9 (15.38)	8.4e−63
Weight loss	157 (3%)	795 (5%)	5.9e–09	36.1 (13.42)	43.3 (15.32)	5.1e−47
Rashes	97 (1%)	257 (1%)	0.190	45.5 (14.96)	43.0 (15.31)	0.002
Runny nose	82 (1%)	344 (2%)	0.012	44.9 (15.64)	43.0 (15.30)	0.010
Blurry vision	266 (5%)	1012 (6%)	0.002	42.6 (14.60)	43.0 (15.35)	0.280
Suicide thoughts	699 (13%)	2560 (16%)	1.9e−05	38.2 (14.24)	43.9 (15.33)	4.7e−87
Suicide attempt	248 (4%)	1090 (6%)	2.4e−07	35.9 (13.41)	43.5 (15.31)	4.5e−69
Other side effects	632 (12%)	1968 (12%)	0.900	41.3 (14.05)	43.3 (15.47)	1.3e−09
No side effects	334 (6%)	1140 (7%)	0.110	43.8 (15.04)	43.0 (15.33)	0.044

aTwo sample Z proportion test.

bTwo sample t-test. Data of each side effect per medication studied are available in Supplementary Data 1.

Table 2

Polygenic prediction of specific side effects across medications.

Antidepressant	BMI PRS predicting weight gain			Insomnia PRS predicting trouble sleeping			Headaches PRS predicting headaches
	OR (95%C.I.)	P-value	Variance explained (%)	OR (95%C.I.)	P-value	Variance explained (%)	OR (95%C.I.)	P-value	Variance explained (%)
Sertraline	1.24 (1.17–1.32)	4.66e-12	1.25	1.06 (0.99–1.14)	0.077	0.09	1.16 (1.08–1.25)	8.17e-05	0.49
Escitalopram	1.21 (1.13–1.29)	3.58e-08	1.01	1.12 (1.03–1.21)	0.008	0.27	1.06 (0.97–1.16)	0.178	0.08
Venlafaxine	1.20 (1.12–1.28)	3.72e-07	0.93	1.03 (0.95–1.11)	0.463	0.02	1.11 (1.01–1.21)	0.025	0.22
Amitriptyline	1.25 (1.12–1.40)	9.84e-05	1.38	1.28 (1.07–1.52)	0.007	1.03	1.09 (0.93–1.28)	0.268	0.15
Mirtazapine	1.20 (1.09–1.32)	2.03e-04	0.95	1.03 (0.88–1.20)	0.745	0.01	1.04 (0.90–1.20)	0.560	0.04
Desvenlafaxine	1.22 (1.11–1.33)	3.01e-05	1.03	1.13 (1.03–1.25)	0.015	0.38	1.08 (0.98–1.20)	0.126	0.16
Citalopram	1.37 (1.24–1.50)	9.06e-11	2.5	1.08 (0.96–1.21)	0.194	0.12	1.11 (0.98–1.26)	0.088	0.23
Fluoxetine	1.26 (1.17–1.36)	5.75e-09	1.42	1.03 (0.94–1.13)	0.523	0.02	1.02 (0.93–1.12)	0.689	0.01
Duloxetine	1.20 (1.08–1.32)	3.60e-04	0.94	1.14 (1.02–1.28)	0.026	0.4	1.03 (0.92–1.15)	0.626	0.02
Paroxetine	1.22 (1.09–1.36)	6.28e-04	1.1	1.16 (1.01–1.33)	0.042	0.47	1.16 (1.00–1.35)	0.053	0.46

Results from logistic regressions predicting weight gain, trouble sleeping and headaches using BMI, Insomnia and headaches PRS respectively. Bolded p-values represent values significant after multiple testing correction (p < 0.005). Results shown for SBayesR, for clumping and thresholding sensitivity results see Supplementary Data.