Identifying Potential Causal Risk Factors for Self-Harm: A Polygenic Risk Scoring and Mendelian Randomisation Approach

Background Multiple individual vulnerabilities and traits are phenotypically associated with suicidal and non-suicidal self-harm. However, associations between these risk factors and self-harm are subject to confounding. We implemented genetically informed methods to better identify individual risk factors for self-harm. Methods Using genotype data and online Mental Health Questionnaire responses in the UK Biobank sample (N = 125,925), polygenic risk scores (PRS) were generated to index 24 plausible individual risk factors for self-harm in the following domains: mental health vulnerabilities, substance use phenotypes, cognitive traits, personality traits and physical traits. PRS were entered as predictors in binomial regression models to predict self-harm. Multinomial regressions were used to model suicidal and non-suicidal self-harm. To further probe the causal nature of these relationships, two-sample Mendelian Randomisation (MR) analyses were conducted for significant risk factors identified in PRS analyses. Outcomes Self-harm was predicted by PRS indexing six individual risk factors, which are major depressive disorder (MDD), attention deficit/hyperactivity disorder (ADHD), bipolar disorder, schizophrenia, alcohol dependence disorder (ALC) and lifetime cannabis use. Effect sizes ranged from β = 0.044 (95% CI: 0.016 to 0.152) for PRS for lifetime cannabis use, to β = 0.179 (95% CI: 0.152 to 0.207) for PRS for MDD. No systematic distinctions emerged between suicidal and non-suicidal self-harm. In follow-up MR analyses, MDD, ADHD and schizophrenia emerged as plausible causal risk factors for self-harm. Interpretation Among a range of potential risk factors leading to self-harm, core predictors were found among psychiatric disorders. In addition to MDD, liabilities for schizophrenia and ADHD increased the risk for self-harm. Detection and treatment of core symptoms of these conditions, such as psychotic or impulsivity symptoms, may benefit self-harming patients. Funding Lim is funded by King’s International Postgraduate Research Scholarship. Dr Pingault is funded by grant MQ16IP16 from MQ: Transforming Mental Health. Dr Coleman is supported by the UK National Institute of Health Research Maudsley Biomedical Research Centre. MRC grant MR/N015746/1 to CML and PFO’R. Dr Hagenaars is funded by the Medical Research Council (MR/S0151132). Kylie P. Glanville is funded by the UK Medical Research Council (PhD studentship; grant MR/N015746/1). This paper represents independent research part-funded by the National Institute for Health Research (NIHR) Biomedical Research Centre at South London and Maudsley NHS Foundation Trust and King’s College London. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care. Research in Context Evidence before this study A search was conducted on PubMed for literature from inception until 1st May 2019 using terms related to suicidal self-harm (SSH) and non-suicidal self-harm (NSSH), as well as polygenic risk scores (PRS), (“self-harm”[All Fields] OR “self-injurious”[All Fields] OR “self-mutilation”[All Fields] OR “suicide”[All Fields]) AND (“polygenic”[All Fields] OR “multifactorial inheritance”[All Fields]). Similar search was done for Mendelian Randomisation (MR), replacing “multifactorial inheritance” and “polygenic” with “Mendelian Randomisation/Randomization”. Evidence was included only if the study had used PRS or MR method to predict self-harm phenotypes using risk factors of self-harm. Ten papers for PRS and no paper for MR were identified. There were mixed results for PRS studies. PRS for MDD predicted SSH in two studies but not in another two studies. PRS for depressive symptoms predicted SSH but not NSSH. PRS for schizophrenia predicted SSH in one but not in another two studies. PRS for bipolar disorder predicted SSH in one study but did not predict SSH nor NSSH in another two studies. Added value of this study By using a large population-based sample, we systematically studied individual vulnerabilities and traits that can potentially lead to self-harm, including mental health vulnerabilities, substance use phenotypes, cognitive traits, personality traits and physical traits, summing up to 24 PRS as genetic proxies for 24 risk factors. We conducted MR to strengthen causal inference. We further distinguished non-suicidal self-harm (NSSH) and suicidal self-harm (SSH). Apart from PRS for schizophrenia, MDD and bipolar disorder, novel PRS were also identified to be associated with self-harm, which are PRS for attention-deficit hyperactivity disorder (ADHD), cannabis use and alcohol dependence. A larger sample size allowed us to confirm positive findings from the previously mixed literature regarding the associations between PRS for MDD, bipolar disorder, and schizophrenia with self-harm. Multivariate analyses and MR analyses strengthened the evidence implicating MDD, ADHD and schizophrenia as plausible causal risk factors for self-harm. Implications of all the available evidence Among the 24 risk factors considered, plausible causal risk factors for self-harm were identified among psychiatric conditions. Using PRS and MR methods and a number of complementary analyses provided higher confidence to infer causality and nuanced insights into the aetiology of self-harm. From a clinical perspective, detection and treatment of core symptoms of these conditions, such as psychotic or impulsivity symptoms, may prevent individuals from self-harming.


Introduction 1
Self-harm is a complex trait that refers to any act of self-injury and self-poisoning carried out 2 by an individual, regardless of intention or motivation. 1 Being a broadly defined term, it can 3 be further categorised into suicidal self-harm (SSH) and non-suicidal self-harm (NSSH), i.e. 4 with or without intention of suicide. According to a meta-analysis, the cross-national 5 prevalence rate for NSSH peaks during adolescence (17.3%), and decreases among adults 6 (5.5%). 2 For SSH, the cross-national prevalence rate is also the highest among adolescents 7 (9.7%) 3 and drops among adults (2.7%). 4 Recently, both SSH and NSSH were included in the 8 fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) as 9 separate conditions for further study. 5 The distinction between SSH and NSSH may facilitate 10 investigations of the aetiology and heterogeneity of self-harm. 11 12 A range of individual vulnerabilities and traits can potentially lead to self-harm, such as 13 psychiatric illnesses, 6 substance use, 7-9 cognitive abilities, 10 personality traits 11 and physical 14 traits. 12 Although associations between these risk factors and self-harm have been shown in 15 numerous observational studies, causality is difficult to infer reliably. Genetically informed 16 designs can help in strengthening causal inference. 13 A polygenic risk score (PRS) is a single 17 individual-level score computed in a given trait, weighted using summary statistics from an 18 independent genome-wide association study (GWAS) for that particular trait. A PRS for an 19 individual risk factor (e.g. schizophrenia) can be regarded as a genetic proxy for this risk 20 factor. 14 To illustrate, if schizophrenia is causally related to self-harm, a PRS for 21 schizophrenia should also be associated with self-harm. A significant association between the 22 PRS for schizophrenia and self-harm can be regarded as an initial indication of a possible 23 causal relationship between the two. The PRS approach can be construed as a first step in a 24 series of genetically informed methods to investigate the aetiology of complex phenotypes, 1 with follow-up steps including Mendelian Randomization (MR) discussed below. 14-16 2 3 In previous studies, a PRS for major depressive disorder (MDD) was found to be associated 4 with SSH in two clinical samples 17,18 and one non-clinical sample. 19 However, this was not 5 replicated in a family-based sample. 20 A PRS for depressive symptoms predicted SSH but not 6 NSSH in a twin sample. 21 On the other hand, a PRS for schizophrenia was positively 7 associated with SSH among offspring of suicide attempters, 20 and a population sample, 22 but 8 not in another clinical sample. 23 A PRS for bipolar disorder predicted SSH in one clinical 9 sample 24 but did not predict SSH nor NSSH among offspring of suicide attempters, 20 and 10 relatives of bipolar disorder patients. 25 11 12 The aforementioned PRS studies with mixed results were limited in several ways. Firstly, 13 these studies focused on PRS for psychiatric disorders or psychiatric symptoms, and did not 14 include potential risk factors from other domains, such as substance use 7-9 , cognitive 15 abilities, 10 personality traits 11 and physical traits. 12  A caveat of the PRS method is its proneness to unmediated (or horizontal) pleiotropy, arising 23 from the inclusion of many thousands of genetic variants. 14 Unmediated pleiotropy exists when a genetic variant associated with an exposure causes the outcome through an alternative 1 pathway, instead of via the exposure. Unmediated pleiotropy can generate associations 2 between PRS and outcome in the absence of a causal relationship between the risk factors 3 indexed by the PRS, and the outcome. Mendelian Randomisation (MR) can more stringently 4 address unmediated pleiotropy and further strengthen causal inference. In MR, individual 5 genetic variants associated with an exposure of interest are used as instrumental variables to 6 infer causality between exposure and outcome. A number of complementary analyses, further 7 detailed in the methods section, can be implemented to account for pleiotropy. 16 To date, 8 there is no published MR study which focuses on any risk factor of self-harm. 9

10
The current study will address the aforementioned limitations by systematically using 24 PRS 11 as proxies for risk factors from different domains to predict both NSSH and SSH, using a 12 population-based sample of 125,925 individuals. We will conduct follow-up MR analyses to 13 strengthen causal inference. 14 15

Methods 16
Participants 17 The participants of the current study are a subset of the UK Biobank 18 To know whether the participants have ever-self-harmed, participants were asked "Have you 7 deliberately harmed yourself, whether or not you meant to end your life?" To ascertain 8 whether their self-harm episodes were NSSH or SSH, they were asked "Have you harmed 9 yourself with the intention to end your life?". In both questions, responses of "Prefer not to 10 answer" (0.43%) were recoded as missing values. A flowchart depicting exclusion of 11 participants and the number of participants who answered each question is shown in Figure 1.

PRS analyses 18
PRS of UK Biobank participants were generated using PRSice-2 28 based on their genotype 19 data and 24 publicly available summary data from GWAS (see Table 1) selected based on the 20 following criteria. First, we selected GWAS indexing individual vulnerabilities and traits that 21 can potentially increase the risk of self-harm, including mental health vulnerabilities (e.g. 22 MDD), 29 cognitive abilities (e.g. education attainment), 30 personality traits (e.g. 23 neuroticism), 31 substance use phenotypes (e.g. cannabis use), 32 and physical traits (e.g. 1 BMI). 12 Second, we selected GWAS which only included participants of European ancestry 2 and did not include UK Biobank participants (to avoid overlapping between discovery sample 3 size and target sample). Finally, we excluded GWAS with effective sample sizes less than N 4 = 15,000 to limit the use of underpowered PRS. 5 6 Each participant had 24 PRS, which were each calculated as the sum of alleles associated 7 with their respective phenotypes, weighted by their effect sizes with p-values less than a 8 threshold pT < 0.3 (selecting and reporting results from a single threshold allowed us to limit 9 multiple testing, as done in previous PRS studies). 22,33 Clumping was used to remove SNPs in 10 linkage equilibrium (r 2 < 0.1 within a 250 kb window). All PRS in the final analytical sample 11 were standardised. 12

13
Single PRS Binomial Logistic Regression 14 For each PRS, a binomial logistic regression was conducted to test whether it predicted self-15 harm (i.e. "Self-harmed" versus "Never self-harmed"). To investigate whether each PRS differentially predicted NSSH versus SSH, we fitted a 2 series of multinomial logistic regression models. We first compared each of the NSSH and 3 SSH groups to the never self-harmed group (i.e. "Never self-harmed" as the reference group). 4 We then directly compared NSSH and SSH by testing a model with "NSSH" as the reference 5 group. 6 7 Covariates and multiple testing 8 All regression models were controlled for sex, age and population stratification (by including 9 assessment centre, genotyping batch and the first 6 principal components as covariates in the 10 models). To control for multiple testing in single PRS binomial and multinomial regressions, 11 we employed the false discovery rate (FDR) method 34 which controls the expected proportion 12 of false positives among the rejected hypotheses. We used q < .05 as the significance 13 threshold. 14 15

MR analyses 16
All MR analyses were conducted using R package TwoSampleMR. 35 Risk factors for which 17 their PRS significantly predicted self-harm were selected for follow-up MR analyses. For 18 self-harm in UK Biobank sample as the outcome for MR analyses, we obtained GWAS 19 summary statistics from Neale Lab (http://www.nealelab.is/uk-biobank). SNPs of the 20 exposures which passed the p-value threshold of p < 5E-5 were selected as instrumental 21 variables for MR analyses. A liberal threshold was used to ensure that enough variants were 22 available for all risk factors, including those with few genome-wide significant SNPs (e.g. 23 ADHD). The strategy entails potential weak instrument bias. In two-sample MR, the resulting 24 bias is towards the null, making estimates more conservative (see below how this was dealt 1 with). 36 Clumping of SNPs with r 2 < .001 within 250 kb was applied. SNPs in exposures and 2 outcomes were harmonized by flipping alleles where possible, and we use allele frequencies 3 to infer strands of ambiguous SNPs. Non-inferable SNPs with minor allele frequency > 0.42 4 were discarded. 5 6 We selected four MR methods which have different strengths and limitations. We conducted 7 univariable MR using: 8 Inverse variance weighted (IVW) method, which is the most powerful method but 9 cannot account for directional pleiotropy; 37 10 (ii) Robust Adjusted Profile Score (RAPS) method, which is used to account for the 11 selection of weak instruments; 33 12 Weighted median method, as it is more robust to directional pleiotropy than IVW 13 and is more robust to individual genetic variants with outlying causal estimates 14 than IVW and MR-Egger; 38 and 15 (iv) MR-Egger regression method, whereby significance of its intercept term can 16 inform on the presence of directional pleiotropy. 39 17 18 In addition, MR Steiger filtering 40 was implemented to address the possibility of reverse 19 causation (i.e. self-harm causing the putative risk factor). For each SNP, we expect that the 20 effect size for the association with the exposure should be larger than the effect size for the 21 association with the outcome. This is because the effect on the outcome is hypothesised to be 22 indirect through the exposure. As such, all SNPs for which the effect size of the association 23 with the outcome was larger than the one with the exposure were filtered out before 1 reimplementing MR. Finally, similar to PRS analyses, exposures which were significant in 2 univariable MR were assessed for their independent effect in a multivariable MR model using 3 the IVW method. 4 5 For PRS analyses, we conducted further complementary analyses excluding cases with MDD 6 and schizophrenia diagnoses to investigate the effect of genetic liability on self-harm with the 7 influence of diagnoses excluded. We also calculated risk ratios for medicated and non-8 medicated cases compared to those with median PRS in the general population (see 9 supplementary materials for definitions of cases and medication). We created a quantile plot 10 separating the participants into three groups: general population (in 20 quantiles), medicated 11 cases and unmedicated cases, and calculated the risk ratios of these groups for self-harm 12 relative to the group in the population with median PRS. 13 14

Results 15
Descriptive statistics 16 Figure 1 shows the number of participants who: never self-harmed, self-harmed, engaged in 17 SSH, and engaged in NSSH. Table 2 shows the gender proportion, and mean age of each 18 subgroup. 19

PRS analyses 20
Single PRS binomial logistic regression 21 Table 1 and Figure 2 show results from 24 single PRS binomial logistic regression tests, 22 using each PRS as a predictor variable. Out of the 24 PRS, 10 PRS were significant 23 had q-value < .05. In order of decreasing effect sizes, they are PRS for: MDD, schizophrenia, 1 ADHD, bipolar disorder, alcohol dependence disorder (ALC), and lifetime cannabis use, with 2 effect sizes ranging from β = 0·179 (95% CI: 0·152 to 0·207) for MDD, to β = 0·044 (95% 3 CI: 0·016 to 0·072) for lifetime cannabis use. Figure S3 shows the pseudo R 2 plots of these 6 4 PRS in accounting for the variance in self-harm. 5 6 Multiple PRS binomial logistic regression 7 In the multiple PRS model, all PRS except the PRS for ALC had an independent effect of 8 self-harm as shown in Table 1 and Figure 2. By controlling for the effects of other PRS, 9 effect sizes of these PRS have diminished slightly compared to those in single PRS binomial 10 logistic regression, ranging from β = 0·144 (95% CI: 0·115 to 0·173) for MDD to β = 0·031 11 (95% CI: 0·002 to 0·060) for bipolar disorder. These PRS were weakly correlated, ranging 12 from r = 0.01 (between bipolar disorder and ADHD) to r = 0.22 (between schizophrenia and 13 bipolar disorder; see Table S3 for all correlations), suggesting that multicollinearity was not 14 an issue. 15 16 Single PRS multinomial logistic regression 17 Table S1 shows results from 24 multinomial logistic regression tests, using PRS as predictor 18 variable for three possible outcomes: "Never self-harmed", "NSSH" and "SSH". When 19 "Never self-harmed" was used as the reference group, PRS for bipolar disorder, lifetime 20 cannabis use and extreme BMI predicted SSH but not NSSH, with q < .05. However, when 21 "NSSH" was set as the reference group in order to directly compare NSSH versus SSH, none 22 of the PRS significantly distinguished between NSSH versus SSH. 23 schizophrenia and ADHD are more predictive of the respective exposures than self-harm, 20

MR analyses 1
suggesting that reverse causation unlikely explained our findings. In PRS complementary analyses which excluded cases, PRS for MDD and schizophrenia still 8 predicted self-harm in a healthy, screened cohort, indicating that genetic liabilities can predict 9 self-harm when influence of diagnoses is excluded (See Table S2). In the quantile plot, cases 10 for schizophrenia and MDD appear to be at much larger risk for self-harm than the rest of the 11 population. Medicated MDD cases were at higher risk of self-harm than non-medicated MDD 12 cases, which was not the case for schizophrenia (See Figure 3). 13 14 Discussion 15 To our knowledge, this is the first study using multiple PRS as genetic proxies to 16 systematically investigate a range of individual vulnerabilities and traits as risk factors for 17 self-harm in a large population sample. In PRS analyses, we identified 6 risk factors (i.e. 18 MDD, schizophrenia, ADHD, bipolar disorder, ALC, and lifetime cannabis use) which 19 predicted self-harm. Five among six (except for ALC) remained significant in a multiple PRS 20 regression. We found no evidence of differential prediction for SSH versus NSSH. In follow-21 up MR analyses, MDD, schizophrenia and ADHD emerged as plausible causal risk factors 22 for self-harm, despite evidence of unmediated pleiotropy for MDD. We discuss in turn: (1) 23 insights into the aetiology of self-harm, and (2) clinical implications. 1

Insights into the aetiology of self-harm 2
Results from our PRS methods corroborated previous observational findings where MDD, 6 3 schizophrenia, 41 ADHD, 42 bipolar disorder, 6 and ALC 8 were phenotypically associated with 4 self-harm. Our results are also consistent with positive associations found in PRS studies for 5 MDD, 17,19,43 schizophrenia, 20 and bipolar disorder 24 . Previous mixed findings for these PRS 6 may have stemmed from lack of power, as sample sizes for those studies varied widely. The 7 current study adds lifetime cannabis use, ADHD, and ALC as novel PRS associated with self-8 harm. However, when controlling for other PRS, the PRS for ALC did not significantly 9 predict self-harm. This finding may suggest that the genetic liability for ALC does not 10 independently predict self-harm when the effect of genetic liability for MDD, bipolar 11 disorder, schizophrenia, ADHD and lifetime cannabis are accounted for. For example, ALC 12 may be a marker for a true predictor such as impulsivity which is more efficiently captured in 13 the PRS for ADHD. 44 . Alternatively, null findings for ALC can also plausibly be due to a 14 lack of power compared to other polygenic scores. Hence, we cannot completely rule out that 15 the PRS for ALC has an independent effect on self-harm and the corresponding causal effect 16 of ALC on self-harm. 17 18 Most of the PRS which predicted self-harm in the current study relate to psychiatric 19 conditions, which confirms the prominence of psychiatric conditions in the aetiology of self-20 harm. 45 Beyond psychiatric conditions, cognitive traits, physical traits, and personality traits 21 were not found to be associated with self-harm using PRS approach, although previous 22 observational findings found significant phenotypic associations for these three domains. 10-12 23 The absence of significant findings in this case is unlikely to be solely due to lack of power, 24 given that GWAS for some of these traits are more powerful than GWAS for psychiatric 1 conditions (e.g. BMI and education attainment). These findings suggest that these traits and 2 vulnerabilities are unlikely to have (strong) causal effects on self-harm. 3 4 Our MR analyses provided further support for the role of MDD, ADHD, and schizophrenia in 5 the aetiology of self-harm. An intriguing finding is the presence of significant pleiotropy in 6 the case of MDD. Rather than signifying that MDD does not have a causal effect on self-7 harm, this may reflect a possible measurement issue. Indeed, one of the diagnostic criteria for 8 MDD is related to having suicidal thoughts and attempts, which could artificially introduce a 9 pleiotropic effect. 5 To deal with this issue, future studies may rely on a GWAS for MDD 10 excluding the diagnostic criteria related to suicidal thoughts and attempts. This might also 11 explain why, in multivariate MR, the effect of ADHD was no longer significant -as we 12 partially controlled for self-harm -whereas it was significant when only considering ADHD 13 and schizophrenia. 14 15 The current study found mixed results for whether there are distinct aetiologies for SSH and 16 NSSH. Most PRS which predicted self-harm also predicted both SSH and NSSH, except 17 bipolar disorder, lifetime cannabis use and extreme BMI, which only predicted SSH but not 18 NSSH from those who never self-harmed. However, in a formal test comparing NSSH and 19 SSH, the estimates of these three risk factors were not significantly different between NSSH 20 and SSH. Hence, our findings do not provide evidence for marked differences in aetiology 21 between SSH and NSSH. 22

Clinical implications 1
The current study suggests that individual vulnerabilities and traits underlying self-harm most 2 likely relate to psychiatric conditions such as MDD and schizophrenia, rather than to other 3 domains such as personality traits. Hence, treatments focusing on the core symptoms of these 4 psychiatric conditions are important in preventing or addressing the risk of self-harm. 5 Findings from PRS analyses suggest that genetic liabilities for these conditions increase the 6 likelihood of self-harm even in those not clinically diagnosed. This may suggest that 7 subthreshold symptoms of these core psychiatric conditions may increase the risk of self-8 harm. Clinicians may want to systematically test for such symptoms in self-harming patients. 9 Future investigations may test whether drugs for such core conditions may be repurposed for 10 treating self-harming patients, with either full blown or subthreshold conditions. For example, 11 prescription of methylphenidate for ADHD treatment was found to be associated with 12 reduction of suicide attempt risk. 46 As a note of caution, treated schizophrenia cases were not 13 at less risk of self-harm than non-treated patients whereas treated MDD patients were at 14 substantial higher risk for self-harm. This could be due to treated patients having more severe 15 symptoms than untreated patients, or it could be due to adverse effects of medication, in In order to avoid the overlapping of discovery and target sample, we excluded GWAS which 21 contain UK Biobank sample, resulting in selecting older GWAS for generating PRS in some 22 cases. This might have led to non-significant findings due to lack of power. The results 23 should be generalised with caution because UK Biobank is not representative of the UK 24 population as they are more educated, older, wealthier, and healthier. 48 The questions asked 25 in MHQ were retrospective and their formulation led to an exclusive dichotomy between 1 NSSH or SSH, whereby some might have engaged in both NSSH and SSH at different times. 2 3

Conclusion 4
Among 24 PRS used as genetic proxies for vulnerabilities and traits possibly associated with 5 self-harm, we found that PRS for MDD, schizophrenia, ADHD, bipolar disorder, ALC and 6 cannabis were statistically significant. After a series of complementary analyses to further 7 strengthen the causal inference, schizophrenia survived as the most plausible causal risk 8 factor, followed by MDD and ADHD. Detection and treatment of core symptoms of these 9 conditions, such as psychotic or impulsivity symptoms, may benefit self-harming patients.