In this study, we used genetic data to stratify statin users by their genetically predicted response to statins and investigated their risk of incident ICH risk. We leveraged data from a GWAS of on-statin LDL lowering from 40 000 statin-treated individuals (75% clinical trial participants) as well as biochemical, lipidomic and primary care data from 225 000 individuals from a population-based study. We found that a higher genetically predicted LDL response to statins associated with steeper LDL lowering, a similar lipidomic signature as high-dose statin use and a lower risk of atherosclerotic cardiovascular outcomes. In addition, this higher genetically predicted LDL response to statins was associated with a higher risk of ICH among statin users only. There was no such association among individuals who were not taking statins. Our results support a causal effect of more aggressive LDL lowering with statins on risk of ICH and highlight the utility of modelling drug response in addition to dose in examining putative causal associations between biomarkers and outcomes.
Our study extends previous findings from genetic[26,29] and observational analyses,[8–13] providing evidence that beyond lifetime variation in LDL levels, genetic variation in statin-induced LDL lowering also influences ICH risk. This result agrees with post hoc analyses of clinical trials supporting a higher risk for haemorrhagic stroke among participants prescribed a high-intensity statin dose. The mechanisms underlying this observation remain poorly understood. It has been speculated that cholesterol is important for vessel integrity, but to date no experimental study has provided evidence for a mechanism connecting low cholesterol levels to vessel damage or loss of vessel structural integrity. As demonstrated in our analyses but also in previous work, statins influence a wide range of lipoprotein particles beyond LDL and thus revealing the main driver of their association with ICH remains a key challenge. Although the follow-up time of existing trials does not exceed 3 years, a meta-analysis did not find aggressive LDL lowering by PCSK9-inhibitors increases ICH risk, even in high-risk patients with previous ischaemic or haemorrhagic stroke, indicating that LDL might not be the sole driver. Because of the widespread lipidomic effect of the genetic score we used, it is not possible from our current analyses to make inferences about which particle class is the causal mediator of this association.
Early clinical trials of statin administration had found a slightly elevated risk for ICH among statin users,[5–7] which was in line with data from prospective observational studies demonstrating that increased serum total cholesterol and LDL levels are negatively associated with ICH risk in a dose-dependent manner.[8–13] Although subsequent meta-analyses of statin trials found inconsistent results for overall statin use and risk of ICH,[12,14–17] high-dose statin use remained associated with an increased ICH risk. However, post hoc analyses from statin trials could not detect statistically significant increases in ICH risk associated with aggressive LDL lowering to <70 mg/dl or <55 mg/dl. These conflicting data about incident ICH among statin users remain a source of concern among medical professionals and are the motivator of the ongoing NINDS-sponsored Statins in Intracerebral Haemorrhage (SATURN) randomized trial (NCT03936361).
By leveraging genetic determinants of response to statin intake, we were able to randomize statin users at the beginning of drug intake, as the prescribing physician is blinded to the genetic variation in statin response. In contrast, when using genetic variants for off-statin LDL or high-density lipoprotein levels in conventional MR approaches,[26–29] randomization is performed at conception and leads to lifelong variations in lipid levels. As such, conventional MR studies have captured lifelong genetically predicted LDL levels and are thus limited in making any inferences about the causal effects of a particular drug prescribed over a shorter timeframe. Our approach overcomes this limitation, facilitating causal inference of the impact of statin intake on ICH using solely observational data. This application could be implemented in other settings as well, and demonstrates the latent utility of additional efforts to develop polygenic predictors of drug response in pharmacogenomic research.
From a methodological perspective, our study also demonstrates the utility of using real-world primary care data for assessing longitudinal trajectories of clinical and biochemical assessments and medication use. Although real-world data are noisier and less standardized than data usually obtained for research purposes, they retain utility to assess drug safety and side-effects, inform clinical trial design and compare drug effectiveness. Leveraging the longitudinal drug prescription and LDL measurement data from primary care data, we were able to track statin prescription and response over a timeframe extending from several years before inclusion of the participants to the study to the end of their follow-up in the UKB. Using data from the rising number of GWAS for drug response,[46,47] future studies could explore in the primary care data from the UKB associations of drug intake with multiple end points. This could allow the detection of previously unreported adverse effects, for which trials are often underpowered, or the investigation of the potential of repurposing opportunities.
Our study has additional specific methodological strengths. Using data from 225 000 participants, including 75 000 statin users and 700 ICH events, we were sufficiently powered to detect meaningful changes in ICH risk by genetically predicted on-statin response. The phenotypic depth of the UKB dataset allowed us to validate the effects of the genetic score statin response on LDL trajectories, lipidomic traits and atherosclerotic end points. Furthermore, we have introduced novel and innovative approaches to leverage GWAS for drug response in large-scale longitudinal population-based datasets. By aggregating data from >4 million drug prescriptions, we were able to precisely phenotype drug intake at an individual level and thus control for statin dose in our outcome models.
Our approach also has limitations. First, the constructed genetic score was associated not only with on-statin LDL lowering but also with off-statin baseline LDL levels. To address this limitation, we introduced an alternative genetic score that was only associated with LDL lowering after statin intake and used that for sensitivity analyses confirming our findings. However, residual confounding due to subthreshold effects of the variants on baseline LDL levels cannot be excluded. Second, we observed a lower incidence of ICH in our study population (27 per 100 000 person-years), as compared to the age-standardized world-wide rate of 42 per 100 000 person-years. This is possibly related to the healthier profile of the UKB population as compared with the general population and necessitates a cautious interpretation of the findings. Third, our study was performed in mainly people of European ancestry and therefore our results cannot be generalized to other populations. Fourth, actual drug intake might also be influenced by poor adherence, which has not been included in our models. Fifth, statins were first introduced in 1988 and prescriptions rose since then, but it was not until 1995 that >90% of the primary care practices in the UK were fully computerized. Sixth, we lacked neuroimaging data from incident ICH events, which would enable stratified analyses by haemorrhage location (lobar versus deep). Seventh, because of the very low number of participants with a prior history of ICH, our study lacked power to explore associations of genetically predicted on-statin LDL response with ICH recurrence. Future studies should focus on exploring the same question among more vulnerable and clinically relevant populations, such as ICH survivors, among whom the balance between the risk of ICH and the prevention of ischaemic cardiovascular events might differ. Finally, because atherosclerotic cardiovascular disease prevention is the main indication for statins, limiting our cohort to statin users might have introduced collider bias for the atherosclerotic endpoints. While we addressed this issue by applying inverse probability weighted models, some relevant bias towards the null might still be present in the measured effect sizes.
In conclusion, we found that higher genetically predicted on-statin LDL response mimics exposure to higher statin doses and increases risk for ICH. These results imply that more aggressive statin-induced LDL lowering might increase risk of ICH and should be balanced against statin benefits in trials of intensive statin treatment. More broadly, our results demonstrate the utility of leveraging genetic data of drug response as a novel method of investigating side-effects and repurposing opportunities of specific drugs with observational data.
GWAS= genome-wide association study; ICH = intracerebral haemorrhage; LDL = low-density lipoprotein; MI = myocardial infarction; MR= Mendelian randomization; PAD= peripheral artery disease; SNP = single nucleotide polymorphism; UKB=UK Biobank
This research has been conducted using data from UK Biobank, a major biomedical database (www.ukbiobank.ac.uk) under the project ID 36993 'Identification of genetic components underlying cerebrovascular diseases and their related outcomes'.
Brain. 2022;145(8):2677-2686. © 2022 Oxford University Press