Emulating biomarker-guided target trials using big data

Stratified medicine has potential to improve the benefit-risk ratio of treatments, and the randomised controlled trial (RCT) is the gold standard for demonstrating the clinical utility of an intervention, including a biomarker-guided ('BM-guided') approach to treatment. Many BM-guided trial designs have been proposed for this purpose, as identified in previous work by our group1,2, which led to our web-based tool for design and analysis of BM-guided trials, BiGTeD3.
However, outcomes of interest in a BM-guided trial are often rare, and the biomarker itself can be rare. Both issues mean that large, unachievable sample sizes are required to achieve a sufficiently powered RCT. In addition, conducting RCTs can take many years making them an unattractive choice in a rapidly advancing field such as stratified medicine 4. Due to these limitations, turning to observational data to demonstrate clinical utility seems appealing. Such data can come from more traditional case-control or cohort studies, which combined in a meta-analysis can provide precise estimates of effect comparable to RCTs5. More recently, observational data are also available from routinely collected sources, such as electronic health records, which may also be linked to genetic and other data. Such data are available in the UK Biobank (UKBB)6. In addition to being less costly, observational studies can produce data more representative of the underlying patient, in the absence of some of the conditions and constraints inherent to being part of an RCT. Further, they are useful in situations where an RCT would be considered unethical, e.g. in the absence of clinical equipoise7.

Despite the many benefits of observational data, a major limitation is controlling for unmeasurable confounding and other biases. However, with careful consideration, inference of causal effects from observational data can be achieved by aiming to replicate the 'ideal RCT' we would otherwise use to address our question of interest. This process is often referred to as 'emulating' a 'target trial'8; in the context of stratified medicine the 'target trial' would be a BM-guided trial. Examples are available of emulating target trials of a personalised approach to treatment based on clinical characteristics9, however given the availability of large databases with data on both patient treatment history and genetic and other biomarkers, it appears sensible to explore how emulating a target trial c be useful in the field of BM-guided trials.

What the studentship will encompass

First the literature will be reviewed and appraised on how observational data are currently used to assess clinical utility of BM-guided treatment. Next, methodologies and guidelines for emulating target trials using observational data will be reviewed, with due consideration to how applicable these are in a BM-guided trial setting. Methods and guidelines for emulating BM-guided trials from observational data will be developed, and these methods applied to real data obtained from UKBB to emulate trials of a BM-guided approach to treatment. The exemplar we propose is a target trial testing clinical utility of a genetic biomarker, SLCO1B1*5, in tailoring statin therapy using UKBB data. Whilst statins are commonly used and generally well tolerated, they are associated with statin-related myotoxicity (SRM) ranging from mild to rare but life-threatening10. Importantly SRM not only causes direct harm to patients, but also leads to statin discontinuation and non-adherence, increasing risk of major cardiovascular events and mortality11. Carriers of SLCO1B1*5 have been found to be at significantly increased risk of SRM from taking simvastatin but not other statins, and testing for SLCO1B1*5 provides an opportunity for tailoring statin therapy based on genetics. Peyser et al12 undertook a RCT of SLCO1B1 guided statin therapy, but failed to show a benefit of a genotype-guided approach. There may be several reasons for this in