Extending the Triangulation Within a Study (TWIST) framework to improve real-world evaluation of genetically driven medication response
Lead Research Organisation:
UNIVERSITY OF EXETER
Department Name: Institute of Biomed & Clinical Science
Abstract
People vary greatly in their responses to medicines, both for therapeutic effects and adverse events. Genetic variation is an important contributor to such variation in many drugs and the science aimed at understanding this is termed pharmacogenetics. A recent UK study using current pharmacogenetic prescribing guidelines estimated that nearly six million UK primary care patients risk drug-gene interactions. For example, Clopidogrel is the most commonly used drug in the UK to reduce the risk of stroke. It requires the CYP2C19 liver enzyme to metabolise it into an active form so that works to its fullest extent. However, it has long been known that about 30% of the population have genetic variants in their CYP2C19 gene region which impacts their ability to metabolise it. When prescribed in a primary care setting it consequently works well for some and not for others.
Existing methods based on the analysis of observational data can struggle to disentangle the effects of the disease for which the medication was given from the effects of drug itself, a problem termed `confounding by indication'. For example, people who take Clopidogrel are more likely to experience a stroke than those who do not, but this does not mean Clopidogrel truly increases stroke risk. Randomised clinical trials (RCTs) provide robust estimates of outcomes to tested drugs within genetic subgroups, but are typically carried out in selective patient cohorts that are free from multimorbidity and not representative of those treated in routine clinical practice. RCTs are also often too small and too short in duration to identify adverse events or assess longer-term outcomes.
This proposal involves developing and extending genetics-based methods to analyse the accumulating wealth of data from observational electronic medical records to (a) discover genetic variants and patient characteristics that influence response to treatment and (b) to quantify the population benefits of personalised prescribing. For example, using data from over 200,000 UK Biobank participants with linked primary care data up to 2017 revealed a potential 13.2% reduction in the total number of strokes is possible if those with a genetically unfavourable CYPC219 genotype could experience the full effect of Clopidogrel, through either dose modification or switching to an alternative therapy.
Our project will build on a recently proposed decision framework termed `Triangulation Within A Study (TWIST) developed by the academic team. This method seeks to address a research question using a range of estimation strategies that are reliant on different sets of assumptions. Statistical tests and expert knowledge is then used to decide how best to combine the estimates to provide the most efficient and reliable estimate. The method is sound, having passed a rigorous peer review process for a respected scientific journal. It shows real promise but requires further research investment to realise its full potential, which is the basis of our proposal.
Our work will be primarily motivated by answering questions about the optimal treatment of patients with hypertension and cardiovascular disease, but has the potential to be applied widely across a whole spectrum of diseases, including patients with multimorbidity.
Enabling pharmacogenetics is one of the core aims of the UK Government's project to genotype 5 million people in the 'Our Future Health' study and there are many large-scale databases internationally linking genotyping to electronic clinical records. Therefore the methods and tools we will develop will have wide application locally and internationally in future years
Existing methods based on the analysis of observational data can struggle to disentangle the effects of the disease for which the medication was given from the effects of drug itself, a problem termed `confounding by indication'. For example, people who take Clopidogrel are more likely to experience a stroke than those who do not, but this does not mean Clopidogrel truly increases stroke risk. Randomised clinical trials (RCTs) provide robust estimates of outcomes to tested drugs within genetic subgroups, but are typically carried out in selective patient cohorts that are free from multimorbidity and not representative of those treated in routine clinical practice. RCTs are also often too small and too short in duration to identify adverse events or assess longer-term outcomes.
This proposal involves developing and extending genetics-based methods to analyse the accumulating wealth of data from observational electronic medical records to (a) discover genetic variants and patient characteristics that influence response to treatment and (b) to quantify the population benefits of personalised prescribing. For example, using data from over 200,000 UK Biobank participants with linked primary care data up to 2017 revealed a potential 13.2% reduction in the total number of strokes is possible if those with a genetically unfavourable CYPC219 genotype could experience the full effect of Clopidogrel, through either dose modification or switching to an alternative therapy.
Our project will build on a recently proposed decision framework termed `Triangulation Within A Study (TWIST) developed by the academic team. This method seeks to address a research question using a range of estimation strategies that are reliant on different sets of assumptions. Statistical tests and expert knowledge is then used to decide how best to combine the estimates to provide the most efficient and reliable estimate. The method is sound, having passed a rigorous peer review process for a respected scientific journal. It shows real promise but requires further research investment to realise its full potential, which is the basis of our proposal.
Our work will be primarily motivated by answering questions about the optimal treatment of patients with hypertension and cardiovascular disease, but has the potential to be applied widely across a whole spectrum of diseases, including patients with multimorbidity.
Enabling pharmacogenetics is one of the core aims of the UK Government's project to genotype 5 million people in the 'Our Future Health' study and there are many large-scale databases internationally linking genotyping to electronic clinical records. Therefore the methods and tools we will develop will have wide application locally and internationally in future years
Technical Summary
Work Package 1 (WP1) sub aims are as follows
1.1 To Conduct a systematic review of triangulation frameworks including Replication and Evidence Factors and Data Fusion.
1.2 To develop methods to control error rates (encompassing FWER and FDR definitions) and statistical power for TWIST.
1.3 To extend heterogeneity testing to allow for correlated estimates (currently only approximately statistically independent estimates are compared)
1.4 To automate confounder selection and adjustment using propensity score and high dimensional confounder adjustment methods
1.5 To develop stand alone software to design, implement and visualise the results of a TWIST analysis
WP2 sub-aims
2.1 To incorporate multiple genetic variants into TWIST to improve power and make the method more pleiotropy robust
2.2 To apply TWIST reliably in conjunction with a genetically instrument exposure for the purposes of safe stratification (i.e. avoiding collider bias) and to test for causal effect heterogeneity across strata
WP3 sub aims
3.1 To apply TWIST at scale across the pharmaco-genome to uncover new pharmacogenetic variants and confirm or refute the status of previously identified variants
3.2 To perform stratified analysis based on sex, age, kidney function and cholesterol levels
3.3 To use TWIST to find the optimal treatment for each individual, taking into account both the benefits and harms, and then to quantify this effect at the population level
3.4 To apply TWIST to investigate the repurposing of diabetes medications for cardiovascular outcomes, and to use clinical trial data to validate the results.
1.1 To Conduct a systematic review of triangulation frameworks including Replication and Evidence Factors and Data Fusion.
1.2 To develop methods to control error rates (encompassing FWER and FDR definitions) and statistical power for TWIST.
1.3 To extend heterogeneity testing to allow for correlated estimates (currently only approximately statistically independent estimates are compared)
1.4 To automate confounder selection and adjustment using propensity score and high dimensional confounder adjustment methods
1.5 To develop stand alone software to design, implement and visualise the results of a TWIST analysis
WP2 sub-aims
2.1 To incorporate multiple genetic variants into TWIST to improve power and make the method more pleiotropy robust
2.2 To apply TWIST reliably in conjunction with a genetically instrument exposure for the purposes of safe stratification (i.e. avoiding collider bias) and to test for causal effect heterogeneity across strata
WP3 sub aims
3.1 To apply TWIST at scale across the pharmaco-genome to uncover new pharmacogenetic variants and confirm or refute the status of previously identified variants
3.2 To perform stratified analysis based on sex, age, kidney function and cholesterol levels
3.3 To use TWIST to find the optimal treatment for each individual, taking into account both the benefits and harms, and then to quantify this effect at the population level
3.4 To apply TWIST to investigate the repurposing of diabetes medications for cardiovascular outcomes, and to use clinical trial data to validate the results.