Scientist Stories: Michael Inouye, Polygenic risk scores and their use in the clinic

Okay, thanks for the chance to talk to you today. I'm Mike Inouye, and I'll be talking to you about polygenic risk scores in the clinic: challenges and opportunities. So I still view polygenic scores as this new thing that's made its way onto the human scene. We've got some prototypes out there, but to a large extent, the world just doesn't quite know what to make of them, nor what to do with them. Kind of like what streets were like when the first automobiles arrived. It was a bit chaotic. They were already filled with people, animals, bicycles, horse-drawn carriages, and trams. So in a clinical setting, polygenic scores face a huge array of challenges as well as opportunities. For the purposes of this presentation, I've selected a few I think are are particularly important and timely. Opportunities such as risk prediction, diagnostic support, and therapeutics, and challenges such as standards, comparability, sources of bias, and what role ultimately polygenic scores may have in society. In terms of risk prediction, polygenic scores have been called the first risk factor. And appropriately so. A polygenic score itself is fixed soon after conception and is not subject to spatiotemporal fluctuations like blood pressure. Together with sex, also genetically determined, a polygenic score can be viewed as the baseline hazard on which everything else to the organism is forward in time. This is already a prism through which we view monogenic disease risk, such as that of LDLR mutations and and familial hypercholesterolemia. As Amit Khera and colleagues have shown, individuals in the tail of a polygenic score distribution can have levels of risk equivalent to a monogenic mutation, and polygenic scores in the tail tend to be more common in a population than the monogenic mutation. According to a consistent set of principles, not a double standard, for how we treat genetic risk, polygenic risk seems to have a clinical role. A key question then becomes what action to take. These include behavior modification, where groups such as those at FIMM have developed exciting tools like CardioCompassy to facilitate behavior change. Other interventions include changes to screening initiation or frequency, independent of or in combination with therapeutic targeting. And for some diseases, we can use polygenic scores to enhance established risk prediction models as exemplified by breast cancer and the Bodecia polygenic score tool. For the use of polygenic scores to enhance existing risk models, my lab has done some work in the area of cardiovascular disease. To overcome regression dilution bias, we decided to design a meta-scoring approach, or what we call a meta-GRS, or score of scores. Using meta-scoring, we published meta-GRSs for coronary artery disease, and ischemic stroke, where these meta-GRSs outperformed previous polygenic scores, as well as any single conventional risk factor at baseline, save blood pressure or save blood pressure and stroke. In each case, as expected, the best risk models integrated polygenic scores and conventional risk factors. However, a major challenge of existing conventional risk models is that they predict the more heterogeneous cardiovascular disease, not coronary artery disease or stroke specifically. As such, we've recently assessed the clinical utility of these meta-GRSs to enhance conventional risk models and guide statin treatment for cardiovascular disease. Importantly, we tried to control for the healthy bias in the UK Biobank sample by using primary care health records, the UK CPRD database, to recalibrate risk to a primary care population. Overall, we found that even modest increases in C-statistics, say 1 or 2%, can indeed translate into meaningful clinical benefits. The polygenic scores correctly increased or decreased 10-year cardiovascular disease risk for 22% of all participants, and notably, we found that an additional 7% of cardiovascular disease events may be prevented when including these polygenic scores. It's not straightforward to estimate exactly how many additional cardiovascular disease events that is, even in the UK, but it may be in the thousands, perhaps 10,000 or more events per 10 years. That's far more than C-reactive protein, or CRP, a biomarker that's now included in some risk prediction guidelines. A second opportunity for polygenic scores is in the area of supporting diagnostics. This is a particularly promising approach for autoimmune diseases, which themselves appear to follow a more common disease-common variant architecture. In type 1 diabetes, Richard Oram and colleagues at Exeter have designed and validated a polygenic risk score with an AUC of 0.92, which shows promise in in guided selection of newborns for autoantibody screening and the classification of type 1 and type 2 diabetics in adulthood, the latter being important to avoid incorrect treatments as well as reduce medical costs and morbidity. In celiac disease, multiple studies have shown that a polygenic risk score can improve upon traditional HLA typing. Back in 2013, we showed that L1-regularized support vector machines are particularly good at highly MHC-dependent disease architectures, and this was certainly the case in celiac disease, where scores externally validated across 4 different datasets with AUCs of roughly 0.87 to 0.9. Similar to type 1 diabetes, polygenic scores for celiac disease have the potential to replace HLA typing and combine with or guide serology. In celiac, it's particularly important to minimize gluten challenges, always a costly and painful exercise, and the small bowel biopsies to follow. More recently, in juvenile idiopathic arthritis, a common cause of disability in children, we've trained polygenic scores again using L1 regularization to predict both JIA and its subtypes. A polygenic score for JIA may have substantial clinical utility because the diagnosis of JIA is currently purely clinical. With no molecular tests to support it. Interestingly, the JIA subtypes where we had the most predictive power corresponded to the oligoarticular subtype, the most common, and the anthracitis-related JIA, which had the longest wait time to secure a diagnosis, a median of 11 months. Those results are currently up on medRxiv. Polygenic scores also present opportunities for research into new therapeutics. I won't say too much on this point as Scott Ritchie, a postdoc in the lab, will be presenting on it on Tuesday, but it's worth highlighting the framework a bit. We of course know that biomolecular pathways mediate polygenic risk of disease. As such, we may be able to use polygenic scores to detect some of these disease pathways. One approach is if we integrate polygenic scores and multi-omics data together for the same individuals in pre-symptom— presymptomatic states, we may be able to associate levels of polygenic risk with biomolecular networks, or ideally, as our data becomes richer, follow polygenic score effects on these networks in the same individuals over time. There's still lots of work to be done in this area, as well as a few notable caveats. First, a polygenic score isn't necessarily a genetic instrument, and to be sure, a polygenic score for disease X doesn't imply causality of pathway Y on disease X. This is still an association study, and other techniques to assess causality, such as triangulation, are even more important here. Second, as disease processes change over time, it's entirely reasonable to expect polygenic score effects to be time and/or age varying as well. Overall though, while it's a bit behind risk prediction and diagnostic support, therapeutic development is quite a promising opportunity for future polygenic score research. First and foremost amongst the challenges for polygenic scores making their way into the clinic is standards. While best practice is evolving, it is difficult to pin down the best approach for a particular task. One thing we can address right now is the poor reporting of polygenic score studies. The NIH ClinGen Complex Disease Working Group, in conjunction with the Polygenic Score Catalog, recently performed a review of 30 representative publications and found there was a glaring lack of reporting for basic elements of the studies. These included statistical model validation, ancestry of the data, and data transparency and availability. Notably, a third of polygenic score papers do not make available the variants and weights of the published polygenic score. This collaboration came up with the polygenic risk score reporting standard, or PRS-RS, a set of 33 criteria needed to establish a polygenic score's analytic validity, transparency, and reproducibility, thus facilitating its clinical translation. Currently, you can read more about PRS-RS in our medRxiv preprint with Hannah Wand, Jen Wasik, and colleagues. Building on reporting standards, the comparability between polygenic scores is a major challenge facing the field. Comparability depends on a few key elements. First, there needs to be an open repository. We've made some progress in this area with the development of the Polygenic Score Catalog, the sister to the NHGRI EBI, GWAS catalog. Sam Lambert, also a postdoc in the lab, will be presenting that on Tuesday. External validation data is of course critical. We need more genotyped cohorts with deep compatible phenotypes. And towards a similar end, for the next generation of GWAS summary statistics and polygenic score training to commonly withhold some samples or cohorts for future validation analyses. And hopefully the direct benchmarking of new or existing polygenic scores in the same individuals will become standard practice. Polygenic score benchmarking, despite its importance, is still unfortunately rare in the literature. Consequently, many analyses and conclusions drawn regarding the clinical utility of polygenic scores for disease X too often suffer from regression dilution bias. As predictors, polygenic scores need to be particularly wary of regression dilution. Unlike noise in a phenotype measurement, if there's noise in the polygenic score, then there is a bias of the model towards the null. Benchmarking is making its way more into literature, to some extent driven by other sources of bias in polygenic score analyses, such as that of ancestry. An excellent example is this recent study from eMERGE, where one is able to assess the relative performance of different polygenic scores for coronary artery disease in large cohorts of European, African, and Hispanic ancestries. This reveals wide variability in PGS performance to the extent that some polygenic scores don't meet the criteria for analytic validity in some ancestries. Sources of bias are major challenges. To reiterate, the European ancestry bias, which was ingrained in cohorts decades ago and subsequently made its way into GWAS and polygenic scores, is particularly pernicious. As Mark Daly, Alicia Martin, and colleagues have pointed out, this raises equity issues that can most fundamentally be solved by more non-European cohorts. Including and beyond ancestry bias, further prospective cohort collection will also need to be cognizant of the common biases in selection, information, and confounding. This is especially true for minorities, low socioeconomic status, and other disempowered groups. Clearly, genetics will need to continue working closely with epidemiology, public health, and related fields to minimize these issues. Implementation biases will also likely plague translation into the clinic. Inequality in society creates inequality in access to healthcare. As a consumable good, polygenic scores can also be tipped to benefit the wealthy and the powerful first and foremost, undermining one of the original virtues of the sequencing of the human genome. Perhaps it's inevitable that what we develop for our health and the clinic also spills out into society. Polygenic score research is frequently caught up in controversy, exemplified in this piece by Amy Harmon, which notes the affinity of white supremacists for polygenic scores and human genetics in general. The vagaries of an infatuation with dairy aside, it's impossible to avoid the wider implications of the polygenic scores we human geneticists develop and the need to stay mindful of how what we develop and what we write can be twisted by those of bad faith. We've made a lot of progress in the last 3 or 4 years or so since polygenic scores really took off, but there's still a long way to go. For polygenic scores, we're still more in the 1890s when the first automobiles hit the streets than in today's highway infrastructure, road laws, and licensing, which have made Tesla's self-driving cars possible, if a little nerve-wracking. Ultimately, a lot of these challenges are themselves opportunities, and we should see them as such. However, to maintain confidence in polygenic scores, the community should come together to address these challenges as much as we can before rolling out polygenic scores in the clinical setting and the wider use that is perhaps inevitable. So that's it. I'd like to acknowledge my lab, especially Gad Abraham, Scott Ritchie, and Sam Lambert, who've done a lot of the heavy lifting for the work that I presented here, as well as colleagues at Cambridge University and the Baker Heart and Diabetes Institute, as well as those right around the world who've who've made this— a lot of this work possible. Thanks again, and thanks for listening.

Thank you very much, Dr. Inouye. That was an excellent overview of the promises and pitfalls of introducing a new technology into a society that may not be ready for it. Our last talk before we move to the panel discussion is from Dr. Amit Khera. He is a clinician and cardiologist at Massachusetts General Hospital and founder of a new preventive genomic clinic there. He's also a human geneticist leading a research group that focuses on using human genetic variation to both uncover new disease biology and enable enhanced clinical care. Dr. Kara will speak about polygenic risk scores and their use in the clinic. Dr. Kara, the floor is yours.