Scientist Stories: Teri Manolio, Polygenic risk scores in the clinic

Welcome to this year's ESHG/ASSG Building Bridges session, which I will be co-moderating together with Dr. Terry Manolio, Chair of the Scientific Program Committee of the American Society of Human Genetics. Let me give you a brief history about these Building Bridges sessions. The goal is that we will learn from each other, see differences, and debate hot topics each year. And we do that both at the ESHG as well as the ASSG. The presentations are short, and we have a lot of time for discussion afterwards. And we've discussed topics such as incidental findings in genome diagnostics, prenatal genome diagnostics, and germline genome editing in the past. This year, it's already the 13th Building Bridges session, and we've chosen to move to the hot topic of polygenic risk scores. There is quite some discussion around it, and there's also a lot of potential to start implementing them in the clinic. What we will have is 4 short presentations by experts in the field, and we will open it up after those 4 for discussions with you. And like in previous sessions, what we ask you is to start asking questions on the chat already from the first presentation on, but we will only answer them in the end. And please show your appreciation of speakers really after the discussion has taken place, because it's difficult for moderators to basically distinguish questions from appreciation. So, if you could only ask questions on the chat while we are going, that would be great. And now I start and I hand over to my co-moderator, Dr. Manolio, to introduce the first speaker.

Thank you, Dr. Veltman. I'm delighted to have the opportunity to introduce the first speaker. He is Mark Daly. He's the director of the Institute of Molecular Medicine Finland, where he was recruited a few years ago from Harvard Medical School and the Broad Institute of MIT and Harvard. While he's primarily based in Finland now, he retains an active role in Boston at his former institutions and is building close partnerships across the seas there. His research focuses on discovery of genes involved in human disease to move toward therapeutic development and improve diagnostics. His talk today is entitled Polygenic Risk Scores: Old Concepts, New Uses. Dr. Daly.

Thanks to the organizers for the invitation to speak in this symposium. While polygenic risk scores are often thought of as new, they actually have a deep and rich history stretching back more than 100 years, back to the early biometric revolution, Galton and Pearson and others, to both empirical and theoretical resolution of Mendelian inheritance and biometric models that demonstrate polygenic inheritance both in empirical plant and soon after fly cross experiments, as well as Fisher's synthesis of the biometric and Mendelian models, demonstrating essentially that 100 years before it happened, we really should have expected the results of GWAS long before we got to them. Fast forwarding a bit, it was still 50 years ago when Gottesman and Shields published published a polygenic theory of schizophrenia, which described very, very accurately what we now understand as the polygenic inheritance of complex traits across the board, but which was remarkably ahead of its time with respect to our understanding of human genetics. And it was still 20 years ago when the basics of how to calculate polygenic risk scores and then how to use them to predict outcomes in untested progeny was first demonstrated in this paper, very much again ahead of its time, but which ushered in widespread use of essentially the same polygenic risk score approaches that we discussed today in humans in animal and plant breeding. And here finally, a seminal paper from Sean Purcell, Pam Slar, and many others that demonstrated empirically the predictions of the previous two papers that I cited, showing for the first time in a complex human disease, and remarkably in the context of a GWAS with no hits, that one could predict case control status to a certain extent in a new sample simply from the additive sum of weights from a previous study. And this was really a, you know, for me and many, many others in the field, mind-opening with respect to the potential of polygenic variation and its ability not to predict, but really to presage great success in GWAS long before it would actually happen. It also presaged something which I'll come back to later, which is that the transferability of these scores was much greater from European to European sample than it was from European to African American. So moving to the present day and the topic of this symposium, risk prediction in common disease, yet despite being 10 years out from that initial schizophrenia finding, doesn't generally utilize genetic risk at all and focuses on, in many cases, high-risk individual identification through clinical risk factors. And these are popular in many cases because they are measurable, directly modifiable, and have a proven causal relationship to end outcomes that we're trying to prevent. For most diseases at this point Genetic risk evaluation is quite limited and focuses really on rarer high-impact variants, and these searches are generally triggered by specific clinical features that predict a strong variant might be at play. For example, unusual early onset, strong family history, or some other aspects of the presentation that make one suggest a significant genetic lesion might be present. And really, outside of pediatrics and development, these are scarcely applied and applied mostly when they are in cancer and and neurology and neurodegeneration and so forth. So the question at hand is, are polygenic risk scores now ready to help? In this talk, I'm going to be drawing some examples and citing some work from the FinnGen project. And FinnGen is a massive public-private partnership, as the name would imply, focused in Finland with a goal of reaching over the next couple of years 500,000+ individuals with a full medical history and genome workup. And this is empowered by unique opportunities in Finland promoted by biobank laws that enable individuals to give a broad consent for research throughout the medical spectrum. And obviously has both new clinic-based recruitment and legacy epidemiology collections that I'll cite in a moment. To give you a sense where the FinnGen project is, this gives some updated numbers from the project. More than 220,000 individuals are already fully genotyped with registry linkage from 50 years of health registry data in Finland integrated into analysis, and frequent updates to the project as well as public releases of summary statistics are occurring on a regular schedule. Several features of FinGen make it uniquely well-suited to PRS evaluation and exploration. One is that there's the inclusion of more than 50,000 individuals who were recruited as part of healthy adult population cohorts more than two decades ago in many cases. And the other is the cross-linkage of registry information through throughout the Finnish medical system includes not only hospitalization and outpatient clinic diagnoses, but also information on prescription medication purchases for the last 25 years. So to the question at hand, are polygenic risk scores ready for clinical use today, and in what ways? Well, one answer, to be perfectly honest, is that there are still some very substantial challenges to the implementation of PRS in clinical settings that need to be addressed before we can be confident to move ahead in this space. And a few of them are listed on this slide. I mean, there are, you know, quite many more. I'll go into a small amount of details. I'm sure you've, you've now heard from Mike on the challenges and opportunities in bringing together clinical risk factors as well as polygenic risk scores. And while obviously there's a lot still to come with algorithm development, that's not so much a challenge, except that it is with respect to introducing a validated pipeline into clinical activities if the state-of-the-art is constantly changing. So a number of papers in the very recent past have started to address this question of how to utilize polygenic risk scores in conjunction with clinical risk scores or clinical risk factors that are measurable and used already. And in Mike's work in 2018 and newer paper from Nina Mars and Samuli Ripati, you can look at the results and you can either be a half-full or half-empty person because realistically there is a statistically meaningful improvement in prediction when you add genetic scores on top of clinical risk scores. Yet in an absolute sense, this is a very modest increase in value, and how to then quantify that and take that forward into, you know, demonstrating clinical utility of the polygenic risk score component continues to be a challenge and a great point of discussion around bars and clinics everywhere. That very modest improvement in overall prediction may mask Some instances, however, in which polygenic risk scores could be more useful and have an outsized contribution to risk assessment, and those are highlighted in Nina and Samuli's paper here, which flagged that, of course, those individuals with early onset disease actually had quite a more substantial contribution from polygenic risk scores relative to many of the clinical factors that themselves are either linked to age or go up with age. Perhaps the key challenge to clinical translation of polygenic risk scores was presented in this meeting last year by Alicia Martin, who described the challenge that emanates from the fact that genomics continues to have a huge diversity problem. She demonstrated this first several years ago in an analysis of 1000 Genomes data in an extensive simulation study study that demonstrated that genetic prediction accuracy would likely decay substantially with increasing genetic divergence between the original GWAS discovery population and populations of other ancestries, and that this doesn't have any mysterious explanation, but simply the drift, the frequency differences, the LD differences, and so forth between variants across populations means that there's a huge discovery bias towards more common variants in populations that you conduct the GWAS in. Alicia, along with Masa Kanai, then has gone on in the past year to demonstrate empirically these predictions using data from the UK Biobank and the Biobank Japan, and showing that indeed GWASes for a number of quantitative traits calculated in European population would predict far less well in non-European populations than they would in other European populations.. And as expected from the earlier simulation study paper, indeed this is a feature simply of what your discovery cohort is determines how well it's going to predict in second cohorts as a function of ancestry. So if the study, original study, is trained in UK or European sample, it's going to predict in new European samples better than in the Biobank Japan. And vice versa. The same— the reverse is true. If you train on a subset of Biobank Japan, you then predict the remainder of Biobank Japan better than the UK Biobank samples. So indeed, what this work has taught us is that there are some substantial steps that we need to undertake before we have a truly representative and equitable PRS in order to bring it into clinical activity, and these will include both methods development, but And importantly, and unachievable any other way, a much larger investment in non-European genome-wide association studies is an absolute prerequisite for effecting a more equitable outcome in polygenic risk score translation. So for balance, though, I don't want to simply highlight the challenges that we have in front of us, but there are also, I think, some very real near-term opportunities that will bring a more positive spin to this talk, perhaps. My most important message of this symposium is to make sure that I tell you all to go to Nina's talk tomorrow and take a look at her medRxiv paper, which has a remarkably comprehensive look at various uses of polygenic risk scores in breast cancer. And in— over and above the simple association of polygenic risk to risk of breast is very easily demonstrated. Nina demonstrates a variety of secondary uses and prediction of outcomes with respect to breast cancer that she will describe in this talk tomorrow. A few of her findings I'll highlight here very briefly just to whet the appetite a bit. Nina is using in this case data from the FinnGen project that I mentioned earlier, and one unique opportunity about working in a bottleneck population such as Finland, is that not only do common variants come easily out of a GWAS conducted at scale with genotyping and imputation, but also rare and high impact variants, for example frameshifts in PALB2 and CHEK2, that would not be normally at this frequency in the general population are observed, and that gives us certain opportunities. One of these derives from the fact that with single easily typed variants representing the vast majority of mutations in a gene such as PALB2, one can very inexpensively survey entire populations, or in this case several hundred thousand individuals from Finland, for this particular variant and get a very, very precise view of its risk to individuals over the lifespan. What that further allows me to do is to then stratify the effect of the PALB2 mutation in the population alongside the polygenic risk score. And what you can see is that individuals who have both the PALB2 mutation and a high polygenic risk score have a truly exceptionally strong risk for breast cancer, whereas those with a low polygenic risk score highlighted in blue down near the bottom are nearly returned to the population average risk for disease. She's also able to perform a very compelling analysis demonstrating the integration of self-reported family history or known family history from clinical records and polygenic risk score, and demonstrating that indeed individuals with a positive family history but a low polygenic risk score are not at demonstrably increased risk, while those with a high polygenic risk score and family history positive have a very, very strong risk of disease. And finally, one of the clearest and most enjoyable, for me anyway, conclusions of this work demonstrates that the polygenic risk score component and the rare mutation component, in this case being a carrier of any one of 7 mutations in BRCA1, 2, PALB2, or CHEK2, is completely and utterly independent and additive with respect to risk. Those individuals with a low PRS, whether they are mutation positive or negative, are attenuated downward. Those with a high PRS are attenuated upward from whatever the baseline is of carrying the mutation or not carrying the mutation. So my prediction is that in the very near future, we will begin first perhaps in family screening of first-degree relatives of new cases, and then soon in population regimes, information from both common and rare variation combined will be used in risk assessment. And this utilization of genetic risk scores is something which is almost guaranteed to save lives and healthcare costs at the same time. So finally, I just close with these few thoughts in summary and look forward to the discussion afterwards. Clinical applications of PRS are as yet unproven. There are decidedly significant challenges regarding integration with standard clinical practice and a strong ancestry bias in today's polygenic risk scores. However, applications towards cancer screening seem to me anyway the most obvious and impactful near-term opportunity that we have, and I expect to see those in place in the not too distant future. And then ultimately, Something I didn't discuss but, but also think is impactful is the use of PRS in defining clinical trial populations that have a high prospective conversion rate to disease could be another very real near-term win for biomedical research. Thanks everyone for your attention, and thanks especially to these folks who I've been privileged enough to work with on these issues and who were kind enough to share ideas and slides for for this presentation.

Thank you very much, Dr. Daly, for this very nice introduction of the PRS scores and their opportunities. People can already start asking questions on the chat, but we will not answer them now. We'll do that at the end after all the four presentations. We now move to the talk of Professor Krista Fischer, who studied applied mathematics in Tartu, did a master's in biostatistics in Belgium, returning to Estonia where she's now a professor of mathematical statistics and the senior statistician in biostatistics at the University of Tartu Institute of Genomics. As you can imagine from her background in mathematics and biostatistics, she's very interested in the technical side of polygenic risk scores and really understanding how they function and how you can improve them. And the title of her talk is Improving polygenic risk scores, but also improving our knowledge about them. Dr. Fischer, the floor is yours.