Scientist Stories: Amit Khera, Improving Polygenic Risk Scores for Clinical Applications

Good afternoon. My name is Amit Khera. I'm a cardiologist and physician at Massachusetts General Hospital and also lead a research group at the Center for Genomic Medicine and the Broad Institute. This afternoon, I'm delighted to discuss with you polygenic risk scores and their use in the clinic. The rise of genome interpretation provides an opportunity to really deliver on the promise of genomic medicine. And at its core, we aim to identify high-risk individuals actually disclose the risk to participants, and ultimately intervene to decrease the risk of developing disease. In order to identify high-risk individuals, it's important to understand how might DNA predispose to a complex disease. And today I'll talk mostly about myocardial infarction or heart attack. There are really two conceptual models. The first is a monogenic model where there's a single mutation and that has a large effect by disrupting a very specific pathway. But the second model that you've heard about is a polygenic model. Here it's not one mutation, but many variants each of which has an individually small effect, but in aggregate the risk may be more substantial. And importantly, these many variants actually tag many different pathways. We recently set out to answer a very simple question: how might we interpret the genomes for risk for myocardial infarction are according to monogenic and polygenic models among 100 patients coming to the hospital with a heart attack. In order to do this, we designed a whole genome sequencing study consisting of over 2,000 individuals presenting to a United States hospital with heart attack at an early age, so age less than 55 years, and we compared their whole genome sequences to almost 4,000 multiethnic controls who had actually remained free of cardiovascular disease. We first used the traditional approach and said, in what percent of these individuals who had a myocardial infarction can we find a monogenic familial hypercholesterolemia mutation? It turns out we could find one in about 1.7% of cases, and this is highly consistent with previous results that we and others have published. On average, these mutations were associated with about a 3.8-fold increase in risk. Now, as some of you know, these mutations actually prevent the clearance of LDL or bad cholesterol from— so here I'm showing you on the y-axis individuals' LDL cholesterol levels and the y-axis stratified whether they did or did not have one of these monogenic mutations. And what you can see is that although there's some overlap, the average LDL cholesterol was 202 for for those who were carriers as compared to 124 for those who did not. So the take-home point is that carriers of these mutations can be distinguished by high LDL cholesterol, albeit imperfectly. We then said, beyond these 1.7% of individuals, can we actually identify additional high-risk individuals instead using a polygenic risk model? And here I co-led the work with Mark Chaffin, a very talented computational biologist in our group, and Sekhar Kathiresan, who was my postdoc mentor at the time. We set out to identify and characterize a genome-wide polygenic score for myocardial infarction. We first started with the results from a previously published genome-wide association study in which each row is a variant, any of 6.6 million variants, and each— and there's a column for the effect size. Was it more common in any of 60,000 cases were more common in the controls. Some variants increased risk and others decreased risk, but on average, the vast majority had a very modest effect on risk for myocardial infarction. We then sought to turn this list into a single number that would reflect inherited susceptibility. And here we used a variety of different computational algorithms that have been proposed by others, 4 different strategies, and we then took these 4 scores into a validation dataset. Here we used 124,000 individuals from the United Kingdom Biobank. Now it turned out that the best score was actually based on the LDpred computational algorithm, and this score actually included information from all 6.6 million variants. We then said, okay, this is our best performing score. Let's then go on and test it in additional independent[No speech] We next went on to test the genome-wide polygenic score in the remaining 300,000 additional European ancestry individuals of the UK Biobank. Here I'm showing you that the score was normally distributed in the population. We then went on to study the relationship of this score to risk for heart attack. And I'm showing you here on the x-axis percentile of the polygenic score.

So each dot actually represents 3,000 individuals.

And on the y-axis is the prevalence of myocardial infarction by the time individuals reach middle age. And what you can see is that there's really a striking, more than tenfold gradient in risk across individuals based on the 0th to 99th percentile. And this is a score that's available from the time of birth.

Now, the traditional approach has actually been to compare carriers to non-carriers for monogenic mutations.

So next, we sought to mimic that approach by arbitrarily saying, well, what if we call the top 5% of polygenic score distribution carriers of a high polygenic risk and the remaining 95% non-carriers? And what we observed was quite striking. These top 5% individuals actually also had greater than threefold risk, and this is risk that's actually equivalent to monogenic familial hypercholesterolemia mutations. Again, that was done in the UK Biobank. We then brought it back to our patients and the whole genome sequencing study. These again were patients who were enrolled in the United States with a myocardial infarction at a young age. And it wasn't 1.7% of patients, but actually 17% of individuals who actually inherited a high polygenic score.

And despite being 10 times as common as monogenic mutations, the risk conferred was almost identical.

Odds ratio was 3.7. Now, as I mentioned, a key advantage of this study was actually that it was a multiethnic study. And we were able to show that a high polygenic score increases risk in all racial and ethnic groups, but the magnitude of the risk conferred was highest in white individuals. So here you can see that the odds ratio conferred was over 5, actually, for white individuals, but between 2 and 3.4 for Black, Hispanic, and Asian individuals respectively. Now, this is actually a predictable consequence of having the vast amount of the data that went in to train the model was from individuals of European ancestry. And so there are a number of issues related to use of non-causal variants in the polygenic score that don't associate with the causal variants based on varying linkage disequilibrium patterns across races, very different minor allele frequencies across different races., and potentially different genetic architecture, meaning gene-gene interactions or potentially a particularly important impact of environmental exposures in various ethnic groups. And so these are aspects that we're hard at work, again, in really trying to level the playing field, so to speak, across these racial groups. So again, the take-home point is that the polygenic score improves risk prediction in all ethnicities, but at present,, it does so with the most pronounced effect size in white individuals.

Here I'm showing you a plot similar to before where the x-axis is individuals stratified by high polygenic score, yes or no, and the y-axis is again the LDL cholesterol. And what I'm showing you is that on average the carriers had a slightly higher LDL cholesterol, 132 versus 124 milligrams per deciliter, but the distributions were largely overlapping. So this biomarker cannot be used to reliably distinguish high polygenic score individuals from the remainder. It turns out it's actually not just LDL cholesterol, but actually any clinical risk factor used in clinical practice actually are inadequate to pinpoint these high polygenic score individuals. So for example, family history is present in 44% of the top 5% of the distribution versus 35% of the remainder. Hypertension 32% versus 28%, hypercholesterolemia 20% versus 13%, and so forth. So many of these risk factors, which as you would expect are causal risk factors for disease, are somewhat enriched for those in the high polygenic score group. None of them are enough, um, enough that you would be able to reliably identify them in my clinical practice. So why is that? Well, if you look at actually the variants that are most strongly weighted in the polygenic score, I've about 20% of them or so relate to LDL or other lipid measurements. Another 20 or 30% actually relate to other pathways that we know are important but are not measured in routine clinical practice currently. Things like inflammation, cellular proliferation, vascular remodeling, vascular tone, nitric oxide signaling. And actually the remaining 50%, we actually aren't sure what the mechanism is. So I think there's a lot of important work to be done in furthering that biologic and mechanistic understanding as well.

We then said, well, is DNA destiny?

To what extent might a healthy lifestyle actually be able to offset high polygenic risk? And we had a very basic rubric, which is defining a healthy lifestyle score according to the American Heart Association criteria. No current smoking, maintaining a normal body mass index, regular exercise, and a healthy diet. And we said if you had at least 3 of these factors, you had a favorable lifestyle. 0 to 1 was unfavorable, and those who had 2 were deemed intermediate. Now here I'm showing you only the high polygenic score individuals, and what we observed is that a favorable lifestyle was actually linked to about a 50% lower risk. So it was about 11% in those who had a high polygenic score but an unfavorable lifestyle versus only 5% in in those who had a favorable lifestyle. And what this meant was that those who had a high polygenic risk could basically get the risk down to the population average with a favorable lifestyle. But of course, they can never do as well as those who had both good genes and a favorable lifestyle. In summary, the genome-wide polygenic score for myocardial infarction identifies 17% of the population of individuals presenting with a heart attack., with an odds ratio of 3.7. These individuals are currently undiagnosed within clinical practice, but we know that a healthy lifestyle and actually a second line of evidence related to cholesterol-lowering medicines are particularly helpful in offsetting the high polygenic risk observed in these individuals.

So this analysis was based on cross-sectional studies, but we then went on to ask the question, can the polygenic score actually predict future risk, especially within individuals binned according to current clinical risk strata? Showing on the X-axis individuals stratified according to current clinical risk strata, and on the Y-axis the 10-year risk of developing a future coronary event. And what you can see is that regardless of clinical risk strata, there is a significant gradient according to the polygenic score, suggesting that it really can refine risk really across the clinical risk spectrum. Importantly, we were able to confirm nearly identical results in the Malmö Diet and Cancer Study, which is a prospective epidemiologic study based in Sweden.

How might this all come together? Specifically, can polygenic background actually modify the penetrance of a—

and here again, I'm showing you a basic framework, monogenic variants, single mutations of We went on to analyze a case-control dataset from the UK Biobank, 6,400 cases and 6,400 controls. Consistent with previous studies, a familial hypercholesterolemia monogenic mutation was associated with about a tripling of risk. Odds ratio was 3.2.

We then stratified these individuals actually by the risk of, their polygenic background.

And the y-axis, I'm showing you the odds ratio for myocardial infarction.

And what you observe with those with a monogenic mutation, seen here in blue, but a low polygenic score, their risk was actually comparable to the population average, odds ratio of only 1.3. We then went on to generalize this to over 48,000 UK Biobank participants and The traditional approach, again, is saying that a small percentage of population actually carries a familial hypercholesterolemia mutation. They may have 41% risk, for example, of having a heart attack by age 75, versus the non-carriers, the vast majority of individuals, have a risk of around 13%. But I'm now showing you what happens when we layer in the importance of the polygenic background. The x-axis, the percentile, the polygenic score. And here we really see a striking gradient of risk even among those with a monogenic mutation. And I think in many ways this explains what we often have called incomplete penetrance, meaning we've known for a long time that not everybody with a monogenic mutation for risk of heart attack, breast cancer, colon cancer, and others goes on to develop disease. And here we provide evidence that the polygenic background is one important driver of this incomplete penetrance. So what does this mean for clinical practice, and how can we actually translate this into improved care? Well, we're just starting to think about that, but at Mass General Hospital, I recently helped launch the Preventive Genomics Clinic, where we assembled a team of genetic counselors, physicians, laboratory geneticists to really think about how we move this. So the examples of types of patients we're seeing so far are guiding asymptomatic individuals who actually have a pathogenic variant, discussing genetic testing and which genetic testing might be appropriate for those with symptoms or family history,, and then interpreting the genome for those who are curious or concerned. We've seen a few patients on a research basis for polygenic score assessment, but are hoping to launch that as a clinical test very soon and are very excited to have the clinical infrastructure to do so already in place. In conclusion, polygenic scores enable identification of high-risk individuals before the onset of disease. In fact, they're available from the time of birth. Polygenic background improves risk estimates within clinical risk strata, even for those with monogenic mutations. And the integration of genomics into routine practice will require new clinical infrastructure and new evidence. Thank you very much and look forward to your questions.

Thank you very much, Dr. Kira, for a very interesting talk and showing us the promise of all of this in the clinic. We now move to our interactive debate. And I hope that you can now see all of the 4 speakers on your screen together with Dr. Manolio and me. So I've been writing already more than 15 questions down that have come up in the chat during the presentation. So the nice thing is that we can already fill the discussion of half an hour. So I will quickly start, and Dr. Manolio, if you have questions in between, let me know. So I'll quickly kick off, and my idea was to start with a question for everybody. So I'll start with Dr. Daly. A question from Idris Mohammed. He asks whether you think PRS scores can incorporate other levels of genetic information such as, or methylation scores to predict the phenotype.

So I guess the question is whether other scores in addition to polygenic scores could help predict the phenotype. It's an interesting and untested thesis, I would say. I mean, I think what has become quite clear is that many different polygenic risk scores can be combined to improve prediction. Although this, you know, has often been demonstrated, I think, you know, in cases where the data has been underutilized. So as you can well imagine, and as from Dr. Kehra's presentation a moment ago shows, many of the known risk factors for cardiovascular endpoints are themselves highly heritable and have polygenic risk scores with them. Adding, you know, simply using the polygenic risk scores together is not an optimal treatment of that. So one can imagine that with many contributing predictors, weighting of the variants, considering all of those independent scores or somewhat independent scores, is an area of methodological development that I know many are looking into now. So I would expect certainly that other components than straight risk endpoint and intermediate predictors can contribute, but as yet I haven't seen evidence of that.

Okay, thank you very much. I'll move quickly on because we have lots of questions to go through. So for Dr. Fisher, a first question: how do you think we should avoid overfitting results from trying different models on the same datasets?

It's a very good question and a good point. I think overfitting could be a serious problem, and that's why I would also repeat the recommendation to use always like different training sets and validation sets. Of course, even then it could be a problem sometimes. So if you really try too many different scores and then pick the ones that replicate in validation set. But still, I think careful use of non-overlapping cohorts for different steps and also the benchmarking ideas proposed by Mike with clear standards could be helpful.

Yeah, this question was actually by Steve Lincoln, just so you know. And I move to Dr. Hinoje. Question from Angus Clark: given important gene-gene interactions buried within PRS scores, how can these be dissected to increase the power of calculated PRS?

Well, it's a great question, and it really depends on the disease or trait. The evidence for, you know, strong or, you know, a high sort of epistatic component of a complex disease is is fairly light on the ground. So I suspect that for a lot of these common traits and diseases that it won't be a huge contributor, but that there might be some small amount of gains by trying to capture that. We've seen instances of that in autoimmune diseases that have a large MHC component. There are interactions there., and several independent groups have shown that, but, you know, it's, um, it's probably not going to be a huge, uh, you know, prediction, uh, source of additional prediction.

Good. I see that we have Dr. Manolio back. Okay, good. Then I go to Dr. Kira for a first opening shot question. Um, clearly there is a bias in European ancestry in all of the scores in the big datasets. And a question from Johanna van Berg is, what's our responsibility to increase diversity and how are you contributing to this? And also perhaps, you know, what sample sizes are needed to reach comparable scores is a question from Michelle, which I wanted to integrate in this.

Yeah, thanks so much. You really hit on one of the most important issues for moving forward, which is that polygenic scores on average do predict risk across all ethnicities because on average, the genetic variants affect in one ethnicity is correlated with its effect in another ethnicity. But really where we see the drop-off is when you look across many polygenic scores and there we've reproducibly seen a drop-off. So I think we as a genetics community have a really important responsibility in one, just to improve the quantity of genetic data in non-European ancestries. And that's something that I think we're all very highly motivated to do. But then secondly, I think in the interim, we and a number of other groups are really looking to solve that using other approaches. So one example approach is using fine-map causal variants, which are more likely to transfer effectively. Use functional genomic annotations in the weightings, which again are biologic rather than factors that vary across different ancestries. And similarly, actually think about using local ancestry, which you can quantify using a different admixture, and thereby take the person's ancestry-weighted effect size and include that in the polygenic score. So I am encouraged overall by the early results of some of those studies. I do think that ideally we'd have representative, globally representative genetic data. I think we hope to get there in coming years.

In the interim, I do think that these, these solutions will help level the playing field. Thank you very much. Dr. Manolio, I'm just looking at you. We are, you know, a little bit apart from each other here in Newcastle and in America.

Any question from your side? We are— I'm having difficulty seeing the chat, so my apologies. That's why I'm trying to log in and log back out and that sort of thing. But one question I did want to ask about had to do with the effect size, probably for Dr. Zenouye and Kara, either of you. In terms of the differences in effect sizes across ethnicities, is there any difference in that if one looks at the very highest risk individuals? So say you took the top 1% of people at risk. Is there more predictive ability in different ancestries with that very rarefied group?

Amit, I'm happy for you to take that. We haven't looked specifically at the top sort of end percentiles yet.

Thanks for your excellent question, Dr. Manolio. We've seen actually that the pattern has been relatively consistent regardless of the endpoint, whether you look at hazard ratio for standard deviation, for example, the data I presented was top 5%. Top 1%, we've really seen a similar pattern. I will say that one of the things that the top 1% versus bottom 99% relative risk estimate doesn't take into account is actually difference in absolute risk estimates. So for example, a twofold gradient in an African-American population who's at higher risk, for example, hypertension, could actually be more pronounced actually than a twofold risk in European individuals where the risk of hypertension is much less. So I think there's an important body of work to think not only about relative risk estimates in the tail, but also how that interplays with absolute risk estimates, which we know very importantly across the population.

Okay, thank you very much. I'll take a next question from Juliana for Dr. Daly. Do you have an explanation for the fact that while PRS scores explain disease risks the scores do not explain phenotypic variation among cases and has to do anything with sample size that we can solve this?

No, no, I don't think that's— actually, I don't think that's mysterious at all. I think what Dr. Khera showed is quite straightforwardly a consequence of there being a considerable risk to heart disease, as there is with many things. That do not directly come from the measurable intermediaries that are themselves quite heritable. I think what you didn't see was, of course, that there is a very strong, much more complete ability to assess polygenically cholesterol levels, BMI, and so forth, and that those in fact do also predict outcomes. And that's where— we get into further methodological developments that the outcomes combining the PRS will be more informative with respect to the ultimate outcome in the patient. So I think it's just a byproduct of the fact that there are many, many, many pathways, processes, and components to those scores based just on one endpoint.

Yeah. And I move to Dr. Fisher. Idris asks the question that probably many people have: which tools do you recommend for calculating PRS scores?

Please unmute. Yes, yes, it's difficult to answer. So which tools we have used? Precise, we have used Stelldipred, but again, it's no one-size-fits-all approach. So I can't really give a very good answer here, as I think it's computationally calculating a polygenic score is not very demanding. So you actually could even calculate the score on a laptop, but then you have to first retrieve the SNPs, so depending on the number of markers you use. So you don't need very sophisticated tools, but of course you You need to know, it's more important to know which method to use to get the score, I think, than the tools outside in many cases.

Yeah, okay. I move to Dr. Inouye. Dr. Jackie Beckman asks, can we calculate the impact of environmental exposures on the PRS score?

And somehow, you know, can we incorporate these kind of non-genetic effects into a PRS score? Yes, so it's not an easy question. I think, you know, the answer though is absolutely. I think that, you know, as our sample sizes get bigger, we will be able to better assess the extent to which environmental interactions might be able to improve the power of polygenic score or indeed to combine with the polygenic score to create something that has better utility. I mean, the end goal here is to better predict disease early. And those additional sources, whether they have effect modification properties on polygenic scores or the individual variants thereof or can be combined, I think is a major area of future research.

Okay, move to Dr. Kieran. Question from Arjan Mensenkamp in Nijmegen. He asks, for breast cancer, a PRS score has been developed. Is this now ready for the clinic, or what needs to be done to get it into the clinic?

Well, you know, a number of folks have developed, developed breast cancer polygenic scores, and I think Mark Daly actually highlighted cancer prevention as an important first opportunity. I would lump it together with actually cardiovascular disease as something that's similar, is getting close to being ready. And I think here, the way I think about it is for a woman who is a 40-year-old woman who's at average risk, the decision about whether to get a mammogram is actually quite controversial because there'll be a number of false positives for every cancer detected. And so this is something that I do think is, in many ways ready for primetime. And I think the key aspects will be, one, there's an education gap. People really need to understand— providers, genetic counselors— around what the meaning of a polygenic score is. There's an equity gap that I think a number of folks here have talked about how we might, how we might address it. And the last, I think, is an evidence gap. And so I think I don't want us to miss our opportunity as a genetics community to really start to implement these things in a very controlled way where we can really convince the world that there's actually utility. So between the education, equity, and evidence gap, I think we really can tackle this with breast cancer potentially being one of the first to do. And I don't think that's a 10-year process. I think that could very easily be a 3- to 5-year process, for example.

Okay.

Move along. Dr. Manolio, just step in if you want to do something, but I have Unfortunately, lots of questions, so I'm going to Dr. Bailey. A question from Dr. Andre Reiss asking, how does the bias of common versus rare variants in PRS score influence predictions, for example, in breast cancer?

Well, I mean, I don't know if it's a bias, but the two need to be combined, I think. You know, obviously a complete, as I tried to point out, assessment of risk needs to involve both common and rare variation. This obviously puts a greater burden on data collection. It also adds some methodological detail that needs to be attended because obviously rare variants have linkage disequilibrium to their surroundings, often quite extensive as in Finland. And I should add that when we do that with the PALB2 and CHEK2 variants, we remove move those regions for quite some distance to the left and right of those genes in calculating the polygenic score from externally, you know, generated common variant data. So there's some work to do with respect to this, but I think it's quite clear that with respect to, you know, something of the nature of, you know, common cancers, there are very strong and meaningful contributions from both that need to be combined, and that's the ultimate target now.

Perfect, thank you very much. A question for Dr. Fischer from Dr. Jakupović: should the predictive performance of the genetic risk score also, alone or combined with other risk factors for a specific trait, always be done by performing a receiver operating characteristics analysis and power calculation? I'm sorry for reading this up and making it a bit complex.

Should I read it again or did you get it? Yes, I got the point. So, I think that it's a very good question again. So what this— what are these ROC and AUC and C-statistics, these calculations should be done. It is difficult. I think the fact is, for instance, if you take type 2 diabetes, you may see that age and BMI alone are very strong predictors. And so you see an AUC of 0. 8 and more depending on the age range you have in your data, and then you see just maybe 2% improvement if you add PRS, and you may conclude that it doesn't add much. But then if you take 2 individuals of the same age and of the same BMI, and then you see a much stronger difference, that really polygenic risk score does make a difference. So it could be that just only looking at AUC could also be quite misleading, because you don't understand the real importance. That's why I also tried to reanalyze the table from this JAMA paper to show that there are different ways to look at the data. And I think we should all, people working on methodology, should clearly give a good thought how to show clearly what is the effect of polygenic score, and what do we gain by adding it to everything else. I'm not sure that AUC is only and the best indicator for that.

Thank you very much. Yes, I thought in that aspect, some of the recommendations by Dr. Inouye on the use of reporting standards and really reporting the underlying data and methods are probably extremely important to push this field in a good way, because yeah, people need to trust that what they get from one, let's say, person doing a score or a company doing the score is exactly the same. So that's probably very important. I have a question— sorry, I'm constantly saying them wrong. I have a question from— let me see, who was that? Oh, I'm sorry, from Anna. Yes, she asks, in the future, which clinician would be the most suitable to discuss a PRS score with a patient?

Healthy person?

Um, I think, uh, that, that's— so I'm a computational biologist, uh, I'm not an MD, no medical training whatsoever, uh, don't let me near any patients. Um, certainly, I mean, I think that's a totally open question. It'll depend on the specific circumstances, um, and, you know, not just the disease, but even potentially what age you know, that information becomes available. So, you know, these are all things that need to be worked out and indeed for me are way above my pay grade.

Well, you know, you may be working together with clinicians in providing these scores.

Who knows?

Oh, no, absolutely. Yeah, I'll echo that. Yeah. Then a question for you from Sarah Pickett. She says, PRS scores are using thousands or millions of SNPs now. How likely are they to suffer from regression dilution bias as they will have more false positives?

Mike?

Sorry, I'm continuing with you now.

Sorry, could you just repeat that real quick? Question from Sarah Pickett asking that PRS scores now use thousands to millions of SNPs and they are more likely to suffer from regression dilution bias as they will have more false positives.

Do you agree on that? Look, it depends on how the polygenic score is constructed. If the architecture of the disease is indeed highly polygenic, infinitesimal, et cetera, then there's actually quite a lot of SNPs that you include, that you could include into the millions, which where their signal, their predictive information outweighs the noise that they introduce into the score. So we've seen that for various diseases, whether it's coronary artery disease or stroke or type 2 diabetes, et cetera.

It's—

if you do, though, let's say that you— if you have an autoimmune disease and you have something sort of smaller, smaller number of variants that are associated, then if you start adding in more and more SNPs, you do start introducing more and more noise into the score, and then you start suffering regression dilution. So it depends, I suppose. I guess the point here, though, is that those sorts of things do need to be assessed and benchmarked. So that you know that the score that you're looking at might have regression dilution bias in it. Currently, those things aren't assessed as nearly as commonly as they should be.

Okay, we go to a question for Dr. Kira from Dr. Luis Perez-Girado.

What's the correlation of PRS with risk derived from family history? So, thanks again for a great question.

So family history is actually a combination of genetic risk, shared genetics, and shared environment.

And so, what we see is that there's a very modest correlation between polygenic score and family history across a whole range of diseases, number one, and number two, they actually contribute additively in some data that was shown today and others that we have unpublished, that there's an important additive contribution of family history and polygenic scores when it comes to a whole range of diseases. And so really from my perspective, I think polygenic scores are a new factor that's measurable very early in life, and in so doing can actually give you a 40-year head start on some various prevention efforts. And, you know, family history, especially in, you know, in my clinic, a parent may not have a heart attack, for example, typically until they're in their 70s or 80s. So for a standard 30, 40-year-old patient, they may not have a family history despite a very high genetic predisposition.

Right. And then another question for you from [INAUDIBLE], I think is the way. He asks, or she asks, could pathway-specific scores be of interest in relation to clinical subphenotypes or patients benefiting from specific treatments?

Yeah, I think that's a really important and relatively unanswered question in the field. We're starting to think about this really using a variety of clustering approaches, both at the variant level as well as at the individual level. Conceptually, it's incredibly attractive because, for example, for cardiovascular disease, we now have anti-inflammatory therapies, anti-cholesterol therapies, blood pressure therapies, and so forth. So, I think to the extent we could deconvolute a set of polygenic variants or clinical risk factors to find those that would benefit the most from that, I think that's a really important clinical need and something that we're actively researching, and I hope others do as well.

Of course, I've picked— well, I picked a lot of questions already. Dr. Fischer, a question for you: how to define the best score for clinical use?

Any further tips for us? Oh yeah, well, I think it starts from really defining the outcome. What What is the phenotype variable you want to predict? How— what do you want to gain? Do you want to, like for breast cancer, you want to get the more, maybe more efficient screening strategy? For type 2 diabetes, probably you would like to find people who would need some better counseling on lifestyle and so on. And then, yes, so how much better to score should be. And really, I think it starts from getting the score using proper techniques and avoiding biases that really gives the best correlation with the phenotype. But then also deciding how to use the score, whether you use some quantiles, whether you divide people to categories, high risk and low risk, and how you combine the score with other clinical risk factors and so on. So it's a complex question and not an easy or short answer there. Should we include PRS as categorical variable in prediction models instead of continuous variable to improve prediction? Well, again, interesting question. So I think, first of all, if you want to assess the effect of PRS, I don't see any any reason to use it as categorical variable. But then if you want to move into risk prediction, sometimes you need to make yes/no decisions whether individual is stratified to a high-risk category or not. And there you need categories, and there you would need to decide what does it mean that you stratify individual to this category. So I think both, again, that you may start with a continuous predictor and combine it then with other predictors, and then you may still end up with categories depending on your question of interest.

Yeah.

Dr. Manolio, do you have further questions? Yes, I'm wondering, did you address— I had to come in and out a couple of times. Did you address the issue about insurance and implications from the polygenic risk score? No, I didn't. So I'm not sure who would be the best suited to maybe the— among the 4 of you, if someone is brave enough to speak up to that.

Well, I would say that that's something we're starting to think about, at least in our preventive genomics clinic. Somebody mentioned who's the best doctor to disclose it. I don't think there's going to be any one person. I do think we're really going to need a team-based approach, both in terms of education and in terms of designing care pathways. In terms of the insurance question, my guess, my supposition is that we'll counsel patients very similarly to the way we do now, which is to talk about, at least in the US, there's protection by the federal laws for at least for health insurance, but this may not be the case for actually long-term disability or life insurance. And I think the one difference that compared to traditional approach with monogenic variants is that monogenic variants, 99% of people won't have one for any given disease. Here everyone has one, and it can be addressed as a continuous trait, as Dr. Fisher mentioned. And so my guess is we'll piggyback on discussions that genetic counselors have been having for some time based on monogenic variants, but I do think there's some innovation in sort of effective disclosure of some of those risks amidst, frankly, a lot of regulatory uncertainty.

But the future may hold. All right, I think we've just identified a topic that we can use for one of the next Building Bridges sessions. It's all about genetics and insurance, which of course is another important thing to discuss. I would like to thank all the four speakers for their presentations as well as for their very open answering of all the questions. And I'd like to thank the audience for a lot of questions. And I would like to give the floor to Dr.

Manolio to say something about the next Building Bridges session coming up.

Great.

Thank you, Dr. Veltman. So we are looking forward to meeting with all of you either in person or virtually in October for the American Society of Human Genetics. As you've seen from a message from our president, we are moving forward with planning it on both tracks, and we'll make a decision soon. Would also note that the abstract deadline is June 11th, which is coming up. Our Building Bridges session next time will be on differing guidelines across the seas, which provide opportunities for research and answering what some of those differences might be. So with that, Dr.

Veltman, I'll turn it back to you. Thank you very much. That leaves me just to thank everybody again and wish everybody a very good rest of the conference.

So thank you very much.

Thanks.

Bye-bye.