Scientist Stories: Valentina Escott-Price, Polygenic Risks in Disorders of Aging

My name is Valentina Eskett-Price. I'm a professor in biostatistics and bioinformatics working in the Division of Neuroscience and Mental Health at Cardiff University and in the UK Dementia Research Institute. Polygenic risk scores are now widely used by the genetic research community for assessing the disease risk, assessing its prediction accuracy, and for identifying actual individuals at high and low disease risk. Today, I will talk about specifics and challenges of polygenic risk score calculation and interpretation in disorders associated with older age. Dementia is defined as a set of symptoms associated with changes in brain function that interfere with the ability to do everyday activities. Alzheimer's disease is the most common cause of dementia and accounts for about 50-75% of all dementia cases. There are other diseases associated with dementia: vascular dementia, Lewy body dementia, frontal temporal dementia, and in fact, rare diseases like Huntington disease, which can lead to develop dementia like symptoms. So I will be focusing today on Alzheimer's disease because, as I said, it's most common, and there is good, powerful genome-wide association studies being conducted for Alzheimer's disease. Alzheimer's disease is a progressive condition, which means that clinical features develop gradually over many years before diagnosis. And the ability to predict disease risk before disease onset is really of great importance for studying people for clinical trials, stratifying people for clinical trials, for selection of candidates for functional experimental studies. However, there is a number of challenges of studying disorders of aging. Focusing again on Alzheimer's disease, firstly, 30% of clinically diagnosed AD cases have no neuropathological or biomarker characteristics. In addition, controls. If the controls are younger than cases, which is quite common case now when we use population-based controls, 30% will become cases given time. In addition, 35% of lifetime risk of dementia is modifiable. For example, nutrition, good healthcare, managing vascular aspects, education may modify the lifetime risk of dementia and delay the age of onset. This potentially has led to decreased incidence of dementia over the last decades. So that actually adds to the problem with controls if they are not age-matched, even if they are age-matched. With the availability of large population samples, like for example UK Biobank, the use of population-based controls become a common practice to boost the statistical power. And for other common genetic disorders like neurodevelopmental disorders like schizophrenia, population controls are okay to use as the disease is rare, about 2% prevalence, and also manifests early in life. We know that penetrance of the phenotype is mostly complete for schizophrenia at about 40 years of age. While, for example, for Alzheimer's disease, even at 80, years old, there are still individuals at risk who have not yet developed Alzheimer's disease. Before moving on to the polygenic risk score, let's briefly discuss heritability of Alzheimer's disease. As I mentioned before, 35% are modifiable or environmental, and between 58% and 79% accounts for genetic heritability. This is a broad sense heritability estimated in twin studies. Looking now at only genetic part. So there is less than 1% heritability attributed to familial early onset Alzheimer's disease. APOE is a very interesting locus or gene. APOE e4 allele is a strong common risk factor for Alzheimer's disease, known already for decades. Also associated with early onset and in fact to other conditions, and consequently with shorter lifespan. Individuals carrying APOE-E33 alleles, A3 is a neutral allele, still develop Alzheimer's disease. What we know about APOE locus, it accounts between 5-9% of heritability. We have also other genome-wide significant loci, which accounts for 3 to 8% heritability, and all the rest we have for other SNPs, which were assessed by genome-wide association studies. So SNP-based heritability is the dark green color, corresponds to SNP-based heritability, narrow sense heritability. And also there is about 14% of other genetic variation. Which are attributed to rare variants, to potential interaction or epistatic effect between SNPs, copy number variants, and everything else which we have not discovered and accounted for yet. Nevertheless, looking at genome-wide association studies for Alzheimer's disease, the first largest genome-wide association study combining 74,000 individuals was published in 2014. And what we did after that study was published, we looked at SNPs not only at genome-wide significance level, but above the genome-wide significance level. So what we did, using GWAS summary statistics, we split SNPs into bins defined by the p-values. For example, the first line here talks about SNPs with with p-values between, well, 0 and 10^-7. It was 1,200 SNPs. Then based upon LD between these SNPs, we identified the number of independent SNPs among these 1,200: 738. Under the null hypothesis, we also calculated the number of expected significant SNPs in this bin. Which was less than 1, and looked at the ratio between observed and expected SNPs, which was high for genome-wide significant SNPs and highly significant. We've done it also for the other SNP threshold bins, and what we observed is that the ratio remained greater than 1 up to the p-value threshold 0.5. Well, of course, we adjusted for genomic control prior to this analysis. So to check whether this effect is attributed to genome-wide associated loss of hearing, we have excluded SNPs and SNPs in OD with those which are genome-wide significant. So definitely in this bin, the number of SNPs dropped to 2 independent SNPs. So the expected one is 0.37 or smaller than 1. The ratio is 3.5, dropped, and P-value was non-significant. However, when we moved down, down the— or up with the P-value thresholds, again we observed exactly the same effects. The ratio estimated independent SNPs to expected was greater than 1 and highly statistically significant. This inspired us to look at the polygenic risk component. So before I move on, polygenic risk, as you know, is defined as accumulation of risk alleles which an individual individual i carries. Each SNP contribution is weighted by the SNP effect size obtained from an independent study. So what we would like to see, we know that polygenic risk scores are normally distributed, we would like to see a clear separation of polygenic risk score in cases, which are blue here, and controls. In reality, these two distributions are overlapping a lot. However, for Alzheimer's disease in our sample, we observed relatively good separation. They're overlapping but still separate. So, white are cases here and gray are controls. And that was achieved at p-value threshold for SNP inclusion 0.5. So, that's the table which demonstrates or accompanies the picture from the previous slide. So when we looked at E4 alleles only and predicted case-control status, the area under the curve for the accuracy of the prediction in our sample was about 68%. When we included E2 genotypes, so it's slightly increased. When we included 20 at that time genome-wide significant, known genome-wide significant SNPs, the accuracy increased to 72%. And then we started including SNPs at more liberal p-value thresholds, and the best accuracy has been achieved at 75% area under the curve. It was highly significant over and above APOE and genome-wide associated locus loci known at that time. We also looked at maximal prediction accuracy which can be achieved in our study. For that, the heritability, narrow sense heritability, has been estimated and published, 24%. Using also lifetime prevalence, 2%, we could estimate the maximal area under the curve which could be achieved to predict AD cases and controls. That's a calculation we have taken from Reytal paper 2010. What we observed looking at all SNPs capturing 24% of narrow sense heritability, we can achieve accuracy up to 82%. In the previous slide, I have shown you that in practice we achieved 75%. So we are relatively close to the maximum in our sample. When we looked at APOE region alone, the heritability captured by the APOE region is about 5%. And based upon the same parameters of 2% of eye prevalence, it's 65% accuracy maximum which we can achieve with APOE genotypes or genotypes from APOE locus., which was approximately the same as I've shown you before. When we include genome-wide significant loci, the accuracy only slightly increased. So polygenic risk score, which is capturing AD risk across the whole genome, does not work uniformly well in all datasets. For example, we have shown that in our study it worked for SNPs with AD association at a liberal p-value threshold, 0.5, Recently we have shown that 0.1 is probably the better one because the genome-wide association studies are now larger than the one which published in 2014. We also showed that it improves AD prediction over and above the genome-wide significant or suggestive SNPs. Whereas others have shown different p-value thresholds for SNP selection. In particular, people were thinking of including only genome-wide significant SNPs or SNPs which are significant at 10���� significance level. This could be explained potentially by specifics of the datasets, including age of the participants, clinical or pathological assessments of cases and controls, etc. So in this paper, which appeared in 2020, authors suggested, according to their findings, that potentially Alzheimer's disease is an oligogenic model, which means that includes only genome-wide significant SNPs or SNPs which has 10^-5, I just said. Having in mind that the cases and controls which we have calculated in this SPRS were age-matched in one study or past confirmed in another study, whereas in the most recent studies, the cases and controls were— well, controls were including population-based controls, younger and not screened, and in some studies, cases were defined via the family history, also well known as UK Biobank proxies. Which means that people were included as cases if their parents had any type of dementia, of course. So looking then at what potentially could go on, why we see these different results, I decided to look at the— to start with APOE allele frequencies depending on age. Perhaps age component here differences in age of cases and controls and between studies may explain this. So looking at a publication in 2011, we can see that APOE allele frequency goes down with age from about 18% to about 8%. Frequency of E2 allele pretty much remains the same, and the frequency of E33 goes up with age. Also, It is known that APOE4 allele associated with age at onset. 91% of homozygous E4E4 carriers will, on average, develop Alzheimer's disease in 68 years, when they're 68 years old. Non-carriers, on average, will develop the disease when they're 40 84 years old. So 16 years difference between age at onset of E4 carriers and non-carriers. Therefore, potentially aging should be or need to be accounted for when we calculate polygenic risk score and when we talk or take effect sizes from genome-wide association studies. So what we've done next, we had an independent sample of about 300 cases and 300 controls. 300 controls from ADNI and ADNP studies. That sample size is so small because we excluded all overlapping individuals with the summary statistics published by Kung et al. 2019. And we looked in the whole sample at distribution of allele frequency with age. So an x-axis is age. So the green one is E2 frequency, it doesn't change, similar as we've seen in the previous picture from the literature. E4 frequency goes down with again, similar to what we've seen before, and E3 frequencies goes up. Then we looked at cases and controls, E4 allele frequencies, which are in red, and also that oligogenic component, excluding APOE, and polygenic component excluding APOE. Then what we see in cases, the oligogenic component basically mimics the behavior of APOE allele frequencies, whereas polygenic component goes up with age. Whereas in controls, basically I would say that they don't really statistically different polygenic and oligogenic, and they don't change too much over time. From this picture, if we select sample young, you can see that polygenic risk score could substantially differ as if we select samples age much older, where polygenic risk score is higher in cases and lower in controls. Potentially, that could explain one of the explanations, one of many explanations, why we observe different results in different studies with different age distributions. How to account for it? What we've done in this smaller sample, 300 cases, 300 controls, we used just standard methodology, climbing pruning and thresholding approach and looked at polygenic risk score as it would be calculated typically. And then we looked at ApoE prediction accuracy, which was 70% to ALU C4+ H2. Then we included the genome-wide significant loci, and that remains the same accuracy. And then we started including more SNPs the accuracy went down, exactly what our colleagues observed when they calculate polygenic risk score as one score. When we excluded ApoE locus, the accuracy with genome-wide significance was about 56% and was increasing up to 0.1 P-value threshold for SNP selection or inclusion. This observation made us think that perhaps if we model two components separately, polygenic risk score excluding APOE and APOE genotypes as two variables for prediction, that should work. And it worked indeed. And that's in fact how we calculated it in our original paper without thinking much about age. At that time, it just simply reviewers were asking us to see if we observe significant enrichment or improvement of prediction accuracy when we include polygenic risk score over the APOE. It just, by the way, So again, starting with 70%, the overall accuracy of prediction of AD status in this small sample increased up to 74%. In fact, I have also managed to simulate this scenario just to confirm that our observations are correct. What I've done, I've simulated 10,000 cases and 10,000 controls. And from controls, assuming that they are population-based controls, I just estimated that 91% of E4 carriers will be cases at 68, 47, at age 84, looking at the previous studies, and I estimated about a third of them will be cases if they reach age 85. And then I have simulated the APOE with frequencies as in the real studies. Then I've taken effect sizes and allele frequencies for another 39 genome-wide significant SNPs, which were published in recent reviews by Andres et al. in 2020. And they simulated 10,000 independent SNPs with minor allele frequency and effect sizes matching the genome-wide association studies. I looked at the R-squared variance explained and prediction accuracy, basically the same but in different units on the y-axis. So APOE alone, the prediction accuracy was 64-65%. Oligogenic risk score including SNPs which are 10^-5 significant, would be slightly better, about 68 or so. Then PRS full, PRS calculated including APOE as a part of the PRS, as what we normally calculate, have shown decreased accuracy, and R squared is also decreased. You can see it better here. But when we calculated the prediction based upon two components, APOE plus PRS, no APOE separately, you can see the best prediction in terms of area under the curve and in terms of R-squared. So it is important to model, perhaps to have in mind the specific disease architecture and model the effects appropriately. The next, what we have looked at, are different methodologies to calculate polygenic risk scores. So far I was talking about standard pruning and thresholding, but we know that there are a number of other approaches to do so. So the first one, we include PRIScISE, which is a software implementation of simple pruning and thresholding method. So that is just when we've done it manually. The second bar is using PRIScISE. The next one, we used LDpred. It's a Bayesian approach which infers the posterior mean effect size of each SNP based on SNP effect sizes of correlated SNPs.. And in fact, we use here LDpred_infinitesimal model, which is optimized for polygenicity parameter. We used next one PRS-CS, again Bayesian approach, which uses a continuous shrinkage prior on SNP effects, and the LD estimates are taken from 10,000 genomes here, so we might lose some SNPs here when we look at the prediction accuracy. SBiasR, is just an approach which scales SNP effect sizes. This is BioSight. BioSight is a Bayesian approach which scales SNP effect sizes using Bayesian multiple regression. And then LDAC is another methodology which adjusts SNP effect sizes for local LD by reducing the contribution of SNPs in regions of high LD. So there are different methodologies available and we tested them all and we looked at the extremes, individuals at the extremes. Because although all approaches have shown similar accuracy, if you look at the AUC here, pretty much similar, we can always see that when we model AD, PRS-AD, which means that APOE is modeled separately in addition to PRS without AD, it predicts the best. In our sample, but nevertheless they all have shown similar accuracy. The next question is, do PRS variables— the next question is, do PRS calculated with different methodologies agree, at least at the extremes of the PRS distributions? So here we looked at the PRS distribution with Boolean thresholding we have taken extremes, negative and positive, and followed them up looking at different other methodologies. And in some individuals, I indeed remain the same, and this is a— I might tell you that it's the E44 carriers, they remain the same. But some individuals in fact end up in the middle of the distribution. So the conclusion here is that the choice of PRS calculation methodology may lead to identification of different sets of individuals with high and low extremes at the distribution curve. The next problem is age. Recently, two well-powered GWASes on longevity have clearly demonstrated that longevity is a genetic trait. We also know that the most popular approach to correct for differential age distribution in cases and controls is to adjust for age when running a SNP association analysis. However, as it's shown in this paper, Ashford et al., adjusting for heritable covariates can introduce bias in the effect size estimates. And it's important difference between environmental and demographic factors, which we use for the adjustment, and heritable human traits which we use for the adjustment. And that made me think about potential collider bias, inspired by this paper. They have showed that the inclusion of height as a covariate created a robust but biologically spurious association between sex and heat-associated variants. So I thought about this and simulated similar Simulated similar situation for AD. Consider the case when Alzheimer's disease and the SNP both associated with age, but Alzheimer's disease is not associated with SNP. That's my simulation design. I simulated 5,000 cases, let's say mean age 78, and 5,000 controls. I have varied age of the controls 74, 70, and 65, just drawn from a normal distribution.. And I simulated, as I said, SNP and age association. See, it's significant everywhere. Also, AD is associated with age, significant highly everywhere. And AD and SNP are not associated, so p-value here is 0.5 approximately. And when I calculated association of SNP to Alzheimer's disease, adjusting for age, when the age distribution was relatively similar, 78 and 74, a SNP has shown a spurious but not so strong association. And when the difference in cases and controls became higher, the association became stronger. I'm not saying that all SNPs are like this, but there might be SNPs in the genome-wide significant studies which show spurious association with Alzheimer's disease whilst being associated rather with age. So in conclusion, While increasing effect size of genome-wide association studies led to identification of more loci, it has also involved the use of non-ID cases, and even more so in proxy case series, and this is having a detrimental— it might have a detrimental effect on PRS prediction of cases. Second, allelic variation at the APOE locus impact survival by both altering the age of onset of AD, by increasing risk of other conditions, and also by the fact that the frequency of APOE-e4 in a population correspondingly goes down with age. Looking at different methodologies, comparing 6 PRS calculation methodologies, we conclude that the prediction accuracies are very similar. However, the individual scores may differ. Another conclusion is that we need to have a clear distinction. What do we calculate polygenic risk scores? What's the purpose of our study? So if a DPRS could be useful to identify individuals at high risk, the accuracy of the current methods perhaps is not as good to use it in clinic. And also, if it's used in clinic, when it's used in clinic, we should really carefully— should be really careful at explaining to the public what polygenic risk scores mean. In terms of using the polygenic risk score to identify extremes for biological studies, for functional studies, at the moment I prefer to use the simplest one, because the biologists ask me simple questions: which exactly SNPs and with which effect sizes contribute to the polygenic risk score? The Bayesian approaches, may, because it uses prior information and calculations, may mask, and it's not easily explainable, at least at the moment. And that is my presentation. I would like to acknowledge the funders: Dementia Research Institute, Dementia Platforms UK, Medical Research Council, Alzheimer's Research Society, and Alzheimer's Association, and JPND, a European funding body. And of course my team who inspired me to work on this topic and explain unexplainable effects.