Questions and Answers – Webinar 6
WCR webinars always plan plenty of time for participants to interact with guest speakers. Find out what advises and recommendations were provided in response to questions and concerns from the audience.
The answers below reflect the opinions and points of view of speakers.
JM: Answer by Dr James McKay
NC: Answer by Dr Nilanjan Chatterjee
LK: Answer by Dr Linda Kachuri
POLYGENIC RISK SCORES (PRSs) – CLINICAL APPLICATION
JM: Genome-wide association studies (GWAS) so far have been focused on identifying susceptibility alleles involved in susceptibility rather than cancer progression and stage. There are a number of studies under way looking at outcomes, but it is a challenge, since there are a lot of other factors involved (progression, stage).
NC: There are a couple of challenges. First, it is sometimes hard to get very large GWAS – on the order of tens of thousands of cases – with cases that have very well characterized outcome data, such as data on stage, grade, survival, etc. We need large studies to find risk scores that would be predictive of those outcomes.
Second, we knew that there was a genetic component to cancer risk; we had good estimates of heritability. We do not have analogous heritability studies for cancer outcomes. The question of how much the genetic component determines the outcome, versus treatment and other factors, remains uncertain. It is more challenging than developing the PRS for risk, but it is very important to do those studies and to figure that out.
LK: The data part has been a big limitation in terms of collecting appropriately large data sets, and also considering the differences in cancer treatment and analysing how those influence the outcomes in different settings. Another potential challenge is to disentangle genetic predictors of disease incidence from outcome. There is potential for bias, and that can also influence our findings.
NC: The ultimate goal is to use PRSs for risk stratification in the general population. While the PRS is becoming an increasingly powerful tool, there are barriers to implementation.
The limited performance of PRSs in non-European-ancestry populations is one of those barriers. We already have a lot of issues related to access to health care and screening for different ancestry groups. If we implement PRSs now in the general population, it could increase health inequities.
There is also a need for more education on how to communicate the risk in an effective way. This is very important and has the biggest impact in terms of intervention uptake.
We also need to better demonstrate the broad utility of PRSs. Once genetic data is collected, it can be used to calculate PRSs across a vast number of cancer types and health outcomes. We need to demonstrate how PRSs could be useful not only for individual diseases and cancers but also in terms of improving broader health indicators, such as overall mortality or disability-adjusted life years. These types of study are needed to convince insurance companies and health-care providers to invest in PRSs.
LK: The examples presented in the two lectures were examples of cancer types for which there are screening programmes that are implemented in many countries. We also need to think about all the other cancer types for which there are no formal screening or intervention programmes. Even if we could develop a well-performing PRS for those cancer types, what could be the relevant intervention? This is another example of a limit of PRSs in terms of clinical utility.
It is important to start deploying a clinical product or practice only when we feel that this product or practice will perform equally well in all groups.
NC: There is a definite application. Because of its side-effects, chemoprevention is something that will not be recommended to the population at large. We really need to identify high-risk people who will benefit the most. The risk–benefit ratio needs to be calculated, and this is where we need predictive risk models that are not only going to be based on genetic scores but that we can combine with other factors, to identify high-risk people.
One of the challenges observed for chemoprevention has been the low uptake, notably because of side-effects. Whether communicating about the risk (e.g. that someone is at a very high risk because of genetics or other factors) is going to increase the uptake is a very interesting question.
Risk communication is very important. Risk is a number. It needs to be communicated in a proper way so that people have a good understanding about what that number means for an individual’s health. Otherwise, it may not have an impact on behaviour changes.
Could the genetic risk make people follow better health recommendations? Results from different studies on this question vary. It depends on how the risk has been communicated. We need more education, more studies of risk communication tools, and to see how this can affect people’s behaviour in terms of screening or lifestyle changes.
Genetic determinants of prostate-specific antigen (PSA) levels
Reference: Kachuri et al. (2023). Genetically adjusted PSA levels for prostate cancer screening. Nat Med. https://doi.org/10.1101/2022.04.18.22273850
LK: In our discovery GWAS, we used data from several data sets. Most of these studies were population-based cohorts, where PSA was measured as part of the routine clinical care. We used formal trials to evaluate our PRS, but we only used the baseline measures. In addition, for the GWAS, we restricted PSA levels: they had to be greater than zero and below 10 nanograms per millilitre. We wanted to look at a pretty large range of PSA variation, while trying to stay below what is typically considered to be indicative of prostate cancer.
LK: For PRSs, we often are not particularly concerned with causality. In this study, it is something that we had to consider, particularly to understand the degree to which the prostate cancer PRS is really modelling prostate cancer risk and actual carcinogenesis, versus just PSA elevation. We need to try to disentangle which pathways are modelling what, because they really need to be separate and orthogonal. We want to use genetic determinants of benign PSA elevation to correct PSA levels, and we want to keep that separate from specifically predicting prostate cancer risk.
LK: My hope is that the approach we used for PSA could be extended to other biomarkers. Of course, PSA is a somewhat unique biomarker because it is highly heritable. It is very specific and so linked to prostate cancer. Finding other similarly strong biomarkers could be a challenge, but CA-125 for ovarian cancer is a great example. If we have enough genetic signals, there is a potential to try to apply a similar genetic correction or adjustment, to hopefully make those biomarkers more accurate.
LK: The question is about the potential for undiagnosed prostate cancer to create some reverse causation in our PSA GWAS. This is something that we can’t rule out conclusively, given the fact that the study sample covered a pretty wide age distribution. In the UK Biobank cohort particularly, although they represent a minority of cohort members, we have PSA levels from men who are in their forties. The combination of (1) trying to skew towards mid to slightly younger men, (2) using a maximum of 10 nanograms per millilitre cut-off, and (3) trying to use registry information to exclude prostate cancer cases, would probably lead me to believe that even if there are undiagnosed prostate cancer within our cohort, these would probably be a minority. There is certainly some potential for that to influence our signals, but I would hope that it certainly wouldn’t create a lot of spurious signals.
LK: The additional benefit of the genetic adjustment is much more for aggressive disease than for low-grade disease. This is a little bit expected because a lot of the genetic predictors, both for prostate cancer risk and other predictors, are really geared towards low-grade disease, because that is overwhelmingly the type of prostate cancer that is included in the GWAS.
What often happens is that we’re predicting PSA elevation, which is not necessarily always due to cancer. The relative importance and the added value of PSA genetic prediction is mostly seen for aggressive disease.
LK: There is substantial overlap between prostate cancer risk loci and genetic determinants of PSA levels (in men without prostate cancer). This has been observed in previous GWAS of prostate cancer and PSA levels, not just in our study. Of the 128 PSA-associated variants, approximately 25% were associated with prostate cancer at the genome-wide significance threshold (p < 5 × 10−8). Some of these shared signals with prostate cancer could simply be due to pleiotropy. However, these associations may occur because of screening bias. For example, men who have seemingly elevated PSA levels because of genetics may be screened more frequently until eventually a tumour is found (most likely a low-grade tumour). So in this oversimplified example, a SNP could appear to be associated with prostate cancer, but only because it is influencing detection, not tumour development.
LK: Yes, KLK3 and other independent KLK variants were included.
LK: Although the performance of the PSA PRS and genetically adjusted PSA is worse in non-European-ancestry populations, offering this only to European-ancestry individuals would be problematic from an ethical and health equity point of view. We are currently working on a much larger more ancestrally diverse GWAS of PSA levels, so we hope to have a much better multi-ancestry PSA PRS in the near future.
Other questions
How will this be possible in the future, given that genetic information is more available in European-descent countries?
NC: The fact that GWAS have been mostly done in predominantly European-ancestry populations is a major challenge. Their performance is not optimal in other populations.
We need much larger studies with non-European-ancestry populations to improve the performance of the PRSs for those populations.
There is also an opportunity to use better statistical methods to improve PRSs in those populations, by borrowing information from larger studies available in European-ancestry populations. This is a very active area of research. We must remember that while there is heterogeneity in genetic risk across populations, there is also homogeneity. A lot of common variants that are present in multiple populations have the same effect across populations. For this reason, when we are trying to develop a genetic score for a population, such as an African-ancestry or an admixture population, we should be able to use not only data from that population but also data from much larger European-ancestry populations. What are the best methods to do that? The question is still in the air. A lot of methods are coming out.
Another question is whether we will be able to make the PRS performance of non-European-ancestry populations equal to the performance of European-ancestry populations. If the requirement is for those populations to have sample sizes as large as for the European-ancestry populations, it will be very challenging.
I hope that with better methods and somewhat bigger GWAS on those populations, we can get much better PRSs for them.
NC: We did some of the calculations for cancer. For breast cancer, for example, we currently have GWAS with almost 250 000 individuals (125 000 cases and controls). A calculation suggested that if we could increase that to 500 000 cases and controls, at least for the European-ancestry populations, we would explain about 80–90% of cancer heritability.
Now, when you think about increasing the sample size of GWAS, it is important to consider increasing the sample size for non-European-ancestry populations. There would be a better value in improving the prediction performance of the genetic score across populations, if we could invest more in these populations.
We need to remember that increasing sample size for any population helps other populations as well. Especially in African populations, there are many unique genetic variants that have not been studied yet. They could help not only the performance of PRSs in African populations but also the performance of PRSs in other populations.
JM: For cervical cancer and for rarer cancers, such as kidney cancer or Hodgkin lymphoma, having a sample size with a couple of hundred thousand cases is problematic. In that context, we need to find clever ways of translating the knowledge that we are learning from the different GWAS from one cancer type to another (e.g. cross-cancer analysis).
Genetic annotations could also contribute to leverage the information that we have from the large scans and try and translate them to the rarer scans.
LK: There are two separate points here. Most GWAS take age into account as a covariate, which means that the coefficients that are used as SNP weights for the PRS are age-adjusted.
This is separate from the point about absolute risk models. Developing such models is important, and ideally these would be integrative models that would include PRS and other clinical predictors.
JM: Yes, PRSs tend to be a continuous distribution (and normally distributed), so can be including regression models (that also include covariates).
NC: The data integration approach allows easily adapting models to new populations (e.g. by updating incidence rates while keeping relative risks constant), but it is always important to perform model validation in the target population.
JM: Yes, we can, and it seems it adds on top of rare genetic variants.