Final Report Summary - GENOMICMEDICINE (Towards evidence-based genomic medicine: filling the evidence gaps through modelling studies)
Genome-wide association studies are rapidly unravelling genetic variants that are implicated in common complex diseases such as coronary heart disease, type 2 diabetes, non-familial forms of cancer and age-related macular degeneration. These discoveries fuel expectations about their future impact of genomic research on prevention and clinical practice. At this moment the predictive ability of polygenic risk models, i.e. models that predict disease risk based on multiple DNA variations simultaneously, is still low for most diseases, but it is expected that genetic testing may become useful for predicting disease risk, targeting pharmacotherapy, and targeting health behaviour recommendations based on individuals’ genetic profiles. A practical problem is that there are so many discoveries and potential applications and not enough resources (finances and personnel) to investigate which applications are promising for use in health care. This project investigated how hypothetical studies, using simulated data, can be used to identify the promising applications.
First, for investigating the potential use of polygenic risk prediction in optimizing population screening programs, we developed a method that can simulate risk data. This method allows to explore ‘what if’ scenarios for the use of polygenic risk prediction, such as: “how predictive does a polygenic risk model need to be for stratifying the breast cancer mammography screening program?” Such questions cannot be answered in real data because the future risk models still need to be discovered. Exploring the potential applications in simulated data informs what can and cannot be expected from polygenic risk prediction. In this project, we investigated the validity of the method.
Second, with regard to the evaluation of pharmacogenetic tests, we focused on how the predictive performance can best be assessed. Currently, the relevance of pharmacogenetics testing is often based on the statistical significance of the association between the genetic marker and the outcome of interest, which often is drug response, side effects or prognosis. Yet, the strength of the association is only one factor that determines predictive performance. We have explained and illustrated that measures of predictive performance can be calculated in the same studies to get a better insight in the potential performance of the test.
And third, we investigated metrics that are used to assess the predictive performance of (polygenic) risk models. Several metrics are available and new ones are being developed. These metrics each assess a different aspect of predictive performance but they are not entirely independent. We helped to clarify which measure should be used for which question, and how metrics should be evaluated when they seem to give contradictory impressions of the predictive performance.
The project also delivered a promising serendipitous finding. In the beginning of the project, we found that it was difficult to find relevant literature on the topics of this study, especially on the topics that were highly detailed. We had tried to search using many different keywords and keyword combinations, but not successfully. When we tried to find related literature via the references of the papers that we knew, we found exactly what we were looking for. Citation searching is not new, but this method is different. Instead of screening the reference lists of known articles, we focus on articles that are cited together with the known articles (co-citations), and instead of doing that for each known article separately, we screen the co-citations of multiple known articles at once. This appeared to be a very effective method, that to our surprise, had never been used or published before. The method is particularly relevant in meta-analyses and systematic reviews, but has numerous other applications.
First, for investigating the potential use of polygenic risk prediction in optimizing population screening programs, we developed a method that can simulate risk data. This method allows to explore ‘what if’ scenarios for the use of polygenic risk prediction, such as: “how predictive does a polygenic risk model need to be for stratifying the breast cancer mammography screening program?” Such questions cannot be answered in real data because the future risk models still need to be discovered. Exploring the potential applications in simulated data informs what can and cannot be expected from polygenic risk prediction. In this project, we investigated the validity of the method.
Second, with regard to the evaluation of pharmacogenetic tests, we focused on how the predictive performance can best be assessed. Currently, the relevance of pharmacogenetics testing is often based on the statistical significance of the association between the genetic marker and the outcome of interest, which often is drug response, side effects or prognosis. Yet, the strength of the association is only one factor that determines predictive performance. We have explained and illustrated that measures of predictive performance can be calculated in the same studies to get a better insight in the potential performance of the test.
And third, we investigated metrics that are used to assess the predictive performance of (polygenic) risk models. Several metrics are available and new ones are being developed. These metrics each assess a different aspect of predictive performance but they are not entirely independent. We helped to clarify which measure should be used for which question, and how metrics should be evaluated when they seem to give contradictory impressions of the predictive performance.
The project also delivered a promising serendipitous finding. In the beginning of the project, we found that it was difficult to find relevant literature on the topics of this study, especially on the topics that were highly detailed. We had tried to search using many different keywords and keyword combinations, but not successfully. When we tried to find related literature via the references of the papers that we knew, we found exactly what we were looking for. Citation searching is not new, but this method is different. Instead of screening the reference lists of known articles, we focus on articles that are cited together with the known articles (co-citations), and instead of doing that for each known article separately, we screen the co-citations of multiple known articles at once. This appeared to be a very effective method, that to our surprise, had never been used or published before. The method is particularly relevant in meta-analyses and systematic reviews, but has numerous other applications.