Researchers have developed a statistical tool called a polygenic risk score (PRS) that can estimate an individual’s risk for certain diseases that have a strong genetic component, such as heart disease or diabetes. However, the data used to construct PRSs are often limited in diversity and scope. As a result, PRS becomes less accurate when applied to populations that are demographically different from the PRS training data.

A new scoring approach was introduced cell genomics Developed by researchers at the Broad Institute of MIT and Harvard University, Massachusetts General Hospital (MGH) uses a comprehensive approach to generate a more accurate and informative PRS.

This approach, aptly named PRSmix due to its ability to “mix” all previously developed PRSs for a particular trait, more accurately estimates a patient’s genetic disease risk than PRSs generated from individual studies. generate a score.

“The big challenge with the PRS is that it comes from one population and is widely published with the assumption that the scores are generalizable,” explained Pradeep Natarajan, corresponding author of the study. Natarajan is an associate member of the Broad’s Cardiovascular Disease Initiative and director of preventive cardiology at MGH. “The overall motivation for this study is to better identify individuals who are at high risk for genetic diseases early on.”

When applied to a set of diseases, PRSmix was on average 20% more accurate in predicting risk for specific traits compared to individual PRSs. This improvement held true across her two ancestral groups, suggesting that PRSmix scores may be more applicable across diverse populations.

Although PRSmix provides a significant improvement in genetic risk assessment, the scope of this tool is limited only to genetic variants that are directly associated with a specific trait.

“Most PRSs are trait-specific. For example, if you want to predict risk of coronary artery disease, most models use data trained only on coronary artery disease. But clinical risk is determined by more than just genetic variation. “We know that it is influenced by more factors. It is associated with cardiovascular disease,” said Vu Truong, lead author of the study. Truong is a computational biologist in Broad in Natarajan’s lab and holds his Ph.D. He is a student in Price’s lab at Harvard University’s TH Chan School of Public Health. “How accurate will PRS become if we aggregate more genetic information?”

To answer this question, the team developed an additional approach called PRSmix+. This aggregates all existing PRSs for a particular trait, plus all PRSs for related traits such as heart disease and lipid levels.

By considering the influence of heterogeneity, PRSmix+ showed even greater accuracy improvement than PRSmix. For example, PRSmix+ estimates of coronary artery disease risk are 3.27 times more accurate compared to previously developed combination methods.

PRSmix+ also takes into account pleiotropic variations, i.e. single mutations that affect multiple traits, helping to uncover associations that are not otherwise recognized. For example, although age at the onset of menopause and ischemic stroke risk are seemingly unrelated, PRSmix+ identified an association between the two.

“If we hadn’t done an unbiased scan, we wouldn’t have known this,” Truong said.

Despite PRSmix+’s predictive power, the team said there are still scenarios where researchers should choose PRSmix instead. PRSmix+ considers more genetic scores than his PRSmix and therefore needs to incorporate much more data, resulting in longer run times. Additionally, PRSmix+ only slightly outperformed PRSmix when it came to predicting highly heritable traits such as height. For researchers studying highly heritable traits, PRSmix can provide similar predictive accuracy in less time than his PRSmix+. For researchers interested in clinical utility, PRSmix+ may be a better option.

Easy-to-access algorithms

To enable PRSmix to continuously aggregate the latest genetic information, the PRSmix framework is published as an R package on AnVIL, a data repository on the biomedical data sharing platform Terra. “This workflow allows anyone to use the platform to generate his or her PRS for their own research,” said Truong. “Users can simply drop files into Terra; they don’t have to do any advanced calculation work.” AnVIL features checkpoints throughout the workflow, allowing users of all skill levels to avoid calculation errors. will do so.

So far, this accessible approach has proven successful. “Several people are already using his R package and reporting that they are getting much better scores than the scores of individual studies,” he said.

Combining the accessible workflow of PRSmix with the potential clinical utility of PRSmix+ makes the team’s PRS framework even easier to use for both researchers and clinicians.

“When considering implementing a PRS in the clinic, we need to anticipate that the score will perform differently across the health system based on different individual groups and practice settings,” said Natarajan. . “We now have a way to provide a framework for training and recalibration to generate the highest possible score within that dataset using all the information currently available. Ta.”