A report says that your polygenic risk score is in the top 5%. It is natural to wonder whether that means you have a 95% chance of developing the disease. It does not. The result places your score within a reference distribution; translating that position into a useful health prediction requires more information.
The most useful questions are what the score predicts, how well it was tested in the intended population, and whether using it would improve a particular decision. This guide shows how to work through those questions without treating a genetic score as a diagnosis.
Start by identifying the number on the report
A polygenic score combines information from many genetic variants. Typically, the calculation weights variants according to their estimated associations with a trait or outcome. It captures part of genetic susceptibility, rather than every influence on health. NHGRI provides an accessible introduction to polygenic scores; its 2020 page is useful for the basic concepts, rather than a current inventory of clinical applications.
| Report item | What to clarify |
|---|---|
| Raw or standardized score | Which calculation and version produced it? What scale is being used? |
| Percentile | Which reference group defines the ranking? |
| Relative association | Compared with whom, and is the measure a risk, odds or hazard ratio? |
| Absolute risk estimate | Which outcome, time period and additional predictors does the model include? |
Example 1: the 95th percentile is a ranking
Fictional teaching example. Mira receives a result at the 95th percentile for a disease-associated score. She reads this as “95 out of 100 people like me will develop the disease.”
The report has not said that. It says her score is approximately higher than those of 95% of the reference group. To estimate disease probability, the service would need an appropriate risk model and the information it requires. The distinction is much like being near the top of a height distribution: a ranking describes position, not the probability of a future event.
A useful question to send back is: “Does this report estimate my absolute risk? If it does, please show the timeframe and explain which information beyond the score was included.” Our absolute and relative risk guide explains how to read the resulting numbers.
Example 2: good ranking can coexist with inaccurate probabilities
Fictional teaching example. A model assigns 1,000 people an average five-year risk of 10%, corresponding to 100 expected events. After complete follow-up, 50 events are observed. Suppose this pattern persists in a sufficiently large, representative validation study, beyond what random variation plausibly explains.
The model has overestimated risk in that setting. It might still rank people with later events above those without them. Ranking performance, often called discrimination, and agreement between predicted and observed outcomes, called calibration, need separate evaluation.
This example is deliberately simplified: a single group’s event count cannot assess calibration throughout the risk range. NCI’s absolute-risk modeling overview describes the importance of independent validation and monitoring calibration. For a real paper, look for performance across relevant risk groups, uncertainty estimates and how incomplete follow-up was handled.
Who was represented in the evidence?
A score developed in one population may perform differently elsewhere. In 2019, Martin and colleagues documented how imbalances in genomic research can produce unequal predictive performance. Their findings should motivate scrutiny of the specific score and validation evidence, rather than a blanket assumption that all scores perform identically today.
Ask about the participants’ ages, clinical setting and ancestry, and how the investigators described those characteristics. A broad racial label is not a complete account of genetic ancestry or of the circumstances that influence disease. The eMERGE Network’s 2024 account of differential performance discusses practical and ethical difficulties in grouping people and communicating these limitations.
For the person reading a report, the question is concrete: “What evidence supports using this model for someone with my characteristics, and where is that evidence weaker?” An honest answer can include uncertainty.
Example 3: moving a number does not prove a benefit
Fictional teaching example. A research model estimates a person’s ten-year risk at 4% using conventional factors and 6% after adding a polygenic score. A study proposes a 5% threshold for an additional assessment. These are invented numbers and an invented threshold, not recommendations for any disease.
The score changes the proposed decision. That observation alone does not show that the change helps. The investigators still need to examine whether the extra assessments identify useful findings, cause avoidable harm, and improve outcomes compared with the existing approach.
For a research discussion, separate three claims: the score is associated with disease; the combined model predicts sufficiently well; and acting on the model improves care. Each requires evidence. A statistically impressive association does not establish all three.
A compact checklist for appraising a paper
The Polygenic Risk Score Reporting Standards provide a structured basis for assessing studies. Record the following before comparing headline performance figures:
- Identify the model. Record its version, outcome and intended use.
- Separate development from validation. Note whether evaluation used independent participants.
- Check the comparator. What does the score add beyond the clinical information already available?
- Read more than one performance measure. Look for calibration as well as discrimination, with uncertainty.
- Trace the proposed action. What decision changes, and what evidence addresses benefit and harm?
The PGS Catalog helps researchers locate published scores, their specifications and evaluation information. Finding a score in a research catalog is not itself evidence that a particular clinical use is appropriate. Keep its identifier with your notes so that similarly named scores do not get mixed up.
Bring the score into the wider clinical conversation
The ACMG’s 2023 points-to-consider statement emphasizes clinical context, including recognition of monogenic forms of disease. Do not use a low score as a reason to dismiss symptoms or a known familial pathogenic variant. If someone proposes changing screening or treatment, ask for the disease-specific evidence supporting that use.
Use our paper-comparison workspace to organize study details. The guide to clinical actionability explores the next question: what makes information useful enough to change care?