This is an explanation of research evidence, not interpretation of anyone's genetic test or advice about screening or treatment. All worked numerical examples are hypothetical. Ancestry labels refer to study descriptions and statistical reference populations, not discrete biological races.
1. The score has a history that the percentile hides
The Polygenic Score Catalog's page checked for this article identified its latest release as October 23, 2026. The resource provides published scores together with information about their development, scoring files and performance evaluations. That infrastructure makes an important distinction visible: a score is a defined research object with a source and a calculation method, rather than a free-floating statement about someone's genetic destiny. PGS Catalog
Yet the result most people encounter may be a percentile, a colored band or a phrase such as elevated genetic risk. Those summaries can be useful, but they leave several questions unanswered. Relative to which population? Calculated from which variants? Predicting which outcome? Evaluated in people whose circumstances resemble the intended users? Integrated with what other information?
The challenge becomes especially important when a score developed in one population is applied in another. The DNA alphabet remains the same, but the statistical relationships used to construct a predictor need not transfer unchanged. A model may rely on patterns of correlation among variants, estimates obtained from unequal sample sizes and outcomes defined through particular healthcare records. Its performance is a property of that entire evidentiary arrangement.
This article examines those links without treating imperfect portability as proof that genetic prediction is useless. The scientific question is how to identify, measure and reduce the gap between a score's development setting and its intended application. The clinical question goes further: does using the resulting information improve a meaningful decision? A convincing answer requires more than a striking percentile or a larger discovery dataset.
2. What is actually being added together
The National Human Genome Research Institute defines a polygenic risk score as a statistical calculation based on multiple genomic variants, using genomic information rather than environmental or other factors. Its glossary also emphasizes that such a score is only one contributor to disease prediction. The definition is narrow in an important way: a polygenic score can be incorporated into a broader risk model, but the score and the combined model are not identical. NHGRI definition
A simple hypothetical calculation illustrates the structure. Suppose three counted variants receive weights of 0.2, minus 0.1 and 0.3, and an individual has dosages of 2, 1 and 0 for the specified effect alleles. The weighted sum is 0.3: two times 0.2, plus one times minus 0.1, plus zero times 0.3. Real scores can involve vastly more variants and more complex development methods, but the example shows why both the dosage and the weight matter.
The number 0.3 has no natural interpretation as a 30 percent disease probability. Its meaning depends on the scale of the weights and the model in which it is used. A report that converts a score into a percentile adds a reference distribution. A report that converts it into absolute risk adds further assumptions and information.
This chain explains why two laboratories can legitimately produce different numbers from the same person's genomic data when they use different scores or reference populations. It also explains why differences can arise from errors. To distinguish those possibilities, the calculation needs an identifier and a documented method rather than only a reassuring label on the final report.
3. Association can support prediction without identifying a mechanism
A variant associated with an outcome need not be the biological cause of that outcome. It may correlate with another variant that has a relevant effect, or its observed association may reflect other features of the study design. Prediction and causal explanation are related scientific goals, but success at one does not automatically establish the other.
Imagine a hypothetical marker that is often inherited alongside a functionally important variant in the discovery population. The marker can help predict an outcome even if changing the marker itself would have no effect. If the relationship between the marker and the functional variant differs in another population, its predictive value can change while the underlying biology remains broadly shared.
This is one reason that portability should not be interpreted as a simple contest over whether one group has more or less genetic susceptibility. A score can perform differently because it is a better statistical measurement in one dataset than another. The failure of a proxy to transfer is not evidence that the condition has an entirely different biological basis in every population.
The same caution applies when a company explains a score through a list of named genes. A compelling biological story can be educational, but it does not validate the aggregate predictor. The score may include many correlated markers whose individual interpretation is uncertain. Conversely, a predictor may have useful empirical performance even when every component's mechanism is not known.
The appropriate report keeps these claims separate. It can describe what the variants predict, how strongly they predict it in specified data, and what is known about mechanism. It should not use an association result to promise that changing a behavior or treatment will produce a particular benefit without separate evidence for that causal claim.
4. Linkage disequilibrium is part of the portability problem
Linkage disequilibrium describes statistical association between variants. A predictor built from markers can partly depend on those associations. If the relationships differ between the discovery and target populations, the same marker weights may capture the underlying signal less well.
A 2020 study by Wang and colleagues developed a theoretical framework for relative prediction accuracy across ancestry-divergent populations and tested it with simulations and UK Biobank data. The framework examined contributions from linkage disequilibrium, allele frequencies, heritability and correlations of causal effects. It provides a mechanistic statistical explanation for why portability can deteriorate even when much causal variation is shared. Its estimates are study- and trait-dependent, not universal conversion factors for every modern score. Wang and colleagues, 2020
A useful analogy is a signpost whose usefulness depends on the road network. The sign may remain readable while the routes it once summarized differ. The analogy is limited, because genomes are not maps designed for prediction, but it captures why copying the same weights does not guarantee copying the same information.
For an evaluation, the practical questions include which reference data informed the correlation structure and how well those data match the intended application. A large discovery study can estimate many associations precisely while still leaving uncertainty about how markers behave in a less represented population.
This is not solved by adding a demographic label to the final report. A label cannot recover information absent from the development data. It may help describe a validation stratum, but the prediction still needs evidence. The important advance is improved modeling and representative evaluation, not a more confident name for an inadequately measured population.
5. Allele frequencies affect what a score can measure
The frequency of a variant influences how much information a dataset provides about its association with an outcome. A variant that is uncommon in one discovery sample may have an imprecise estimated weight there, even if it is more common in a target population. The reverse situation can also occur. Portability therefore depends partly on the distribution of observations available for estimation.
Consider a hypothetical study of 10,000 people in which a relevant allele is extremely rare. The total sample size sounds substantial, but the number of people carrying the allele may still be too small to estimate its effect precisely. Increasing the sample with many more people who do not carry it may help less than a headline about total enrollment suggests.
A score usually combines many such estimates. Their uncertainty and correlation matter collectively. Counting the number of included variants is therefore a poor stand-alone measure of quality. A million weakly estimated or poorly matched contributions do not necessarily outperform a more carefully constructed set for a particular task.
This also explains why representation should be reported at more than one level. The overall ancestry composition of a consortium is informative, but trait-specific case counts, variant coverage and validation samples determine the evidence for a particular score. A population can be present in a biobank yet still be sparsely represented in the cases relevant to one disease.
The editorial lesson is to ask what the sample contains, not simply how large it is. A news story about a record-setting genetic dataset should distinguish total participants from the participants contributing usable information to the specific prediction. That distinction makes a scientific advance more comprehensible without diminishing the work needed to build the resource.
6. Ancestry categories are approximations, not natural boxes
Genetic ancestry describes patterns of shared inheritance relative to reference data and analytical choices. It should not be equated mechanically with race, nationality or a person's social identity. Those categories can be relevant to research for different reasons, but they answer different questions.
A broad continental label can contain substantial genetic diversity and many environmental circumstances. An individual can have ancestry from multiple sources, and different genomic regions can reflect different histories. Assigning that person to one category may be convenient for an analysis while losing information important to a prediction method.
This creates a reporting obligation: explain how categories were obtained and what they are being used to represent. Self-identification, geographic recruitment and genetically inferred similarity are not interchangeable variables. If a study uses one, a secondary article should not quietly replace it with another because the replacement sounds more familiar.
A hypothetical validation labeled Asian, for example, would be difficult to interpret without further information about recruitment and composition. It might describe one regional population, several populations combined, or a broad administrative category. Its performance estimate should not automatically be applied to every person who uses that identity.
Careful language avoids two opposite mistakes. One is treating categories as fixed biological essences. The other is pretending that patterns of genetic relatedness have no relevance to the portability of genetic predictors. The scientifically useful position is more specific: ancestry-related structure can matter for model development and validation, while the categories used to summarize it remain imperfect and context-dependent. A responsible score report describes the uncertainty introduced by that simplification instead of turning it into an identity claim about the user.
7. A better method can help without closing every gap
The 2022 PRS-CSx study by Ruan and colleagues developed a method that combines genome-wide association summary statistics from multiple populations. It reported improved cross-population prediction in simulations and empirical analyses. The method shares information while allowing population-specific patterns, rather than simply assuming one set of estimates transfers unchanged. These results are evidence of methodological progress, not a guarantee that every score built with the method is clinically useful in every population. Ruan and colleagues, 2022
This distinction matters because discussions of disparities can freeze the field in an outdated picture. It would be wrong to repeat an old limitation as though no subsequent work had improved prediction. It would be equally wrong to announce that a multi-ancestry method has solved representation merely because its average benchmark is better.
The relevant comparison is task-specific. Which discovery samples were available? Which target populations were evaluated? What was the baseline method? How much improvement occurred in absolute terms? A large relative improvement from a weak starting point may still leave limited predictive power. A modest relative improvement can be meaningful if it changes an important decision with acceptable consequences.
An evaluation should also identify where the method has not been tested. The word multi-ancestry describes a development strategy, not universal coverage. A study incorporating several large populations can still leave many others poorly represented.
For readers, the constructive question is not whether portability is possible in principle. It is how far a particular method and dataset have demonstrated portability for the intended use. That framing recognizes real advances while preserving the evidence needed before extending them beyond the populations and outcomes actually examined.
8. Development, tuning and testing need separate roles
A score is often selected from several candidates or developed through choices about variant inclusion and statistical parameters. Those choices can adapt the final predictor to the data used for selection. A credible performance estimate needs an assessment process that does not merely repeat the same optimization exercise.
The distinction is easy to see in a hypothetical competition. A team tries twenty scoring methods on one dataset and announces the best result from that dataset as its expected real-world performance. Even if all twenty methods were honestly implemented, the winning score may partly reflect chance features of the selection sample. Independent testing asks whether the advantage persists in observations that did not determine the winner.
The relevant unit of independence also matters. Related people, overlapping biobank participants or repeated records can connect datasets that appear separate by filename. A study should explain how it assessed overlap between discovery and evaluation samples rather than relying on the existence of different project names.
The 2021 Polygenic Risk Score Reporting Standards provide a framework for describing study populations, development, validation and limitations. Their purpose is to make the evidence interpretable and reproducible. A reporting checklist does not make a weak design strong, but it helps readers identify what design was actually used. Wand and colleagues, reporting standards
For a purchaser or clinician reviewing a score, an important question is whether the presented evaluation was genuinely external to the development choices. A polished report should not require the reader to infer that distinction. The score's performance, the conditions of evaluation and any remaining overlap should be stated plainly enough to support comparison with alternatives.
9. The outcome definition travels with the score
A score for a disease is built against a definition of that disease. That definition may use clinical examinations, laboratory thresholds, hospital codes, self-report or combinations of records. These choices affect who counts as a case and who counts as a control.
Imagine a hypothetical discovery study that identifies cases through hospital admissions. It may overrepresent people with more severe disease or greater access to the participating system. A target application aimed at earlier, milder disease in primary care is asking a related but different question. The score's association may persist, weaken or behave differently; the answer requires validation rather than a change in label.
Control definitions matter as well. A person without a recorded diagnosis may be unaffected, undiagnosed, too young to have developed the condition or treated elsewhere. Treating all such people as biologically equivalent can introduce uncertainty. Follow-up and age distribution help clarify what the comparison actually represents.
This is especially important when a risk report uses a lifetime phrase. A study of prevalent disease in middle-aged volunteers does not automatically establish a precise lifetime probability for a young person. Translating between those settings requires assumptions about incidence, survival and competing events.
A clear evidence summary should identify the outcome, ascertainment method and time frame. It should also explain whether the intended application predicts existing disease, future disease or a related trait. Those distinctions often disappear when a research result becomes a consumer message, yet they determine the meaning of the number. A score cannot be fully understood without understanding the event it was trained to recognize.
10. Percentiles describe a reference distribution
A percentile tells a person where a score lies relative to a specified reference set. It does not directly state the probability of disease. A score in the ninetieth percentile is higher than roughly ninety percent of the reference distribution under the stated convention; it does not mean a ninety percent chance of the outcome.
The reference set is therefore part of the result. If it changes, the percentile can change without any change in the person's DNA or the underlying raw score. That can be entirely legitimate, but the report should explain why the comparison changed and whether the new reference is more appropriate.
Suppose a hypothetical score is standardized using a mean of 1 and a standard deviation of 0.5. A raw value of 2 lies two standard deviations above that mean. If another reference uses a mean of 1.5 and a standard deviation of 0.5, the same raw value lies one standard deviation above its mean. These are different relative positions, not evidence that the person's genome changed.
Standardization can improve comparability within an analysis, but it cannot on its own establish equal predictive validity across populations. Making two distributions look similar does not prove that the score-outcome relationship is the same. The evaluation must still examine prediction in the target setting.
This is why reports should avoid presenting an ancestry-adjusted percentile as a complete solution to transferability. Adjustment may address one feature of the score's distribution. Clinical interpretation requires more. Readers should be told which reference was used, how it was selected and what evidence connects the resulting percentile to the risk statement being communicated.
11. Relative risk needs a baseline and a time horizon
A relative association can sound dramatic while leaving the practical probability unclear. Doubling a small risk and doubling a large risk have different absolute implications. The underlying measure also matters: an odds ratio is not generally identical to a risk ratio, especially when the outcome is common.
Consider an explicitly hypothetical risk-ratio example. If the relevant baseline probability over a defined period is 2 percent, a risk ratio of two corresponds to 4 percent under that simplified model. If the baseline is 20 percent, the same ratio corresponds to 40 percent. Neither calculation can be performed responsibly without specifying the baseline, outcome and period.
An odds-ratio example differs. Starting from a probability of 20 percent gives odds of 0.25. Doubling those odds gives 0.5, corresponding to a probability of one third, not 40 percent. The arithmetic demonstrates why a report should not substitute one association measure for another because both are described conversationally as twice the risk.
A polygenic score can be incorporated into a model that estimates absolute risk using age and other information. That combined model needs its own validation. The genetic component's association alone does not establish the accuracy of the final probability.
For public reporting, the best practice is to state the measure and comparison in ordinary language while preserving the numerical distinction. A headline need not contain every statistical term, but the article should explain what the estimate means before readers use it to imagine an outcome. An impressive relative result is a starting point for interpretation, not a self-contained forecast of an individual's future.
12. Calibration tests whether probabilities mean what they say
A model can rank individuals reasonably well while assigning probabilities that are too high or too low. Calibration evaluates the relationship between predicted probabilities and observed outcomes. It becomes especially important when a score is used to place people above or below a decision threshold.
Suppose a hypothetical model assigns approximately 10 percent risk to a group of 1,000 people for a defined period. Observing about 100 events would be consistent with calibration for that group, subject to uncertainty and complete follow-up. Observing 200 would suggest underestimation for that group, even if people assigned lower probabilities experienced fewer events.
Calibration should be assessed over relevant ranges and populations rather than inferred from one overall average. A model could predict the total number of events correctly while overestimating risk in one subgroup and underestimating it in another. Those errors can cancel in the aggregate.
Recalibration can sometimes improve probability estimates without changing the score's underlying ranking. But it is not a substitute for a useful signal. If a score barely distinguishes outcomes in a target population, adjusting its average level cannot create strong discrimination from nothing.
The relevant evidence therefore includes both how the model orders people and how its probabilities correspond to observed outcomes. A report should say whether recalibration was performed in the target data and whether a separate evaluation assessed the result. Otherwise, apparent agreement may partly reflect fitting the model to the very observations used to judge it. A probability is a stronger promise than a relative score, and its validation should reflect that stronger promise.
13. The performance metric must fit the question
Different metrics describe different aspects of prediction. A measure of explained variation for a continuous trait does not directly tell readers how well a test identifies future disease. An odds ratio per standard deviation does not reveal every person's probability. A discrimination measure does not by itself establish that acting on the score improves health.
The EMBL-EBI training material on evaluating polygenic scores distinguishes performance measures for quantitative and binary traits and emphasizes the context of evaluation. That technical distinction is important when comparing studies whose headline numbers use different scales. A larger number is not necessarily a better result if the numbers measure different properties. EMBL-EBI evaluation tutorial
An explicitly hypothetical example shows the danger. One study reports that a score explains 10 percent of variation in a measured trait. Another reports an odds ratio of two for disease in a selected high-score group. Neither result can be ranked against the other without considering the outcome, population and intended application. They answer different questions.
A useful comparison holds as much as possible constant: the target outcome, evaluation sample, baseline model and metric. It also reports uncertainty. Small differences between point estimates may not support a confident ranking, especially when evaluation samples differ.
The practical question is incremental information. Does the score add useful prediction beyond information already available for the intended decision? A genetic score may be associated with disease yet add little to a strong existing model. In another setting, particularly earlier in life, its contribution could differ. The relevant comparison must be demonstrated rather than inferred from association alone.
14. Technical calculation errors can masquerade as biology
Before debating ancestry transfer or clinical benefit, the calculation itself must be correct. Scoring requires matching variants, effect alleles, genomic coordinates and weights. Differences in genome builds, missing variants or allele coding can alter the result. An unexpected score may reflect a technical mismatch rather than an unusual biological profile.
The PGS Catalog's score-page tutorial explains that individual entries include development metadata, variant counts, source publications and performance information. This structure helps users connect a scoring file to the evidence describing it. It also makes clear why a filename or a disease label alone is insufficient identification. EMBL-EBI score-page guide
A hypothetical laboratory transferring a score between platforms should verify that the intended alleles are counted consistently. If an effect is attached to one allele but the software counts the other, the contribution can change direction. With many variants, a collection of small technical mismatches can become a substantial aggregate problem.
Missingness also needs a defined policy. Omitting unavailable variants, substituting estimates or using imputed dosages are not interchangeable choices. Their effect depends on which variants are missing and how much information they contribute. Reporting only that most variants were available can hide the importance of the remainder.
Technical validation and clinical validation therefore answer separate questions. The first asks whether the implemented pipeline calculates the intended score reliably. The second asks whether that score supports the intended clinical interpretation. A technically flawless pipeline can calculate a poorly portable predictor, and a promising predictor can be undermined by an incorrect implementation. Both links need attention.
15. Imputation brings information and uncertainty
Genotyping platforms do not necessarily measure every variant used in a score. Imputation estimates unobserved genotypes using observed data and reference patterns. This can expand the information available for analysis, but the estimates have uncertainty and depend on the suitability of the reference data.
A hypothetical dosage of 0.8 is not the same kind of observation as a confidently measured integer count. It can represent an expected allele count under an imputation model. Whether using such dosages is appropriate depends on the scoring pipeline and its validation, not on a blanket assumption that fractional values are errors.
The relevant quality question is variant-specific and population-specific. A reference panel can support strong estimation in some genomic regions and weaker estimation in others. A high average quality measure may conceal uncertainty at variants that contribute meaningfully to a particular score.
A report should therefore describe the input platform, imputation procedure where used, filtering and handling of uncertainty. This is especially important when a score is calculated from data originally collected for another purpose. A consumer ancestry file and a clinically validated genomic assay may differ in coverage and intended use even if both contain familiar variant identifiers.
The broader point is that the score begins before the weighted sum. Sample handling, measurement, quality control and genotype processing shape the data that enter the calculation. Portability across populations and portability across technical platforms can interact. A comparison that changes both at once should avoid attributing every performance difference to ancestry. Careful design separates those sources of variation where possible and clearly describes what remains entangled.
16. Selection into a study can change apparent performance
A biobank is not automatically a random sample of the population in which a score may later be used. Participation can depend on age, health, geography, healthcare access and willingness to contribute data. These features can influence both the observed outcomes and the relationships available for prediction.
Consider a hypothetical volunteer cohort with unusually complete follow-up and high engagement with preventive care. A model evaluated there may encounter different event rates and measurement patterns in a busy community clinic. Even if genetic associations are similar, the complete risk model can require further evaluation because its non-genetic context differs.
Selection can also affect the apparent relationship between a score and an outcome. Conditioning on participation, survival or receipt of testing may create patterns that are not identical to those in the intended target population. The details depend on the causal structure; selection should not be invoked as a universal explanation for every result. It should be examined as a plausible source of bias when the design supports that concern.
A strong validation report describes recruitment, exclusions and follow-up rather than relying only on a broad ancestry label. Matching ancestry does not ensure matching age, healthcare setting or outcome ascertainment. A study can be genetically relevant yet clinically unrepresentative for a proposed application.
The implication for readers is practical: ask who could enter the study and who was likely to be absent. Those questions often reveal more about transportability than the total sample size. A large, selected sample can provide excellent evidence for some purposes while leaving uncertainty about others. The task is to name that boundary rather than label the entire study either definitive or unusable.
17. Clinical utility requires a decision and a consequence
A predictor can be statistically valid without having demonstrated clinical utility. Utility asks whether using the information improves a meaningful decision or outcome compared with an appropriate alternative. The answer depends on what actions are available, their benefits and harms, and how the score changes who receives them.
A hypothetical screening program could use a score to identify people for additional evaluation. To assess benefit, researchers would need to consider more than how many extra cases are found. They would also examine unnecessary investigations, missed cases, resource demands and whether earlier detection leads to better outcomes under the proposed pathway.
The threshold matters. Moving it changes the number of people classified as eligible and the balance of consequences. A score cannot choose that balance without values and clinical evidence entering the decision. A statistical maximum is not automatically a socially or medically appropriate policy.
There is also a difference between offering information and ensuring access to the response it implies. A program that identifies elevated risk but leaves people unable to obtain follow-up may produce anxiety without the anticipated benefit. That is a feature of the implementation, not a failure of genetic arithmetic, but it belongs in the assessment of real-world utility.
This article does not recommend a screening schedule or treatment based on a score. The evidentiary point is that such recommendations require more than an association study. A complete claim should identify the score, population, decision pathway and outcome being improved. Without those components, the phrase clinically useful can become a broad endorsement unsupported by the specific research presented.
18. Family history and rare variants remain separate evidence
A polygenic score summarizes one form of genetic information. It should not be assumed to replace family history, clinical assessment or evaluation of particular rare variants when those are relevant. Different kinds of evidence can overlap while still carrying information that the score does not capture.
A hypothetical person with an average polygenic score could still have a clinically important rare variant that the score was not designed to evaluate. Conversely, a person without a known high-impact variant could have an elevated polygenic score. These statements do not establish what either person should do; they show why a report should define the test's scope.
Family history also contains more than shared common variants. It can reflect shared environment, rare genetic factors, patterns of diagnosis and incomplete knowledge about relatives. Its absence may mean no affected relatives, limited family size, young relatives or unavailable information. A model should handle those possibilities according to a defined method rather than treating every blank field as the same biological observation.
Integrating evidence requires validation of the combined approach. Simply adding several risk labels can double-count information or create inconsistent interpretations. If a report presents a combined risk estimate, it should explain which components were included and how their relationships were modeled.
For public communication, the most important boundary is that a low or average polygenic score is not a general certificate of health. It concerns a particular predictor and reference framework. Likewise, an elevated score is not a diagnosis. Preserving those distinctions helps prevent genetic information from crowding out other evidence that may be more immediately relevant to a person's clinical situation.
19. Better representation must include the benefits of participation
Improving genetic prediction requires broader and better characterized data, but recruitment alone is not the whole equity question. Communities contributing data should have a meaningful role in how research is conducted and how its benefits are distributed. A dataset can become more diverse without automatically producing equitable access to the resulting tools.
A hypothetical consortium might recruit from an underrepresented population while performing analysis elsewhere and returning little useful infrastructure or knowledge to the contributing institutions. Its scientific models could improve, yet the communities involved might remain unable to access validated applications. The fairness of that arrangement cannot be judged solely by the ancestry percentages in a publication.
Data stewardship also matters because genomic information is enduring and potentially informative about relatives. The consent and governance appropriate for one study may not automatically cover every future commercial or clinical use. Specific legal obligations vary, but research reporting can still make the intended uses, access arrangements and participant protections visible.
Scientific benefits should be described concretely. Did the work improve prediction in the population that contributed? Were the relevant outcomes measured well? Are there independent evaluations? Can local researchers scrutinize the methods? Those questions distinguish substantive inclusion from a decorative statement about diversity.
The portability problem is therefore both technical and institutional. Better statistical methods can share information more effectively, while better partnerships can improve the quality and relevance of the data being shared. Neither substitutes for the other. A durable advance combines accurate prediction with a credible account of who participated, who can evaluate the result and who is positioned to benefit from its use.
20. Read the next score as a defined prediction, not a destiny
A careful reading begins with the score's identity and purpose. Find the original publication or catalog entry. Identify the outcome, scoring method, development population and independent evaluations. Distinguish a raw score from a percentile and a percentile from an absolute-risk estimate. Check whether the stated interpretation has been tested in the intended population and setting.
Then ask what the number adds. An association may be scientifically important without changing a clinical decision. A new method may improve portability without eliminating every gap. A larger dataset may strengthen evidence while leaving sparse information for a particular disease or community. These are not contradictions; they are different stages in moving from discovery to useful application.
The research described here supports a constructive conclusion. Portability problems have identifiable statistical causes, and methods such as multi-population modeling can improve performance. Reporting standards and shared scoring resources make the evidence easier to inspect. Those advances deserve recognition precisely because the remaining questions are specific enough to investigate.
What should be resisted is the shortcut from a single genetic summary to a comprehensive statement about a person's future. Health reflects many interacting factors, and the meaning of a score depends on the reference and model surrounding it. A scientifically responsible report makes those dependencies understandable rather than treating uncertainty as an inconvenience to hide.
The strongest promise of polygenic prediction is conditional: for this outcome, in this population, using this validated method and decision pathway, the information may improve what can be predicted or decided. Establishing those conditions is the work. A percentile without them is easy to print. Evidence that travels with the score is much harder, and much more valuable.