Gene Action Modeling Tightens Polygenic Prediction for Select Traits
Polygenic scores are built on a deceptively simple premise: sum the small effects of many variants and estimate an individual’s genetic liability or trait value. In practice, the standard framework assumes additive inheritance alone, which means each risk allele contributes in a linear, independent fashion. That assumption is often convenient, but it can miss biology when dominance shapes how alleles interact within a genotype.
This preprint tests whether incorporating dominance effects into score construction improves prediction. The investigators developed PGS a,d, a scoring framework rooted in quantitative genetics that explicitly includes both additive and dominance terms. They then compared it with conventional GWAS-based polygenic scoring in simulations and in real-world data from UK Biobank.
The core result is straightforward: when dominance mattered more, the revised score performed better. In simulation, PGS a,d consistently produced lower root mean squared error and progressively higher R² as dominance contributed more to the genetic architecture. The pattern fits basic genetic theory. Additive models are efficient when trait biology is largely linear, but they lose precision when heterozygous genotypes behave differently from what an additive model would predict.
Lower Error Emerged Where Dominance Was Stronger
The UK Biobank analysis extended that logic to 16 blood cell traits, height, and body mass index. Across the full set of traits, PGS a,d delivered comparable R² to the standard score, including for height and BMI, two traits whose architecture is predominantly additive. Even without broad gains in explained variance, the new framework still showed a practical advantage in prediction accuracy: after strict Bonferroni correction, four traits had significantly lower root mean squared error with PGS a,d.
One sentence captures the bedside logic of the result: when heterozygous effects are real, a purely additive score leaves measurable prediction error on the table.
The improvement tracked with the magnitude of estimated dominance effects. Traits with larger dominance signals were the same traits in which PGS a,d produced greater reductions in RMSE. That relationship makes biological sense because dominance captures departures from additivity at the genotype level, whether through masking, nonlinear protein function, or developmental buffering. For traits that sit close to additive expectations, the benefit is modest. For traits with stronger nonadditive structure, the gain becomes easier to detect.
This is an important distinction for genetic prediction. R² measures how much variation a score explains across a population, but RMSE reflects the size of individual-level prediction error. A scoring method can leave R² nearly unchanged and still improve how close predicted values come to observed values in individual cases. In clinical genetics, that difference matters for risk stratification, phenotype calibration, and any setting where a single patient’s estimate needs to be as close to reality as possible.
Dominance Signals Reappeared Across 39 Genomic Regions
The same analysis also surfaced 141 variants in 39 genomic regions with significant dominance effects after Bonferroni correction across traits. Several of those regions recapitulated known associations, while others represented novel signals. This is not merely a statistical footnote. Detecting dominance effects at the variant level suggests that some loci exert influence through more than a simple additive dosage model, and those loci can be carried forward into more realistic trait architectures.
In quantitative genetics, dominance often receives less attention than additive effects because additive effects are easier to estimate and use in prediction. Yet dominance can reflect biologically meaningful mechanisms, including altered protein complexes, dosage sensitivity, or nonlinear interactions within a pathway. In a trait controlled by many loci, even modest nonadditivity at a subset of variants can distort a score built on additive assumptions alone.
The blood cell traits are especially informative in that respect. Hematologic phenotypes are shaped by tightly regulated developmental programs and lineage-specific thresholds, conditions under which a heterozygous genotype may not behave as a simple midpoint between two homozygous states. Height and BMI, by contrast, are widely treated as highly polygenic and mostly additive. Their comparable R² between models suggests that explicit dominance modeling does not necessarily harm performance in predominantly additive traits, even if the payoff is larger in more nonadditive settings.
Why the Method Matters for Future Genetic Scores
The practical appeal of PGS a,d is that it does not require abandoning GWAS-based scoring. Instead, it extends the framework by adding a dominance component directly into the score construction. That makes the approach conceptually familiar while better aligned with the biology of gene action. In settings where dominance is negligible, the added term may do little. Where dominance is present, the same term can trim prediction error without forcing a wholesale redesign of polygenic scoring pipelines.
The study is also a reminder that the assumption of pure additivity is a modeling choice, not a law of nature. Polygenic scores have become central to genetic risk research because they compress massive genotype data into a clinically usable statistic. Their value depends on how faithfully that statistic mirrors the underlying architecture of the trait. If a trait includes meaningful dominance, then a score that ignores it is mathematically tidy but biologically incomplete.
Limits of the Preprint and the Remaining Validation Gap
This work is still a preprint and has not undergone peer review. Its findings should therefore be treated as an early technical advance rather than a clinical tool. The data come from UK Biobank, which offers scale and quality but also a specific ancestry and recruitment context that may not translate evenly to other populations. The article also focuses on prediction performance and dominance detection, not on clinical decision thresholds, disease outcomes, or prospective patient benefit.
Even so, the direction of the results is clear. Explicitly modeling additive and dominance effects improved individual-level prediction for some complex traits and identified a set of loci with significant dominance signals. The next question is not whether dominance exists in human trait architecture; it does. The harder task is determining for which traits, populations, and use cases dominance-aware scores meaningfully outperform standard additive models in a way that survives external validation and future clinical testing.