Genetic risk scores fail to predict survival after diagnosis

By Vijay Kumar MalesuReviewed by Lauren HardakerOct 2 2025

New research across seven global biobanks shows that the DNA driving disease onset does not determine survival; instead, lifespan-linked genes and cross-trait scores hold the real clues to prognosis.

Study: Limited overlap between genetic effects on disease susceptibility and disease survival. Image credit: Natalia Kirsanova/Shutterstock.com

In a recent study published in Nature Genetics, researchers tested whether genetic determinants of disease risk also predict post-diagnosis survival across nine diseases, and compared susceptibility versus longevity polygenic scores (PGSs) for prognosis.

Background 

Two neighbors can share the same diagnosis yet live for dramatically different lengths of time. Genetics that nudge a person toward a disease may not be the same genetics that shape what happens after the first clinic visit. For years, genome-wide association studies (GWASs) have mapped thousands of variants for who gets a disease, but far fewer for how fast it progresses or whether it proves fatal.

Clinicians and families care about the latter because it guides treatment intensity, follow-up, and planning. Emerging biobanks and electronic health records make survival analyses possible at scale, yet signals look sparse. More research is needed to understand which genetic factors truly predict prognosis.

About the study

Researchers pooled seven biobanks (primary analysis) and registry-linked cohorts to study nine high-mortality conditions: Alzheimer’s disease, breast cancer, colorectal cancer, coronary artery disease, type 2 diabetes mellitus, chronic kidney disease, heart failure, prostate cancer, and stroke. Disease definitions and causes of death were standardized using the International Classification of Diseases, Tenth Revision (ICD-10). The main endpoint was disease-specific mortality, with all-cause mortality in sensitivity analyses.

Within-patient GWASs of disease-specific mortality used Cox proportional hazards models implemented in Genome-wide Analysis of Time-to-Event (GATE) or Saddlepoint Approximation Cox (SPACox), adjusting for age at diagnosis, birth year, sex, principal components (PCs), and study covariates. Eligible patients required ≥3 months follow-up. Summary statistics passed quality control (imputation information (INFO) score > 0.7; minor allele count ≥ 20), were aligned to human genome build 38 (hg38) via LiftOver, meta-analyzed with fixed-effect models in Meta-Analysis Helper (METAL), and assessed for heterogeneity with Cochran’s Q.

PGSs were constructed with Mega Polygenic Risk Score (MegaPRS) under Baseline Linkage Disequilibrium-Linkage Disequilibrium Adjusted Kinships (BLD-LDAK) assumptions; a general longevity PGS used the Linkage Disequilibrium Adjusted Kinships-Thin (LDAK-Thin) model. Associations with diagnosis and post-diagnosis survival were tested via logistic or Cox models. Sensitivity analyses addressed survivor bias, follow-up truncation (2/5/10 years), age-at-diagnosis strata, and relatedness. For type 2 diabetes mellitus, macrovascular and microvascular complication endpoints were analyzed, with matched GWASs in non-diabetic populations to probe shared architecture.                  

Study results

Across nine diseases, only one genome-wide significant locus for disease-specific mortality emerged: rs7360523 near Sulfatase 2 (SULF2) for heart failure mortality. Notably, that locus did not show a comparable effect on heart failure susceptibility and even had the opposite direction of effect in susceptibility analyses. When the team compared 804 lead susceptibility variants against mortality, none remained significant after multiple-testing correction; about half shared the same effect direction, no more than expected by chance.

These patterns matched lower heritability estimates for mortality versus susceptibility. When researchers equalized sample sizes and methods in a down-sampling test, susceptibility GWASs still uncovered many more loci than mortality GWASs, suggesting the lack of mortality signals was not simply a power issue.

Disease-specific PGSs strongly predicted who developed each disease (hazard ratios per standard deviation from ~1.17 to ~1.90), yet they were weak predictors of disease-specific mortality after diagnosis. In heart failure, the susceptibility PGS had only a modest association with heart failure mortality, while in chronic kidney disease and prostate cancer, susceptibility PGSs even trended toward protective effects on mortality.

In contrast, a general longevity PGS, derived from lifespan GWAS, was significantly associated with disease-specific mortality in seven of the nine diseases and outperformed susceptibility PGSs in most settings. Notably, the longevity PGS beat susceptibility PGSs in seven of nine diseases. At the same time, in FinnGen, a composite mortality PGS edged out longevity for coronary artery disease and type 2 diabetes mellitus, highlighting the value of cross-trait information.

Because mortality may be an imprecise proxy for progression in some diseases, investigators examined type 2 diabetes mellitus complications. A locus on chromosome 9 achieved genome-wide significance for macrovascular complications among individuals with type 2 diabetes mellitus, but was not associated with type 2 diabetes mellitus susceptibility. Prior cardiovascular disease was excluded when defining macrovascular complications to ensure a cleaner phenotype. In similar GWASs of cardiovascular traits in people without diabetes, the same signal appeared: it was stronger in the general population but weaker in those with diabetes, suggesting common biology shaped by disease-specific modifiers.

Moreover, the PGS for coronary artery disease predicted macrovascular complications in type 2 diabetes mellitus far better than the type 2 diabetes mellitus susceptibility PGS; for microvascular outcomes, only the age-related macular degeneration PGS showed a small nominal association, while the chronic kidney disease PGS did not.

Age at diagnosis also mattered: for Alzheimer’s disease, the susceptibility PGS showed a stronger association with mortality in younger patients but not in older ones. Simulations under a liability-threshold framework showed that conditioning on cases can induce index-event bias. Still, bias correction changed little in a context where progression heritability appears low and mortality is highly heterogeneous.

Together, the data imply that biological mechanisms governing who gets a disease and who dies from it overlap only modestly. Cross-trait information, like lifespan or cardiovascular risk, can better capture survival risk after diagnosis than disease-specific susceptibility genetics alone.

Conclusions 

This large multi-biobank analysis finds limited overlap between genetic effects on disease susceptibility and disease-specific mortality. Lead susceptibility variants rarely influence survival, susceptibility PGSs perform poorly for prognosis, and a general longevity PGS better stratifies post-diagnosis mortality across many diseases.

Clinically, this cautions against using disease susceptibility scores to counsel patients about survival and highlights the potential of cross-trait or longevity-informed models for risk discussions and trial enrichment.

Methodologically, more power, refined progression phenotypes, and integration of related general-population traits are needed to reveal progression biology and actionable targets, especially where care access and treatments strongly shape outcomes.

Download your PDF copy now!