Polygenic Risk Scores in the Early Detection of Cancer and Autoimmune Disease in Patients with Non-Specific Symptoms: Insights from the SCAN Cohort

Technical Summary: Polygenic Risk Scores in the Early Detection of Cancer and Autoimmune Disease in Patients with Non-Specific Symptoms

Problem Statement
Early detection of cancer and autoimmune diseases (AID) in patients presenting with non-specific symptoms (e.g., fatigue, weight loss, pain) remains a significant clinical challenge. Primary care clinicians often struggle to distinguish serious underlying pathology from benign causes, leading to diagnostic uncertainty. While Polygenic Risk Scores (PRS) have demonstrated associations with disease incidence in large population cohorts, their clinical utility in symptomatic, referred populations—specifically within rapid diagnostic pathways—is unclear. The central question is whether PRS can provide incremental predictive value when integrated with standard clinical variables (demographics, symptoms, routine blood tests, and disease-specific biomarkers) to refine risk stratification in a real-world diagnostic setting.

Methodology
The study leveraged the Oxfordshire Suspected CANcer (SCAN) pathway, a rapid diagnostic service for individuals with non-specific symptoms.

• Cohort: Germline DNA was analyzed from 2,000 consecutively biobanked and consented participants (1,996 after quality control). The primary analysis was restricted to 1,891 individuals of European ancestry due to current limitations in PRS transferability across diverse ancestral groups.
• Genotyping & PRS Calculation: DNA was genotyped using a custom Thermo Fisher Axiom array. PRS were calculated for 13 conditions: five cancers (breast, colorectal, ovarian, prostate, melanoma) and eight autoimmune diseases (coeliac disease, Crohn's disease, multiple sclerosis, psoriasis, rheumatoid arthritis, systemic lupus erythematosus, type 1 diabetes, and ulcerative colitis). Scores utilized "Enhanced" weights trained on GWAS data including UK Biobank subsets. Two additional PRS (lung and pancreatic cancer) were developed specifically for this project.
• Statistical Modeling:
  • Univariable Analysis: Compared PRS distributions across diagnostic groups (New Diagnosis, Historical Diagnosis, Non-Diagnosis) using t-tests with False Discovery Rate (FDR) correction.
  • Multivariable Analysis: Employed Bayesian regularised logistic regression (using rstanarm with Cauchy priors) to assess the additive value of PRS. Models included clinical/demographic features, presenting symptoms, routine blood biomarkers, and disease-specific markers (e.g., PSA, CA125, FIT).
  • Model Validation: Performance was evaluated using approximate leave-one-out cross-validation (LOO-CV). Metrics included sensitivity, specificity (fixed at 95%), positive/negative predictive values, and likelihood ratios.
  • Clinical Utility: Decision curve analysis (DCA) was used to quantify net clinical benefit across decision thresholds, comparing "Clinical + PRS" models against "Clinical" models alone and default strategies ("treat all" or "treat none").

Key Results

• Cancer Detection:
  • Incidence: 171 patients (9.0%) received a new cancer diagnosis.
  • Discrimination: Adding cancer-specific PRS to clinical models modestly improved out-of-sample discrimination for prostate, breast, and pancreatic cancer. These models showed higher sensitivity and improved likelihood ratios compared to clinical models alone.
  • Null Results: PRS added no significant predictive value for lung and colorectal cancer, where models were driven primarily by existing biomarkers (e.g., FIT) and symptoms.
  • Clinical Utility: Decision curve analysis indicated a small but consistent net clinical benefit for prostate, breast, and pancreatic cancers. This corresponded to detecting an additional 5 to 20 cancers per 10,000 individuals triaged, without increasing the false-positive burden beyond the clinical model alone.
  • Absolute Performance: Despite relative improvements, absolute positive predictive values (PPV) remained low, reflecting the low prevalence of cancer in the referred population.
• Autoimmune Disease (AID) Detection:
  • Incidence: 360 patients (19%) had a current or prior AID.
  • Association: Disease-specific PRS were significantly associated with their corresponding conditions across all eight examined diseases (after multiple-testing correction).
  • Risk Stratification: Higher PRS were consistently associated with greater odds of disease. Predicted absolute risk increased monotonically across PRS percentiles, with marked enrichment in the upper tail. The effects were strongest for coeliac disease and type 1 diabetes, but evident across all conditions.
  • Magnitude: Individuals in the upper tail of the PRS distribution exhibited several-fold increases in disease risk compared to cohort prevalence.

Significance and Claims
The authors conclude that in patients referred with non-specific symptoms, PRS are associated with both cancer and autoimmune disease but offer limited standalone predictive value. The study supports a specific, modest role for PRS:

1. Risk Stratification, Not Diagnosis: PRS should not be viewed as diagnostic tools. Instead, they function as "adjunctive risk enrichers" that modestly improve risk stratification when integrated into multivariable models alongside phenotypic and laboratory data.
2. Context Dependence: The utility of PRS is highly context-dependent. It is most pronounced for cancers with well-characterized polygenic architectures and established biomarkers (e.g., prostate, breast, pancreatic) and less effective where strong existing predictors (e.g., FIT for colorectal cancer) dominate the signal.
3. Autoimmune Utility: The consistent and strong associations observed for AID support the potential integration of PRS alongside standard diagnostic tests when autoimmune disease is part of the differential diagnosis, potentially helping clinicians consider non-oncological explanations earlier.
4. Limitations: The authors acknowledge that sample sizes for specific cancer types were small, limiting power, and that findings are currently restricted to European ancestry. They emphasize that PRS performance in this symptomatic cohort does not necessarily extrapolate to population screening or primary care settings without further prospective validation.

In summary, the paper argues that while PRS capture inherited susceptibility, their clinical translation in symptomatic populations requires integration with standard clinical workflows to provide marginal, yet potentially meaningful, improvements in risk stratification for specific conditions.