Calibration of polygenic risk scores is required prior to clinical implementation: results of three common cancers in UKB

Abstract

BackgroundSNP-based polygenic risk scores have recently been adopted in the clinic for risk assessment of some common diseases. Their validity is supported by a consistent trend between their percentile rank and disease risk in populations. However, for clinical use at the individual level, the reliability of score values is necessary considering they are directly used to calculate remaining lifetime risk.ObjectivesWe assessed the reliability of polygenic score values to estimate prostate cancer (PCa), breast cancer (BCa) and colorectal cancer (CRC) risk in three incident cohorts from the UK Biobank (n>500 000).MethodsCancer-specific Genetic Risk Score (GRS), a well-established population-standardised polygenic risk score, was calculated.ResultsA systematic bias was found between estimated risks (GRS values) and observed risks; β (95% CI) was 0.67 (0.58–0.76), 0.74 (0.65–0.84) and 0.82 (0.75–0.89), respectively, for PCa, BCa and CRC, all significantly lower than 1.00 (perfect calibration), p<0.001. After applying a correction factor derived from a training data set, the β for corrected GRS values in an independent testing data set were 1.09 (1.05–1.13), 1.00 (0.88–1.12) and 1.08 (0.96–1.21), respectively, for PCa, BCa and CRC.ConclusionAssessing the calibration of polygenic risk scores is necessary and feasible to ensure their reliability prior to clinical implementation.

Canalization of the Polygenic Risk for Common Diseases and Traits in the UK Biobank Cohort

“…S3, Supplementary Material online. Our approach produces almost identical calibration to that proposed by Wei et al (2020) in their study using Platt scaling to provide appropriate calibration for clinical implementation of cancer PGS. Prima facie, these results imply widespread canalization, but they are actually an artifact of the effect of prevalence on PGS estimation under the liability threshold model.…”

Polygenic risk scores for the prediction of common cancers in East Asians: A population-based prospective cohort study

“…As highlighted by Wei et al, the reliability of score values is necessary for application at the individual level [36]. Even when the PRS have adequate discrimination, estimated risks can be unreliable [37].…”

Utility of family history in disease prediction in the era of polygenic scores

“…Family history is considered a risk enhancing factor, however, we advocate for formal inclusion of both family history and PGS to risk estimation models given that both variables are significant and independent predictors of CAD in HUNT and UK Biobank. Currently PGSs are limited by trans-ethnic portability 31,32 , sensitivity to population stratification 33 , and miscalibration 34 among other considerations. Future iterations of PGSs may integrate genetic risk for clinical risk factors such as genetic prediction of LDL cholesterol or body mass index or other multi-trait risk models that improve prediction.…”