Skip to content

PRS evaluation

PGS performance was evaluated using two main metrics:

  1. PGS Correlation
    Pearson’s correlation between PGS derived from imputed SNP array or low-pass WGS data and those obtained from high-coverage (30×) WGS.

  2. ADPR (Absolute Difference in Percentile Ranking)
    The absolute difference in percentile ranking between PGS obtained from array-imputed or low-pass WGS data and those derived from the high-coverage WGS gold standard.

These evaluations were conducted across multiple p-value thresholds to ensure unbiased comparison, based on the method from Nguyen et al., 20221.

PRS formula

For an individual \(i\), the polygenic score at p-value threshold \(P_{T}\) is calculated as:

\[ PGS_i(P_T) = \sum_{j=1}^{M} {1}_{\{P_j < P_T\}} \, x_{ij} \, \hat{\beta}_j \]
  • \(P_T\) : p-value threshold
  • \(M\) : number of SNPs after clumping
  • \(x_{ij}\) : allele count for SNP \(j\) in individual \(i\)
  • \(\hat{\beta}_j\) : marginal effect size from GWAS for SNP \(j\)
  • \(1_{\{P_j < P_T\}}\) : indicator function for p-value filtering

PGS correlation

Pearson’s correlation quantifies the degree of linear agreement between PGS derived from imputed SNP array or low-pass sequencing (LPS) data and those from high-coverage (30X) whole-genome sequencing (WGS), providing a direct measure of score concordance.

PGS correlation was quantified using Pearson's correlation coefficient between the array/LPS-based PGS (\(X\)) and the 30X WGS-based PGS (\(Y\)):

\[ r = \frac{\sum_{i=1}^N (X_i - \bar{X}) (Y_i - \bar{Y})}{\sqrt{\sum_{i=1}^N (X_i - \bar{X})^2} \, \sqrt{\sum_{i=1}^N (Y_i - \bar{Y})^2}}, \]

with \(r^2\) used for reporting. Correlations were calculated separately for each population, phenotype, and feature type (array or LPS), and visualized in bar chart.

Absolute Difference in Percentile Ranking (ADPR)

Absolute Difference in Percentile Ranking (ADPR) complements this by assessing the stability of individuals’ relative positions between platforms, which is critical for downstream applications such as risk stratification and clinical decision-making.

To quantify concordance in relative standing between platforms, raw PGS values were transformed to within-platform percentiles using the empirical cumulative distribution function (ECDF):

\[ p^{(m)}_i = \frac{\mathrm{rank}_m(i)}{N}, \]

where \(p^{(m)}_i\) denotes the percentile of individual \(i\) in method \(m\) (array/LPS or WGS), \(\mathrm{rank}_m(i)\) is the individual’s rank, and \(N\) is the sample size.

The Absolute Difference in Percentile Ranking for individual \(i\) was defined as:

\[ ADPR_i = \left| p^{(\mathrm{array/LPS})}_i - p^{(\mathrm{WGS})}_i \right| \times 100, \]

expressed in percentage points. An ADPR of \(0\%\) indicates identical rankings, whereas higher values denote greater rank displacement between platforms. ADPR values were summarised across populations, phenotypes, and feature types, and visualised to assess platform-specific deviations in ranking concordance.

  1. Dat Thanh Nguyen, Trang TH Tran, Mai Hoang Tran, Khai Tran, Duy Pham, Nguyen Thuy Duong, Quan Nguyen, and Nam S Vo. A comprehensive evaluation of polygenic score and genotype imputation performances of human snp arrays in diverse populations. Scientific Reports, 12(1):17556, 2022. ↩