ASCVD Integrated Polygenic Risk Score (iPRS)

ASCVD Integrated Polygenic Risk Score (iPRS) - EU Grant Task 1.2

README

OVERVIEW

This repository contains the optimized Polygenic Risk Score (PRS) for Atherosclerotic Cardiovascular Disease (ASCVD) prediction, developed as part of EIC project MIRACLE Task 1.2. The model integrates genetic risk across three major ASCVD phenotypes: Coronary Artery Disease (CAD), Ischemic Stroke, and Peripheral Artery Disease (PAD).

Best Performing Model: C+T SetD Pan-UKBB CorrectedPerformance: AUC = 0.7383 [95% CI: 0.7358-0.7409]Improvement: +2.0% over base clinical modelValidation Cohort: UK Biobank (N = 435,502 individuals)

Files Included

Core Model Files

panukbb_setD_weights.txt - SNP weights (9,552 variants with rs IDs, alleles, beta coefficients)

Documentation Files

model_performance_summary.csv - Complete performance metrics across all 46 tested models

bias_correction_analysis.csv - UK Biobank overlap bias quantification results

top_contributing_snps.csv - Leading genetic variants and their effect sizes

Methodology

1. GWAS Source Data

Training Data Sources:

CAD: Aragam et al. discovery GWAS (N = 921,279)

Ischemic Stroke: Mishra et al. discovery GWAS (N = 1,330,540)

PAD: Klarin et al. discovery GWAS (N = 719,538)

Quality Control Applied:

SNP-level filtering: INFO > 0.8, MAF > 0.01

Variant harmonization across phenotypes

Allele frequency consistency checks

Genomic inflation factor correction (λ < 1.1)

2. PRS Construction Methods

We systematically evaluated 4 state-of-the-art PRS methods:

A. Clumping + Thresholding (C+T)

Implementation: PLINK 1.9 clumping algorithm

Parameters: r² < 0.1, 250kb window

P-value thresholds: 4 sets tested (5×10⁻⁸, 1×10⁻⁶, 1×10⁻⁴, 0.01)

Rationale: Establishes baseline performance, computationally efficient

B. PRS-CS (Bayesian Shrinkage)

Implementation: PRS-CS-auto for automatic global shrinkage

LD Reference: 1000 Genomes European panel

Approach: Continuous shrinkage of effect sizes

Advantages: Handles LD structure, reduces overfitting

C. LDpred2 (Bayesian Framework)

Implementation: LDpred2-auto with sparse architecture

LD Matrix: UK Biobank-derived European LD

Parameters: Automatic heritability and architecture estimation

Features: Accounts for LD, sparse variant selection

D. Lassosum (Penalized Regression)

Implementation: Cross-validation for optimal λ parameter

Regularization: L1 penalty with automatic tuning

LD Correction: Incorporated via correlation matrix

Benefits: Built-in variable selection, prevents overfitting

3. Variant Set Definition

Four variant sets tested systematically:

Set

P-value Threshold

Description

Typical SNP Count

A

p < 5×10⁻⁸

Genome-wide significant

~500

B

p < 1×10⁻⁶

Suggestive associations

~800

C

p < 1×10⁻⁴

Lenient threshold

~2,400

D

p < 0.01

Very lenient

~22,000

Rationale: Systematic evaluation from highly stringent (Set A) to inclusive (Set D) approaches, capturing different aspects of polygenic architecture.

4. Bias Correction Framework

Problem Identification

UK Biobank participants overlap between GWAS training and PRS validation cohorts creates systematic bias, inflating apparent performance.

EraSOR HapMap 3 Correction Method

Concept: Remove overlapping UK Biobank individuals from training GWAS

Implementation: Participant-level exclusion before PRS derivation

Limitation: Reduces effective sample size and statistical power

EraSOR Pan-UKBB Correction Method

Concept: Retain full GWAS power while adjusting for overlap bias

Implementation: Post-hoc statistical correction maintaining sample size

Advantage: Preserves discovery power while eliminating bias

Critical Finding: Without correction, Set D models show 1.5-3% AUC inflation, potentially misleading clinical implementation.

5. Model Selection Process

Systematic Evaluation Framework

Total Models Tested: 46 combinations

Design: 4 methods × 4 variant sets × 3 bias corrections - 2 incomplete = 46

Validation: Consistent evaluation pipeline across all models

Metrics: AUC with 95% confidence intervals, calibration assessment

Performance Hierarchy

Best Overall: C+T SetD Pan-UKBB (AUC = 0.7383)

Method Ranking: C+T ≥ PRS-CS ≥ LDpred2 ≈ Lassosum

Variant Set Ranking: Set D >> Set C > Set B ≈ Set A

Bias Correction Impact: Essential for Set D, minimal for Sets A-C

Selection Criteria

Primary: Bias-corrected AUC performance

Secondary: Clinical interpretability and implementation feasibility

Robustness: Consistent performance across evaluation metrics

Key Scientific Findings

1. Bias Correction is Essential

Set D models require mandatory bias correction for clinical validity

Pan-UKBB approach preferred over HapMap3 for maintaining statistical power

Quantitative framework established for future UK Biobank PRS studies

2. Polygenic Architecture Insights

More variants = better performance when properly bias-corrected

Diminishing returns beyond ~20,000 variants for current GWAS sizes

Method convergence after bias correction suggests robust signals

Technical Specifications

Computational Environment

Primary Analysis: R 4.3.0 with data.table, dplyr, pROC packages

PRS Software: PLINK 1.9, PRS-CS v1.0.2, LDpred2 v1.0.6, lassosum v0.4.5

Infrastructure: High-performance computing cluster (64 cores, 512GB RAM)

Quality Assurance

Reproducible pipeline: Version-controlled analysis scripts

Cross-validation: Consistent results across independent runs

Sensitivity analysis: Robust to parameter variations

Code availability: Analysis pipeline documented and available upon request

Statistical Framework

Base Model: age + sex + 10 genetic PCs

Evaluation: Logistic regression with ROC analysis

Confidence Intervals: Bootstrap resampling (n = 1,000)

Significance Testing: Paired t-tests for cross-method comparisons

Usage Instructions

Loading the PRS Model

r

# Load required libraries

library(data.table)

library(pROC)

# Load PRS scores

prs_scores <- fread("clump_thresh_panukbb_setD.sscore")

# Load phenotypes and covariates 

phenotypes <- fread("combined_ascvd_pheno.txt")

covariates <- fread("combined_ascvd_covariates.txt")

# Merge datasets

analysis_data <- merge(merge(prs_scores, phenotypes, by = c("FID", "IID")),

                      covariates, by = c("FID", "IID"))

Model Implementation

r

# Standardize PRS

analysis_data$PRS_std <- scale(analysis_data$SCORE1_AVG)[,1]

# Fit prediction model

model <- glm(ASCVD_INCIDENT ~ age_bl + sex + PC1 + PC2 + PC3 + PC4 + PC5 +

             PC6 + PC7 + PC8 + PC9 + PC10 + PRS_std,

             family = "binomial", data = analysis_data)

# Calculate performance

predictions <- predict(model, type = "response")

roc_result <- roc(analysis_data$ASCVD_INCIDENT, predictions)

auc_value <- auc(roc_result)

Applying to New Data

r

# Load SNP weights

weights <- fread("weights/clump_thresh/panukbb_setD_weights.txt")

# Calculate PRS for new individuals (requires genotype data)

# Implementation depends on your genotype format (PLINK, VCF, etc.)

Validation and Performance

Primary Validation Cohort

Dataset: UK Biobank European ancestry participants

Sample Size: 435,502 individuals

ASCVD Cases: 31,854 incident events

Follow-up: Median 12.3 years

Age Range: 37-73 years at baseline

Performance Metrics

C-statistic (AUC): 0.7383 [0.7358-0.7409]

Improvement over base model: +0.0202 AUC units

Net Reclassification Index: 0.0156 [0.0134-0.0178]

Calibration: Hosmer-Lemeshow p = 0.23 (well-calibrated)

Risk Stratification

PRS Percentile

ASCVD Risk

Hazard Ratio

95% CI

<10th

5.2%

1.00 (ref)

-

10th-50th

7.1%

1.37

[1.31-1.44]

50th-90th

8.9%

1.72

[1.64-1.80]

>90th

12.4%

2.39

[2.25-2.54]

Acknowledgments

Data: ASCVD iPRS Model v1.0. Zenodo. DOI: 10.5281/zenodo.17077087

Funding

This project has received funding from the European Union’s Horizon Europe (European Innovation Council) programme under grant agreement No 101115381 MIRACLE.

Computing Resources

Analysis performed on Leibniz Rechenzentrum high-performance computing infrastructure.

Contact Information

Principal Investigator: Martin Dichgans, Rainer Malik, Ling Li, Heribert SchunkertInstitution: Institute for Stroke and Dementia Research, Munich, GermanyEmail: [e-mail] Website: https://sites.uef.fi/miracle/

Version History

v1.0 (2025): Initial release

46 models systematically evaluated

Bias correction framework established

Best model: C+T SetD Pan-UKBB (AUC = 0.7383)

Comprehensive validation in UK Biobank

License

This dataset is made available under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You are free to share and adapt the material for any purpose, provided you give appropriate credit.