← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / OpenAI4S / skills / bioskills / bio-clinical-databases-polygenic-risk / references / usage-guide.md

Polygenic Risk Scores - Usage Guide

Overview

Construct and validate polygenic risk scores using the 2026 method landscape (LDpred2-auto, SBayesRC, MegaPRS, PRS-CS, PROSPER, MUSSEL, BridgePRS, JointPRS, PRSmix) and the PGS Catalog Calculator. Covers cross-ancestry portability with Ding 2023 continuous-ancestry framing, Hingorani 2023 BMJ Medicine performance critique, PRS-RS reporting standards, EraSOR sample-overlap detection, and the 11-item reviewer audit. Regulatory state: no general FDA PRS guidance exists as of 2026; August 2025 Federal Register Cancer Predisposition Risk Assessment System device classification is the operative text.

Prerequisites

# Production multi-PGS pipeline
nextflow pull pgscatalog/pgsc_calc

# LDpred2 (R)
# install.packages('bigsnpr')

# PRS-CS / PRS-CSx
git clone https://github.com/getian107/PRScs.git
git clone https://github.com/getian107/PRScsx.git
pip install scipy h5py

# SBayesRC (GCTB)
# Download from https://cnsgenomics.com/software/gctb/
# Need UKB eigenvalue-decomposed LD + baselineLD_v2.2 annotations

# Multi-ancestry
# PROSPER: https://github.com/Jingning-Zhang/PROSPER
# MUSSEL: https://github.com/Jin93/MUSSEL
# BridgePRS: https://github.com/clivehoggart/BridgePRS
# JointPRS: https://github.com/LeqiXu/JointPRS

# Sample overlap
# Choi 2023 EraSOR: https://github.com/choishingwan/EraSOR
# LDSC: pip install ldsc

Quick Start

Tell the agent what to do: - "Compute LDpred2-auto PRS for CAD using Aragam 2022 GWAS sumstats. Run EraSOR for sample-overlap detection first." - "Use pgsc_calc to compute PGS000004 + PGS001775 + PGS000019 on this cohort with auto-liftover and ancestry projection" - "Multi-ancestry breast-cancer PRS using PRS-CSx with EUR + AFR + EAS sumstats. Tune phi for sparse vs polygenic architecture." - "For my T2D PRS, apply Ding 2023 continuous-ancestry framing: report R^2 vs PC distance, NOT vs discrete ancestry boxes." - "Audit this PRS manuscript against the 11-item reviewer checklist (PRS-RS items 13, 16, 19, 21 priority)."

Example Prompts

Single-Ancestry Construction

"Compute LDpred2-auto PRS for breast cancer in this UKB-NFE cohort. Use UKB-LD reference (not 1KG-EUR); pin allow_jump_sign=FALSE, shrink_corr=0.95; report s LD-mismatch parameter."

"Use SBayesRC with baselineLD v2.2 annotations + UKB eigen-LD for this lipid panel cohort. Run --impute-summary first."

"Run MegaPRS on T2D sumstats with the BLD-LDAK heritability model; select the best auto-model on training set."

Multi-Ancestry

"Compute PRS-CSx joint posterior for CAD using EUR (n=200k) + AFR (n=30k) + EAS (n=50k) sumstats. Phi=1e-2 for highly polygenic; report cross-ancestry R^2."

"For this admixed cohort, run MUSSEL with multivariate spike-and-slab + super-learner ensemble. Compare to BridgePRS for low-h^2 settings."

"JointPRS-meta with small tuning set; compare to PRS-CSx auto-tuning."

PGS Catalog Production

"Run pgsc_calc on this cohort VCF with PGS000004 (CAD-Khera-2018), PGS001775 (CAD-Aragam-2022), PGS003725 (BC-Mavaddat-2019). Auto-liftover GRCh37 -> 38; project to HGDP+1KG references for ancestry."

Calibration and Reporting

"For my CAD PRS, integrate over BOADICEA-like age-conditional incidence to produce absolute lifetime risk; flag the Hingorani 2023 detection-rate-vs-FPR caveat."

"Apply Ding 2023 continuous-ancestry framework: for each test sample, compute genetic distance to discovery cohort and report R^2 binned by quintile of distance."

Audit / Reproducibility

"Audit this PRS paper against PRS-RS 22-item checklist. Priority items: cohort independence (#13), confounder adjustment (#16), absolute-risk reporting (#19), ancestry composition of validation (#21)."

"Run EraSOR / bivariate LDSC intercept on this PRS application; report whether |intercept| > 0.05 flags substantial overlap."

What the Agent Will Do

  1. Choose method by data: SBayesRC for EUR + functional annotations; LDpred2-auto for EUR generic; PRS-CSx / PROSPER / MUSSEL / BridgePRS / JointPRS for multi-ancestry; PRSice-2 for legacy C+T comparison.
  2. Match variants strand-aware (LDpred2 snp_match with frequency-flip; PRSice-2 drops A/T C/G).
  3. Use UKB-LD reference (not 1KG-EUR) for EUR; ancestry-matched for non-EUR.
  4. Run EraSOR / bivariate LDSC intercept to detect sample overlap; require |intercept| < 0.05 with target n >= 1000.
  5. Exclude HLA region (chr6 28-34 Mb) from main PRS; model classical HLA alleles separately for autoimmune traits.
  6. Recalibrate ancestry-conditional Z using TEST cohort PCs (not discovery PCs).
  7. Transform PRS percentile to absolute risk via external age-conditional incidence curve.
  8. Apply Hingorani 2023 caveats: HR/OR per SD (~1.3) is comparable to family history alone; PRS does not meet population-screening utility threshold.
  9. Cite methods correctly: PRSmix and MUSSEL are Cell Genomics; PROSPER is Nat Commun; Mavaddat 2023 is CEBP; Mullins 2021 is bipolar (not MDD).

Tips