---

A Guide to Performing Polygenic Risk Score Analyses (2019)

1. Problem & Tasks

• Practical tutorial for constructing, validating, and reporting polygenic risk scores (PRS).
• Relevant for defining covariate controls and baseline genetics features when combining with MRI.

2. Datasets

• Examples drawn from UK Biobank and PGC cohorts; outlines required inputs: GWAS summary stats, target genotypes.
• Emphasizes ancestry-matched target sets and QC thresholds (MAF, INFO, Hardy–Weinberg).

3. Model / Method Details

3.1 Pipeline

• QC target genotypes → LD clumping/thresholding or more advanced methods (LDPred, PRS-CS).
• Calculate PRS per subject, standardize, residualize vs covariates.
• Evaluate predictive performance with logistic regression or linear regression depending on phenotype.

3.2 Confound Handling & Evaluation Discipline

• Always include ancestry PCs (≥10), sex, age, batch in regression models.
• Use nested CV when tuning PRS hyperparameters to avoid overfitting.
• Report incremental R² / pseudo-R² and calibration metrics.

4. Results & Tables

• Provides illustrative tables showing how PRS performance varies with clumping parameters and ancestry mismatches (drops of 30–50% accuracy when mismatched).

5. Limitations & Cautions

• Focused on SNP array + European ancestry; does not directly cover WES/WGS nuance.
• Methods (clumping + thresholding) may be superseded by PRS-CS, but workflow guidance still valid.

---

6. Hooks into Neuro-Omics KB

Relevant KB assets

• kb/paper_cards/prs_guide.yaml
• kb/datasets/ukb_manifest_stub.yaml (covariate list, PC requirements).

Configs / recipes informed

• Future baseline where PRS is an additional modality alongside gene embeddings and MRI.
• Provides justification for including PCs, age, sex, site covariates in configs/experiments/*.

Concrete guidance for our project

• Whenever PRS features are added, include at least 10 ancestry PCs + batch covariates in modeling (per guide).
• Use nested CV for PRS hyperparameters; avoid leaking test folds.
• Document QC thresholds and summary-stat sources in dataset cards, referencing this guide.