tooluniverse-polygenic-risk-score — скилл для Claude Code: PRS по GWAS

---
name: tooluniverse-polygenic-risk-score
description: Build and interpret polygenic risk scores (PRS) for complex diseases using GWAS summary statistics. Covers PRS construction (clumping/thresholding, PRS-CS), validation in independent cohorts, ancestry-aware adjustment, and clinical interpretation (population-relative risk, not absolute prediction). Use for PRS-based risk stratification.
disable-model-invocation: true
---

# Polygenic Risk Score (PRS) Builder

Build and interpret polygenic risk scores for complex diseases using genome-wide association study (GWAS) data.

## Reasoning Strategy

A polygenic risk score predicts genetic risk, not disease. A high PRS means elevated risk relative to the population — it does not mean the person will develop the condition, and a low PRS does not confer immunity. PRS performance varies dramatically across ancestries: a European-derived PRS applied to a West African population can lose 50–70% of its predictive power because the underlying GWAS was trained on European allele frequencies and LD patterns. Effect sizes from discovery GWAS are subject to winner's curse (overestimation in single studies); always prefer weights from large meta-analyses or validated PGS Catalog models. PRS should always be interpreted in the context of non-genetic risk factors — for most complex diseases, environmental factors contribute as much or more than genetics.

**LOOK UP DON'T GUESS**: Do not assume effect sizes, allele frequencies, or which SNPs are genome-wide significant for a trait — always query GWAS Catalog (`gwas_get_associations_for_trait`) for actual data. Do not assume a validated PRS model exists for a trait; check PGS Catalog via PubMed search.

## Overview

**Use Cases:**
- "Calculate my genetic risk for type 2 diabetes"
- "Build a polygenic risk score for coronary artery disease"
- "What's my genetic predisposition to Alzheimer's disease?"
- "Interpret my PRS percentile for breast cancer risk"

**What This Skill Does:**
- Extracts genome-wide significant variants (p < 5e-8) from GWAS Catalog
- Builds weighted PRS models using effect sizes (beta coefficients)
- Calculates individual risk scores from genotype data
- Interprets PRS as population percentiles and risk categories

**What This Skill Does NOT Do:**
- Diagnose disease (PRS is probabilistic, not deterministic)
- Replace clinical assessment or genetic counseling
- Account for non-genetic factors (lifestyle, environment)
- Provide treatment recommendations

## Methodology

### PRS Calculation Formula

A polygenic risk score is calculated as a weighted sum across genetic variants:

```
PRS = Σ (dosage_i × effect_size_i)
```

Where:
- **dosage_i**: Number of effect alleles at SNP i (0, 1, or 2)
- **effect_size_i**: Beta coefficient or log(odds ratio) from GWAS

### Standardization

Raw PRS is standardized to z-scores for interpretation:

```
z-score = (PRS - population_mean) / population_std
```

This allows comparison to population distribution and percentile calculation.

### Significance Thresholds

- **Genome-wide significance**: p < 5×10⁻⁸ (default threshold)
- This corrects for ~1 million independent tests across the genome
- Relaxed thresholds (e.g., p < 1×10⁻⁵) can include more SNPs but may add noise

### Effect Size Handling

- **Continuous traits** (e.g., height, BMI): Beta coefficient (units of trait per allele)
- **Binary traits** (e.g., disease): Odds ratio converted to log-odds (beta = ln(OR))
- Missing effect sizes or non-significant SNPs are excluded

## Data Sources

This skill uses ToolUniverse GWAS tools to query:

1. **GWAS Catalog** (EMBL-EBI)
   - Curated GWAS associations, 5000+ studies
   - Tools: `gwas_search_associations` (param: `disease_trait`, `size`; also `gwas_get_associations_for_trait`), `gwas_get_snps_for_gene` (param: `gene_symbol`), `dbsnp_get_variant_by_rsid`
   - Note: `disease_trait` search returns associations where the trait is one of potentially several linked EFO traits. For precise filtering, use EFO IDs via `efo_trait` param.

2. **Open Targets Genetics**
   - Integrated genetics platform with fine-mapped credible sets
   - Tools: `OpenTargets_search_gwas_studies_by_disease`, `EnsemblVEP_annotate_hgvs` (for variant consequence/frequency)

3. **Variant Annotation**
   - `gnomad_search_variants` + `gnomad_get_variant` — population allele frequencies (ancestry-specific via VEP colocated_variants)
   - `MyVariant_query_variants` — CADD, SIFT, PolyPhen, ClinVar, gnomAD in one call
   - `gnomad_get_gene_constraints` — gene constraint metrics (pLI, oe_lof) for target prioritization

4. **PGS Catalog** (pgscatalog.org, EMBL-EBI) — **check this FIRST, before building anything from raw GWAS associations**
   - This skill's own Best Practices already say "use validated PRS from PGS Catalog when available" — these are the tools that do it. No API key needed.
   - `PGSCatalog_search_traits(query=...)` — resolve a free-text trait name to its real EFO/MONDO `trait_id` (e.g. "type 2 diabetes" -> `MONDO_0005148`) and see how many published scores exist. **Always use the `trait_id` the tool returns** — don't guess or reuse a remembered EFO/MONDO id, since the search can resolve to a different (but correct) ontology term than expected.
   - `PGSCatalog_get_scores_by_trait(trait_id=...)` — list every published PGS for that trait: PGS id, variant count, development method (LDpred, P+T, genome-wide-significant-variants, etc.), genome build, and source publication. A trait can have 50+ scores of wildly different size and quality (e.g. type 2 diabetes: from a 62-variant 2014-era score to a 6.9-million-variant LDpred model) — don't assume the first result is the best one.
   - `PGSCatalog_get_score(pgs_id=...)` — full metadata for one score, including its `ancestry_distribution` for both the development (GWAS) and evaluation cohorts. **Read this before trusting a score**: a score developed on 100% European samples (`ancestry_distribution.dev.dist: {"EUR": 100}`) is exactly the transferability risk this skill warns about above, even if it was later evaluated in more diverse cohorts.
   - `PGSCatalog_get_performance_metrics(pgs_id=...)` — the actual published validation numbers per evaluation cohort: AUROC/C-index, hazard/odds ratios (including "top 10% vs. rest" comparisons), R², **each tied to the specific ancestry and sample size it was measured in**. This is how you judge whether a score is trustworthy for a given population — a high AUROC measured only in a European cohort tells you nothing about performance in other ancestries; look for an evaluation row matching your population of interest, or note its absence.

## Key Concepts

### Polygenic Risk Scores (PRS)

Polygenic risk scores aggregate the effects of many genetic variants to estimate an individual's genetic predisposition to a trait or disease. Unlike Mendelian diseases caused by single mutations, complex diseases involve hundreds to thousands of variants, each with small effects.

**Key Properties:**
- **Continuous distribution**: PRS forms a bell curve in populations
- **Relative risk**: Compares individual to population average
- **Probabilistic**: High PRS doesn't guarantee disease, low PRS doesn't guarantee protection
- **Ancestry-specific**: PRS accuracy depends on matching GWAS and target ancestry

### GWAS (Genome-Wide Association Studies)

GWAS compare allele frequencies between cases and controls (or correlate with trait values) across millions of SNPs to identify disease-associated variants.

**Study Design:**
- **Discovery cohort**: Initial identification of associations
- **Replication cohort**: Validation in independent samples
- **Sample size**: Larger studies detect smaller effects (power ∝ √N)
- **Multiple testing correction**: Bonferroni-type correction for ~1M tests

### Effect Sizes and Odds Ratios

- **Beta (β)**: Change in trait per copy of effect allele
  - Example: β = 0.5 kg/m² means each allele increases BMI by 0.5 units
- **Odds Ratio (OR)**: Multiplicative change in disease odds
  - OR = 1.5 means 50% increased odds per allele
  - Convert to beta: β = ln(OR)

### Linkage Disequilibrium (LD) and Clumping

Nearby variants are often inherited together (LD). To avoid double-counting:
- **LD clumping**: Select independent variants (r² < 0.1 within 1 Mb windows)
- **Fine-mapping**: Statistical methods to identify causal variants
- This skill uses raw associations; production PRS should include LD pruning

### Population Stratification

GWAS and PRS are most accurate when ancestries match:
- **Population structure**: Different ancestries have different allele frequencies
- **Transferability**: European-trained PRS perform worse in non-European populations
- **Solution**: Train PRS on diverse cohorts or use ancestry-matched references

## Applications

### Clinical Risk Assessment

PRS can stratify individuals for:
- **Screening programs**: Target high-risk individuals (e.g., mammography, colonoscopy)
- **Prevention strategies**: Lifestyle interventions for high genetic risk
- **Drug response**: Pharmacogenomics based on metabolism genes

**Example**: Khera et al. (2018) showed PRS identifies 3× more individuals at >3-fold coronary artery disease risk than monogenic mutations.

### Research Applications

- **Gene discovery**: PRS-based phenome-wide association studies (PheWAS)
- **Genetic correlation**: Compare PRS across traits — but for a rigorous, GWAS-summary-statistics estimate of cross-trait genetic correlation (rg), use `run_ldsc_genetic_correlation` (LD Score regression), which needs only summary stats (no individual genotypes) and corrects for sample overlap. Far more principled than correlating PRS values.
- **Causal inference**: Mendelian randomization using PRS as instruments
- **Simulation studies**: Model polygenic architecture

### SNP-heritability and genetic correlation (LDSC)

Before or alongside building a PRS, quantify how much of the trait is captured by common SNPs and how traits relate — directly from GWAS summary statistics:

- `run_ldsc_heritability` — SNP-based heritability (h²_SNP) from one GWAS's summary stats; the intercept also flags confounding/inflation vs. true polygenicity. This sets the ceiling a PRS can reach (the "heritability gap" below is exactly h²_SNP minus PRS R²).
- `run_ldsc_genetic_correlation` — genetic correlation (rg) between two GWAS, for shared-aetiology and cross-trait PRS questions.

Both are remote tools (LD Score regression engine + reference LD-score panels). Use them to ground heritability/rg claims in data rather than citing literature point estimates.

### Personal Genomics

Consumer genetic testing (23andMe, Ancestry DNA) provides raw genotypes. Users can:
- Calculate PRS for traits not reported
- Compare to published PRS models
- Understand genetic contribution vs. lifestyle factors

**Caution**: Personal PRS should not replace medical advice. Results may cause anxiety if not properly contextualized.

## Limitations and Considerations

- **Heritability gap**: PRS explains only a fraction of genetic heritability (T2D: ~50% heritable, PRS explains ~10–20%). Rare variants, epistasis, and gene-environment interactions are not captured.
- **Ancestry bias**: European-derived PRS performance drops substantially in non-European populations. Use multi-ancestry GWAS weights when available.
- **Winner's curse**: Discovery effect sizes are overestimated; use meta-analysis weights or PGS Catalog validated models.
- **Not diagnostic**: High PRS does not guarantee disease; low PRS does not guarantee protection. Environmental factors contribute equally or more for most complex diseases.
- **Actionability varies**: Alzheimer's PRS has limited actionable interventions; cardiovascular PRS can guide statin or lifestyle decisions. Always consider what the person can do with the information.
- **Ethical**: Genetic data is permanent and familial. GINA protects employment/health insurance in the US, but not life insurance. Provide genetic counseling context.

## Workflow

### 0. Check PGS Catalog Before Building From Raw GWAS

Before extracting associations and building a PRS from scratch (steps 1-5 below), check whether a validated score already exists:

```
PGSCatalog_search_traits(query="<trait>")            # resolve to a real trait_id
PGSCatalog_get_scores_by_trait(trait_id="<from above>")   # list candidate scores
PGSCatalog_get_score(pgs_id="<candidate>")                 # check dev-cohort ancestry
PGSCatalog_get_performance_metrics(pgs_id="<candidate>")   # check validated AUROC/OR/R2 by ancestry
```

If a well-validated score exists for the target population's ancestry, use its published weights (`scoring_file` URL) directly rather than re-deriving a cruder PRS from raw GWAS Catalog associations — this is exactly the "winner's curse" and LD-clumping problem the raw-GWAS path (steps 1-5) has to work around manually. Only fall back to building from raw associations when PGS Catalog has no score for the trait, or when every existing score was developed and evaluated in an ancestry that doesn't match the population you're scoring.

### 1. Trait Selection

Identify the disease or trait of interest:
- Use standard terminology (e.g., "type 2 diabetes" not "T2D")
- Check GWAS Catalog for availability
- Verify sufficient GWAS studies exist (n > 10,000 samples ideal)

### 2. Association Collection

Query GWAS databases for genome-wide significant associations:
```python
prs = build_polygenic_risk_score(
    trait="coronary artery disease",
    p_threshold=5e-8,  # Genome-wide significance
    max_snps=1000
)
```

**Considerations:**
- P-value threshold: 5e-8 is conservative, 1e-5 includes more variants
- LD clumping: Production systems should prune correlated SNPs
- Study quality: Prefer large meta-analyses over small studies

### 3. Effect Size Extraction

Extract beta coefficients or odds ratios:
- Beta for continuous traits (direct use)
- OR for binary traits (convert to log-odds)
- Handle missing values (exclude or impute from meta-analysis)

### 4. SNP Filtering

Quality control filters:
- **MAF filter**: Exclude rare variants (MAF < 0.01) for robustness
- **Genotype QC**: Remove SNPs with high missingness (> 10%)
- **Hardy-Weinberg**: Exclude SNPs violating HWE (p < 1e-6)
- **Ambiguous SNPs**: Remove A/T and G/C SNPs (strand ambiguity)

### 5. Score Calculation

Calculate weighted sum of genotype dosages:
```python
result = calculate_personal_prs(
    prs_weights=prs,
    genotypes=my_genotypes,
    population_mean=0.0,
    population_std=1.0
)
```

**Genotype Sources:**
- 23andMe raw data export
- Ancestry DNA raw data
- Whole genome sequencing (VCF files)
- SNP array data (Illumina, Affymetrix)

### 6. Risk Interpretation

Convert to percentiles and risk categories:
```python
result = interpret_prs_percentile(result)
print(f"Percentile: {result.percentile:.1f}%")
print(f"Risk: {result.risk_category}")
```

**Risk Categories:**
- **Low risk**: < 20th percentile (genetic protection)
- **Average risk**: 20-80th percentile (typical genetic predisposition)
- **Elevated risk**: 80-95th percentile (moderately increased risk)
- **High risk**: > 95th percentile (substantially increased risk)

**Clinical Interpretation:**
- Percentiles assume normal distribution
- Relative risk vs. average (not absolute risk)
- Combine with family history, clinical risk factors
- PRS is NOT diagnostic - many high-risk individuals never develop disease

## Best Practices

- Use validated PRS from PGS Catalog when available (externally validated, includes LD clumping and ancestry-specific weights)
- Match ancestries between GWAS and target population; use multi-ancestry GWAS when available
- For highly polygenic traits (height, education), relaxed p-value thresholds capture more signal; for oligogenic traits (IBD, T1D), strict thresholds are better
- Combine PRS with clinical risk scores (Framingham, QRISK) for integrated prediction
- In research: document SNP selection criteria, LD clumping parameters, and ancestry of GWAS; validate in held-out cohorts; report R² or AUC stratified by ancestry

## Disclaimer

**This skill is for educational and research purposes only.**

- **Not for clinical diagnosis or treatment decisions**
- **Not validated for clinical use** - use PGS Catalog models for clinical-grade PRS
- **Requires genetic counseling** - interpretation requires expertise
- **Does not account for family history, environment, or lifestyle factors**
- **Ancestry-specific** - accuracy depends on matching GWAS ancestry

**For clinical genetic testing, consult:**
- Genetic counselors (certified by ABGC/ABMGG)
- Medical geneticists
- Healthcare providers with genomics training

PRS is a rapidly evolving field. Guidelines and best practices will continue to change as research progresses.

**Regulatory Status:**
- FDA does not currently regulate PRS (as of 2024)
- Some countries restrict direct-to-consumer genetic risk reporting
- Check local regulations before clinical implementation