bio-clinical-databases-polygenic-risk

What this skill does

## Version Compatibility

Reference examples tested with: bigsnpr 1.12+ (LDpred2; Privé 2020), PRSice-2 2.3.5+, PRS-CS 1.0.0+ (Ge 2019), gctb 2.5+ (SBayesR/SBayesS/SBayesRC; Zheng 2024), LDAK 6.0+ (MegaPRS; Zhang 2021), pgsc_calc 2.0+ (nf-core; Lambert 2024), Hail 0.2.130+, numpy 1.26+, pandas 2.2+. No general FDA PRS guidance document exists as of May 2026; the operative regulatory text is the August 2025 Federal Register notice on Cancer Predisposition Risk Assessment Systems (Class II device with special controls).

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- R: `packageVersion('<pkg>')` then `?function_name`
- CLI: `<tool> --version` then `<tool> --help` to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying. The LDpred2-auto `snp_ldpred2_auto()` signature changed in bigsnpr 1.11+; pin `allow_jump_sign = FALSE` and `shrink_corr = 0.95` explicitly.

# Polygenic Risk Scores; Construction, Calibration, Reporting

**'Compute a PRS for my cohort using these GWAS summary statistics'** -> Match variants to target genotypes, choose method by data availability + trait architecture, derive LD-aware effect estimates, score, normalize by ancestry, transform to absolute risk.

- CLI (recommended one-stop): `pgsc_calc --target target.vcf --pgs_id PGS000001` (nf-core, Lambert 2024)
- R (SOTA single-ancestry): `bigsnpr::snp_ldpred2_auto()` (Privé 2020)
- CLI (multi-ancestry SOTA): `PRS-CSx`, `PROSPER`, `MUSSEL`, `BridgePRS`, `JointPRS`
- CLI (SBayesRC with functional annotations): `gctb --sbayes-rc --bfile target --gwas-summary sumstats.ma`
- CLI (legacy baseline): `PRSice_linux` (clumping + thresholding; still cited for some clinical scores)

## Method Landscape: 2026 Operational Ranking

| Method | Approach | Best for | Fails when |
|--------|----------|----------|-----------|
| **SBayesRC** (Zheng 2024 *Nat Genet*) | Bayesian + 96 functional annotations | EUR; sparse traits | Sumstats LD-incoherent with reference; chain divergence (run --impute-summary first) |
| **MegaPRS** (Zhang 2021 *Nat Commun*) | BLD-LDAK heritability model | EUR; sparser traits | Lacks GCTA-model assumptions; legacy GCTA pipelines |
| **LDpred2-auto** (Privé 2020) | Bayesian + auto-tuning | Polygenic EUR | LD ref mismatch (s > 0.05); allow_jump_sign default True (must pin FALSE) |
| **PRS-CS-auto** (Ge 2019 *Nat Commun*) | Continuous-shrinkage prior | Polygenic EUR | Sparse-trait architecture; HapMap3-restricted variants only |
| **lassosum2** (Privé 2022) | Penalized regression | EUR alternative | Highly polygenic (Bayesian methods better); requires tuning data |
| **C+T** (PRSice-2; Choi 2019) | Clumping + thresholding | Legacy clinical scores (PRS313) | Highly polygenic; Bayesian methods dominate |
| **PROSPER** (Zhang 2024 *Nat Commun*) | Ensemble penalized regression | Multi-ancestry, AFR + others | Single-ancestry; tuning set < 1000 |
| **MUSSEL** (Jin 2024 *Cell Genomics*) | Spike-slab + super-learner | Multi-ancestry; admixed AFR | Single-ancestry; lacks tuning data |
| **JointPRS** (Hu S et al 2025 *Nat Commun*) | Data-adaptive Bayesian | Multi-ancestry; sumstats only | Single-ancestry; very small target. (Verify exact volume/page reference in the latest Nat Commun citation; the earlier "16:59243" attribution appears implausible.) |
| **PRS-CSx** (Ruan 2022 *Nat Genet*) | PRS-CS multi-ancestry extension | Multi-ancestry with EUR + non-EUR sumstats | Low causal-variant overlap across ancestries |
| **BridgePRS** (Hoggart 2024 *Nat Genet*) | Ridge-bridge sharing | Low-h^2 AFR / low causal overlap | Standard scenarios (PROSPER/MUSSEL win) |
| **PolyPred / PolyPred+** (Weissbrod 2022) | BOLT-LMM + PolyFun-SuSIE | Multi-ancestry; biobank-scale | Small individual-level data; expensive |

**Citation traps caught by senior PIs:**
- **PRSmix** (Truong 2024); *Cell Genomics*, NOT *Nat Genet*.
- **MUSSEL** (Jin 2024); *Cell Genomics*.
- **PROSPER** (Zhang 2024); *Nat Commun*, NOT *Nat Genet*.
- **Hingorani 2023**; *BMJ Medicine* (NOT main *BMJ*).
- **Mavaddat 2023 BOADICEA update**; *Cancer Epi Biomark Prev*, NOT *Nat Genet*.
- **Mullins 2021** is bipolar disorder, NOT MDD (Howard 2019 *Nat Neurosci* = MDD; Mullins 2021 = BD).

## Multi-Ancestry: The Big Problem

Martin 2017 *Nat Genet* established the 4.5x R^2 attenuation between EUR and AFR. Updates:
- **Martin 2019** *Nat Genet*: "Clinical use of current polygenic risk scores may exacerbate health disparities".
- **Mostafavi 2020** *eLife*: PRS accuracy varies even within a single ancestry due to age, sex, SES, GxE.
- **Ding 2023** *Nature* 618: PGS accuracy decays *continuously* along genetic-ancestry continuum (Pearson r = -0.95 vs PC distance from training data). **The 2026 standard is to report PRS performance vs continuous PC distance, NOT discrete ancestry boxes.**
- **Hou 2023** *Nat Commun*: theoretical decay bounds in MAF/h^2/effect-correlation across ancestries.

Multi-ancestry method choice:
1. Individual-level non-EUR training data + tuning set >= 1000: **PROSPER** or **MUSSEL** (top performance).
2. Summary statistics only + small tuning: **JointPRS** or **PRS-CSx**.
3. Low h^2 / very polygenic / low causal overlap: **BridgePRS**.
4. Functional annotations critical + EUR-dominant: **SBayesRC** (cross-ancestry via `--ldm-eigen`).

## Calibration: The Hingorani Reframing

Khera 2018 *Nat Genet* established the clinical-PRS narrative; **Hingorani 2023** *BMJ Med* is the operative critique:

- HR per SD typically lies between 1.3 and 1.7; **similar to family history alone**.
- Among individuals who develop disease, only ~11% are detected at conventional high-risk PRS threshold; 5% false-positive rate.
- CAD top-2.5% PRS captures 7% of cases; breast-cancer top-quintile captures 6%.
- Wald-Hingorani detection-rate / false-positive-rate ratios approach 10:1 for screening; current PRS achieve 2-3:1.

**Calibration mechanics:**
- Cross-ancestry calibration breaks for the *variance* of the PRS distribution, not just the mean (Sun 2021 *Genome Med*).
- Recalibration: subtract conditional mean given first 4-10 PCs; divide by conditional SD; convert to percentile. **Compute PCs in the test cohort, NOT discovery-cohort PCs**.
- For absolute risk: integrate over external incidence curve (BOADICEA v6 / CanRisk for breast cancer; FOS for CAD).

## PRS-RS Reporting Standards (Wand 2021 *Nature*)

22-item checklist. Reviewer-priority items: cohort independence between development + evaluation (item 13); confounder adjustment in evaluation (item 16); absolute-risk reporting (item 19); ancestry composition of validation (item 21). *Nature Genetics*-tier manuscripts without PRS-RS adherence are rejected at review.

## Decision Tree by Scenario

| Scenario | Recommended path | Why |
|----------|------------------|-----|
| EUR cohort + individual-level data | LDpred2-auto or SBayesRC | SBayesRC integrates functional annotations |
| EUR cohort + sumstats only | MegaPRS or SBayesRC | LDpred2 also viable |
| Highly polygenic trait (height, BMI, education) | LDpred2-auto, PRS-CS-auto | Continuous-shrinkage priors well-suited |
| Sparse trait (lipids, AMD) | MegaPRS, SBayesRC | BLD-LDAK or SBayesR mixture priors |
| Multi-ancestry, large tuning set | PROSPER or MUSSEL | Top performance per benchmarks |
| Multi-ancestry, sumstats + small tuning | JointPRS or PRS-CSx | Joint Bayesian framework |
| Multi-ancestry, target = AFR | MUSSEL or BridgePRS | Best non-EUR performance |
| Combining multiple PGS Catalog scores | PRSmix (single trait) or PRSmix+ (cross-trait) | Elastic-net combination |
| Production score for biobank | `pgsc_calc` Nextflow nf-core | Handles liftover, ancestry, normalization automatically |
| No tuning data available | LDpred2-auto, PRS-CS-auto, JointPRS-auto | Bayesian auto-tuning |
| Clinical