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

name: bio-clinical-databases-polygenic-risk description: Constructs and validates polygenic risk scores using LDpred2-auto, SBayesRC, MegaPRS, PRS-CS, PROSPER, MUSSEL, BridgePRS, JointPRS, PRSmix, or PGS Catalog Calculator with ancestry-aware reference panels (HapMap3, UKB-LD), ancestry-conditional calibration, and PRS-RS reporting standards. Use when computing PRS for cohorts, applying absolute-risk transformation, assessing cross-ancestry portability (Martin 2017 / Ding 2023 continuous ancestry), or auditing PRS manuscripts against the 22-item PRS-RS reviewer checklist. origin: openai4s category: bioskills/clinical-databases metadata: tool_type: mixed primary_tool: PGS Catalog Calculator third_party: name: GPTomics/bioSkills repository: https://github.com/GPTomics/bioSkills commit: d91ed3d563019e649dc854c56ccd62551359488a license: MIT

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.

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 (Xu L et al 2025 Nat Commun 16:3841) Data-adaptive Bayesian Multi-ancestry; sumstats only Single-ancestry; very small target
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 2019 Nat Genet established the ~4.5x R^2 attenuation between EUR and AFR (Martin 2017 AJHG first showed demographic history drives it). 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 Genet: causal effects are similar across local ancestries within admixed individuals (radmix ~0.95), supporting cross-population transfer.

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:

Calibration mechanics: - Cross-ancestry calibration breaks for the variance of the PRS distribution, not just the mean. - 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 v5 / 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 reporting Apply Hingorani-aware framing: absolute-risk transform via external incidence curve HR/OR per SD (~1.3) alone is inadequate for screening

Standard Workflow: LDpred2-auto

Goal: Compute LDpred2-auto PRS from sumstats + target genotypes with appropriate LD reference and sample-overlap detection.

Approach: Use bigsnpr R package; match variants strand-aware; compute LD or use prebuilt UK Biobank LD; run snp_ldpred2_auto() with shrinkage and jump-sign guards; assess s parameter for LD mismatch.

library(bigsnpr)
library(data.table)

# Load target genotypes (PLINK .bed/.bim/.fam -> bigSNP object .rds)
# snp_readBed('target.bed', 'target.rds')
obj <- snp_attach('target.rds')
G <- obj$genotypes
map <- obj$map

# Load GWAS summary stats (standardized format)
sumstats <- fread('gwas_sumstats.txt')
# Required columns: chr, pos, a0 (ref), a1 (effect), beta, beta_se, n_eff, p

# Match variants strand-aware (snp_match handles A/T C/G ambiguity)
df_beta <- snp_match(sumstats, map, strand_flip = TRUE)

# Compute LD correlation matrix (in-sample) OR use prebuilt UKB LD.
# For a 3 cM window, pass the cM positions via `infos.pos = CHR_POS_CM` and set
# `size = 3`. Writing `size = 3/1000` is silently broken because 0.003 rounds to 0.
corr <- snp_cor(G, ind.col = df_beta[['_NUM_ID_']],
                infos.pos = df_beta[['cM']],
                size = 3,  # 3 cM window when infos.pos is in cM
                ncores = 8)

# LDSC heritability estimate + LD mismatch (s) diagnostic
ldsc_res <- snp_ldsc2(corr, df_beta)
h2_est <- ldsc_res[['h2']]
ldsc_s <- ldsc_res[['int']]  # intercept; large => sample overlap

# LDpred2-auto; run multiple chains in parallel
multi_auto <- snp_ldpred2_auto(
    corr, df_beta,
    h2_init = h2_est,
    vec_p_init = seq_log(1e-4, 0.2, 30),
    burn_in = 500, num_iter = 200,
    allow_jump_sign = FALSE,        # CRITICAL: pin to FALSE to avoid sign artifacts
    shrink_corr = 0.95,             # LD-shrinkage guard
    ncores = 8
)

# Filter divergent chains
beta_auto <- sapply(multi_auto, function(x) x$beta_est)
range_auto <- sapply(multi_auto, function(x) diff(range(x$corr_est)))
keep <- range_auto > (0.95 * quantile(range_auto, 0.95))
beta_final <- rowMeans(beta_auto[, keep, drop = FALSE])

# Score
pred <- big_prodMat(G, beta_final, ind.col = df_beta[['_NUM_ID_']])

SBayesRC Workflow (Functional Annotations)

Goal: Compute PRS integrating ~96 functional annotations (baseline-LD v2.2) for +14% R^2 over SBayesR.

Approach: GCTB CLI with the SBayesRC algorithm; preprocess sumstats with --impute-summary.

# 1. Impute missing variants in sumstats against reference panel
gctb \
    --sbayes-rc \
    --impute-summary \
    --gwas-summary gwas_sumstats.ma \
    --ldm-eigen ukb_eigen_ld.eigen \
    --annot baselineLD_v2.2 \
    --out sbayesrc_preprocessed.ma

# 2. Run SBayesRC main step
gctb \
    --sbayes-rc \
    --gwas-summary sbayesrc_preprocessed.ma \
    --ldm-eigen ukb_eigen_ld.eigen \
    --annot baselineLD_v2.2 \
    --num-chains 4 --chain-length 25000 --burn-in 5000 \
    --out sbayesrc_run

# 3. Score target genotypes
plink2 --bfile target \
       --score sbayesrc_run.snpRes 2 5 8 header \
       --out sbayesrc_scores

PGS Catalog Calculator (Production Pipeline)

Goal: Compute multiple published PGS scores in one pipeline with automatic liftover, ancestry projection, and normalization.

Approach: Nextflow nf-core pgsc_calc workflow handles GRCh37/38 liftover, PC-projection-based ancestry assignment, ambiguous-SNP filtering, and mean/variance normalization.

# Calculate multiple PGS for cohort
nextflow run pgscatalog/pgsc_calc \
    --input samplesheet.csv \
    --target_build GRCh38 \
    --pgs_id PGS000004,PGS000019,PGS001775 \
    --run_ancestry resources/pgsc_HGDP+1kGP_v1.tar.zst \
    --outdir pgsc_output/ \
    -profile docker

Multi-Ancestry PRS-CSx

Goal: Compute multi-ancestry PRS using ancestry-specific GWAS sumstats jointly via continuous-shrinkage Bayesian framework.

Approach: PRS-CSx jointly models multiple ancestries with shared shrinkage prior; outperforms per-ancestry PRS-CS for non-EUR.

python PRScsx.py \
    --ref_dir=ldblk_1kg \
    --bim_prefix=target \
    --sst_file=eur_sumstats.txt,afr_sumstats.txt,eas_sumstats.txt \
    --n_gwas=200000,30000,50000 \
    --pop=EUR,AFR,EAS \
    --out_dir=prscsx_out --out_name=cohort \
    --phi=1e-2  # tune by trait architecture: 1e-2 polygenic, 1e-4 sparse

# Score each ancestry-specific posterior, then combine on tuning set
for pop in EUR AFR EAS; do
    plink2 --bfile target \
           --score prscsx_out/cohort_${pop}_pst_eff_a1_b0.5_phi1e-02.txt 2 4 6 \
           --out scores_${pop}
done

Score Normalization and Ancestry Recalibration

Goal: Convert raw PRS to ancestry-conditional percentiles for clinical interpretation.

Approach: Subtract conditional mean given test-cohort PCs (NOT discovery PCs); divide by conditional SD.

import numpy as np
from scipy import stats
import statsmodels.api as sm

def ancestry_conditional_normalize(prs, pcs, n_pcs=10):
    '''Recalibrate PRS by removing ancestry effects (Ding 2023 continuous-ancestry).

    pcs: matrix of principal components computed in the TEST cohort (not discovery).
    Returns: ancestry-conditional Z scores.
    '''
    X = sm.add_constant(pcs[:, :n_pcs])
    mean_model = sm.OLS(prs, X).fit()
    expected = mean_model.predict(X)
    residuals = prs - expected

    log_var_model = sm.OLS(np.log(residuals ** 2 + 1e-12), X).fit()
    expected_log_var = log_var_model.predict(X)
    sd = np.sqrt(np.exp(expected_log_var))

    return residuals / sd


def prs_to_percentile(z_scores):
    '''Convert ancestry-conditional Z to population percentile.'''
    return stats.norm.cdf(z_scores) * 100


def absolute_risk_transform(percentile, incidence_curve_age, age):
    '''Integrate over external age-conditional incidence curve to get absolute risk.

    Khera 2018 used FOS for CAD; Lee 2019 BOADICEA v5 for breast cancer.
    The HR-per-SD-only framing (Hingorani 2023 critique) is insufficient for screening.
    '''
    base_risk = incidence_curve_age(age)
    z = stats.norm.ppf(percentile / 100)
    relative_risk = np.exp(z * 0.5)  # placeholder; trait-specific log HR
    return base_risk * relative_risk

Sample Overlap Detection (EraSOR / Bivariate LDSC Intercept)

Goal: Detect overlap between PRS-discovery cohort and target cohort, which inflates apparent PRS performance.

Approach: Run bivariate LDSC; |intercept| > 0.05 with target n >= 1000 is the typical alarm.

# bivariate LDSC for sample overlap
ldsc.py \
    --rg target_sumstats.sumstats.gz,discovery_sumstats.sumstats.gz \
    --ref-ld-chr eur_w_ld_chr/ \
    --w-ld-chr eur_w_ld_chr/ \
    --out overlap_check

# Inspect *.log: gcov_int is the sample-overlap-driven intercept (after gencov correction)
# |gcov_int| > 0.05 with target n >= 1000 => substantial overlap

Per-Operation Failure Modes

1. Discovery + tuning + test overlap (the inflated R^2 trap) - Trigger: GWAS run on UKB; PRS evaluated in UKB; tuning on UKB subset. - Mechanism: Discovery samples leak into test set; PRS-derived effects fit the test data perfectly. - Symptom: Reported R^2 inflated 2-5x; performance does not replicate in external cohort. - Fix: Three-way disjoint sample partition (Wray 2014 J Child Psychol Psychiatry; Choi 2020 Nat Protoc); use EraSOR (Choi 2023 GigaScience) or bivariate LDSC intercept; threshold |intercept| > 0.05 with target n >= 1000.

2. Treating discrete ancestry boxes as ground truth - Trigger: Report "PRS for AFR cohort" without quantifying genetic distance. - Mechanism: Ancestry is continuous (Ding 2023 Nature); per-PC distance from training data drives R^2 with r = -0.95. - Symptom: Reviewers reject for incomplete ancestry characterization. - Fix: Project test cohort onto reference PCs; report performance as a function of continuous PC distance.

3. Strand-ambiguous SNPs (A/T, C/G) handled wrong - Trigger: Drop or trust strand annotation across cohorts. - Mechanism: ~10-15% of SNPs are strand-ambiguous; cross-platform/cohort scoring requires explicit handling. - Symptom: Strand-flipped variants reverse effect direction; PRS uncorrelated with phenotype. - Fix: PRSice-2 default drops them; LDpred2 snp_match() frequency-matches with 0.4-0.6 MAF tolerance.

4. LD reference mismatch (high s parameter) - Trigger: Use 1KG-EUR LD for UK Biobank target; LDpred2 silently uses mismatched LD. - Mechanism: snp_ldsc2() returns LD-mismatch parameter s; values > 0.05 indicate problematic mismatch. - Symptom: Posterior effect estimates inflated; PRS doesn't replicate. - Fix: Use UKB LD panel (n=40k+) instead of 1KG-EUR (n=489); inspect s routinely.

5. PCs in derivation but not in evaluation - Trigger: GWAS uses 10 PCs; PRS evaluation regression omits PCs. - Mechanism: Population stratification effects re-enter the PRS-phenotype association without PC control. - Symptom: Apparent PRS association is partially ancestry confounding, not biology. - Fix: Include same PCs in evaluation regression; compute PCs in TEST cohort (not discovery).

6. Treating HR per SD = 1.5 as clinically actionable - Trigger: Report "top 5% PRS = 1.5x risk -> screening criterion". - Mechanism: HR/OR per SD (~1.3) is similar to family history; Wald-Hingorani DR-to-FPR ratios approach 10:1 for true screening utility; PRS achieve 2-3:1. - Symptom: Clinical implementation produces high false-positive rates. - Fix: Integrate over external incidence curve for absolute-risk reporting; benchmark against family history specifically.

7. PRSmix / MUSSEL / PROSPER cited in wrong journal - Trigger: Manuscript cites these as Nat Genet. - Mechanism: PRSmix and MUSSEL = Cell Genomics; PROSPER = Nat Commun. - Symptom: Senior PIs catch in review. - Fix: Verify citations against PubMed before submission.

8. HLA region included in PRS without explicit handling - Trigger: Compute autoimmune-disease PRS with HLA region included naively. - Mechanism: HLA region (chr6 25-35 Mb) has extreme LD; SNP-based PRS does not capture classical HLA allele effects. - Symptom: Reviewer flags incomplete HLA modeling. - Fix: Exclude HLA region from main PRS; model classical HLA alleles separately via HIBAG / SNP2HLA / HLA-LA; report HLA-excluded + HLA-augmented versions.

9. Sex chromosomes mishandled - Trigger: PRS includes chrX without sex-specific dosage coding. - Mechanism: Males 0/1; females 0/1/2 unless XCI-corrected. - Symptom: Cross-sex comparison invalid. - Fix: Drop chrX OR apply explicit sex-stratified dosage encoding.

Reconciliation: When Methods Disagree

Pattern Likely cause Action
LDpred2-auto vs PRS-CS-auto disagree on top decile Trait sparsity differs; method assumption mismatch Use SBayesRC as tiebreaker; ensemble via PRSmix
Multi-ancestry methods rank differently Tuning set size; ancestry composition Report ensemble; document hyperparameter sensitivity
EUR PRS R^2 << non-EUR target R^2 Expected; Martin 2017 transferability Switch to PROSPER / MUSSEL / BridgePRS
PRS performance differs across age strata Mostafavi 2020; within-ancestry heterogeneity Report age-stratified PRS performance
PRS unrelated to phenotype despite high h^2 Sample overlap; strand-ambiguous SNPs; PC issues Run EraSOR; check snp_match() flips; recompute PCs in test
PRS percentile shifts with array platform Different ascertainment of common variants Recalibrate per-platform; use ancestry-conditional Z

Quantitative Thresholds and Conventions

Threshold Convention Source
HapMap3 SNPs ~1.1M variants; required for PRS-CS, LDpred2 HapMap project
LDpred2 LD ref s < 0.05 indicates well-matched LD Privé 2022
EraSOR intercept
MAF range for ambiguous SNPs Drop or frequency-match 0.4-0.6 LDpred2 default
INFO score (imputed) >= 0.8 for inclusion QC convention
Kinship coefficient cutoff KING > 0.0884 = 3rd-degree or closer; remove KING documentation
HLA region exclusion chr6 28-34 Mb (some use 25-35) Convention
HR/OR per SD typical ~1.3 per SD Hingorani 2023 BMJ Med
Top 2.5% PRS CAD detection rate ~7% of cases captured Hingorani 2023
Mavaddat PRS313 313 SNPs, breast cancer Mavaddat 2019
Khera CAD PRS ~6.6M variants Khera 2018
Patel CAD PRS GPS_Mult; multi-ancestry SOTA 2023 Patel 2023

Common Errors

Symptom Cause Solution
PRS R^2 << expected on held-out data Sample overlap with discovery Run EraSOR; rebuild three-way disjoint splits
LDpred2-auto chains diverge LD mismatch (s > 0.05) or wrong h^2 init Use UKB LD panel; check snp_ldsc2() output
Strand-flip warnings ignored snp_match defaults Set strand_flip = TRUE; review flipped variants
HLA region produces extreme PRS values Long-range LD Exclude chr6 28-34 Mb; model HLA separately
Non-EUR PRS at chance Method not multi-ancestry-aware Switch to PROSPER/MUSSEL/BridgePRS
Different ranking with same data Stochastic MCMC Seed; report posterior CIs; check convergence
PGS Catalog liftover failure Incompatible build Use pgsc_calc auto-liftover; check log
Top 1% PRS appears benign Conditional mean drift across ancestries Apply Ding 2023 ancestry-conditional Z

Anticipated Reviewer Pushback

Pushback Standard response
"PRS HR 1.5 per SD is similar to family history" Acknowledged (Hingorani 2023). We report absolute-risk integrated over age-conditional incidence (Wald-Hingorani framing); we do NOT recommend population-level screening on PRS alone.
"Why use Bayesian methods over C+T?" LDpred2-auto / PRS-CS / SBayesRC capture polygenic architecture better; ~16-18% higher prediction correlation (r) vs C+T per Pain 2021 / Privé 2022 benchmarks.
"Why not use the largest published GWAS?" We checked sample overlap between discovery and target with EraSOR / bivariate LDSC intercept;
"Discrete ancestry boxes don't capture genetic structure" Agreed; we report PRS performance vs continuous PC distance (Ding 2023 Nature); discrete labels are summary only.
"Why exclude HLA?" HLA region extreme LD makes SNP-based posteriors unstable; we model classical HLA alleles separately via HIBAG / SNP2HLA.
"PRS-CS phi parameter; how chosen?" We use --phi=auto with multiple chains; for sparse traits (lipids) 1e-4; for highly polygenic (height) 1e-2.
"Where is the absolute risk?" Reported per Wand 2021 PRS-RS item 19; integrated over age-conditional incidence using external curve.
"FDA PRS guidance?" No general FDA PRS draft guidance exists as of 2026; August 2025 Federal Register Class II classification for Cancer Predisposition Risk Assessment Systems is the operative regulatory text.
"Why three different references for SBayesRC?" Zheng 2024 Nat Genet is the primary paper; baseline-LD is Gazal 2017; UKB LD panel is the bigsnpr/Privé 2020 release.

References