Genomic Claims You Can Trace

 Quick Analysis Plan Copied

 Long Analysis Plan

Skeptical Verdict: Portability Is Real but Fragile, Ancestry-Asymmetric, and Poorly Quantified

What the evidence actually shows. The strongest quantitative test of cross-ancestry PRS generalization in this corpus is X-Wing, which reports median relative R² improvements of 4.6–35.2% in East Asians (vs PRS-CSx and XPASS) and up to 119.1% in Africans, across 31 traits in East Asians but only 13 traits in admixed Americans and Africans (X-Wing gains by ancestry). Critically, the 119.1% African gain comes from a very low baseline (target N=6,490 for Africans vs ~360k Europeans in training), so large relative gains can coexist with clinically negligible absolute accuracy — a framing bias the skeptical reader must catch.

Why the baseline is untrustworthy for non-European populations. Every training cohort in this corpus is ancestry-narrow: PRS-PGx uses only European IMPROVE-IT data (n≈5,661) and explicitly limits generalizability (PRS-PGx European-only limitation); STREAM-PRS trains on only N=404 non-Finnish Europeans and acknowledges admixed populations are unsupported (STREAM-PRS small training set). The high-accuracy schizophrenia PRS (ACC ≥ 0.95) is restricted to Ashkenazi Jewish ancestry with leave-one-out cross-validation that risks optimism in a founder population (Ashkenazi schizophrenia PRS study).

Counterpoints the field under-weights. (1) The PLCO gene–environment study found significant interaction P-values (0.009–0.031) for selenium, aspirin, ibuprofen, and vegetables modifying a 20-SNP prostate risk score — but across 48 tests with no multiple-comparison correction reported, in non-Hispanic white men only (PLCO gene-environment interaction study). (2) Even the myopia ExGRS, spanning Han Chinese and European cohorts, assumes fixed allelic effects across ancestries — the exact assumption portability studies repeatedly falsify (ExGRS myopia multi-ancestry study). (3) A blindspot almost never discussed: PRS themselves leak ancestry-correlated genotype information (94.6% median genotype recovery), and identifiability varies by ancestry group (PRS privacy leakage study).

What would change this verdict. X-Wing's authors supply the falsification criterion themselves: gains vanishing once LD-reference mismatches are properly handled would collapse the portability claim (X-Wing falsification criteria). Missing from all records: absolute R²/AUC for non-European targets on matched traits, heritability differences across ancestries, and any pre-registered multi-ancestry clinical validation. Confidence: moderate-to-high that relative gains are genuine; low confidence they translate to clinically useful absolute risk in non-European populations.