gscore software validation
Current scoring uses gscore::gscore(weights, Glist); weight learning belongs to glma, gbayes or another statistical model owner, and conventional C+T uses gscore::clump_threshold().
These recorded simulation results assess selected behavior on ten fixed datasets from the shared design. They do not revalidate the statistical method or cover every setting. The use snippets show the current package interfaces. This report step summarizes retained artifacts and does not rerun analyses. The gscore guide describes current inputs and options.
Use
ct_weights <- gscore::clump_threshold(statistics$A, reference_Glist,
p_threshold = 5e-8, r2_threshold = 0.1, window_bp = 250000)
prediction <- gscore::gscore(gbayes_fit$weights, held_out_Glist)Results
| Weights | Mean r with genetic value | Mean r with phenotype | Mean calibration slope |
|---|---|---|---|
| ridge | 0.158 | 0.091 | 0.046 |
| bayesc | 0.495 | 0.274 | 1.488 |
| bayesr | 0.512 | 0.282 | 1.106 |
Applying the true effects reproduced centered simulated genetic values within 7.33e-15 at worst. This checks marker order and coding. Prediction quality depends on the fitted weights and the deliberately difficult polygenic architecture.
The ridge fit uses a fixed penalty of 1 without tuning. Its low held-out correlation and calibration slope are reported as observed; this setting is an execution check, not a recommended prediction recipe.
