Out-of-fold prediction (Module 03)

genboostgpu lsp is the secondary, translational endpoint: can a region be predicted from local SNPs in held-out donors? It ports 03_local_snp_prediction under its PI-locked config/prediction.yml:

  • run-level donor fold table (5 outer folds × 5 repeats, seeded by seed_for(run_id, region, repeat_i));

  • inside each outer split, everything is fit on training donors and applied to held-out donors: covariate residualization and scaling, variant missingness/MAF filters, mean imputation, genotype scaling;

  • a Haseman–Elston screen calibrated with 1000 within-fold permutations (one-sided, p = (1 + #{null ≥ observed}) / (1 + B), α = 0.05);

  • inner 5-fold cv.glmnet (standardize = FALSE) over alpha {0.1, 0.5, 0.9, 1}, lambda.min;

  • the prespecified null prediction (training mean) for held-out donors when a fold’s screen fails — failed folds are never dropped.

Metrics are computed from pooled OOF predictions: r2_pred_oof = 1 − SSE/SST (negative values are kept), cor2_oof, RMSE, MAE, calibration intercept and slope (undefined when the prediction spread is degenerate), screening pass frequency, and per-repeat spread. combine stops if the median r2_pred_oof exceeds the leakage tripwire (0.5).

Replaying an accepted run reproduces its donor folds and screening decisions exactly. Per-donor predictions are not bit-for-bit. The in-fold residualized phenotype already differs from R’s at ~1e-14, from BLAS order and TSV precision. glmnet’s thresh=1e-7 convergence then magnifies that difference in lasso fits over SNPs in near-perfect LD, where the solution is not unique.

On 500 random loci of lsp-AA-caudate-20260925-a, 92% of the 12,300 outer folds agreed to 1e-6. Of the folds that differed, 90% had alpha = 1 in both runs. r2_pred_oof had max |Δ| = 6e-4 and Spearman 1.000000, and the median was identical.

genboostgpu lsp init --run-dir runs/lsp-replay --run-id gbg-lsp-replay \
    --replay <repo>/03_local_snp_prediction/_m/runs/lsp-AA-caudate-20260925-a
genboostgpu lsp run --run-dir runs/lsp-replay --shard 0/8 --device gpu
genboostgpu lsp combine --run-dir runs/lsp-replay

--from-lgv-run <GENBoostGPU lgv run> predicts the loci of a Module 02 run (e.g. CpH regions) with a new run-level fold table.