Add OTTERS pipeline, check_ld(), and fix pval_acat()

[gaow]

Summary

Implement the OTTERS framework (Zhang et al. 2024) for omnibus TWAS, plus LD quality checking and a bug fix in pval_acat().

New functions

• otters_weights(sumstats, LD, n, methods, p_thresholds, check_ld_method) — Stage I: train eQTL weights via P+T + any *_weights() learner. Defaults match original OTTERS exactly (phi=1e-4, thin=1, lambda from 1e-4, s grid search). Extensible via do.call dispatch.
• otters_association(weights, gwas_z, LD, combine_method) — Stage II: per-method TWAS z-scores + ACAT/HMP combination. Includes dimension validation.
• check_ld(R, method, r_tol, shrinkage) — LD matrix QC with three modes: "check" (diagnostic), "shrink" (regularization), "eigenfix" (set negative eigenvalues to 0). Integrated into otters_weights() by default.
• lassosum_rss_weights() now searches over s = c(0.2, 0.5, 0.9, 1.0) grid (was fixed s=0.9).
• lassosum_rss() lambda range extended to 1e-4 (was 1e-3), matching OTTERS.

Bug fix

pval_acat() had a sign error: qcauchy(p) = tan(π(p-0.5)) but ACAT needs T = mean(tan(π(0.5-p))). Combined p-values were near 1.0 instead of near 0.

Audit results

All 17 OTTERS default parameters verified matching. Systematic comparison of every formula, default, and edge case against the original Python/R pipeline. Key fixes: PRS-CS phi, SDPR thin, lassosum lambda/s, cor>=1 safeguard, P+T weight clamping.

Validation

         method   twas_z    twas_pval n_snps
1      PT_0.001 4.921729 8.578e-07      3
2       PT_0.05 4.770889 1.834e-06      5
3  lassosum_rss 4.440683 8.967e-06     46
4        prs_cs 2.775733 5.508e-03     50
5          sdpr 2.808055 4.984e-03     50
6 ACAT_combined       NA 2.743e-06     NA

Test plan

• Unit tests for otters_weights(), otters_association(), check_ld()
• End-to-end integration test with simulated data
• Vignette with real data loading pattern + simulation demo
• devtools::test(filter="otters") and devtools::test(filter="ld")
• R CMD check

🤖 Generated with Claude Code