Methylation gene-body units, poolable SEs, and three first-run fixes (the part of #67 that missed its merge)

[sr320]

What and why

#67 merged on 2026-10-16 with only its first commit (the limma/DSS pooled-tile engine). Five later commits were pushed to its branch after the merge, so they're not on main, although #67's description lists them. They're rebased onto main here, unchanged. All five came from running CIBNOR2021_HEAT_RRSS (#14) end to end, and #14 needs them.

1. Cap the methylation report's region table at 500 rows (8188da3). Embedding all 319,088 tiles in the self-contained HTML exhausted Quarto's 8 GB JavaScript heap (exit 133). The report shows the 500 smallest q-values, and the full table is the standardized TSV.
2. Capture Bismark's version in deduplication and extraction (ea00d0f). Those tools print "Deduplicator Version" and "Bismark Extractor Version", so versions.yml had an empty bismark field.
3. Carry a standard error and p-value to evidence (6dd5425). The methylation harmonizer hard-coded standard_error and p_value to None, so no methylation study could enter inverse-variance pooling.
   • DMR_POOLED_TILES now writes pvalue, se_meth_diff_percent and analysis_method.
   • Standardizing passes them through when present, and the harmonizer reads them.
   • methylKit tables are unchanged and stay labelled methylKit.
   • New tests: tests/test_methylation_harmonize.py.
4. Gene-body units, methylation.dmr_features = 'genes' (f4a7554). Genome-wide tiles give many records per gene, and once harmonized to the gene they collide in meta-analysis. build-evidence-cards then fails for every study. Gene mode pools each gene's body and strand-aware promoter CpGs into one unit, with intergenic CpGs in tiles. ANNOTATE_REGIONS keeps the gene and context those units arrive with. tiles stays the default.
5. Make limma's SE reproduce limma's p-value under a Wald test (9b14413). Using |diff / t| dropped limma's small-sample df. At 2 vs 2 families, one CIBNOR gene went from limma p = 2.7e-4 to pooled p = 4e-27, and 383 features became "significant", 53 of them promoted to multi_omics_convergence. The SE is now |diff| / qnorm(1 − p/2).

Testing

• Code matches the run: the DMR, annotation, standardize, harmonizer and main.nf code here is byte-identical to what produced [study] CIBNOR2021_HEAT_RRSS (PRJNA690951): first reanalysis #14's results (git diff 7c8664e HEAD on those files is empty).
• End to end: a full Pages-sequence rebuild with CIBNOR's evidence runs register → demo → harmonize_committed_results.py → gene meta-analysis → cards → top candidates → dashboard → all-feature meta-analysis:
  • no errors;
  • CIBNOR gives 30,732 emerging candidates, none significant, as in limma;
  • 0 multi-omics candidates;
  • 889 high-priority candidates, unchanged from before CIBNOR's evidence.
• Synthetic data:
  • gene-body units: a two-gene GTF places + and − strand promoter windows correctly, recovers the true-signal gene (+26.6 points), and gives unique feature IDs after annotation;
  • Wald p from the SE matches limma's p within output rounding.
• Checks: pytest -q: 644 passed. ruff: clean. The methylation demo (CI's command) passes.

Checklist

• pytest -q and ruff check src tests pass
• New logic has a test (tests/test_methylation_harmonize.py). The R inside the Nextflow modules is checked on synthetic data, as for Pool condition levels per arm, and fit DESeq2 on tanks with --replicate_unit #12, Fit methylKit DMRs on families, not libraries (methylation.replicate_unit) #15 and Test pooled family tiles with limma (or DSS), not per-region methylKit MN #67.
• Relevant docs/ page updated (workflows/methylation/README.md)
• Study branches that need this change are listed below

Study branches waiting on this

• study/PRJNA690951/CIBNOR2021_HEAT_RRSS (#14): its config already sets dmr_features = 'genes'. Merge this first; merging [study] CIBNOR2021_HEAT_RRSS (PRJNA690951): first reanalysis #14 alone would make main's evidence cards fail.

🤖 Generated with Claude Code