feat(methylation): EPIC v2 replicate table and cross-reactive list

[ESCRI11]

What

• inst/extdata/epicv2-replicates.rds (new, 48 KB). Every EPIC v2 probe whose cg id is replicated: 11,616 probes over 5,222 ids. Columns are probe_id (full IlmnID), cpg (bare id), recommended (exactly one per cpg) and evidence (the Peters 2024 verdict). attr(, "manifest") = "20a1", and attr(, "source") names the inputs. playbase.epigenetics::collapse_epicv2_replicates() reads this table and refuses one whose manifest doesn't match its annotation package.
• inst/extdata/cross-reactive-probes.rds gets a fourth source, peters2024_epicv2, with 37,034 bare ids. The three existing lists are byte-for-byte the same rows.
• data-raw/build-epicv2-reference.R builds both files from IlluminaHumanMethylationEPICv2manifest / IlluminaHumanMethylationEPICv2anno.20a1.hg38 (Bioconductor 3.19; both install on R 4.3 / minfi 1.48) and the published Peters et al. 2024 Additional file 4, pinned by md5.

How the recommended replicate is chosen

Illumina's manifest doesn't designate a preferred replicate. Rep_Num is only a synthesis counter. The only per-probe evaluation that has been published is Peters et al. 2024 (BMC Genomics 25:251), which labels every replicated probe in Rep_results_by_NAME. The script ranks those labels, and the first match wins:

1. superior probe / superior by WGBS
2. best sensitivity
3. best precision
4. no verdict, "insufficient evidence", or a verdict saying only the probe-set mean is best
5. inferior

Ties go first to a probe the annotation can place, then to the lowest replicate number, then to the probe id. This is DMRcate's filter.strategy = "sensitivity" (same author), made deterministic. We never average replicates. Here is how the 5,222 recommendations break down: 684 superior probe, 1,181 superior by WGBS, 2,714 best sensitivity, 271 "group mean" sets and 372 "insufficient evidence" sets, the last two decided by tie-break. The 684 + 1,181 = 1,865 undisputed sets match the paper's count.

Cross-reactive definition

The list takes Peters' CH_BLAT == "Y" (30,627 in-silico cross-hybridising probes) plus Illumina's unmapped chr0 probes, which is the paper's "37,346". Each is recorded by its bare id, and for a replicated id the flag comes from the replicate the collapse keeps. SNP-affected probes aren't added here, because the Methylome app masks those from the manifest's SNP columns for each array. Consumers should select sources by array: peters2024_epicv2 describes v2 probe sequences.

Note for pinning

The monorepo images currently ship playdata 4d53ff0, which has neither cross-reactive-probes.rds nor ewas-catalog.rds. This branch is based on main (71fd13c), which has both.

🤖 Generated with Claude Code