kemalandClaude Opus 5.5
cb3f5fac58
beta r 0.99998 vs sesame 1.28.1 (max |dbeta| 0.011), detection calls 99.1% identical, noob within 0.5% median; replicate r 0.997; all 21 control metrics pass. Spec deviations recorded (5-field control lines, pOOBAH formula, minfi sheet keys). User guide section. Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01814GKVzZVL3FuF3DeiZXhv
3.2 KiB
v2 module validation — methylation control QC (simurg methyl qc)
Date: 2026-09-29. Development on the Raspberry Pi; oracle runs on AWS (c7g.xlarge, five short runs, ≈ $0.025). Spec: MET-001 … MET-018 (and QC-004/QC-006, QC-062).
Data
- minfiDataEPIC (Bioconductor, Artistic-2.0): 3 EPIC v1 arrays of NA12878 (IDATs + minfi-style sample sheet).
- Infinium MethylationEPIC v1.0 B5 manifest (CSV): 865 918 probes (142 158 Infinium I: 92 212 red, 49 946 green), 635 control beads. Read in 8.6 s on the Pi.
Delivered
simurg methyl qc, simurg methyl convert-manifest, simurg qc controls --array-type methylation, methylation
datasets in simurg qc report; modules formats/methyl_manifest.py, signal/methyl.py, pipeline/methyl.py;
the 21 methylation metrics in qc/controls.py. User guide: "Methylation arrays".
Results on minfiDataEPIC
| Check | Result |
|---|---|
| Run time, 3 samples, 3 workers (Pi) | 25 s (2.9 s signal processing per sample) |
| 21 control metrics | all evaluable, all pass the methylation_epic profile (e.g. staining green 57, BS conversion I green 13, BS conversion II 5.3, specificity II 24) |
| Probes passing detection (p < 0.05, analytical CG + SNP probes) | 90.9 %, 91.1 %, 91.3 % |
| β of the three NA12878 replicates | Pearson r 0.997 – 0.9974 between replicates; bimodal distribution |
Sheet Sample_Group → PCA colours; outputs of MET-013 |
present; folder loads in qc report |
Against the sesame oracle (sesame 1.28.1, same probe set via probes.tsv/controls.tsv)
scripts/v2/sesame_oracle.R + scripts/v2/methyl_oracle_compare.py, bead-count masking off on both sides (sesame's
min_beads masks every Infinium II probe because their M address is NA), 865 918 probes per sample:
| Quantity | Agreement (3 samples) |
|---|---|
| β values | Pearson r 0.99998; |Δβ| median 0.0009, 99th percentile 0.008, max 0.011 |
| noob-corrected M/U signals | relative difference median 0.5 %, max 5 % |
| Detection calls at p < 0.05 | 99.0 – 99.1 % identical; passing fraction 91.0 – 91.4 % (simurg) vs 92.0 – 92.3 % (sesame) |
| Detection p-values | not identical (median |Δp| 0.0008) |
The remaining difference is systematic (same size in all three samples) and is not the negative controls: sesame did
not attach the supplied control table (negControls failed, so its background is out-of-band Infinium I signal only),
and simurg without negative controls is not closer (noob difference 0.56 % instead of 0.53 %). The exact cause is
open (K-026). For QC use — pass/fail per sample, detection rate, β for a PCA — the agreement is sufficient; simurg's
methylation numerics are documented as its own [OURS] implementation of the published methods (MET-010), not as a
bit-exact port.
Spec deviations (K-026)
- MET-002: the vendor EPIC manifest has a 5th field (
AVG) on someNEGATIVEcontrol lines; ≥ 4 fields are accepted and the first four used. Fewer than 4 still givesE_METH_CONTROLS_FORMAT. - MET-010: the detection p-value follows sesame's pOOBAH (p = 1 − max(F(M), F(U)), each in its own channel) instead of the ambiguous "1 − ECDF(M + U)", so an open-source oracle exists.
- MET-004:
Sentrix_ID/Sentrix_Position(minfi/sesame sheets) accepted as a third key-column pair.