APEX-05 — Provenance of the cax23 sample-labelling anomaly

This document reconstructs, from the archived analysis files, the sample mislabelling that affected the APEX-05 RNA-seq dataset and how it was detected and corrected. It is the authoritative record for reviewers and re-users of the data package. All paths are relative to the repository root; the historical files themselves are preserved under archive/targets_file_versions/ and archive/diagnostics/.

1. Executive summary

The APEX-05 growth plate carried cax2-2 (cax22) and cax2-3 (cax23) — two independent mutant alleles of the same gene, CAX2 (a Ca²⁺/H⁺ exchanger), not two different genes — in physically adjacent wells (well pairs A3↔A4, B4↔B3, B8↔B7, A7↔A8). That same-gene relationship is what makes the divergence diagnostic: two alleles of one gene should respond alike, so a cax2-3-specific anomaly stands out sharply (quantified in cax2_allele_concordance.md). A quality-control screen of differentially-expressed-gene (DEG) overlap — archived as the "CAX23oddity" diagnostic — showed cax23 behaving anomalously: its DEG set did not partition cleanly against the other genotypes, the classic signature of an adjacent-well sample swap / cross-labelling between cax22 and cax23.

Two corrective actions are visible in the file trail:

  1. During diagnosis (V12–V13): the single odd replicate — cax23 Ground-Control replicate 1 (GCRoot_CAX23_1 + GCShoot_CAX23_1) — was dropped, and a dedicated cax22-vs-cax23 contrast was built to test whether the two lines were separable at all.
  2. In the final published master: the entire cax22 well A7 (4 libraries: A7_FL_Root/Shoot, A7_GC_Root/Shoot) was removed, taking the study from 64 → 60 samples, and the master was re-keyed to plate-well sample IDs so that any future genotype↔well mismatch is auditable.

2. Version-by-version trail

File (in archive/… unless noted) Samples Genotypes Notes / cax23-relevant change
targets_file_versions/HiSat_APEX5_Targets_File_V3.txt 64 all four (16 ea.) Baseline; MiSeq-era counts_out_V3; retains a Location/well column — the adjacency evidence
…_V4.txt 64 all four Column trim (Location dropped)
…_V5.txt 64 all four Dot-delimited labels; Location restored
APEX5_HiSat_Targets_File_V6.txt / …_V6.csv 64 all four Identical design to V5
APEX5_Hiseq_Hisat_Targets_V11.txt 64 all four Switched to deeper HiSeq counts_out_V8; re-indexed FL/GC_Root/Shoot_<geno>_1..4 labels — the re-processed data in which the oddity surfaced
diagnostics/apex05_roots_v11_cax23-oddity.csv / .png (+ shoot twin in results/) — Root{CAX22,CAX23,COL00,RBOHD} + all intersections The diagnostic that named the anomaly: jVenn/UpSet DEG set-membership showing cax23's odd overlap
…_V11_cax22_v_cax23.txt (+ transposed) 48 COL00, CAX22, CAX23 Purpose-built cax22-vs-cax23 contrast (rbohD excluded) to test separability; cax23 still full 16
…_V12_cax23_test.txt 16 CAX23 only Isolates cax23 for standalone inspection
…_V12_cax23_removed.txt 61 all four; cax23 = 14 Drops the odd GCRoot_CAX23_1 + GCShoot_CAX23_1 pair (name is misleading — only the GC rep-1 pair is removed)
…_V12_WTvRBOHD.txt 32 COL00, RBOHD Splits off the cax-independent WT-vs-rbohd analysis so it proceeds unaffected
…_V13_cax22_v_cax23_2023.csv (+ transposed variants) 46 COL00 (16), CAX22 (16), CAX23 (14) Formalised fix: cax22-vs-cax23 contrast with cax23 GC rep-1 permanently excluded; adds explicit tissue/genotype columns
…_V13_just_cax23.csv / …_JUST_cax23_transposed.csv 14 CAX23 only cax23-only, GC rep-1 already excluded
data/metadata/apex05_hisat_targets_master.txt + data/metadata/apex05_sample_metadata.csv 60 COL00 (16), CAX22 (12), CAX23 (16), RBOHD (16) Published master: counts_out final; well-ID labels (A1_FL_Root…); cax22 well A7 removed

Filename caveat. The cax23_removed names are inconsistent between epochs: …_V12_cax23_removed.txt keeps 14 cax23 libraries, but a same-named .csv that once sat in data/ was actually a COL00-only subset, and …_V12_NOcax23_JustGC.csv is COL00 ground-control only. Do not treat the .csv and .txt "cax23_removed" files as equivalent.

3. The anomaly, in plain English

The MiSeq-era design (V3–V6) was a clean 4×4 layout — four genotypes × four biological replicates × two organs × two conditions = 64 libraries. The Location column of HiSat_APEX5_Targets_File_V3.txt records the source well, and it shows that cax22 and cax23 were pipetted from interleaved, adjacent wells:

cax22  →  B8, B4, A7, A3
cax23  →  B7, B3, A8, A4        # each cax22 well sits beside a cax23 well

This is exactly the plate geometry in which a hand-off swap or index cross-talk between the two mutants is most likely. When the libraries were re-processed against the deeper HiSeq counts (counts_out_V8, from V11 on), a DEG-overlap QC export flagged the cax23 "oddity" (archive/diagnostics/apex05_roots_v11_cax23-oddity.*): cax23's DEG membership did not separate cleanly from cax22/COL00/rbohD, indicating at least one "cax23" library was not genotypically pure cax23.

4. Implicated samples

(a) Odd replicate dropped during diagnosis (V12–V13): GCRoot_CAX23_1 and GCShoot_CAX23_1 — present in V11 (and in …_V11_cax22_v_cax23.txt) but absent from every V12/V13 cax23 file, leaving cax23 with 14 libraries in those revisions.

(b) Well removed in the published master: A7_FL_Root, A7_FL_Shoot, A7_GC_Root, A7_GC_Shoot — the whole cax22 well A7, physically adjacent to cax23 well A8. A7 appears in the V3/V6 Location columns but is absent from both apex05_sample_metadata.csv and apex05_hisat_targets_master.txt. Its removal is why the master has cax22 = 12 samples (versus 16 for each other genotype) and why the final design is 60 samples.

These are two different corrective actions at two stages, not a contradiction: the cax22↔cax23 adjacency investigation traced the "cax23 oddity" back to the neighbouring cax22 well (A7).

5. Resolution

  1. Detect — HiSeq re-alignment (V11) + the CAX23oddity DEG-overlap diagnostic flag cax23 as misbehaving.
  2. Isolate & test — build a dedicated cax22-vs-cax23 contrast (…_V11_cax22_v_cax23, then the V13 family) and a cax23-only view (…_V12_cax23_test); split off the cax-independent WT-vs-rbohd comparison so it is unaffected.
  3. Trim — exclude the suspect cax23 GC replicate 1 from V12 onward; V13 bakes this in (cax23 = 14) and adds explicit tissue/genotype columns.
  4. Finalise — re-key the master to plate-well sample IDs, drop the problematic cax22 well A7, and publish the clean 60-sample design. Keying samples on well ID is the durable fix: it makes any genotype↔well mismatch auditable.

6. How the machine-learning package relates to this

analysis/ml/apex05_ml_anomaly_detection.py provides an independent, data-driven check on the class of error described here:

See ../results/ml/ml_metrics.json for the exact figures produced by the current data.