[study] DUPERRET2025_POMS_STARVE (MTBLS13313): reanalyzed from Table S12, QC failed

[sr320]

Replaces sr320/AREE#70. That PR was stacked on #69, and GitHub closed it instead of retargeting it when #69 merged and its branch was deleted. The branch and commits are unchanged. The workflow options it needs are now on main through sr320/AREE#69.

Study

• study_id: DUPERRET2025_POMS_STARVE, registered in [study] DUPERRET2025_POMS_STARVE (MTBLS13313) #22
• Accession: MetaboLights MTBLS13313; Duperret et al. 2025, iScience, doi:10.1016/j.isci.2025.114333
• Mode: raw_reanalysis from the paper's Table S12 peak areas. The MetaboLights MAFs have no abundance values.
• Machine: Hete, -profile local

What was run

• Source data: Table S12 (mmc13.xlsx) is CC BY-NC-ND 4.0, so it is not committed. build_feature_table.py fetches it from the Europe PMC supplementary-files API, checks its SHA-256 and writes the workflow inputs. table_s12_provenance.json records what it did: 310 metabolites, plus 10 internal standards (the rows the authors filled green, checked against the names in the Methods), across 48 samples.
• Workflow run: the 0 hpc starved-vs-fed contrast, with:
  • --metabolomics.normalization internal_standard, the authors' own first step;
  • --metabolomics.detection_limit 0.001, so the 0.0001 and 0 placeholders count as below detection;
  • --metabolomics.replicate_unit tank, giving 3 vs 3 tanks, as in DUPERRET2025_STARVE_RNASEQ;
  • --metabolomics.acquisition targeted.
• Result: 122 of 310 metabolites were tested and 68 have padj < 0.05. The other 188 are in the missingness report, 149 of them never detected.

Why it failed QC

Diet is fully confounded with MS batch: fed and starved samples ran months apart. This run shows the consequence:

• The internal standards shift with diet. On GM the per-sample offset is about +0.8 log2 in fed samples and −0.9 in starved ones. On nucleotides, which use the same MES spike, the shift goes the other way.
• Normalization decides direction. Median normalization ranks metabolites identically within GM, NUCs and TCA (r = 1.00), but moves every log2FC by about 1 log2 per platform. On PUFAs the two methods disagree outright (r = 0.40). Only 61% of tested metabolites keep their direction (normalization_sensitivity.tsv).

Direction is what an evidence record asserts, and here it can't be separated from batch.

QC decision (sr320):

• qc_status: failed;
• results_file stays null;
• nothing is harmonized or pooled.

As with the IOCAS studies, the run's outputs are committed as provenance for the decision: the standardized table, missingness report, manifest, report and sensitivity table. On the dashboard the study moves from "harmonize" to "resolve QC failures".

Checklist

• aree validate-study registry/studies/DUPERRET2025_POMS_STARVE.yaml passes
• Source checksum verified before the run (table_s12_provenance.json)
• Workflow manifest committed, with parameters, tool versions and input checksums
• Standardized result table and the HTML report committed; no source table, work/ or results/
• qc_status and analysis_status reflect what was actually run (failed / in_progress)
• Limitations and curation notes record the batch confounding, the sensitivity check and the MSI level 2 choice
• Curation tests in tests/test_duperret2025_study.py pin the decision and its evidence
• No edits to shared code, vocabularies or docs; aree check-study-scope reports 9 files, all owned by the study

Full suite: 613 passed; ruff clean.

Follow-ups, not in this PR:

• a batch_confounded quality flag (the registry notes AREE has none);
• a metabolite-to-enzyme-gene map, without which metabolomics evidence can't converge with RNA-seq genes in the multi-omics tier.

🤖 Generated with Claude Code