This repository accompanies the PLOS Computational Biology submission:
Recorded metadata predicts the phenotype label across public single-cell cohorts and changes what internal validation can establish
The current submission snapshot was prepared on 2026-09-08 and sealed on 2026-09-13.
It is archived as release v3.0.0, Zenodo version DOI
10.5281/zenodo.22733028; the concept DOI
10.5281/zenodo.20813922 always resolves to the latest version. It supersedes the earlier
v2.0.0-rc3 identifiability-framed manuscript while retaining that historical release
for provenance.
Patient-level single-cell classifiers can report very high internal AUC even when the recorded details of recruitment, processing and sequencing also predict the phenotype label. The current study asks a narrower empirical question: how strongly does recorded metadata predict the phenotype label in a cohort, which variable groups carry that association, and what can common validation procedures establish when collection structure and the label are entangled?
The manuscript does not claim a general identifiability theorem and does not present the checks as a required validity standard. They are complementary diagnostics with explicit scope and failure modes.
The main screen contains nine phenotype contrasts from five public datasets (809 donors), spanning lupus, COVID-19, influenza, cytomegalovirus serostatus, and a sepsis-versus-COVID-19 contrast.
Key results:
- recorded metadata alone predicts the phenotype label significantly in eight of nine comparisons, in all five split seeds and after Benjamini-Hochberg correction;
- collection variables alone are significant in four of seven comparisons that record a varying collection block;
- a prespecified CMV cohort is the negative control, with metadata-only prediction at chance while the expression classifier remains strong;
- adding expression to the metadata raises AUC by 0.087 to 0.354 in seven comparisons and by nothing measurable in sepsis-versus-COVID-19 and influenza;
- in an influenza comparison every collection stratum is label-pure, so the collection-stratified permutation is uninformative (p = NA);
- free and collection-stratified complete-pipeline permutations disagree in the sepsis-versus-COVID-19 comparison;
- two COVID-19 cohorts with similar overall metadata predictability are affected through different variable groups, sample quality in one and recruiting hospital/sub-study in the other;
- downstream lupus analyses distinguish representation-level batch information from actual classifier use of that information;
- residualisation and restriction answer different questions and cannot be interpreted as bounds on a biological effect;
- external validation only tests design confounding to the extent that the target acquisition mechanism is genuinely different.
The simulation study covers seven data-generating regimes, seven levels of design-label association, binary and continuous outcomes, and 19,600 simulated cohorts.
The research object for this submission is under ploscb_submission_2026-09-08/:
the manuscript source, cover letter and line-numbered submission PDF, all 15 figures,
response to the previous review, Supporting Tables S1-S11, the analysis and figure scripts, every result table behind
Figures 2-8, the five per-seed screen outputs, the cohort registry and the donor-level
interface tables, and the QA records. See its REPRODUCE.md.
Public expression matrices are not mirrored there; they stay at their original GEO and
CELLxGENE Discover accessions, which cohorts/registry.yaml lists.
To re-run the manuscript checks against those files:
bash ploscb_submission_2026-09-08/verify.shThat re-derives 24 numbers from the released outputs and requires each verbatim in the
manuscript, then runs ~30 structural checks over the manuscript, the rendered figures
and the released tables. SHA256SUMS covers every file in the directory.
All supervised analyses use the donor as the independent unit. Cells from one donor do
not cross training/test partitions. ploscb_submission_2026-09-08/ records the cohort
metadata, design variable inventories, analysis scripts, simulation outputs, per-seed
results and figure source code for the current study.
The checks only see recorded structure. A low metadata-only AUC does not rule out unrecorded collection effects. A significant collection-stratified permutation result is evidence of discrimination beyond what the recorded collection strata retain; it does not rule out technical, demographic or sample-quality explanations outside those strata. None of the internal checks alone establishes biological attribution or clinical usefulness.
release_v2.0.0-rc3/- previous PLOS Computational Biology review object, retained unchanged for provenance.release_v1.0.2/- earlier SLE representation benchmark snapshot.
The public package name rheumlens is retained for repository continuity; it is not the
name of a method introduced by the current manuscript.