# Protein abundance in ischemic failing versus non-failing human hearts

## Result
In this exploratory analysis of **12 ischemic cardiomyopathy (ICM) and 10 non-failing (NF) male left-ventricular samples**, two protein groups pass 5% Benjamini–Hochberg FDR on the archived MaxLFQ scale: **TF/serotransferrin** and a **CP/ceruloplasmin-associated group**. Both are fully observed and remain significant across all 100 missing-data stress tests. Neither remains below 5% FDR after sample median centering. CP additionally has unresolved protein-group identity because its group includes a contaminant-database accession.

The evidence supports a small, normalization-sensitive set of associations, not a robust general heart-failure signature. **TF is the cleaner follow-up candidate**, but it is not robust to every preprocessing/influence check.

| Protein group | ICM/NF geometric-mean ratio (95% pointwise CI) | Log2 difference | Primary BH q | Median-centered BH q |
|---|---:|---:|---:|---:|
| TF, P02787 | 1.85 (1.46–2.34) | +0.890 | 0.0346 | 0.0968 |
| CP-associated, P00450;CON__ENSEMBL:ENSBTAP00000031900 | 1.74 (1.38–2.19) | +0.799 | 0.0401 | 0.0968 |

These are relative LFQ effects, not absolute concentration differences. CIs are not simultaneous or selection-adjusted. No lower-abundance group meets primary 5% FDR.

## Cohort recovery and reconciliation
The first audit stopped because archive sample labels were incomplete. The authorized recovery obtained the original supplementary archive from the **official PMC Open Access S3 mirror**; a longer Europe PMC request independently returned the same supplement, byte for byte. Table S5, sheet `MatrixForLFQ_RawValues`, explicitly supplies the diagnosis/experiment mapping.

- ICM: **116, 23, 33, 6, 615, 618, 619, 632, 637, 640, 71, 9**.
- NF: **155, 172, 176, 18, 26, 54, 69, 70, 72, 73**.
- DCM: **621, 641, 648**, excluded from the ICM contrast.
- Archive-only experiments: **11, 24, 65, 115, 128**, absent from published Table S5 and excluded. Their diagnoses, relation to other donors and original omission reasons remain unknown; they are not asserted to be replicates.

Thus the archive's 30 experiment columns reconcile to **25 publication-listed samples plus five unexplained extra experiments**, and the selected contrast has **22 samples**. All 1,500 S5 protein-group IDs join the archived matrix. Every one of its 25 sample columns has identical zero/nonzero patterns and Spearman correlation 1.0 among jointly observed values with its named archive column. This corroborates the explicit labels; diagnoses were not inferred from abundance or sample-number patterns.

Each included experiment has six fraction runs already aggregated into its LFQ column. Fraction runs are not independent donors. The article describes the 25 samples as separate hearts; no repeated included donor is documented. Non-failing controls are donor hearts not accepted for transplantation, not a random healthy population.

The original submission describes two measurement batches, but no authoritative experiment-to-batch mapping was recovered. The submission's cohort counts also differ from the final paper. We use the final article/S5 labels, do not invent batch membership, and retain batch confounding as an unresolved limitation. The earlier audit, feasibility report and failed/incomplete review history remain preserved; their “mapping unresolved” conclusion is superseded by this recovery.

## Methods and quality control
Start from `proteinGroupss_LFQ.txt`, whose SHA-1 matches PRIDE: `02149d5c49eea651694c4f2207a650aa92f452cc`. The paper's 1,500-row table was used for labels and validation, not to restrict the analysis or import published p-values.

Remove the union of MaxQuant contaminant, reverse and site-only '+' flags: 217 of 2,917 groups, leaving **2,700**. Treat zero LFQ and NA as missing, log2-transform positive LFQ, and retain protein groups rather than collapsing shared accessions to genes. The primary analysis uses archived MaxLFQ normalization with no additional centering.

Across the 22 selected samples, **36.54%** of retained-group measurements are missing. Requiring at least **9/12 ICM and 7/10 NF observations** retains **1,327 groups**, with **4.26%** missing entries. **824** groups are observed in all 22 samples. The remaining 1,373 groups are not primary abundance tests, rather than negative findings.

Estimate ICM-minus-NF observed log2 means using Welch unequal-variance standard errors, Satterthwaite degrees of freedom and two-sided tests, with BH correction across the 1,327-group family. Limma was unavailable; these are **unmoderated** estimates. Individual clinical covariates and batch were not adjusted because sample-level information was unavailable. The frozen specification was saved before group testing.

The design has full rank. Primary p-values agree with SciPy's independent Welch-test implementation. Standardized within-group residuals and variance ratios are included in the result tables; the largest standardized residual is 3.10. No samples were removed after examining results.

## Sensitivity to missing measurements
The protein universe and group labels stay fixed.

| Analysis | Tested groups | Discoveries at BH q < 0.05 | Interpretation |
|---|---:|---|---|
| Available observations, no imputation | 1,327 | TF, CP-associated | Primary analysis |
| Fully observed groups, BH over 824 | 824 | Same two | Their q values are 0.0215 and 0.0249 |
| Fully observed groups, conservative original-family correction | 824; family size 1,327 | Same two | Other groups are untested, with p=1 only for correction |
| Low-intensity imputation, 5 parameter settings × 20 seeds | 1,327 per run | Same two in all 100 runs | No new FDR discoveries |
| Separate observed/missing Fisher tests | 2,700 | None | Separate BH family; low power for sparse groups |

Imputation draws from per-sample normal distributions shifted below the observed log2 mean, with SD-based downshift/width settings **(1.0,0.3), (1.8,0.3), (2.5,0.3), (1.8,0.2), (1.8,0.5)** and an upper truncation at the sample's observed fifth percentile. Seeds are 20260928–20260947 per setting. Observed measurements never change. These are conditional low-intensity stress tests, **not calibrated multiple-imputation inference**; p-values were not pooled.

TF and CP have no missing values, so their estimates and p-values remain unchanged; their BH q values also remain unchanged in these runs. This supports missing-data stability of this particular shortlist, not identification of the true missingness mechanism.

Incomplete proteins can still have materially different estimates: across runs, effect-rank correlation with the primary analysis is **0.707–0.860**, and **85.9–90.8%** of effects keep their direction across all 1,327 groups. For example, NQO2 changes from an observed-data log2 effect of −0.126 to a median imputed effect of +1.703, while remaining non-significant (imputation q at least 0.56). Ranking incomplete proteins as biological leads without this sensitivity check would be misleading.

There are **408 retained groups with an entire group unobserved**. No finite abundance effects are claimed for them; detection counts and separately corrected detection tests are provided. No significant detection test does not establish equal detectability.

## Other limits on robustness
**Normalization:** subtracting each sample's observed-protein median and adding the grand median reduces TF and CP effects to +0.767 and +0.676 log2, respectively; both have q=0.0968, and no protein passes 5% FDR in that sensitivity analysis. This is not batch correction. Median centering assumes broad proteome stability and is itself sensitive to which proteins are detected. Neither normalization is demonstrated to remove the unknown batch effects.

**Donor influence:** removing each donor in turn preserves positive effect directions for both groups in all 22 refits. TF retains q<0.05 in only **8/22** refits and CP in **5/22**. Reduced sample size also lowers power, so these counts are not proof of an outlier artifact. They do show that threshold-based conclusions are fragile.

**Protein identity:** the flag-filtered data still contain 25 groups whose accession strings include `CON__`, 19 of them primary-testable. CP's majority accession is P00450, but its group also contains a contaminant-database entry; peptide-level attribution was not reprocessed. In a disclosed **post hoc** sensitivity excluding all 19 such primary groups and recalculating BH, TF remains significant (q=0.0341); CP is excluded rather than “confirmed human CP.”

**Biological interpretation:** bulk-tissue composition, residual blood/plasma, donor selection, medications and unknown batches could contribute to TF/CP-associated signals. The data do not establish cardiomyocyte regulation, causality, circulating biomarkers, or applicability to women, earlier disease, DCM or HFpEF. No pathway enrichment was attempted on this two-group shortlist. Absence of FDR significance elsewhere is not evidence of no difference.

## Figures and reproducibility
The QC figure displays per-sample missingness, observed log2 LFQ distributions and detection-count coverage. Its boxplots show median/IQR and conventional 1.5-IQR whiskers; outlier markers alone are omitted from this overview, not from tests. Bubble area scales with the number of protein groups at each detection-count coordinate.

The sensitivity figure shows primary BH-adjusted evidence, observed-versus-imputed effects, and all 22 raw sample values for TF and CP. Circles/triangles and labels distinguish NF/ICM; black short lines mark means. It explicitly labels CP as an ambiguous group. PNG and vector PDF exports are general-purpose reports, not certified journal figures.

The complete analysis was rerun in a fresh Python namespace into a separate output directory. **Ten result/summary files matched byte-for-byte**, and all 132,700 per-protein imputation results matched after gzip decompression. Figure dimensions and labels were inspected; an initially clipped QC bubble was corrected. A validation-cell syntax error was corrected before the successful rerun, without altering primary results.

The reproducibility ZIP includes raw processed input, recovered original supplements, the 30-experiment manifest, all 2,700-group results with testability/identity flags, primary/sensitivity/detection tables, observed matrix and missingness mask, all imputation runs, figures, software versions, checksums, script, README and frozen specification. Run `python reproduce_analysis.py` from the extracted directory. No network or raw-spectrum processing is needed. Prior artifacts are not overwritten.

## Sources checked during recovery
- [Original article via PubMed](https://pubmed.ncbi.nlm.nih.gov/33670142/), [DOI](https://doi.org/10.3390/ijms22041787): cohort and methods.
- [PRIDE PXD021261](https://www.ebi.ac.uk/pride/archive/projects/PXD021261): processed matrix, SDRF, summary, parameter metadata and complete 185-file inventory reviewed across two pages. The 150 SDRF disease values are unavailable; the LFQ PDF is a field dictionary.
- [Publisher supplementary index](https://www.mdpi.com/1422-0067/22/4/1787/s1): prior attempt returned 403.
- [Publisher static ZIP candidate](https://mdpi-res.com/d_attachment/ijms/ijms-22-01787/article_deploy/ijms-22-01787-s001.zip): returned 403; no access-control bypass.
- [PMC article](https://pmc.ncbi.nlm.nih.gov/articles/PMC7916846/): supplied supplementary filename. Its `/articles/instance/7916846/bin/ijms-22-01787-s001.zip` route returned a browser challenge, which was not solved or bypassed.
- [Europe PMC supplementary service](https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7916846/supplementaryFiles): first request timed out; longer request succeeded and contained the original ZIP.
- PMC OA lookup routes at `https://www.ncbi.nlm.nih.gov/pmc/utils/oa/oa.fcgi?id=PMC7916846` and `https://pmc.ncbi.nlm.nih.gov/utils/oa/oa.fcgi?id=PMC7916846`: returned 404.
- [Official PMC public S3 supplement](https://pmc-oa-opendata.s3.amazonaws.com/PMC7916846.1/ijms-22-01787-s001.zip): public prefix listing located the file; download succeeded (3,707,431 bytes), identical to Europe's embedded ZIP. Table S5 supplies the verified mapping; the supplementary PDF supplies no extra clinical mapping.

Recovery resolved the essential group labels within the authorized 20-minute bound. PXD018678 was briefly assessed as a possible alternative from existing archive metadata and [its PubMed-indexed article](https://doi.org/10.1038/s41467-020-16584-z); it was not substituted or analyzed.
