Heart Failure Proteomics Module Discovery

What this research found

Heart failure has several distinct underlying causes, so a protein signature that only holds for one of them is of limited use. The human left-ventricular proteome from a published mass-spectrometry dataset of 34 hearts was re-analysed to find groups of proteins that move together, separate decompensated heart failure from both non-failing references, and behave consistently in ischemic, dilated and hypertrophic disease. One module of 565 proteins met all three tests and is the only one recommended for experimental validation — with the explicit caveat that part of its signal may come from blood retained in congested tissue rather than from heart cells.

• One of five co-abundance modules qualified for validation. Module M3, with 565 proteins, is coordinately raised in decompensated failure against both non-failing references (false-discovery rate below 0.001), shows no evidence of divergence between etiologies (I-squared = 0%, Cochran's Q p = 0.43 and 0.74), and survives leave-one-heart-out resampling at a mean absolute correlation of 0.82.
• Ranked by their smallest absolute fold change across all six group contrasts, a deliberately conservative robustness criterion, 31 proteins reach the Validate tier and a further 44 are Replicate-First. Eighteen of the top 20 Validate candidates sit in M3, led by COL14A1, SERPIND1, SERPINA4, HRG, AHSG, CLEC3B, PLG and SYNM.
• The second module is reproducible but etiology-divergent, at an I-squared of roughly 82%, so it is held back for replication before any validation spend; the remaining three modules were deprioritised as unstable or uninformative.
• The dominant confound is surfaced rather than buried. Several top M3 candidates are abundant plasma and acute-phase proteins whose elevation could reflect blood retained in congested, oedematous myocardium, so the validation plan adds perfusion and haemoglobin covariate controls and prioritises the resident-cell candidates COL14A1, SYNM and FLOT1.
• CRP, the obvious inflammatory candidate, was quantitatively deprioritised on an I-squared of about 58% and retained only as a reference marker rather than a validation target.

How it was done

Left-ventricular mass-spectrometry data for 34 hearts — 7 normal donors, 6 with compensated hypertrophy, and 6 ischemic, 6 dilated and 9 hypertrophic failing hearts — were retrieved from PRIDE and taken through a missingness cascade that reduced 3,850 protein groups to 2,878, followed by down-shifted imputation of low-abundance missing values. Signed weighted co-abundance network analysis produced five modules, and group differences were modelled with age- and sex-adjusted empirical Bayes. Consistency across the three failing etiologies was tested with Cochran's Q and I-squared, and module stability with 34 leave-one-heart-out iterations plus 100 bootstrap resamples, feeding a three-tier triage into Validate, Replicate-First, and Deprioritize. The output is a 23-page report with 8 figures, 7 tables and 41 references, plus a tiered validation plan from targeted mass spectrometry through antibody assays to functional work, restricted throughout to total protein abundance with no inference about post-translational modifications.

Data sources

• ProteomeXchange and PRIDE accession PXD008934 — human left-ventricular proteome from Chen et al., Nature Medicine 2018; 34 hearts, 2,878 protein groups after quality control

Figures from this analysis