Proteomics + lipidomics integration via Similarity Network Fusion (SNF) to stratify Stage C/D heart failure patients and validate cluster-based risk in an independent asymptomatic cohort.
Data access: Raw patient data cannot be shared due to privacy restrictions. Aggregated intermediate objects and summary statistics are provided in
/data. Contact the corresponding author for data access requests.
| Discovery cohort | Stage C/D HF |
| Validation cohort | Stage A/B HF |
| Omics layers | Proteomics (Olink) + lipidomics (4D-LC-TIMS lipidomics) |
| Primary outcome | Worsening heart failure (WHF), time-to-event |
-
Data preprocessing — Imputation, near-zero-variance removal, z-score scaling, and decorrelation (Pearson r > 0.9) applied independently to each omics panel.
-
Network fusion (SNF) — Per-panel affinity matrices fused via Similarity Network Fusion; spectral clustering tested over K = 3–10.
-
Cluster quality evaluation — Harrell's C-index, log-rank tests, and NMI used to select the optimal K and compare SNF against single-modality clusterings.
-
Cluster characterisation — Kruskal-Wallis + BH FDR on all features; cluster-wise z-scores; UpSet plots for exclusive markers; LASSO multinomial regression for clinical predictors.
-
Stability & comparison — Bootstrap subsampling (B = 100, 80% of samples) reporting ARI and NMI; MOFA2 run as a comparison multi-omics method.
-
Validation cohort transfer — Per-cluster Random Forest classifiers trained on the discovery cohort project cluster labels onto Stage A/B patients; survival and stage-progression outcomes tested in the transferred clusters.