Graph-Based Multi-Omics Integration Improves Subtype Recovery and Survival Prediction Over Classical Integration Strategies in TCGA-BRCA

TL;DR

Graph-based multi-omics fusion recovers breast cancer subtype biology more faithfully than feature concatenation and outperforms the weakest unimodal baselines in survival prediction.

Abstract

Background. Breast cancer comprises at least five molecular subtypes with distinct prognoses, yet PAM50 classification relies on transcriptomics alone. Whether integrating DNA methylation and copy number data improves subtype recovery and survival prediction over single-omic baselines remains an open question. Methods. We applied Similarity Network Fusion (SNF) to n = 644 TCGA-BRCA patients with matched RNA-seq, 450k DNA methylation, and GISTIC2 copy number profiles. Per-modality patient similarity networks were iteratively fused (K = 20, T = 20, u = 0.5) and partitioned by spectral clustering; k = 2 was pre-specified on eigengap and silhouette criteria. SNF was benchmarked against RNA-only, CNV-only, methylation-only, and early concatenation baselines using PAM50 NMI for subtype recovery and out-of-fold concordance index (OOF C-index) from a Ridge Cox model with N = 1,000 bootstrap CIs for pairwise comparisons. Results. SNF produced a stable two-cluster partition (stability ARI = 1.00, silhouette = 0.228), with NMI = 0.495 versus PAM50, exceeding RNA-only (0.428) and early concatenation (0.175). IHC receptor data confirmed cluster biology independently (ER+: 92.8% vs 15.6%; triple-negative: 1.0% vs 45.4%; both p < 10^-100). SNF achieved an OOF C-index of 0.681 (95% CI 0.610-0.760), significantly outperforming CNV-only (Delta = +0.122, CI 0.020-0.211); the advantage over RNA-only (Delta = +0.049, CI -0.036-0.144) did not exclude zero. Conclusion. Graph-based multi-omics fusion recovers breast cancer subtype biology more faithfully than feature concatenation and outperforms the weakest unimodal baselines in survival prediction. The improvement over RNA-seq alone is positive in direction but not yet statistically conclusive at this cohort size, pointing to the trade-off between integration complexity and the sample sizes needed to quantify its marginal benefit.

Figures (14)

• Figure 1: PC1 vs PC2 of the ComBat-corrected methylation matrix, coloured by TCGA tissue source site (TSS). No residual plate-driven clustering is visible, confirming that batch effects were successfully removed before downstream analysis.
• Figure 2: Overview of the SNF $k = 2$ partition. Top-Left: UMAP of the fused 50-dimensional spectral embedding, coloured by cluster assignment (cluster 0 = Luminal-enriched, cluster 1 = Basal-enriched). Top-Right: same UMAP coloured by PAM50 intrinsic subtype label. Bottom-Left: Kaplan-Meier overall survival curves for the two clusters (log-rank $p = 0.144$).
• Figure 3: Fused patient similarity matrix ($W_{\text{fused}}$, $644 \times 644$) after $T = 20$ SNF diffusion iterations ($K = 20$, $\mu = 0.5$). Patients are sorted first by $k = 2$ cluster assignment (cluster 0: $n = 503$; cluster 1: $n = 141$), then by descending within-cluster affinity sum. Colour encodes pairwise similarity. Cluster boundaries are marked with white lines; the PAM50 annotation bar is shown above.
• Figure 4: Per-modality affinity matrices alongside the fused matrix, all using the same patient ordering and colour scale (vmax $= 0.08$). Inter-cluster contrast is strongest in RNA-seq and methylation and weakest in CNV, which carried the least PAM50 information (NMI $= 0.037$; Table \ref{['tab:baselines']}).
• Figure 5: Mean silhouette coefficient for the fused SNF embedding at $k = 2$--$6$. The maximum at $k = 2$ ($s = 0.228$) provides one of the two pre-specified criteria for selecting the primary partition number; all other values are below 0.06.