Skip to main content
Back to timeline
arXivSource publication:

Survival analysis stops regressing outcomes: SOSM lays multi-omics data on a low-curvature manifold and separates risk out-of-fold in 184 kidney cancer cases

Synopsis

Survival analysis has long treated censored survival times as labels for supervised regression, which overfits in high-dimensional, low-sample settings. SOSM decouples representation learning from survival labels: a contrastive loss aligns RNA, DNA methylation and miRNA views, and a pseudotime-weighted graph Laplacian penalty pulls patient embeddings onto a low-curvature manifold, with survival labels used only for out-of-fold evaluation. On a 184-patient clear cell renal cell carcinoma multi-omics cohort under five-fold cross-validation, frozen embeddings with a penalized Cox model still separated high- and low-risk groups significantly on held-out test folds.

Source-provided article image: Prognosis Without Prediction: The Intrinsic Geometry of Survival
arXiv

Interpretation

SOSM projects each patient's multi-omics profile into a 16-dimensional latent space so that patients with similar prognosis sit close together geometrically, instead of having the network regress survival time directly. Earlier survival models, from proportional hazards to deep networks, fit survival time as an external label and overfit in high-dimensional, low-sample settings with thousands of features and few events; unsupervised factor analysis and network fusion reduce dimensionality but typically discard temporal dynamics. On the 184-patient fully matched TCGA-KIRC cohort, frozen encoder embeddings with a penalized Cox model still discriminated prognosis on held-out test folds of five-fold cross-validation, and the authors report clearly better discrimination than K-Means partitioning of the latent embeddings.

The training objective has two parts: one aligns the three omics views of the same patient in latent space, the other penalizes curvature among neighboring patients so trajectories line up smoothly. The alignment term is a normalized temperature-scaled cross-entropy loss with temperature 0.1; the curvature term builds a Gaussian kernel from a prognostic pseudotime obtained by single-sample gene set enrichment analysis, then penalizes the squared Frobenius norm of the normalized graph Laplacian projection. The pseudotime proxy aggregates proliferation and immune infiltration marker genes from the RNA modality and has near-zero linear correlation with raw clinical survival time, so the initial orientation is not a trivial stand-in for outcome labels; training loss decreased monotonically over 500 epochs without representational collapse.

High-risk and low-risk patients differ in local dissipation on the latent manifold, so risk stratification corresponds to geometric separation on that manifold. After splitting at the 75th percentile of out-of-fold Cox partial hazard scores, high-risk patients in the concatenated full cohort had markedly worse overall survival, with significant separation under the Mantel-Cox log-rank test. On the strictly held-out test split of the best fold, high-risk and low-risk groups still separated significantly; the authors also report higher mean local curvature dissipation in the high-risk group than the low-risk group, reading that difference as structural collapse on the manifold.

Perspective

The demonstrated scope is a 184-patient fully matched TCGA-KIRC multi-omics cohort with RNA-seq, DNA methylation and miRNA-seq, an aggregate event rate of 26.09%, and evaluation by five-fold cross-validation with a fixed seed, where survival labels are used only for out-of-fold evaluation and one-bit polarity calibration on training folds.

The initial manifold orientation depends on a pseudotime proxy built by single-sample gene set enrichment analysis of the transcriptome, aligned in sign by one-bit polarity calibration on training folds; that proxy has near-zero linear correlation with clinical survival time and is an uncalibrated heuristic. The authors propose replacing this domain-specific heuristic with intrinsic entropic mappings derived from optimal transport theory, and list validation on larger multi-center clinical cohorts and on non-biological industrial and financial systems as future directions.