Do Omics Add Early Risk? A Clinically Anchored, Time-Varying Cox Test

Multi-omic survival models, which predict how long cancer patients live from gene expression, DNA methylation, copy-number and mutation data, report impressive accuracy in their own papers. Yet the most careful neutral benchmark found that across 18 cancer types none of them reliably beat a plain Cox regression on age, sex and stage. We propose one specific explanation for this mismatch and a direct test of it. The explanation is that molecular risk may be time-limited: aggressive tumor biology shows up as excess deaths in the first year or two, while stage keeps predicting risk for the whole follow-up. A Cox model (the standard regression for time-to-event data, which assumes each variable multiplies risk by the same factor at every point in time) averages a short-lived molecular effect over the whole follow-up and dilutes it.

The proposal fits a model in two parts. First, the clinical Cox model is fit and then frozen as a fixed term. Second, a small learned term lets a compact molecular summary raise or lower risk by an amount that is allowed to change smoothly over time. If the dilution hypothesis is right, the molecular term should help most at short time horizons, and in cancers where the clinical model already does well it should stay close to zero. If the hypothesis is wrong, a constant-effect version of the same model will match it, and the field learns that the omics shortfall has some other cause.

Core Claim

The claim to be tested is that much of the prognostic value of multi-omic data lies in effects that change over time, and that two design choices lose it. The first is the proportional hazards (PH) assumption built into every deep multi-omic survival head we found. The second is letting molecular features compete with clinical covariates for weight. Concretely, we fit

$$h(t \mid c, x) = h_0(t)\,\exp\!\big(\hat\beta^\top c + \gamma(t)^\top z(x)\big),$$

where $h(t \mid c, x)$ is the instantaneous risk at time $t$ for a patient with clinical covariates $c$ and molecular profile $x$. The baseline risk over time, $h_0(t)$, is left unspecified as in the standard Cox model. The term $\hat\beta$ holds clinical coefficients estimated beforehand by a clinical-only Cox fit and held fixed. The vector $z(x)$ is a low-dimensional molecular summary of about 8 to 16 numbers. Finally, $\gamma(t)$ is a vector of time-varying coefficients written as a smooth spline in $t$ with few knots. Setting $\gamma(t)=\gamma$ recovers an ordinary PH residual model. The difference between the time-varying and constant versions is the quantity the study measures.

Relation to Prior Work

Three deep integration models motivate the gap because each uses a standard Cox head and never tests the PH assumption. SAEsurv-net (Stacked Autoencoder Based Multi-Omics Data Integration for Cancer Survival Prediction) compresses mRNA and copy-number data with stacked autoencoders and reports C-indices of 0.6212 (glioblastoma) and 0.6451 (breast). DeepSigSurvNet (Investigate the relevance of major signaling pathways in cancer survival using a biologically meaningful deep learning model) routes genes through KEGG pathway nodes and reports 0.7438 on lung adenocarcinoma. MOJO (Bimodal masked language modeling for bulk RNA-seq and DNA methylation representation learning) pretrains a joint RNA and methylation model and reports a pan-cancer C-index of 0.771. We reuse the MOJO encoder as one source of $z(x)$. Unlike all three, we do not let the molecular block share capacity with clinical variables, and we let its effect vary over time.

Large-scale benchmark study of survival prediction methods using multi-omics data constrains the design. In that benchmark, clinical-only Cox was rarely beaten, unstructured pooling of features underperformed, and CoxBoost was competitive only when clinical variables were protected from shrinkage. That last finding is why we freeze $\hat\beta$ rather than learn it jointly. The benchmark also leaves a gap we fill, because it never evaluated foundation-model embeddings.

Deep Cox Mixtures for Survival Regression supplies the evidence that PH violations cause miscalibration: expected calibration error was 0.0992 for Cox against 0.0200 for mixtures on FLCHAIN. Deep Cox Mixtures relaxes PH by mixing latent subgroups over all inputs. We instead relax it only for the molecular term, which keeps a readable clinical model and yields one interpretable curve $\gamma(t)$ per molecular axis.

Block Forests: random forests for blocks of clinical and omics covariate data matters because it was the only method to edge out clinical Cox on average in the neutral benchmark. Being tree-based, it never assumes PH, so it is the natural rival to our explanation.

Research Program

The study runs in three stages.

The first stage is a diagnostic pass across cancer types. In every cohort we fit a clinical-only Cox model and then an omic risk score inside the training fold. We then test whether that score's effect is constant over time, using Schoenfeld residuals (a standard check of whether a coefficient drifts over follow-up) and landmark analysis (refitting from survivors at fixed later times). This stage alone tells us how common time-varying molecular effects are before any new model is trained.

The second stage fits the anchored time-varying model with three choices of $z(x)$. The first is principal components computed within each fold on each modality, which needs no pretraining. The second is frozen MOJO embeddings taken from the released code instadeepai/multiomics-open-research. The third is the frozen unimodal BulkRNABert embedding (BulkRNABert: Cancer prognosis from bulk RNA-seq based language models), projected to the same dimension. MOJO's reported tolerance of missing methylation lets us score patients who lack that modality. We fit $\gamma(t)$ by penalized partial likelihood with time-varying covariates, and the roughness penalty is chosen by inner cross-validation.

The third stage interprets the result. For each cancer type we report the fitted $\gamma(t)$ curves with bootstrap bands, and the time beyond which molecular effects are no longer distinguishable from zero.

One caveat applies to the pretrained encoders. MOJO and BulkRNABert were pretrained on TCGA profiles without survival labels, so test patients' molecular profiles were seen during pretraining. The in-fold principal-component version of $z(x)$ is the leakage-free control for this.

Baselines

The naive reference point is the clinical-only Cox model on age, sex and stage, the model the neutral benchmark found so hard to beat. We will run it in the same folds; the file reports no numeric per-dataset C-index for it on Lacuna. Beating it means a better integrated Brier score in a majority of cohorts, together with a fold-paired improvement in time-dependent AUC at one and two years that survives multiple-testing correction.

The strongest penalized competitor is clinically protected CoxBoost together with Priority-Lasso, which fits clinical then molecular blocks one after another. Both were run in the neutral benchmark and both already anchor on clinical covariates, so they directly test whether our anchoring adds anything. They receive the same raw features, the same folds, and a tuning grid of matched size. Here too no reported per-dataset number was found on Lacuna.

The strongest non-PH competitor is Block Forest, which we run on the raw blocks with its standard block-sampling scheme and an equal tuning budget. No reported per-dataset C-index for Block Forest was found on Lacuna. The only figures available are its adjusted p = 0.027 against random survival forests and its marginal average gain over Cox.

For the deep model we fine-tune MOJO with its own Cox head on clinical features plus omics. Its reported pan-cancer C-index of 0.771 comes from its own split, so we re-run it in our folds rather than borrow that figure.

Two further comparisons cover relaxed PH and time variation. Deep Cox Mixtures on clinical features plus $z(x)$ is the general non-PH alternative. A landmark Cox model with a clinical offset fits separate constant effects in a few time windows and is a cruder way to capture time variation.

The ablation that isolates the key ingredient holds everything fixed and sets $\gamma(t)$ to a constant. A second ablation releases $\hat\beta$ so that it is fit jointly, which isolates the value of anchoring.

Why not just…?

Why not just use clinically protected CoxBoost or Priority-Lasso, which already anchor on clinical covariates? They are the closest existing designs, and on their own they did not reliably beat clinical Cox in the neutral benchmark. Our constant-$\gamma$ ablation with raw-feature principal components is close to Priority-Lasso. If that ablation matches the full model, anchoring gives no new benefit and time variation is the only contribution left to judge. If the full model also fails to beat Priority-Lasso, the hypothesis is refuted.

Why not just use Block Forest, which needs no PH assumption at all? This is a real threat, because a forest can pick up time-limited effects without any special modeling. Two outcomes are informative. If Block Forest wins mainly at short horizons, that independently supports the time-variation explanation. If it wins evenly across all horizons, its advantage comes from something else, such as interactions. Our model's claim then shrinks to interpretability, an explicit $\gamma(t)$ curve the forest cannot give, and we would say so.

Why not just use Deep Cox Mixtures over all features? Mixtures fix calibration on low-dimensional cohorts, but here they would combine clinical and molecular inputs into a single high-capacity model. The neutral benchmark warns that molecular noise in such combined models tends to drown out clinical signal. Running Deep Cox Mixtures on the same compact $z(x)$ separates the effect of the PH relaxation from the effect of anchoring.

Why not just do landmark analysis? If piecewise landmark fits match the spline $\gamma(t)$, we will report that, and the practical recommendation becomes the simpler method. The spline is chosen only for its smoothness and its single-pass estimation.

Proposed Experiments

Data come from the CMOB/MLOmics benchmark (CMOB: Large-Scale Cancer Multi-Omics Benchmark): 8,314 TCGA patients across 32 cancer types and four modalities. We restrict to cohorts with enough events and follow the neutral benchmark's protocol of repeated five-fold cross-validation with Uno's C-index and integrated Brier score. To these we add time-dependent AUC at one, two and five years and calibration error at each horizon.

The decisive checks are three.

The first check is the diagnostic. Molecular-score PH violations should be common, concentrated in cancers with strong molecular drivers, and should show effects that decay over time. The neutral benchmark noted that low-grade glioma did benefit from omics, so that cohort is a positive control.

The second check is the ablation. The full model should beat the constant-$\gamma$ version on short-horizon AUC and calibration, while long-horizon metrics stay level.

The third check is the cancer-type pattern. Across cohorts, the gain over clinical Cox should track how strongly the omic score violates PH. A null correlation across 30 or so cohorts would falsify the dilution explanation, whatever happens to aggregate C-index.