Most cancer prognostic signatures are built the same way. A researcher takes tumor RNA levels from a pooled cohort of men and women, keeps the genes associated with survival, and fits a sparse Cox regression (a survival model in which each gene multiplies the patient's hazard by a fixed factor). The weighted sum of the selected genes becomes a risk score. This recipe assumes a gene means the same thing for prognosis in both sexes. Two network studies in colon and lung cancer suggest this assumption fails in a specific way.
In those studies, male and female tumors had essentially identical RNA levels outside the sex chromosomes. What differed was how strongly transcription factors appeared to regulate genes. The regulation of drug-metabolism genes was markedly stronger in women, and it predicted survival after chemotherapy only in women. Here "regulation" means a patient-specific network estimate of how many transcription-factor inputs each gene receives, called its targeting score.
We propose to test whether this is a general rule that can be built into prognostic models. The model would pool information across sexes where the biology is shared and allow sex-specific deviations where it is not. It would choose among two kinds of features: RNA levels and regulatory targeting. The prediction is that the sex-specific part of prognosis should concentrate in the targeting features. It should also improve calibrated survival probabilities for each sex, not just the ranking of patients.
Core Claim
Our hypothesis comes in two parts. First, in non-reproductive cancers treated with chemotherapy, a partially pooled survival model will predict survival better than either a pooled model or two separate sex-specific models. Better here means a lower Integrated Brier Score (IBS, the time-averaged squared error of predicted survival probabilities) within each sex. Second, when the model can use both expression and targeting features, its sex-specific deviation terms will land mostly on targeting features, especially drug-metabolism pathways.
If the second part fails, the abundance-versus-regulation story is wrong. The failure would look like one of two outcomes: the deviations land on expression features, or targeting features add nothing once expression is in the model. In that case sex-aware signatures are just interaction models on RNA levels. That would still be a useful, cheaper conclusion for the field.
Grounding in Prior Work
Two studies motivate the regulation half of the claim.
- Colon cancer. Gene regulatory network analysis identifies sex-linked differences in colon cancer drug metabolism processes built an aggregate network with PANDA, which combines transcription-factor binding motifs, protein interactions and co-expression. It then derived per-patient networks with LIONESS. In 445 TCGA colon tumors there were no autosomal expression differences between sexes, but drug-metabolism genes were more heavily targeted in women. Among stage III women given adjuvant chemotherapy, 10-year survival was 89% with high targeting versus 61% with low targeting. Men showed no such effect.
- Lung adenocarcinoma. Gene regulatory networks reveal sex difference in lung adenocarcinoma found the same female-specific drug-metabolism targeting effect after chemotherapy (p = 0.016).
Neither paper built a multi-gene risk model, evaluated calibration, or asked whether men have their own regulatory predictors. We reuse their network method but change the downstream question: from "is one pathway score associated with survival" to "where in feature space does sex-specific prognostic signal live, and does modeling it improve calibrated prediction."
Prognostic Significance of a 9-Gene Signature Related to Tertiary Lymphoid Structures in Male Stomach Adenocarcinoma Patients supplies both motivation and a cautionary design lesson. Its immune-structure signature was adverse only in men (HR 2.68, p = 0.02 in TCGA) and turned protective in an all-adjuvant validation cohort (HR 0.61). It also contains the Y-linked gene EIF1AY, so it cannot be scored in women at all. Our model handles this by allowing sex-chromosome features only in the sex-specific deviation terms, never in the shared weights.
Two papers shape the method rather than the motivation:
- The standard expression pipeline (univariate Cox screening, then sparse Cox, then a risk score) of Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma is our pooled reference recipe.
- Stop Chasing the C-index: This Is How We Should Evaluate Our Survival Models is the evaluation constraint. It shows that discrimination alone hides miscalibration and that censoring is often not random. We therefore treat IBS and calibration within each sex as primary metrics.
Two further results constrain how much we can expect:
- Comparison of prognostic gene expression signatures for breast cancer found that signatures sharing almost no genes still agreed on risk. That points to a common proliferation signal that could swamp any sex-specific component.
- Genome-Wide Sex and Gender Differences in Cancer stresses that splitting a cohort by sex halves the sample. This is the main reason we pool partially rather than stratify.
Research Program
For each cohort we compute two feature blocks on the same tumors:
- Expression: standardized RNA levels.
- Targeting: LIONESS per-patient in-degree for each gene, summarized to pathway-level scores to reduce noise.
Both blocks are standardized within cohort so that weights can transfer across sequencing and microarray platforms. The survival model for a patient of sex $s \in \{F, M\}$ with features $x$ is
$$h_s(t \mid x) = h_{0,s}(t)\,\exp\!\big(x^\top(\beta + \delta_s)\big),$$
where the terms are:
- $h_{0,s}(t)$ is a separate baseline hazard for each sex, so absolute survival levels can differ.
- $\beta$ is a shared weight vector.
- $\delta_s$ is a sex-specific deviation.
We fit by maximizing the Cox partial likelihood minus $\lambda_1\lVert\beta\rVert_1 + \lambda_2\sum_s \lVert\delta_s\rVert_1$. The two tuning weights are chosen by nested cross-validation stratified by sex and event status. When $\lambda_2$ is very large, the model collapses to a pooled model with sex-specific baselines. When $\lambda_1$ is very large, it becomes two separate models.
Adjuvant chemotherapy status enters as a stratification variable, and the main analyses use chemotherapy-treated patients. The two network papers suggest the female drug-metabolism signal appears specifically after chemotherapy. We also check whether censoring depends on sex or features by fitting a censoring model. Where it does, the IBS is computed with inverse-probability-of-censoring weights.
The scientific readout is where the nonzero $\delta_s$ fall. We measure this as the fraction of total deviation magnitude carried by targeting versus expression features, with a bootstrap interval. We compare it against a null distribution obtained by permuting sex labels within event strata.
Baselines
Pooled signature with a sex covariate. This is the naive reference. We run the six-step pipeline of the ccRCC six-gene signature on expression, with sex as a covariate. Training splits, feature budget and tuning folds are identical to ours. It is what the field ships by default, so beating it within each sex is the minimum bar.
Two fully separate sex-specific sparse Cox models. This is the natural sex-aware competitor. Each is fit on expression plus targeting with its own cross-validated penalty. It is strongest when sex differences are large. We must beat it in the smaller sex stratum, where we expect its variance to hurt.
Single pathway targeting score. This is the strongest existing sex-specific method on our cohorts. It is the drug-metabolism targeting score of the colon network study, used as the sole predictor in a Cox model within each sex. It reported 89% versus 61% 10-year survival in chemotherapy-treated stage III women (p = 0.034), with validation in the GEO meta-cohort (FDR = 0.009). The lung analogue reported p = 0.016. No reported C-index or IBS for any sex-specific signature on these cohorts was found on Lacuna. We will therefore recompute this baseline's C-index and IBS ourselves on identical folds. Beating it means a lower female IBS at 5 and 10 years, with a non-overlapping bootstrap interval, plus any gain at all in men, where it is expected to be uninformative.
Expression-only partial pooling. This is the key ablation. It is the same model with the targeting block removed. If it matches the full model, regulatory features are not the ingredient that matters.
Multimodal deep model. For context, we include the multimodal model of Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis, restricted to its transcriptomic branch, if compute allows. It reported an average 2.9% C-index improvement across TCGA cohorts. No per-cohort or per-sex figure is in the file. It gets the same tuning budget in GPU-hours as our grid search, and is judged on IBS within each sex.
Why not just…?
Why not just add sex as a covariate to a pooled signature? A covariate shifts the baseline hazard but keeps the same gene weights for both sexes. The stomach immune signature (adverse in men, null in women) and the female-only drug-metabolism effect are both differences in slope, which a shifted baseline cannot represent. The pooled baseline above quantifies this directly. If it matches our model on IBS within each sex, the claim fails.
Why not just fit separate models for men and women? Halving the sample is costly in cohorts where chemotherapy-treated women number in the low hundreds, as the pan-cancer review notes. Separate models also cannot tell shared prognostic biology from sex-specific biology. This is a real threat in cancers with large sex effects. Our model contains separate fitting as a special case, so cross-validation will select it if the data favor it, and we report when that happens.
Why not just fit sex-by-gene interactions on expression, or use pathway expression scores, and skip networks entirely? This is the most serious alternative. Targeting scores are computed from expression data plus prior knowledge, so a flexible expression model might recover the same signal. The colon study's finding of identical expression but divergent targeting argues against this, but it tested group differences, not prediction. The expression-only ablation settles the question. We will also give the expression arm pathway-level enrichment scores computed on the same gene sets as the targeting pathways, so the comparison isolates regulation rather than aggregation. If targeting adds nothing over this arm, we will say the network step is unnecessary.
Why not just let a deep multimodal model learn the interactions? Such models are judged mostly on C-index. As Stop Chasing the C-index argues, a model can rank well overall while being miscalibrated within subgroups. A pooled model also cannot separate shared from sex-specific signal in a form a clinician can inspect. We include it as a baseline. If it wins on IBS within each sex, then interpretability becomes our only remaining argument, and we will say so.
Proposed Experiments
Discovery and validation cohorts.
- Colon: discovery on TCGA colon adenocarcinoma (445 patients: 238 men, 207 women). External validation on the five-study GEO Affymetrix meta-cohort (1,193 patients: 621 men, 572 women), refitting only the baseline hazards.
- Lung adenocarcinoma: discovery on TCGA, validation on GSE68465 and LGRC.
- Stomach: TCGA (362 patients) for discovery, GSE26253 (432 patients, all adjuvant-treated) for validation. This pair tests whether the model recovers a male-specific immune deviation without depending on Y-linked genes.
Primary decisive check. In colon validation, the full model must beat the strongest baseline on female IBS at 10 years. In addition, at least half of the female deviation magnitude must fall on targeting features, above the sex-permutation null. Both results must hold for the claim to stand.
Secondary checks.
- Lung replication: the same targeting-dominance pattern in lung adenocarcinoma.
- Calibration: calibration slopes within each sex near 1 in external cohorts, where the pooled baseline should drift.
- Proliferation control: a deliberate attempt to break our own result. We add a proliferation-pathway score to the shared part and ask whether the sex deviations survive. This answers whether the proliferation common denominator documented in the breast signature comparison masks sex-specific drivers or explains them away.
Negative control. All analyses are repeated in chemotherapy-naive patients. There, the two network papers predict the female drug-metabolism deviation should shrink. Finding it unchanged would suggest the effect reflects generic prognosis rather than drug handling.
References
- Gene regulatory network analysis identifies sex-linked differences in colon cancer drug metabolism processes
- Gene regulatory networks reveal sex difference in lung adenocarcinoma
- Prognostic Significance of a 9-Gene Signature Related to Tertiary Lymphoid Structures in Male Stomach Adenocarcinoma Patients
- Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma
- Stop Chasing the C-index: This Is How We Should Evaluate Our Survival Models
- Comparison of prognostic gene expression signatures for breast cancer
- Genome-Wide Sex and Gender Differences in Cancer
- Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis