A prognostic gene signature is a weighted sum of a few genes' expression levels. It is meant to rank cancer patients by risk of death or relapse. Antenna engineers would recognize this object at once: it is a beamformer, a set of weights applied to many sensors so that the combined output responds strongly to one direction and weakly to the others. A signature built to track tumor proliferation will also respond, unintentionally, to how much stroma, immune infiltrate, or platform artifact a sample contains. In antenna terms these are side lobes, meaning energy picked up from directions the design never aimed at. This proposal measures each signature's side-lobe pattern against a fixed library of known nuisance axes. It tests whether that pattern, computed without any survival data, predicts which signatures will fail in external cohorts. It then fits signatures under an explicit side-lobe budget, the way array designers cap side-lobe level, rather than either ignoring off-target pickup or nulling it completely.
The opportunity matters because the field's best-documented failures look like side-lobe failures. Signatures with nearly disjoint gene lists agree on most patients because they share a proliferation "main lobe" (Comparison of prognostic gene expression signatures for breast cancer). Panels selected for immune or tumor-specific states visibly absorb background stromal and immune signal (Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma, An Immune-Related Prognostic Signature Predicts Overall Survival in Stomach Adenocarcinomas). External validation is often much weaker than internal validation: the stomach signature goes from p=0.001 in its training cohort to p=0.044 externally. If signal picked up off target is what fails to transport, it should be measurable before anyone looks at outcomes.
Core Claim
The claim has two parts. First, a signature's normalized response to cohort-variable nuisance axes predicts how much its concordance index falls from internal to external validation, and it does so better than gene count or training performance. The concordance index (C-index) is the probability that the model ranks a random pair of patients in the correct survival order. The nuisance axes are tumor purity, stromal score, immune score, and platform or batch indicators.
Second, a Cox model fitted with a side-lobe budget transfers better across cohorts than both plain LASSO Cox and hard removal of the nuisance axes. LASSO Cox is the standard sparse survival regression. Hard removal regresses the nuisance out before fitting. The budget is not uniform across axes. It is tight on nuisance axes whose association with survival is unstable across training cohorts, and absent on axes whose association is stable.
The second part resolves the tension the research file identifies. Immune infiltration can carry real prognostic value, as in the "hot versus cold" stomach result. In those cases it should be treated as part of the main lobe and not as leakage.
The Wireless Ingredient, Stated Honestly
What transfers from the wireless literature is a design stance and a diagnostic condition, not a theorem.
Gaussian 1-2-1 Networks with Imperfect Beamforming models leakage with an explicit side-lobe gain $\beta$. It shows that treating leakage as noise costs only a bounded amount when the main-to-side-lobe ratio dominates the network degree, a diagonal-dominance condition. This paper motivates our diagnostic, but its capacity bound does not carry over to survival likelihoods, and we make no such claim.
Elevation Beamforming With Full Dimension MIMO Architectures in 5G Systems: A Tutorial shows that abstractions which ignore side lobes, by treating each antenna port as an idealized mask, overstate signal-to-interference ratios. Modeling individual antenna elements corrects this. We read it as a warning that applies directly here: a signature summarized only by its internal AUC is a port-level abstraction.
Sectoring in Multi-cell Massive MIMO Systems shows that directional filtering, using main-lobe and back-lobe gains, cuts contamination roughly threefold. The point we take from it is that bounded rejection, not perfect nulling, captures most of the gain.
The formal object is simple. Let $x$ be a patient's standardized expression vector and $w$ the signature weights, so the risk score is $s = w^\top x$. Let $a_k$ denote nuisance score $k$, for example ESTIMATE purity. We define the side-lobe gain $g_k(w) = |\mathrm{corr}(w^\top x, a_k)|$, and the main-lobe gain $g_0(w)$ as the score's correlation with the training-pooled prognostic direction. Both quantities can be computed on any cohort without survival labels.
By analogy with diagonal dominance, a signature is "leakage-dominated" when $\sum_k g_k(w) \ge g_0(w)$. The budgeted fit is
$$\hat w = \arg\max_w \; \ell(w) - \lambda \|w\|_1 \quad \text{s.t.} \quad g_k(w) \le \epsilon_k \text{ for } k \in \mathcal{U},$$
where $\ell$ is the Cox partial log-likelihood, $\lambda$ the sparsity penalty, and $\mathcal{U}$ the set of nuisance axes whose hazard ratio is heterogeneous across training cohorts by a Cochran Q test. The budget $\epsilon_k$ plays the role of a Dolph–Chebyshev side-lobe level. It is tuned by leave-one-cohort-out validation, never on the test cohort.
Research Program
The work has three stages.
The first is a leakage atlas. We reimplement published signatures from their reported genes and weights: the ccRCC 6-gene model, the stomach 3-gene formula, the LUAD 3-gene model from Genomic and immunogenomic analysis of three prognostic signature genes in LUAD, and the 70-gene, 76-gene, and Gene Expression Grade Index signatures. To these we add several hundred LASSO signatures obtained by bootstrap resampling of each training cohort. Bootstrap resampling is precisely the source of the gene-list instability documented in the breast comparison. For each signature and cohort we compute the side-lobe vector $g(w)$.
The second stage tests the diagnostic claim. We regress each signature's internal-to-external C-index drop on its side-lobe gains, with gene count and internal C-index as competing predictors.
The third stage fits budgeted signatures and compares how well they transfer. A side analysis on METABRIC asks whether pairwise disagreement among the three breast signatures is explained by differences in their side-lobe vectors. This is a falsifiable reading of the finding that the 70-gene signature and GGI agree on 88% of patients despite almost no gene overlap.
Baselines
The naive reference is clinical staging alone, meaning a Cox model on age, stage, and grade. The research file reports that molecular scores add independent information beyond it, so any method that fails to beat staging externally is uninformative.
The main competitor is plain LASSO Cox. It is the method behind the ccRCC and LUAD signatures and is the field's default. Reported numbers are internal to their cohorts: time-dependent AUC "consistently above 0.65" in TCGA-KIRC, and 1-, 3-, and 5-year AUCs of 0.916, 0.90, and 0.863 in TCGA-LUAD. For the breast signatures, the TRANSBIG comparison reports a 10-year C-index of 0.90 for the 70-gene signature and 0.80 for GGI and the 76-gene signature. We will refit LASSO on the same training cohorts, with the same gene universe, the same leave-one-cohort-out tuning, and the same number of hyperparameter evaluations as the budgeted model.
The strongest existing method in the file is discretized topic modeling with multi-task logistic regression, from Gene Expression based Survival Prediction for Cancer Patients: A Topic Modeling Approach. It reports C-index 0.7202 on METABRIC and 0.8495 on KIPAN and passes the D-calibration test, which checks whether predicted survival probabilities match observed outcomes. We run it on the same splits. To keep the comparison fair, the C-index we report for it is computed on held-out cohorts rather than its original internal splits.
Two ablations isolate the key ingredient. The first, zero-forcing, residualizes all nuisance axes out of expression before LASSO. The second applies the uniform budget, the same $\epsilon$ on every axis without the heterogeneity-based choice of $\mathcal{U}$.
Beating the baselines means a higher mean external C-index than LASSO and zero-forcing across held-out cohorts, with a bootstrap confidence interval excluding zero, and calibration no worse than the topic model. Matching the topic model's discrimination with a 10–20 gene linear score would also count, because that score is far cheaper to deploy.
Why not just…?
Why not just add purity and immune scores as covariates in the Cox model? Covariate adjustment changes how the score is tested inside one cohort, but it leaves the deployed weights unchanged. The fixed score still carries its side lobe into a new cohort with a different purity mix. Adjustment also requires measuring the nuisance at deployment. We include "LASSO plus nuisance covariates" as an additional baseline. If it transports as well as the budgeted score, the method reduces to a reporting recommendation, and we will say so.
Why not just regress the nuisance out first, as in zero-forcing? This is the real threat, and it is the main ablation. The stomach immune signature shows that infiltration can be prognostic, so removing it blindly should cost discrimination wherever that association is stable. The decisive comparison is per axis: zero-forcing should lose on axes outside $\mathcal{U}$ and tie on axes inside it. If zero-forcing ties everywhere, the heterogeneity-based budget adds nothing.
Why not just treat off-target expression as noise, as LASSO implicitly does? The wireless result that justifies treating leakage as noise depends on the main gain dominating. The atlas directly measures how often published signatures violate that condition. If most signatures are not leakage-dominated and the diagnostic has no predictive power, then treating leakage as noise is adequate here and the project's premise fails.
Proposed Experiments
The cohorts are those in the file: TCGA-LUAD trained, then tested on GSE50081, GSE11969, and GSE42127; TCGA-KIRC paired with KIPAN; TCGA-STAD paired with its reported GEO validation cohort, whose identity we will verify because the listed accession is usually a lung dataset; and METABRIC for the breast analyses. Nuisance scores come from ESTIMATE, the purity, stromal, and immune scoring tool already used in these papers, together with cohort and platform indicators.
The first decisive check is pre-registered. It requires the side-lobe gains to explain significantly more variance in external C-index drop than gene count plus internal C-index, across bootstrap signatures pooled over cancer types.
The second check is leave-one-cohort-out external C-index, plus D-calibration, comparing the budgeted fit against all baselines.
The third check is the METABRIC disagreement decomposition.
The project is falsified if the first check fails. In that case side-lobe gain would carry no information about transport beyond what reviewers already look at, and the beamforming lens would be decoration rather than mechanism.
References
- Comparison of prognostic gene expression signatures for breast cancer
- Identification of a Six-Gene Prognostic Signature Characterized by Tumor Microenvironment Immune Profiles in Clear Cell Renal Cell Carcinoma
- An Immune-Related Prognostic Signature Predicts Overall Survival in Stomach Adenocarcinomas
- Gaussian 1-2-1 Networks with Imperfect Beamforming
- Elevation Beamforming With Full Dimension MIMO Architectures in 5G Systems: A Tutorial
- Sectoring in multi-cell massive MIMO systems
- Genomic and immunogenomic analysis of three prognostic signature genes in LUAD
- Gene Expression based Survival Prediction for Cancer Patients: A Topic Modeling Approach