Exposure-Free Sex Contrasts: Cancelling Unrecorded Smoking and Drinking from Sex Differences in Tumors

Men and women get many of the same cancers at different rates and respond to different drugs. Researchers want to know how much of that gap is biology (sex chromosomes, hormones, immune wiring) and how much is behavior, since men in most cohorts have historically smoked and drunk more. The trouble is that the large tumor databases rarely record those exposures well, so a "sex-biased gene" may simply be a smoking-biased gene. We borrow one specific trick from satellite navigation to separate the two. A GPS receiver cannot measure the delay the ionosphere adds to a signal, but it knows that this delay scales differently across two radio frequencies. It therefore forms a weighted combination of the two measurements in which the delay cancels exactly, without ever estimating it. Tumors carry something similar: carcinogen exposure leaves footprints, in the form of mutation patterns and exposure-responsive gene programs, that appear in different molecular readouts with different strengths. If the ratio of those strengths can be calibrated, a weighted combination of readouts cancels the exposure for every patient, whatever their unrecorded dose. What remains is a sex contrast that behavior cannot explain.

The core claim is narrow and testable. When the confounding exposure is observed only through noisy tumor-derived proxies, the usual fix of adding the proxy as a covariate leaves residual confounding in proportion to how noisy the proxy is. An exposure-cancelling combination, with its weight calibrated by within-sex instrumental variation, removes that residual bias. Because the calibration pools both sexes, it does so without halving the sample, and the price is a predictable increase in variance. We will build this estimator, check it against tumors where smoking history is recorded, and re-test published sex-specific findings in lung adenocarcinoma, liver and bladder cancer.

Where the two directions meet

In navigation, the ionosphere-free combination works because the nuisance delay is known to scale as one over frequency squared. The companion "geometry-free" combination isolates the delay itself, and regional ionosphere maps are built from it (Wen Jung Li et al.; Yang Wang 0077 et al.). The field also knows the cost: the combination amplifies measurement noise because its coefficients are large ratios. Practitioners accept that cost because the alternative, modeling the delay, leaves systematic error. The comparison of conventional ionosphere-free positioning with temporal self-differencing in Kaihui Lv et al. shows how much the choice of combination matters. In their dual-system tests, their self-differencing method reached 0.06 m 3D RMS in 152 s, against 0.132 m in about 700 s for conventional ionosphere-free positioning. We take the navigation work as the source of the logic: cancel a nuisance through calibrated cross-channel scaling rather than estimating it per observation, and budget explicitly for noise amplification. We do not claim any physical analogy beyond that.

On the oncology side, the motivating evidence is that sex effects in tumors are real but tangled with exposure. The pan-cancer whole-genome analysis in Sex differences in oncogenic mutational processes used regression adjusted for age, ancestry and subtype on 1,983 genomes. It found higher mutation density in males and male enrichment of the liver-associated signature SBS16, and it explicitly names missing exposure data (pack-years, alcohol, occupation) as a limitation. In lung adenocarcinoma, Multi-omics protein signaling networks identify sex-specific therapeutic candidates reported 16 sex-specific drug candidates, among them female-biased AURKA and male-biased MAPK14. Robert Hsu et al. found that hormone receptor expression in lung cancer is higher in women and tracks driver mutations such as EGFR and KRAS, whose frequencies are themselves tied to smoking. In bladder cancer, men have threefold higher incidence but women fare worse once disease is muscle-invasive (Keiichiro Mori et al.), and smoking is a dominant risk factor there.

Some findings are almost certainly biology and serve as our positive anchors. These include the androgen receptor–CCRK–β-catenin loop in liver cancer (Hai Feng et al.), the sex-opposite immune drivers of glioblastoma shown by bone-marrow chimeras (Defne Bayık et al.), and X-linked effects such as KDM6A and escape from X-inactivation (Camila M. Lopes-Ramos et al.). A method that erased these would be wrong.

The research file found no prior work applying ionosphere-free or differencing logic to sex differences in cancer. The obvious naive transfer is "tumor minus normal, then male minus female." That is ordinary difference-in-differences, and a sex-by-tissue interaction term already does it. We deliberately do not propose that.

Core Claim

Write a molecular feature (a gene's expression, a protein, a mutation count) for patient $i$ as

$$Y_i = \beta S_i + \gamma E_i + \epsilon_i, \qquad Z_{1i} = \lambda_1 E_i + u_{1i}, \qquad Z_{2i} = \lambda_2 E_i + u_{2i}.$$

Here $S_i$ is sex, $E_i$ is the unrecorded exposure dose, $\beta$ is the biological sex effect we want, and $\gamma$ is the feature's sensitivity to exposure. $Z_1$ and $Z_2$ are two tumor-derived exposure footprints with sensitivities $\lambda_1, \lambda_2$ and independent noise $u_1, u_2$. Examples of footprints are the activity of a carcinogen-associated mutational signature and the score of an exposure-response expression program. If we knew $\kappa = \gamma/\lambda_1$, the combination $Y_i - \kappa Z_{1i}$ would contain no exposure term at all. This is the analogue of the ionosphere-free observable.

We estimate $\kappa$ by the instrumental-variable ratio within sex, $\hat\kappa = \mathrm{Cov}(Y, Z_2 \mid S)/\mathrm{Cov}(Z_1, Z_2 \mid S)$. This uses the second footprint to cancel the noise in the first, and it pools the within-male and within-female covariances. It plays the role of the geometry-free calibration step, and because it pools both sexes it does not halve the cohort.

The exposure-free sex contrast is then the sex difference in $Y - \hat\kappa Z_1$, with the usual adjustments for age, ancestry and subtype. By contrast, adjusting for $Z_1$ as a covariate leaves bias of roughly $\gamma\,(1-r_1)\,\Delta E$. Here $r_1$ is the reliability of $Z_1$ (the fraction of its variance due to $E$), and $\Delta E$ is the male–female exposure gap. This is the textbook attenuation from measurement error. The combination removes this bias at the cost of variance inflated by about $\kappa^2 \mathrm{Var}(u_1)$, which is the same noise-amplification trade the navigation literature reports.

The estimator belongs to the family of proxy and negative-control methods in causal inference. What is new is the instantiation: using the tumor genome's own exposure footprints as the paired "frequencies." We also add a validation design that measures how much of published sex dimorphism survives.

Research Program

The first component builds the footprints. For tobacco, these are the activity of tobacco-associated mutational signatures fitted to each tumor's mutation catalog, plus an expression score for smoking-responsive genes trained only on normal-adjacent or external data, so that it carries no tumor sex signal. For alcohol and liver exposure, we will use SBS16 activity and a hepatic exposure-response program. Each exposure needs at least two footprints whose measurement noise plausibly comes from different assays, such as DNA versus RNA. Unrelated noise is the key assumption that makes the instrumental ratio valid.

The second component is the estimator, with uncertainty from a patient-level bootstrap, applied feature-wide with false-discovery control. The third component handles the main threat to identification: footprint yield could itself differ by sex. For example, sex-biased DNA repair could change how many mutations a given dose produces, and the PCAWG analysis found the aging signature SBS1 female-biased. To handle this, we will fit $\kappa$ separately by sex as a diagnostic, test equality, and report a sensitivity analysis in which the sex contrast is recomputed across a plausible range of sex differences in $\lambda_1$.

Baselines

The do-nothing reference is the unadjusted sex contrast with standard age, ancestry and subtype adjustment. This is essentially the regression design of Sex differences in oncogenic mutational processes. It anchors how much signal exists before any exposure handling.

The strongest existing competitor is proxy-covariate adjustment: the same regression with the footprint $Z_1$, or both footprints, added as covariates. It is the strongest because it uses exactly the same information as our method, and any careful analyst would try it first. It will run on identical patients, features and covariates, with no extra tuning budget on either side. We count our method as beating it if three things hold in cohorts with recorded smoking history. First, our contrast falls closer to the contrast adjusted for the recorded history. Second, its bootstrap intervals cover that reference at their nominal rate. Third, it does both while the covariate approach shows systematic residual bias toward the naive contrast. For this baseline, no reported result was found on Lacuna for exposure-adjusted sex contrasts. The closest reported numbers concern the claims under test, not the method: Danusertib sensitivity elevated in female LUAD lines, reported in the file as both p = 0.0021 and p = 0.021 for the LUAD multi-omics study. We will recompute rather than adopt either.

Sex-stratified modeling, fitting male and female models separately, is included because Camila M. Lopes-Ramos et al. advocate it. It does not address confounding, and it is expected to lose power, consistent with Roman Hornung et al.. Their benchmark found simple clinical Cox models with fewer than 20 features competitive with complex models across 18 TCGA cancers.

The ablation that isolates the key ingredient keeps the combination but estimates $\kappa$ by ordinary regression of $Y$ on $Z_1$ within sex, with no second footprint as instrument. If this ablation performs as well as the full method, the instrument is doing nothing and the contribution collapses to covariate adjustment. A second ablation substitutes recorded pack-years for the footprints and serves as the oracle ceiling.

Why not just…?

Why not just add sex and a sex-by-gene interaction to a Cox or linear model? That approach estimates the sex effect well only if exposure is measured. Here exposure is not measured, so the interaction absorbs whatever exposure gap exists between sexes. Roman Hornung et al. show that simple clinical models are hard to beat for prediction. Our question, though, is attribution, not prediction, and a sex covariate cannot tell smoking from biology. Our model is in fact this regression with one extra calibrated term, so the comparison is direct.

Why not just use batch correction such as ComBat or mixed models? These remove variation tied to known discrete labels. An unrecorded continuous exposure has no label to correct against. Treating the leading latent factor as noise would also remove genuine sex biology whenever it loads on the same factor. The positive-anchor check below, which tests whether X-linked and androgen-axis effects are preserved, would expose that failure.

Why not just adjust for the mutational-signature activity as a covariate? This is the real threat. If footprints are nearly noise-free, covariate adjustment and our combination coincide, and our method only adds variance. We therefore estimate footprint reliability directly from the agreement between the two footprints. If reliability is high in a cohort, we will say so and recommend the simpler method there. The proposal is worthwhile only where reliability is moderate, and the oracle comparison settles which regime each cancer is in.

Why not just stratify by sex? Stratifying halves the sample and still leaves exposure confounded within each stratum, because heavy and light smokers exist among both men and women. The pan-cancer genome study already needed pooled regression to get enough power for subtypes such as biliary cancer (Sex differences in oncogenic mutational processes).

Proposed Experiments

The first experiment is semi-synthetic. On real TCGA lung and bladder expression data, we inject a known sex effect and a known exposure confounder whose sex gap and footprint reliability we vary. We then measure bias, interval coverage and variance for every baseline. This maps exactly when the combination's variance cost is worth paying. The few-shot framing of The TCGA Meta-Dataset Clinical Benchmark, on which logistic regression reached 68.78% average accuracy and did best on sex-labelled tasks, suggests that linear sex signal is strong enough for this stress test to be informative.

The decisive real-data check uses the subset of TCGA lung adenocarcinoma, lung squamous and bladder patients whose tobacco history is recorded; the extent of this subset must be confirmed per cohort. We hide the history, run all methods, and compare each to the oracle contrast adjusted for recorded history. Success means our estimate tracks the oracle while proxy-covariate adjustment sits measurably between the naive contrast and the oracle.

The scientific payoff comes from three applications. We re-test the 16 lung adenocarcinoma sex-specific candidates and the hormone-receptor and driver associations of Robert Hsu et al., separating those that survive exposure cancellation from those that do not. We ask how much of the male SBS16 and mutation-density excess in liver cancer remains once alcohol footprints are cancelled. We test whether the bladder survival reversal reported by Keiichiro Mori et al. persists after cancellation.

Two built-in falsification checks frame these results. Features with strong prior biological grounding, namely XIST, KDM6A and other X-escape genes, and the androgen receptor–CCRK targets in liver cancer, must keep their sex contrast essentially unchanged. Exposure-response genes used in neither footprint must lose most of theirs. If the anchors move or the exposure genes do not, the method is mis-specified, and we will report that rather than the downstream claims.