Sex-Dimorphic Microenvironmental Noise as a Stabilizer of Clonal Equilibria in Tumors

A recurring finding in algorithmic game theory is that randomness, usually treated as a defect, can instead be a regulator. In large auctions and capacity markets, when the environment is deterministic, big players find the exact thresholds at which a small strategic move causes a large price distortion, and they exploit them; when supply or demand is noisy, those sharp thresholds smear into smooth hills, exploitation stops paying, and the market settles near the efficient outcome (The Price of Anarchy in Large Games, The Price of Anarchy of Large Walrasian Auctions). Cancer offers a striking, unexplained instance of what this theory predicts. In biliary adenocarcinoma, female tumors are frequently polyclonal (80%) while male tumors collapse toward a single dominant clone (only 26% polyclonal) (Sex differences in oncogenic mutational processes), and the two sexes also run qualitatively different immune-shelter machinery: males shield tumors with localized, proliferating monocytic suppressor cells inside the tumor, while females rely on systemically circulating, secretory granulocytic suppressor cells (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner).

The proposal here is that these two facts are connected through the game-theoretic mechanism. Tumor subclones play a competitive game for finite metabolic and immune-shelter capacity. The claim is that the female microenvironment — systemic, fluctuating, spatially distributed immune suppression — imposes high-variance capacity noise that smooths the competitive landscape and stabilizes polyclonal coexistence, while the male microenvironment — a localized, low-variance suppressive umbrella plus a direct androgen-receptor-driven growth circuit (Cell cycle–related kinase is a direct androgen receptor–regulated gene that drives β-catenin/T cell factor–dependent hepatocarcinogenesis) — acts as a deterministic incubator in which one clone can win outright. If this is right, it changes what therapy should do: not only kill cells, but shape the noise regime the clones compete in, differently by sex.

Core Claim

The sex difference in clonal architecture is not only a downstream consequence of different mutation rates or drivers; it is a property of the noise topology of the competitive game, and it is therefore manipulable. Concretely: (i) a clonal capacity game with a localized, low-variance shelter regime converges to monoclonal dominance, while the same game with a systemic, high-variance shelter regime sustains polyclonal equilibria, quantitatively matching the 26%/80% split observed in PCAWG biliary tumors; (ii) swapping the noise regime between virtual sexes reproduces the phenotype flip seen in reciprocal bone-marrow chimeras, where female mice given male marrow acquire male-like glioma dynamics (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner); and (iii) dosing policies that deliberately introduce or preserve variance in the shelter capacity can delay competitive release of resistant clones in the male regime, where existing deterministic controllers leave the landscape sharp.

Research Program

The theoretical backbone is the Cournot game with capacity uncertainty and shortfall penalties of Zhang, Johari, and Rajagopal (Competition and Efficiency of Coalitions in Cournot Games With Uncertainty), reinterpreted biologically. Each subclone $i$ commits to a proliferation "quantity" $q_i$; the realized shelter capacity $k_i$ it actually receives is random, drawn from a sex-indexed noise distribution $\sigma_s$ ($s \in \{M, F\}$); if committed growth exceeds realized shelter, the excess is killed by immune predation at penalty rate $\kappa$. The clone's fitness is the expectation over that noise:

$$u_i^{s}(q_i) = \mathbb{E}_{\sigma_s}\!\left[\, p(Q)\, q_i \;-\; \kappa\, \max\!\big(0,\, q_i - k_i(\sigma_s)\big)\right], \qquad Q = \sum_j q_j,$$

where $p(Q)$ is the inverse "price" of shared resources (perfusion, glucose), falling in total burden $Q$. Zhang et al. prove that individual risk hedging under this kind of uncertainty causes underproduction, while pooling into coalitions of size $\Omega(\sqrt{N})$ restores efficiency — the biological reading being that high-variance predation pressure pushes clones toward conservative, slow-cycling phenotypes and shared-burden polyclonal structure, while a low-variance regime licenses aggressive over-expansion by whichever clone holds the strongest driver. The male androgen-receptor–CCRK–β-catenin axis (Cell cycle–related kinase is a direct androgen receptor–regulated gene that drives β-catenin/T cell factor–dependent hepatocarcinogenesis), amplified metabolically in obesity-associated HCC (An inflammatory-CCRK circuitry drives mTORC1-dependent metabolic and immunosuppressive reprogramming in obesity-associated hepatocellular carcinoma), is modeled as an internal capacity amplifier — a commitment device that partially overrides the noise, exactly the "bullying" strategy that Immorlica et al. show uncertainty neutralizes (The Price of Anarchy in Large Games). The female regime adds protective hormonal volatility via estrogen-receptor pioneer-factor programs (Na Liu et al.) as a second variance source.

The computational vehicle extends the Stackelberg mean-field-game adaptive-therapy simulator of Mehrjou (Stackelberg Mean-Field Games for Adaptive Cancer Therapy), which tracks sensitive, partially resistant, and fully resistant phenotypes in twenty virtual patients and solves the clinician-leader/tumor-follower problem through coupled Hamilton-Jacobi-Bellman and Fokker-Planck equations. That model supplies the scaffolding but is sex-neutral and deterministic in its follower dynamics; we graft the two noise regimes onto it, with parameters calibrated from measured quantities: Fludarabine extending male survival from 20.5 to 30.5 days with zero effect in females, and anti-IL-1β producing long-term cures only in females (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner); clonal-architecture frequencies and TERT-promoter skews from the PCAWG cohort (Sex differences in oncogenic mutational processes); and OLR1/IL1B expression versus survival in the 188-specimen GBM cohort and TCGA-GBM (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner).

Baselines

The naive reference point is maximum tolerated dose, which applies the highest tolerable dose continuously; on the twenty-virtual-patient benchmark it triggers competitive release of resistant clones with time to resistance of only 180 days and full cumulative drug exposure (Stackelberg Mean-Field Games for Adaptive Cancer Therapy). A middle baseline is heuristic adaptive therapy — dosing on and off against gross tumor volume — which reaches 310 days on the same benchmark. The strongest competitor is the Stackelberg mean-field-game controller itself, which reports time to resistance exceeding 500 days at 50% of the MTD cumulative dose (Stackelberg Mean-Field Games for Adaptive Cancer Therapy). We will run all four on the identical extended benchmark: the same twenty virtual patients, each instantiated in a male and a female noise-parameterization, with equal compute, the same random-search budget over controller hyperparameters, and the same cumulative-dose cap, so the comparison isolates the modeling ingredient rather than tuning effort. Because the sex-blind Stackelberg controller already posts >500 days, beating it on aggregate time to resistance is not the right bar; we count success as (a) strictly longer per-sex time to resistance than the sex-blind controller in at least one sex with no regression in the other, (c) qualitative reproduction of the cross-sex efficacy pattern of Fludarabine and anti-IL-1β, and (c) predicted polyclonal fractions within roughly ten points of the observed 26% and 80% for the two regimes. The key ablation is the same model with the two noise regimes collapsed into one — this isolates whether the gain comes from noise topology rather than from merely adding parameters; a second ablation keeps the hormonal driver asymmetry but flattens the noise, separating the amplifier effect from the variance effect. No reported result found on Lacuna for clonal architecture under sex-specific noise manipulation, so these ablations are genuinely new measurements.

Why not just…?

Why not just add sex as a binary covariate in a survival model? This is the cheapest version of "sex-aware" prediction, and it works as far as it goes: Bayesian neural survival models reach a C-index of 0.71 and an Integrated Brier Score of 0.12 on TCGA-BRCA and METABRIC, and flag out-of-distribution demographics (Uncertainty Estimation in Cancer Survival Prediction). But a covariate is a label, not a mechanism: it cannot say why female biliary tumors stay polyclonal, cannot represent feedback circuits like the CCRK–β-catenin–AR loop, and offers no dosing policy. Our proposal answers a mechanistic question those models cannot pose.

Why not just run the existing Stackelberg controller and ignore sex? The existing controller is the strongest baseline and a real threat on aggregate metrics. But sex-blind control has a documented failure mode at the pre-clinical level: anti-proliferative therapy produced zero benefit in female glioma models and anti-IL-1β produced zero benefit in males (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner). A controller that treats the immune channel as one scalar cannot distinguish these regimes. Experiment 2 below directly tests whether the sex-blind controller, retrained on our mixed-sex cohort, loses time to resistance in one sex relative to the noise-aware controller; if it does not, the proposal's practical value shrinks to explanation and we will say so.

Why not just inject generic noise and make the controller robust? This is the closest existing alternative: hierarchical control that reshapes the fitness landscape and maintains an immune-influx stability margin already achieves 86% cure under stochastic perturbations, with 89% eradication in melanoma and 76% in colorectal priors (Steering the Evolutionary Game: Hierarchical Control of Therapeutic Resistance in Cancer Treatment). That work treats stochasticity as a disturbance to be suppressed. Our claim is the opposite: the noise is the regulated quantity, and its topology — localized versus systemic, independent versus correlated, low versus high variance — is what sets the equilibrium, a sensitivity that Immorlica et al. explicitly flag since supply-side and demand-side noise behave differently (The Price of Anarchy in Large Games). The head-to-head within Experiment 2 — a noise-suppressing robust controller versus a noise-shaping controller under identical assumptions — settles this empirically rather than by argument.

Why not just do static bioinformatic regressions on the genomic data? Logistic regressions already found the static sex skews, including the 64%-versus-11% TERT-promoter split in thyroid cancer (Sex differences in oncogenic mutational processes). They cannot, by construction, capture thresholds, bifurcations, or temporal feedback; they describe the disparity without explaining its stability conditions. We reuse their outputs as calibration targets, not competitors.

Proposed Experiments

Experiment 1 tests the architecture claim. We simulate the clonal capacity game under the two calibrated noise regimes and measure the equilibrium clonal structure as a function of clone count and noise variance, comparing the predicted polyclonal fractions against the PCAWG biliary observations of 80% (female) and 26% (male) (Sex differences in oncogenic mutational processes). The decisive sub-check is the in-silico chimera: swap the noise parameters between virtual sexes and verify the clonal dynamics flip, mirroring the bone-marrow transplant result that shifted female glioma outcomes to male-like patterns (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner). Theory runs alongside: we adapt the shortfall-penalty efficiency bounds of Zhang et al. (Competition and Efficiency of Coalitions in Cournot Games With Uncertainty) and the smoothness-in-the-large machinery (The Price of Anarchy in Large Games) to characterize when the variance needed for polyclonal stability is achievable given realistic clone counts, and how the deterministic AR-CCRK amplifier shifts that threshold — including the known open problem that correlated (non-independent) noise breaks the standard proofs.

Experiment 2 tests the control claim. Within the extended virtual-patient benchmark we compare four dosing policies per sex: the sex-blind Stackelberg controller (Stackelberg Mean-Field Games for Adaptive Cancer Therapy), a noise-suppressing robust controller in the spirit of the hierarchical-control work (Steering the Evolutionary Game: Hierarchical Control of Therapeutic Resistance in Cancer Treatment), our noise-shaping controller that deliberately schedules variance into the shelter channel, and a constant-dose reference. The hypothesis is that variance injection delays competitive release in the male (sharp) regime, while in the female regime the controller's job is to preserve existing stabilizing variance rather than add it. Success is measured per-sex against the >500-day sex-blind baseline, and the qualitative check is that the model ranks Fludarabine-like and anti-IL-1β-like interventions correctly by sex, as the mouse data do (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner).

Experiment 3 tests the claim against human data rather than simulation. Using the 188-specimen GBM cohort and TCGA-GBM expression of the gMDSC markers OLR1 and IL1B, plus Ki-67-based mMDSC prominence (Myeloid-derived suppressor cell subsets drive glioblastoma growth in a sex-specific manner), we infer each patient's noise-regime parameters and ask whether inferred variance — not the sex label — predicts survival and subclonal structure, with the Bayesian uncertainty machinery of Loya et al. (Uncertainty Estimation in Cancer Survival Prediction) supplying calibrated posteriors so we can flag where inference is out of distribution. A positive result here would show the mechanism generalizes beyond the two canonical regimes.

Decisive Checks

Three results would make or break the proposal. First, the chimera flip: if swapping noise parameters between virtual sexes does not flip clonal architecture, the noise story is not the mechanism, and the observed dimorphism is better attributed to drivers alone. Second, the noise-shaping delta: if deliberately injecting shelter variance does not outperform noise-suppressing control in the male regime at equal cumulative dose, then the therapeutic half of the claim fails and the work reduces to an explanation. Third, the variance-over-label test in human GBM: if inferred noise variance carries no predictive weight beyond the binary sex label, the model has not earned its parameters. No reported result found on Lacuna for any of these comparisons, which is precisely the gap this program occupies.