When a targeted cancer drug fails, it often fails without any new mutation: cells under treatment shed their original identity — the very identity the drug was aimed at — and slide into a tolerant state, a phenomenon called lineage plasticity. The best quantitative account of this, in triple-negative breast cancer, models the tumor as a set of compartments with cells flowing between them under drug pressure Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer. That model is deterministic: it tracks average counts and nothing else. Queueing theory spent decades learning exactly when such deterministic averages are trustworthy and when they break. They break at bottlenecks — regimes where a system is pushed to capacity and the quantities that matter become small enough that randomness, not the average, decides the outcome The number of overlapping customers in Erlang-A queues: an asymptotic approach. The drug-tolerant persister reservoir that seeds relapse is precisely such a bottleneck: a small population whose extinction or survival is a random event that a mean-field model is structurally blind to.
This proposal imports the full queueing-theoretic scaling toolkit — fluid limits, diffusion limits, state space collapse, and heavy-traffic optimal control — into drug-driven cell-state dynamics. A fluid limit is the deterministic equation you get by letting the population grow large, so that discrete random cell events average into smooth flows; a diffusion limit is a Brownian-motion description of the fluctuations around that average; state space collapse is the proved phenomenon that systems under heavy load stop wandering the full high-dimensional space and settle onto a low-dimensional manifold. Cancer biology currently uses the first of these informally (and only the first), while the clinically decisive questions live in the second and third.
Core Claim
Consider a stochastic population model in which each of $N$ cells occupies one of $d$ phenotypic states, and each cell independently divides, dies, or switches state, with all rates modulated by a time-varying drug concentration vector $u(t)$. I claim three things. First, as $N \to \infty$ this process has a fluid limit — a functional law of large numbers — whose limit equation is exactly the compartmental ODE already fitted empirically in breast cancer Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer, which means that model gains, for free, validity conditions and an explicit $O(1/\sqrt{N})$ error guarantee. Second, the $\sqrt{N}$-scale diffusion correction around the fluid path, analyzed near the absorbing boundary at zero persisters, yields something no existing plasticity model produces: the probability that the persister reservoir goes extinct under a given dosing schedule, as opposed to a deterministic yes-or-no. Third, treating drug doses as controls in this diffusion limit and solving the associated control problem produces adaptive dosing regimens that, in the asymptotic regime, provably dominate both static drug combinations and any policy optimized on the deterministic fluid model alone — the same ordering proved for multiclass parallel-server systems, where the diffusion-based joint policy achieved an operational cost of 4.0850 versus 4.9069 for no control and 4.9074 for a partial policy optimized without the diffusion structure Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic.
The fluid limit itself is simple enough to state. Let $x(t) \in \mathbb{R}^d_{\ge 0}$ be the scaled population vector (counts divided by $N$), $Q(u)$ the matrix of state-to-state switching rates, and $\Lambda(u)$ the diagonal matrix of net proliferation (division minus death) per state. Then
$$\dot{x}(t) = \big(Q(u(t)) + \Lambda(u(t))\big)\, x(t),$$
and the extinction quantity comes from the diffusion limit: with $\tau$ the first time the persister-equivalent mass of the population hits zero, the hitting probability $h(x_0) = \mathbb{P}(\tau < \infty \mid x_0)$ solves a linear boundary-value problem for the diffusion generator, in direct analogy to hitting probabilities for reflected Brownian motion in queues. The non-mass-conserving term $\Lambda$ is what makes this a genuinely new limit theorem, not a re-derivation of existing queueing results: births and deaths break the conservation structure that classical fluid-limit proofs rely on, mirroring the unbalanced-mass difficulty flagged for cell dynamics in Integrating Dynamical Systems Modeling with Spatiotemporal scRNA-Seq Data Analysis.
Research Program
The first work package builds the stochastic model and proves the fluid limit. The discrete process is a multiclass birth-death-transition chain; the functional law of large numbers follows the population-scaling route of Analyzing Distributed Join-Idle-Queue: A Fluid Limit Approach, with the new technical burden being the unbalanced growth term. The payoff is that the compartmental ODE of Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer — five states, division rates, death rates, transition gains — stops being a plausible curve fit and becomes the $N \to \infty$ limit of an explicit process, with known error rates and known failure regimes. Its treatment of discrete marker cutoffs is retained for now, but a continuous-state extension along the lines of fluid queues that track continuous workloads Size-Aware Dispatching to Fluid Queues is a planned generalization.
The second package derives the diffusion limit and the boundary analysis. The central limit theorem gives Gaussian fluctuations around the fluid path while populations are large; the scientifically interesting regime is the bottleneck, where persisters number in the hundreds or fewer and drug pressure drives non-tolerant states toward zero. There, linearization fails and the analysis becomes a small-population hitting problem — the biological twin of the queue-starvation events that pure fluid models are known to miss, an explicitly open problem in the fluid-approximation literature. Validation of division and death rates against EdU proliferation assays and YO-PRO-1 death staining in Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer gives the experimental hooks for parameterizing the diffusion.
The third package uses state space collapse to answer a standing tension in the field: single-cell data shows cell states on continuous high-dimensional manifolds, yet tumors under therapy reliably collapse into a handful of attractor subtypes — four transcription-factor-defined subtypes in small cell lung cancer Intratumoral heterogeneity and potential treatment strategies in small cell lung cancer, four subtypes in neuroendocrine prostate cancer Multi-omics analysis constructs a novel neuroendocrine prostate cancer classifier and classification system. Queueing theory explains exactly this kind of dimensional reduction: under heavy load, switched networks collapse onto invariant manifolds even without the usual symmetry conditions, via Bramson's cluster-point method Switched networks with maximum weight policies: Fluid approximation and multiplicative state space collapse, and layered networks collapse onto a one-dimensional manifold fixed by their bottleneck Heavy-traffic approximations for a layered network with limited resources. The hypothesis: strong therapeutic pressure is the biological analog of heavy traffic, and the effective dimension of cell-state dynamics should measurably shrink under drug stress relative to untreated or microenvironmentally-driven dynamics. This turns the discrete-compartment-versus-continuum argument into a testable prediction rather than a philosophical dispute.
The fourth package closes the loop with control. Drugs modulate the entries of $Q$ and $\Lambda$ — trametinib blocking exits from the basal state, PI3K inhibition forcing luminal transitions Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer, BET inhibition blocking entry into the persister state Differentiation-state plasticity is a targetable resistance mechanism in basal-like breast cancer. Doses become the control variables of a diffusion control problem; the cost integrates tumor burden, toxicity, and a penalty on non-extinction of the persister reservoir; the optimum solves a Hamilton–Jacobi–Bellman equation — a partial differential equation whose solution directly yields the optimal feedback dosing rule — following the program that proved asymptotic optimality for multiclass queues Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic.
Baselines
The naive reference point is the untreated (or fixed-dose monotherapy) trajectory generated by the fitted rates — the "do nothing" against which any scheduling claim must show improvement, since the entire literature of therapy-as-active-driver, from barcoding studies Focal adhesion kinase-YAP signaling axis drives drug-tolerant persister cells and residual disease in lung cancer to drug screens Differentiation-state plasticity is a targetable resistance mechanism in basal-like breast cancer, establishes that untreated dynamics differ from treated ones only in their rates, not in needing a different model class.
The strongest existing competitor for trajectory prediction is the deterministic compartmental ODE with alternating-minimization fitting and wild-bootstrap confidence intervals Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer, run on the same HCC1143 time-course data (DMSO, trametinib, BEZ235, and combination arms, with K14/VIM/K19 markers), with the same parameterization, the same fitting budget, and held-out time points reserved identically for both models. No reported result found on Lacuna for a numerical trajectory-prediction error for that model, so the comparison metric — held-out marker-frequency error — will be established on our own runs. Beating it does not mean lower mean-trajectory error: our fluid limit is that model's equation, so means should agree within its bootstrap bands. Beating it means calibrated bottleneck variance and a well-defined extinction probability where it offers only a resampling interval around a deterministic path.
For the control question the strongest empirical benchmark is static combination therapy: co-treatment with the PI3K/mTOR inhibitor BEZ235 and the BET inhibitor JQ1, which produced nearly 100% cell death across basal-like cell lines in vitro and significant xenograft regression in vivo Differentiation-state plasticity is a targetable resistance mechanism in basal-like breast cancer. An adaptive policy beats this benchmark if it reaches equal or lower expected cost (tumor burden plus toxicity) while raising the model-predicted extinction probability of the persister reservoir — the static regimen's near-100% death is a population-average figure that says nothing about the tail of surviving persisters. The key ablation is the fluid-only policy: the identical optimization run on the deterministic ODE with the diffusion term deleted, plus a variant with fixed (non-feedback) rates. The expected gap between the diffusion-based and fluid-only policies is the quantity of interest; the parallel-server benchmark, where the diffusion-based joint policy cost 4.0850 against 4.9069 for the uncontrolled system and 4.9074 for a partial policy optimized without the full diffusion structure Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic, sets the scale of gap — roughly 17% — that would make the ingredient worth its cost. For the state-identification subtask, the Random Forest NEP100 classifier at AUC 0.986 in pseudobulk validation over 196,309 cells and 70 samples Multi-omics analysis constructs a novel neuroendocrine prostate cancer classifier and classification system is the reference: our collapse-based state description must not degrade state-identification accuracy below that level while adding what a classifier cannot — dynamics.
Why not just…?
Why not just use the existing compartmental ODEs? Because that model is our fluid limit, and the interesting claims are all about where fluid limits fail. The queueing literature documents that pure fluid models drop extinction and starvation events at boundaries — a listed open problem — and the persister reservoir is exactly a small-count boundary regime; deterministic prediction there is not approximately right, it is answering a different question. The control argument is sharper still: in the one setting where both options have been compared with numbers, optimizing on the deterministic system was strictly dominated by the diffusion control solution (4.0850 versus 4.9069, and partial deterministic-informed control at 4.9074, barely better than nothing) Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic. Our first decisive check is designed to settle this on biological terms: if, at the persister bottleneck, the ODE's deterministic path stays within empirical variance bands and extinction outcomes are tightly concentrated, the simpler model wins and the diffusion package is unnecessary — we would report that honestly.
Why not just run discrete Markov chain methods (CellRank, PAGA) or optimal transport on the single-cell snapshots? These are real tools and we use them for preprocessing, but they face two limits the scaling approach removes. Single-cell sequencing is destructive, so these methods see cross-sectional snapshots, not trajectories, and must interpolate — the continuum-limit review acknowledges this and points at Fokker–Planck equations but establishes no scaling theorem, no boundary analysis, and no control results Integrating Dynamical Systems Modeling with Spatiotemporal scRNA-Seq Data Analysis. Worse, they assume a fixed transition kernel, while the central experimental finding of the field is that therapy switches the kernel: trametinib and BEZ235 produce different transition structures Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer. The queueing side adds a computational argument: discrete-chain approaches suffer state-space explosion, while fluid limits reached roughly 0.1% relative error at load 0.99 where prior discrete approximations erred up to 6% Analyzing Distributed Join-Idle-Queue: A Fluid Limit Approach. Experiment two, the switched-kernel prediction test, will show directly whether snapshot-interpolation methods mispredict state distributions across a dose change.
Why not just train a machine-learning classifier on the multi-omics data? NEP100 already achieves 0.986 AUC for identifying neuroendocrine state Multi-omics analysis constructs a novel neuroendocrine prostate cancer classifier and classification system, and nothing here competes with that. But a classifier is a static map from a molecular profile to a label: it cannot output an extinction probability, cannot represent time-varying dosing as an input, and cannot produce a control policy. It also has no mechanism for the microenvironment: the proportion of basal-like pancreatic cells varies linearly with oxygen gradient depth in the TRACER organoid system at $R^2 = 0.78$ Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes, which in our framework is simply a spatially varying rate — a case a classifier handles only by conditioning on more inputs, not by explaining the gradient. This alternative is a genuine complement, not a threat.
Decisive Checks
Three experiments decide the proposal. First, fluid validation on the HCC1143 time courses: fit rates on training arms, predict held-out marker-frequency trajectories under combination therapy, and confirm agreement with the compartmental ODE where $N$ is large — establishing that we lose nothing at the mean. Second, extinction calibration at the bottleneck: in a ground-truth stochastic simulation protocol mirroring the Euler–Maruyama multiclass benchmark of Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic, check that diffusion-predicted extinction probabilities match empirical hitting frequencies, and that the deterministic model gives no calibrated answer at all; then apply the same test to persister dynamics in the lung-cancer residual-disease setting, where persister formation is an active, widespread reprogramming across clones rather than selection of rare mutants Focal adhesion kinase-YAP signaling axis drives drug-tolerant persister cells and residual disease in lung cancer. Third, the collapse test: measure the effective dimension of cell-state dynamics in the neuroendocrine prostate cohort and the SCLC subtype structure under therapy versus untreated conditions, predicting a measurable dimensional drop under drug stress, with the TRACER hypoxia gradient ($R^2 = 0.78$) serving as the no-therapy spatial control where no collapse should occur. If all three hold, the final step — an HJB-derived adaptive schedule benchmarked against static BEZ235+JQ1 co-treatment in organoids — is a principled design rather than a heuristic search, and the field gains what it currently lacks: a mathematical account of when the simple compartment picture is valid, what it cannot see, and what to do about dosing when it cannot see it.
References
- Modeling differentiation-state transitions linked to therapeutic escape in triple-negative breast cancer
- The number of overlapping customers in Erlang-A queues: an asymptotic approach
- Asymptotic optimality of a joint scheduling–control policy for parallel server queues with multiclass jobs in heavy traffic
- Integrating Dynamical Systems Modeling with Spatiotemporal scRNA-Seq Data Analysis
- Analyzing Distributed Join-Idle-Queue: A Fluid Limit Approach
- Size-Aware Dispatching to Fluid Queues
- Intratumoral heterogeneity and potential treatment strategies in small cell lung cancer
- Multi-omics analysis constructs a novel neuroendocrine prostate cancer classifier and classification system
- Switched networks with maximum weight policies: Fluid approximation and multiplicative state space collapse
- Heavy-traffic approximations for a layered network with limited resources
- Differentiation-state plasticity is a targetable resistance mechanism in basal-like breast cancer
- Focal adhesion kinase-YAP signaling axis drives drug-tolerant persister cells and residual disease in lung cancer
- Analysis of an engineered organoid model of pancreatic cancer identifies hypoxia as a contributing factor in determining transcriptional subtypes