Extinction at the Nadir: Fluid–Diffusion Limits for Drug Pulsing Against Costly Resistance

Resistance has a price. A bacterium or cancer cell that mutates to survive a drug usually pays for it in slower growth, delayed development, or weaker communal behavior — a rifampicin-resistance insertion in M. tuberculosis measurably slows growth (A Glutamine Insertion at Codon 432 of RpoB Confers Rifampicin Resistance in Mycobacterium tuberculosis), and resistant wheat aphids reproduce more slowly than their susceptible cousins (Fitness Cost of the Field-Evolved Resistance to Sulfoxaflor and Multi-Insecticide Resistance of the Wheat Aphid Sitobion miscanthi (Takahashi)). That price can be exploited therapeutically: if drug pressure is relaxed at the right moments, the faster-growing sensitive cells outcompete the resistant ones for shared resources. Exploiting ecology in drug pulse sequences in favour of population reduction made this quantitative with coupled logistic ODEs, showing that pulse timing controls how low the total population sinks — its nadir — and finding concrete structure: an optimal pulse skewness boundary and a critical high-stress duration ($t_r \approx 30$ in their units) beyond which timing stops mattering.

The catch is that a small nadir is not a cure. Cure is extinction — a discrete, random event in which a population of a few dozen cells happens to die out before regrowth. A deterministic ODE cannot represent this: a "nadir" of $10^{-3}$ cells is a fractional fiction that deterministically regrows. Queueing theory spent decades on exactly this shape of problem — enormous discrete stochastic systems whose mean behavior is a smooth flow but whose fate is decided by rare boundary events such as a queue emptying or a buffer overflowing. Its answer is a two-tier approximation: a fluid limit (a functional law of large numbers turning the discrete process into ODEs) plus a diffusion correction that tracks fluctuations near the boundary, which is how The number of overlapping customers in Erlang-A queues: an asymptotic approach captures variance and transient behavior that the fluid path alone is blind to. This proposal builds that two-tier approximation for populations evolving costly resistance under drug pulsing, proves it against the individual-based process it approximates, and uses it to redesign pulse schedules whose objective is the probability of extinction rather than the depth of a deterministic nadir.

Core Claim

The claim has two parts. First, for the stochastic analogue of the competitive pulse model of Exploiting ecology in drug pulse sequences in favour of population reduction — wild-type and resistant strains with shared resources, drug pulses parameterized by total duration $\tau$, high-stress duration $t_r$, and skewness $s$, and a fitness cost $c$ of resistance — the extinction probability at the nadir can be computed by a fluid-plus-boundary scheme: the fluid limit carries the population to the nadir, and a linearized branching analysis of the fluctuations there yields the absorption probability, with a $O(N^{-1/2})$-style error control in the spirit of the finite-player bound proven in Stackelberg Mean-Field Games for Adaptive Cancer Therapy. Second, and this is the part that would be new: the schedule that minimizes the deterministic nadir is generally not the schedule that maximizes extinction probability. Extinction is a joint event — the last cells die regardless of genotype — so it depends on the total count, its variance, and the composition at the nadir, not just the mean nadir depth. Competitive suppression deliberately keeps wild-type cells alive as a weapon against resistant ones, which raises the total count at the nadir; a diffusion-aware objective prices that trade-off and can shift the optimum toward later, harder pulses that accept a more resistant composition in exchange for a smaller, noisier total population. The claim is falsifiable: if extinction probability turns out to be a monotone function of nadir depth alone across the schedule space, part two is dead, and we will report that.

Research Program

Phase one defines the individual-based model: a continuous-time birth–death process with mutation, logistic resource competition, and a drug stress function $S(t)$ shaped by $(\tau, t_r, s)$, where the resistant strain's growth rate is reduced by the cost $c$ at low stress. We prove the functional law of large numbers showing that, as the carrying capacity $N$ grows, the scaled process converges to the ODE system of Exploiting ecology in drug pulse sequences in favour of population reduction. This formalizes what they used phenomenologically, and it follows the established pattern — Analyzing Distributed Join-Idle-Queue: A Fluid Limit Approach showed such limits can be strikingly accurate against discrete ground truth (about 0.1% relative error at load 0.99 in their system), and Modeling Buffer Occupancy in bittide Systems showed continuous-flow approximations matching cycle-accurate discrete hardware.

Phase two layers on the fluctuations. Writing $x = (w, r)$ for the scaled wild-type and resistant counts, the diffusion correction takes the form

$$dx_t = f(x_t, S(t))\,dt \;+\; N^{-1/2}\,\Sigma(x_t, S(t))^{1/2}\,dB_t,$$

where $f$ is the fluid drift, $\Sigma$ the per-capita birth–death covariance, and $B_t$ standard Brownian motion. Near the nadir, when the unscaled count $N(w+r)$ falls to the order of tens, the diffusion itself is no longer the right tool; there we freeze the per-capita rates at nadir conditions and analyze absorption for the resulting two-type branching process, obtaining a semi-closed-form extinction probability $P_{\text{ext}}(\tau, t_r, s, N, c)$. We will characterize the scaling window in which the approximation is controlled and the window (nadir counts of order unity) in which it degrades, since Switched networks with maximum weight policies: Fluid approximation and multiplicative state space collapse is explicit that pure fluid limits are blind exactly at such boundaries. The Lyapunov-stability machinery of Heavy-traffic approximations for a layered network with limited resources — fluid-level guarantees lifting to discrete systems — supplies the template for the error analysis.

Phase three re-optimizes the pulse schedule against $P_{\text{ext}}$ instead of nadir depth, and maps the region of parameter space $(N, c)$ where the two objectives select different schedules. Our hypothesis is that divergence grows as $N$ shrinks and as the cost $c$ weakens, and vanishes for large populations or strong costs. Phase four extends from two strains to the sixteen-genotype beta-lactamase landscapes from Mira et al. used in Inverted topographies in sequential fitness landscapes enable evolutionary control, where resistance is multi-step with heterogeneous and collateral costs — a setting their discrete Markov decision process treatment handles exactly but does not scale in, and which our continuum limit is designed for. No reported result found on Lacuna for numerical fitness values on those landscapes, so they serve as a parameter source, not a performance benchmark.

Baselines

The naive reference point is continuous maximal drug pressure — the analogue of maximum tolerated dose, which eradicat[es] the sensitive competitors and thereby removes the competitive suppression of costly resistant cells. In the cancer cohort of Stackelberg Mean-Field Games for Adaptive Cancer Therapy, twenty virtual patients under maximum tolerated dose reached resistance dominance (>75% resistant cells) in 180 days. In our bacterial setting this baseline will be run as permanent high stress in the same Gillespie simulator, same population sizes, and same cost parameters as every other arm; beating it means any schedule with higher extinction probability at equal or lower cumulative drug exposure.

The strongest existing competitor is the deterministic nadir-minimizing pulse optimization of Exploiting ecology in drug pulse sequences in favour of population reduction, which reports a critical high-stress duration threshold of $t_r \approx 30$ (their units) beyond which pulse timing becomes irrelevant, and an optimal skewness boundary. It is the strongest because it is the incumbent theory of exactly this ecological mechanism, fitness cost exploited through competition. To make the comparison fair, we will run their ODE optimization over the identical three-parameter schedule space $(\tau, t_r, s)$, with the same grid resolution and compute budget as our own optimizer, and evaluate both policies on the same stochastic simulator with matched cumulative drug. Beating it means a strict increase in extinction probability across a pre-registered grid of $N$ from $10^2$ to $10^7$ and cost $c$, with the margin predicted to grow as $N$ decreases; if their schedule is already extinction-optimal everywhere on the grid, our central claim fails.

In the cancer-therapy instantiation, the competitor is the Stackelberg mean-field game controller of Stackelberg Mean-Field Games for Adaptive Cancer Therapy, which reports time to resistance dominance of 310 days for heuristic adaptive therapy and more than 500 days for its own controller, at 50% of the cumulative dose required by maximum tolerated dose, on twenty virtual patients with 15% parameter perturbation. Our fluid–diffusion controller must exceed the 310-day heuristic benchmark and approach the >500-day figure at comparable dose to be competitive there. Finally, two ablations isolate the proposal's key ingredients: a fluid-only controller — same stochastic model, same search, but optimizing deterministic nadir — isolates the contribution of the boundary/extinction analysis, and a diffusion-informed controller that optimizes nadir depth instead of extinction probability isolates the contribution of the objective change itself.

Why not just…?

Why not just use the deterministic ODE and treat a small nadir as success? Because the quantity of clinical interest is a probability, not a depth. The deterministic model treats cell counts as continuous and cannot distinguish a nadir of 5 cells from 500 — both regrow — while extinction probability differs enormously between them. More decisively, our own first experiment is designed to answer this: we will construct pairs of schedules with identical deterministic nadir (within 1%) but different nadir composition and variance, and measure whether their Gillespie extinction probabilities differ by an order of magnitude. If they do, the deterministic proxy is provably lossy; if they never do, the deterministic baseline wins and we will say so.

Why not just run Gillespie Monte Carlo simulations and estimate extinction probabilities directly? This is the most serious alternative, and for small systems it would plausibly work — we therefore use Gillespie as ground truth, not as a rival. The gap appears at realistic pathogen population sizes ($10^8$–$10^{12}$ cells, as noted in the scaling discussion around Approximating Discrete Stochastic Systems as Continuous Flows): simulating a single trajectory down from $10^9$ cells is expensive, extinction events near the nadir can have probabilities below $10^{-4}$ so that each schedule evaluation needs $10^4$–$10^6$ replicates, and optimizing over a schedule space multiplies that cost further. The approximation supplies an explicit formula that makes schedule optimization a cheap calculus problem and enables a hybrid scheme — fluid fast-forward, stochastic boundary phase. We will measure the compute crossover honestly rather than assert it: if Monte Carlo with $10^5$ replicates per schedule already optimizes the three-parameter space at $N = 10^4$ in reasonable time, our advantage is confined to large-$N$ and real-time regimes, and the paper will say exactly where the line sits.

Why not just train a deep reinforcement learning agent to find good pulse schedules? Reinforcement learning for adaptive control of phenotypically heterogeneous bacterial populations achieved 100% extinction in simulation when jointly controlling antibiotic and nutrients — but against phenotypic tolerance, not genetic resistance, and its authors report that the learned policies degrade when populations accumulate genuine resistance mutations, which is precisely our setting. It also offers no analytical guarantee that a policy avoids competitive release, whereas the Lyapunov-style lifting from fluid to discrete guarantees (Heavy-traffic approximations for a layered network with limited resources) is exactly the kind of certificate this proposal aims to carry over.

Why not just stick with maximum tolerated dose or simple on/off cycling? The numbers are against it: in the virtual cohort of Stackelberg Mean-Field Games for Adaptive Cancer Therapy, maximum tolerated dose reached resistance dominance in 180 days versus 310 for heuristic cycling and more than 500 for optimized dosing at half the cumulative drug; and Exploiting ecology in drug pulse sequences in favour of population reduction showed mistimed high-stress pulses prematurely eliminate the wild-type competitors and raise the nadir. These remain in the comparison as the naive reference and the heuristic arm respectively, but they are baselines to beat, not methods to adopt.

Proposed Experiments and Decisive Checks

The first experiment validates the fluid layer: reproduce the threshold structure of Exploiting ecology in drug pulse sequences in favour of population reduction — the $t_r \approx 30$ saturation and the optimal-skewness "white line" — as exact properties of the proved limit, and confirm that individual-based trajectories concentrate on the fluid path at large $N$. The second validates the boundary layer: across a grid of $N$, $c$, and schedules, compare the approximated $P_{\text{ext}}$ against Gillespie ground truth, report calibration error, and locate the scaling window where the branching approximation degrades. The third is the decisive check of the core claim: the equal-nadir pairs described above, plus the full re-optimization showing whether extinction-optimal and nadir-optimal schedules diverge, and where in $(N, c)$ space the divergence lives. The fourth runs the cancer translation on a twenty-patient virtual cohort matching the 15%-perturbation protocol of Stackelberg Mean-Field Games for Adaptive Cancer Therapy, scoring time to resistance dominance against their reported 180, 310, and >500-day figures at matched cumulative dose. The fifth stress-tests the assumptions: parameterizing costs from the sixteen-genotype beta-lactamase landscapes behind Inverted topographies in sequential fitness landscapes enable evolutionary control, and testing robustness to frequency-dependent ecology using the PC-9 coculture interactions documented in Frequency-Dependent Ecological Interactions Increase the Prevalence, and Shape the Distribution, of Preexisting Drug Resistance — their finding that ecological interactions shift extinction times by five orders of magnitude is exactly the effect size that could break a two-strain Markov model, so it is the right adversarial test. If frequency dependence overturns the objective divergence, that is a result worth reporting in its own right, and the natural sequel is the spatially structured extension toward the deme refugia of Spatial structure facilitates evolutionary rescue by drug resistance, where Gillespie-only methods are already at their computational limit and hydrodynamic limits have never been applied.