Cohort Scaling Laws for Steering Steroid-Refractory Immune Evolution

Between 30% and 60% of patients with autoimmune and chronic inflammatory disease fail to achieve control on standard corticosteroid and biologic regimens, and the dominant clinical answer is escalation: push the dose to the maximum tolerated level, then switch mechanisms when the patient stops responding (Binvignat et al.). This proposal starts from a different premise, borrowed from evolutionary oncology: aggressive suppression of an immune response can cause the refractory state, because it removes the steroid-sensitive cell populations that were otherwise outcompeting the resistant ones. The one formal demonstration of this "competitive release" effect exists only in cancer — Mehrjou showed that game-theoretic bounded dosing extended time-to-resistance from 180 days under maximum tolerated dose to over 500 days — and it rests on assumptions (Nash-equilibrium cells, a fixed three-type discretization, a single optimally-controlled patient) that do not hold in autoimmune disease.

The second half of the problem is that even if we knew the right dosing policy, we could not find it in a trial. Complex interventions live on high-dimensional protocol manifolds — dose, pulse timing, taper slope, adjunct pathway blockade — and pooling trials that differ on these axes inflates heterogeneity and masks real effects (Hesam-Shariati et al.); in stroke rehabilitation, timing alone flips the sign of the effect, so aggregation produces false nulls (Chengshuo Wang et al.). Adaptive trial designs such as the Vibe Up multi-arm bandit study pick batch sizes heuristically ($n \approx 120$ per mini-trial) with no guarantee that the cohort is large enough to resolve the parameter space (Huckvale et al.). Wireless engineering solved an exactly analogous problem: multi-cell networks provably fail unless infrastructure and scheduling pools scale at a precise rate with the user count and signal quality (Aggarwal et al.; Grokop et al.; Jung et al.).

Core Claim

We claim two things. First, that corticosteroid resistance in autoimmune inflammation can be modeled as bounded-rationality replicator dynamics — the evolutionary-game-theory model in which the share of each strategy in a population grows in proportion to how far its payoff exceeds the population average — over competing immune phenotypes (steroid-sensitive Th2/lymphocytic versus refractory Th17/neutrophilic and non-canonical-pathway populations), with the dosing protocol acting as the control input that reshapes the payoff landscape. Second, that identifying a steering protocol across a heterogeneous patient population requires an analytically derived cohort scaling law — a minimum relationship between enrolled patients, protocol dimension, biomarker noise, and within-patient evolutionary drift — analogous to the user scaling laws of wireless networks. The deliverables are the dynamical model, the scaling law, and an offline validation on real registry and cohort data.

The Model

Within a patient, let $x_i$ be the share of immune phenotype $i$ in the inflammatory compartment, and let $p$ denote a point on the protocol manifold (dose intensity, pulse timing, taper slope). The phenotype shares evolve as

$$\dot{x}_i = x_i\left[f_i(p, x) - \bar{f}(p, x)\right], \qquad \bar{f} = \sum_j x_j f_j$$

where $f_i(p,x)$ is the fitness of phenotype $i$ under protocol $p$, including the competitive suppression term: high-dose continuous steroid crushes sensitive phenotypes and thereby raises the relative fitness of refractory ones. This differs from Mehrjou in two ways that matter: cells follow imitation-and-reproduction dynamics rather than Nash equilibrium (his own stated limitation, and the regime biological mutation actually suggests), and the phenotype set is data-driven rather than a fixed three-type split. The biology comes from the steroid-resistance literature: IL-6 as an independent risk factor for steroid refractoriness (Meng et al.), NIK-driven non-canonical signaling that takes over when canonical TNF pathways are blocked (Wang), and the Ceramide 24:1 / EP2 / JAK2-STAT3 axis that licenses Th17 neutrophilic inflammation in steroid-recalcitrant asthma (Endogenous Ceramide 24:1 Constrains Th17-Driven Neutrophilic Inflammation). Each documented escape route becomes a phenotype in the state vector, with fitness parameters fit to longitudinal biomarkers.

The trial layer translates the wireless analogy directly. In multi-cell MIMO uplink networks, Jung et al. proved that a distributed two-threshold rule — admit a user only if its signal strength exceeds $\eta_{tr}$ and its interference leakage stays below $\eta_\iota$ — achieves optimal multiuser diversity without central coordination, provided the active user pool scales as $N = \omega(\text{SNR}^{1-\epsilon})$. In a multi-site adaptive trial, each enrolled patient is an "opportunity": a two-threshold rule allocates a protocol point only when the patient's expected benefit exceeds one threshold and the patient's contribution to cross-site protocol variance stays below the other. The scaling-law deliverable is the clinical analog of $N = \omega(\text{SNR}^{1-\epsilon})$: the minimum per-arm enrollment as a function of protocol dimension $d$, biomarker noise $\sigma$ (the SNR analog), and replicator drift rate $\nu$ (which has no wireless counterpart and is where the derivation is genuinely new). Our working hypothesis, by analogy with Grokop et al.'s result that bandwidth must scale at least as $O(K/\log K)$, is that the resolvable number of protocol arms grows sublinearly in cohort size, with a multiplicative penalty for drift; the exact form is to be derived, not assumed.

Research Program

The program has three components. We build a calibrated replicator simulator of steroid-refractory immune evolution, with phenotypes and fitness maps anchored to the asthma neutrophilic-endotype data (173 patients, 53 controls, Cer24:1 and IL-17A measured in exhaled breath condensate, plasma, and sputum; no scalar benchmark performance is reported for the underlying single-cell datasets GSE135779, GSE142637, GSE162577) and to the IL-6 colitis and NIK-escape findings. On top of it we derive the cohort scaling law for two-threshold opportunistic protocol allocation, and finally we validate offline: the simulator against the SCQM rheumatoid arthritis registry, and the scaling law against simulated multi-site trials at varying cohort sizes. Shnayder et al., who used replicator dynamics to show that a mechanism design can expand the basin of attraction of truthful reporting, supplies the mathematical template for our basin-of-attraction analysis: the clinical question is whether a protocol family keeps the phenotype distribution inside the basin where sensitive competitors dominate, and how large a trial is needed to confirm a protocol does so across a heterogeneous population.

Baselines

The naive reference point is standard treat-to-target practice: escalate steroid dose to the maximum tolerated level, then switch salvage mechanisms (JAK inhibitor, IL-6 blockade, FcRn targeting) upon failure, as catalogued across the refractory-disease literature (Binvignat et al.; Zhao et al.; Zheng and Chen). We implement it as the escalation controller in our simulator at matched cumulative-dose budgets. The strongest existing competitor is Mehrjou's Stackelberg mean-field game controller, which in its own oncology testbed reported time-to-resistance (defined as >75% resistant cells) of 180 days for maximum tolerated dose, 310 days for heuristic adaptive therapy, and over 500 days for the S-MFG controller, while using 50% of the cumulative drug dose of maximum tolerated dose — at the cost of a higher average burden (1.19 versus 0.50). We re-implement it on our immune testbed with the same tuning budget and compute as our controller, and count ourselves as beating it only if we achieve longer time-to-refractoriness at matched cumulative steroid exposure and matched disease-activity ceiling, specifically under conditions where the follower population is boundedly rational and drifts rather than converging to equilibrium. A second controller baseline is the LSTM-augmented PPO agent of Reinforcement Learning for Adaptive Tacrolimus Dosing, which achieved 93.1% time in therapeutic window under mechanistic simulation (outperforming a proportional controller by 13.9 percentage points, with 84.5% for observed real-world practice) but treats the patient as a stationary compartmental system; running it on our evolving testbed isolates what the evolutionary state estimate contributes. Finally, for the registry validation the competitor is static deep embedded clustering: Patient groups in Rheumatoid arthritis identified by deep learning clustered 3,516 SCQM patients into 5 groups with pronounced response hazard ratios (Tocilizumab HR = 5.46 for patients with ≥2 prior csDMARDs and prednisone at initiation; Golimumab non-response HR = 5.27 for seropositive women with higher disease burden). The decisive ablation removes the replicator feedback: the same trial machinery and the same two-threshold allocator run with a stationary-response patient model, so any performance gap is attributable to the evolutionary state estimate rather than to the scheduling machinery.

Why not just…?

Why not just use classical pharmacokinetic/pharmacodynamic compartment models? These are the field default, and the file contains direct evidence of where they fall short. In the tacrolimus study, model predictive control with exact access to the true PK equations underperformed the adaptive learning agent because population-average models fail on individual variation; and in Mehrjou, maximum-dose regimens grounded in static pharmacodynamics produced competitive release and resistance dominance at 180 days. More fundamentally, a compartment model has no state variable for phenotype shares, so it cannot express the phenomenon this proposal is about.

Why not just run a standard contextual bandit like Vibe Up? This is the most serious alternative. If the protocol manifold were low-dimensional and patient responses nearly stationary, a Gaussian-process UCB bandit would likely match our allocator, and the scaling law would be decoration. We do not argue around this; our simulated multi-site experiment runs the GP-UCB design of Huckvale et al. at, above, and below the derived cohort threshold $N(d, \sigma, \nu)$, with non-stationarity turned on and off. If the bandit matches our method at smaller cohorts under drift, the scaling law is miscalibrated and we revise it — that outcome would itself be informative, since Hesam-Shariati et al. show unstructured pooling over heterogeneous protocols masks efficacy entirely.

Why not just stratify patients once with deep clustering and assign fixed regimens? The SCQM deep-clustering result shows static stratification works at baseline, and it is a real threat to the registry component. But its own authors note the clustering required manual visual grouping and uses static snapshots; it cannot represent the emergence of Th17/neutrophilic escape during treatment (Endogenous Ceramide 24:1; Wang). The head-to-head test — time-updated replicator state features versus static baseline features predicting drug discontinuation on SCQM — settles this empirically, and we commit to reporting both.

Why not just escalate dose and switch drugs when patients fail (treat-to-target)? This is the clinical gold standard, and we treat it as the primary naive baseline rather than dismiss it. The case against it is cumulative: 30–60% refractory rates despite treat-to-target (Binvignat et al.), a therapeutic ceiling for broad cytokine blockade (Zheng and Chen), escape via non-canonical pathways under suppression (Wang), and hypoxia-driven steroid insensitivity (Xiu et al.) — plus the oncology result that maximum-intensity suppression accelerates the failure it is meant to prevent.

Proposed Experiments and Decisive Checks

Three experiments. First, the calibrated simulator: twenty virtual patients with parameters perturbed 15%, mirroring Mehrjou's protocol so numbers are comparable; metrics are time-to-refractoriness (refractory share > 75%), cumulative steroid dose, and disease-activity burden; the decisive check is whether replicator-steered dosing beats the S-MFG controller's oncology-style performance pattern under bounded-rationality drift where its Nash assumption fails. Second, offline validation on the SCQM registry (3,516 patients with longitudinal DAS28 and drug-cycling records): test whether dynamic replicator-state features improve prediction of biologic discontinuation over static clustering, evaluated by time-dependent C-index, with the SCQM-reported hazard ratios (5.46, 8.44, 3.64 for Tocilizumab subgroups; 2.36 and 5.27 for Golimumab) as the reference stratification structure. Third, the scaling-law experiment: simulated multi-site adaptive trials over a protocol grid of controlled dimension, varying cohort size and biomarker noise, run with our two-threshold allocator, the GP-UCB bandit, and uniform random allocation; the decisive check is a measured phase transition in protocol-identification accuracy at the predicted $N(d, \sigma, \nu)$, and whether the Vibe Up heuristic of $n \approx 120$ per mini-trial sits above or below it for realistic clinical effect sizes — for which the trial itself reports no outcome numbers yet, since it is a protocol publication.