When a cytotoxic T cell or an engineered CAR-T cell enters a solid tumor, it does not explore the tumor at random. It climbs chemokine gradients, piles into the most signaling-rich niches, and — crucially — does what thousands of other identically programmed effectors do at the same moment. Distributed-systems research has a name and a cure for exactly this failure mode: deterministic greedy selection by many concurrent agents creates a "thundering herd" that overwhelms the chosen resource while starving the rest, and randomized selection — turning "go to the best server" into "sample from a probability distribution over servers" — fixes it without any central coordinator. I propose to treat immune-cell infiltration of the immunosuppressive tumor microenvironment (TME) as precisely this problem: each effector cell is a client, each tumor niche is a server with a queue, and the extracellular matrix, suppressive myeloid shields, and hypoxic cores are congestion that slows service. The core intervention is a decentralized, stochastic migration policy — the biological analogue of probability-matched load balancing — evaluated in agent-based tumor simulators where deterministic gradient-following demonstrably fails.
This matters because the leading mechanistic explanations of CAR-T failure in solid tumors are spatial and distributional, not just molecular. The tumor microenvironment excludes and exhausts effectors through physical matrix barriers, suppressive cell shields, and metabolic deprivation Tumor microenvironment review by Rui Zheng et al., and engineered cells that do reach the tumor dissipate systemically and exhaust rapidly Spatiotemporal profiling of CAR-T and CAR-NKT cells by Li et al.. Existing interventions attack the barriers (matrix digestion Human Hyaluronidase PH20 potentiating mesothelin-specific CAR-T cells by Zhao et al., vessel normalization and checkpoint knockdown Recombinant oncolytic adenovirus by Yuan Meng et al.) or the suppressive signaling axes (CCL2-driven macrophage recruitment MYBL2–CCL2 axis in ovarian cancer by Baoyue Pan et al., PPARγ-driven MDSC recruitment Zhewen Xiong, MTA-driven M2 polarization Landon J. Hansen et al.). None of them asks whether the routing policy of the effector population itself — where each cell chooses to go, given what it senses — is a binding constraint. The distributed-systems literature says it should be: deterministic selection is provably fragile under concurrency, while randomized policies carry explicit performance guarantees in analogous settings On the Power of Randomization for Scheduling Real-Time Traffic in Wireless Networks by Tsanikidis et al..
Core Claim
The claim, in two sentences: a population of identically programmed effectors that all chase the strongest chemokine signal self-organizes into a localized "immune herd" — a dense, suppressible cluster at one or a few niches — leaving most of the tumor undefended; replacing the deterministic choice with a sampled choice from a probability distribution over niches (weighted by signal quality minus local congestion and suppression) spreads the same cell dose across more niches, delays exhaustion, and kills more tumor. Randomization here is not noise added to biology; it is the known antidote to synchronized congestion, imported into a domain where the concurrent agents happen to be cells.
Concretely, each effector $i$ at time $t$ evaluates candidate niches $j$ (local tumor regions within its sensing radius) with a dual cost, directly adapting the historical-plus-real-time cost structure of the cloud probability-matching algorithm CDLBA-PM Towards a Content Delivery Load Balance Algorithm Based on Probability Matching in Cloud Storage by Chen et al.:
$$ c_{ij}(t) \;=\; \beta_h\, h_{ij}(t) \;+\; \beta_r\, r_{ij}(t), $$
where $h_{ij}$ is a slowly updated historical cost (has this effector lineage or local cohort been suppressed, exhausted, or starved near $j$?) and $r_{ij}$ is a real-time cost combining the chemokine signal deficit, the local crowding of effectors already in or queued at $j$, and the local density of suppressive shield cells (M2-like macrophages, MDSCs). The effector then samples its next niche rather than taking the argmin:
$$ P(i \to j) \;\propto\; \exp\!\left(-c_{ij}(t)/\tau\right), $$
where the temperature $\tau$ controls stochasticity; $\tau \to 0$ recovers greedy gradient-following, which serves as the built-in ablation. Dominated niches (worse signal, more crowded, more suppressed) are pruned before sampling, following the pruning logic of AMIX-ND, which discards links with both lower deficits and looser deadlines On the Power of Randomization for Scheduling Real-Time Traffic in Wireless Networks. I call the resulting policy family PIM, for Probabilistic Infiltration Mixing. The biological plausibility lever — what would make this more than a simulation curiosity — is that CAR-T receptor affinities and chemokine-receptor expression can be engineered to be heterogeneous across a manufactured batch, giving a physical substrate for sampling-like behavior without requiring any cell to compute a softmax.
Research Program
Three things get built. First, an effector-routing layer on top of PhysiGym, the existing bridge between the PhysiCell agent-based simulator and the Gymnasium reinforcement-learning interface PhysiGym by Bertin et al.. PhysiGym's toy tumor model (512 tumor cells, 128 immune-like cells, four diffusing chemicals) currently treats immune-like cells as passive rule-bound agents; I will upgrade them to policy-driven mobile agents that sense local fields, maintain per-cell historical cost state, and select migration targets via PIM or its baselines. Suppressivity will be modeled as service-rate degradation: niches with high M2/MDSC density or dense extracellular matrix process arriving effectors slowly (delayed killing, accelerated exhaustion), so a niche behaves exactly like a congested server in queueing terms.
Second, a harder synthetic environment with non-stationarity and adversarial remodeling: niches that respond to effector influx by upregulating suppressive recruitment (a spatial, mechanistic caricature of the CCL2 and PPARγ recruitment axes Baoyue Pan et al., Zhewen Xiong), so that a herd triggers its own suppression. This is the setting where deterministic policies should fail worst, because the environment punishes synchronization rather than merely ignoring it. Tumor growth dynamics and latency calibrations will follow the hexagonal cellular-automaton model of Deng et al., whose microenvironment index $m \in [0,1]$ and stochastic hourly fate sampling I will reuse as the substrate the effectors act on Modeling tumor progression in heterogeneous microenvironments: A cellular automata approach; growth-curve calibration will use the longitudinal lung-cancer volume dataset that paper employed (no calibrated parameter values are reported on Lacuna, so we re-estimate).
Third, a theoretical analysis of when randomized routing provably beats greedy routing in this setting. The wireless-scheduling result that greedy Largest-Deficit-First collapses to an efficiency ratio of 0.5 while randomized AMIX-ND guarantees at least 0.63, and degrades as $O(1/\sqrt{K})$ on star topologies On the Power of Randomization for Scheduling Real-Time Traffic in Wireless Networks, suggests a direct analogue: with $K$ niches and a suppressive shield whose strength grows with local effector density, greedy routing should suffer a herding penalty that grows with population size, while probability-matched routing should retain a constant-factor guarantee. I will state and prove this under a Markovian niche-dynamics assumption, flagging honestly that the real-tumor regime — heavy-tailed, non-Markovian remodeling — sits outside the proof, which is exactly the open limitation the wireless authors themselves flag for their own setting.
A note on the closest prior work, so the delta is unambiguous. The cellular automaton of Deng et al. models stochastic cell fates but has no mobile effectors and no routing Deng et al.. PhysiGym adds a controller, but a centralized one that doses drugs while immune cells remain passive Bertin et al.. SWARM parallelism demonstrates stochastic wiring and adaptive rebalancing among compute nodes, including the finding that a baseline without rebalancing loses over 50% of throughput under churn, but never leaves the digital cluster SWARM Parallelism by Ryabinin et al.. The neurosymbolic multi-agent work synthesizes communication policies with an explicit random operator, but on 2D navigation toy tasks with no biology Neurosymbolic Transformers for Multi-Agent Communication. No work on Lacuna models effector infiltration as decentralized stochastic load balancing against a remodeling, adversarial microenvironment. The proposal's novelty is the policy layer plus the adversarial-remodeling environment, not any single ingredient.
Baselines
The naive reference point is passive and uniform-random migration: PhysiGym's default rule-bound immune-like cells (unchanged), and a policy that picks uniformly among sensed niches. Uniform random is a genuine control because it separates "randomization helps" from "smart weighting helps"; in the cloud setting, probability matching beat uniform random selection on mean job time at high load Chen et al., and we expect the same ordering here.
The strongest obvious competitor is greedy chemotaxis: every effector migrates toward the strongest chemokine signal or largest tumor cluster it senses, which is the argmin of the real-time cost — the policy most CAR-T migration models implicitly assume, and the direct biological analogue of Largest-Deficit-First. It is the strongest competitor because it is the default: it uses all the same sensory information as PIM and differs only in being deterministic. The wireless-scheduling literature gives a calibrated expectation for this comparison: greedy LDF provably drops to an efficiency ratio of 0.5 while AMIX-ND guarantees at least 0.63, and in periodic-traffic simulations AMIX-ND achieved delivery ratios near 0.99 while LDF variants suffered severe drops On the Power of Randomization for Scheduling Real-Time Traffic in Wireless Networks. We will run greedy with the identical sensing radius, time step, and cell dose as PIM, tuning nothing extra — the two policies share every parameter except $\tau$. Beating it means higher tumor-cell kill fraction and greater infiltration depth at fixed dose, with the gap growing in population size.
The strongest existing system baseline is the centralized Soft Actor-Critic controller trained in PhysiGym, which learns dynamic drug-burst schedules from spatial observations of the TME Bertin et al.. No reported quantitative result is found on Lacuna for its tumor-burden reduction, so we will retrain it ourselves on our environment under an identical interaction budget (same number of environment steps, same observation channels, same compute) and compare on kill fraction and coverage. Beating it requires matching or exceeding its kill fraction while additionally scaling to agent populations where its central observation breaks, and without systemic drug dosing — a decentralization claim, not just a performance claim. The SWARM churn result — a no-rebalancing baseline losing over 50% of throughput under constant node churn Ryabinin et al. — sets the ablation expectation: when we remove PIM's rebalancing component (the historical-cost updates), we should see a comparable large fractional loss under effector attrition.
The key ablation is $\tau \to 0$: PIM with the sampling replaced by the argmax, holding the dual cost, the pruning, and all sensing identical. This isolates stochasticity itself as the active ingredient. A second ablation drops the historical cost term ($\beta_h = 0$), isolating the pheromone-like memory that lets the population learn which niches suppress arrivals, in the spirit of the ant-colony-inspired updates of Chen et al. Chen et al.. A third drops dominated-niche pruning, testing whether the AMIX-ND pruning step matters biologically.
Why not just…?
Why not just use the centralized RL controller on a bigger simulation? It is a real threat on raw kill fraction, and we treat it as the primary system baseline rather than arguing it away. But it optimizes systemic drug pulses over passive cells; it cannot program the migration of each of a million effectors, and its observation channel scales badly with population size, which is the regime solid tumors actually occupy. Decisively, the cellular-automaton results show timing and locality dominate outcomes: complete blockade of microenvironmental remodeling starting at day 10 achieved an efficiency of 0.8, while delaying to day 50 fell to 0.55 Deng et al. — a central controller that must aggregate state and act globally will always lag this window in a heterogeneous tumor. Our experiments will settle the head-to-head under matched budgets, and if the centralized controller wins even at large population scale, the proposal is wrong and we will report that.
Why not just digest the barriers — more hyaluronidase-secreting CAR-T, more vessel normalization? This is the strongest biological alternative, and it is complementary rather than competing: PH20 co-expression widens specific bottlenecks Zhao et al., and anti-VEGF secretion plus PD-L1 knockdown attacks vascular and checkpoint barriers simultaneously Meng et al.. But barrier digestion increases capacity, not dispersion: deterministic chemotaxis will still herd the freed cells into the first high-signal niche, where suppressive shields concentrate — exactly the multi-shield redundancy that makes single-axis checkpoint interventions fail, as Li et al. showed with TIGIT-versus-CD96 divergence between CAR-T and CAR-NKT cells Li et al.. Our simulations include barrier-digestion and vessel-normalization conditions (as service-rate increases) crossed with both routing policies; the prediction is that digestion alone saturates — the queue simply re-forms at the next bottleneck — while digestion plus dispersion combines super-additively. If digestion alone saturates kill fraction at the population sizes we test, the routing question is moot at those scales, and we will say so.
Why not just increase the dose? Because the failure mode is not effector scarcity but effector distribution and survival, and dose escalation attacks the wrong constraint. The CAR-T single-cell atlas shows systemic dissipation, rapid exhaustion, and toxicity subsets (the ICANS-associated cells) whose burden grows with the infused population An open CAR-T single-cell atlas by Camara-Peña et al., and spatiotemporal profiling shows PBMC-derived CAR-T cells dissipating rather than accumulating in tumors Li et al.. In queueing terms, adding clients to a congested server does not clear the queue; it deepens it. The experiment that settles it: a dose-scaling sweep at fixed routing policy, which should show kill fraction saturating (congestion-limited) under greedy routing and continuing to climb under PIM. If kill fraction instead scales linearly with dose under greedy routing across our range, congestion is not the binding constraint in the model, and the proposal's premise fails empirically.
Why not just let cells random-walk and let gradients do the work? Uniform randomness loses targeting: the cloud experiments showed probability matching over dual costs beats uniform random selection on mean job time and load variance at high load Chen et al., and top-$k$ candidate pooling shows that fully unconstrained random routing bottlenecks quality in high-concurrency systems Tang et al.. Pure diffusion also wastes the dose outside the tumor — the systemic-dissipation problem the atlas documents. We include the uniform-random baseline explicitly to test this, and we expect it to underperform both greedy and PIM on kill fraction.
Proposed Experiments and Decisive Checks
The primary testbed is the PhysiGym toy tumor environment Bertin et al., upgraded with our effector-routing layer and run across cell-dose sweeps and population sizes from the native 128 immune-like agents up to tens of thousands, on the hexagonal-automaton growth substrate with its stochastic fate sampling at hourly steps Deng et al.. Metrics are tumor-cell kill fraction, infiltration depth (fraction of tumor volume within effector reach), spatial dispersion of effectors (variance across niches, the direct analogue of the node-access-variance metric of Chen et al. Chen et al.), exhaustion counts, and time-to-tumor-dominance. The adversarial-remodeling environment is the second setting, where influx-triggered suppressive recruitment punishes herding; the third is a churn setting with stochastic effector attrition, where the SWARM no-rebalancing result (over 50% throughput loss Ryabinin et al.) predicts PIM's historical-cost updates carry most of the benefit.
Three checks are decisive. First, the herding observation itself: under greedy routing with realistic population sizes, effectors must measurably concentrate (high niche-occupancy variance) with suppression-triggered exhaustion localized exactly where they concentrate, and kill fraction must fall below what uniform random achieves — the biological rendering of the greedy trap that drops LDF to 0.5 efficiency Tsanikidis et al.. If greedy routing does not herd in these simulators, the premise is falsified and the project stops there — that would be a useful negative result about the limits of the load-balancing analogy. Second, the stochasticity ablation: PIM at $\tau \to 0$ must collapse toward greedy performance, with the gap growing in population size, confirming randomization rather than the dual cost or pruning does the work. Third, the combination test: barrier digestion (in the manner of Zhao et al.'s PH20 secretion Zhao et al.) crossed with routing policy must show digestion-only saturation and super-additive gains with dispersion.
The translational path, kept deliberately modest given the acknowledged sim-to-real gap for PhysiGym-style platforms Bertin et al., is not to implant a softmax in a T cell. It is to use the simulator to design heterogeneous CAR-T products — batches with deliberately varied chemokine-receptor expression and antigen-receptor affinities, so the population behaves statistically like a sampler rather than a deterministic herd — and to pre-register, from simulation, which heterogeneity profiles should beat homogeneous batches before any wet-lab test in organoid or xenograft models of the kind used in the spatiotemporal CAR-T profiling work Li et al.. The single-cell atlas already shows that product composition predicts response and toxicity Camara-Peña et al.; our claim is that composition should be engineered for dispersion, not just for potency.
References
- Frontiers and future of immunotherapy for pancreatic cancer: from molecular mechanisms to clinical application
- Spatiotemporal profiling reveals distinct dynamics and checkpoint regulations of CAR-T and CAR-NKT cells against solid tumors
- Human Hyaluronidase PH20 Potentiates the Antitumor Activities of Mesothelin-Specific CAR-T Cells Against Gastric Cancer
- RCAd-LTH-shPD-L1, a double-gene recombinant oncolytic adenovirus with enhanced antitumor immunity, increases lymphocyte infiltration and reshapes the tumor microenvironment
- The MYBL2–CCL2 axis promotes tumor progression and resistance to anti-PD-1 therapy in ovarian cancer by inducing immunosuppressive macrophages
- IDDF2020-ABS-0201 Targeting hepatoma-intrinsic pparγ signaling overcomes immune checkpoint therapy resistance by inflaming the tumor microenvironment
- MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization
- On the Power of Randomization for Scheduling Real-Time Traffic in Wireless Networks
- Towards a Content Delivery Load Balance Algorithm Based on Probability Matching in Cloud Storage
- PhysiGym: bridging the gap between the Gymnasium reinforcement learning application interface and the PhysiCell agent-based model software
- Modeling tumor progression in heterogeneous microenvironments: A cellular automata approach
- SWARM Parallelism: Training Large Models Can Be Surprisingly Communication-Efficient
- Neurosymbolic Transformers for Multi-Agent Communication
- An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products
- Dominance-Based Service Selection Scheme with Concurrent Requests