Inflammatory Admission Control: Timing Cytokine Load to Rescue Progenitor Maturation

In chronic inflammatory disease, the failing tissue is usually not out of repair cells — the repair cells are stuck. In multiple sclerosis, stem cells from patients, including those with progressive disease, still become myelin-producing oligodendrocytes at 60–80% efficiency in a calm culture dish, but collapse below 20% when exposed to a single inflammatory cytokine, interferon-gamma (IFNγ), at the wrong time (iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu). The same story holds in bone marrow failure after CAR-T cell therapy: the culprit is not death of blood-forming progenitors but a paracrine barrage of IFNγ, TNFα, and IL-6 that freezes them in an immature state (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma) — and blocking just one cytokine (the IL-6 receptor antagonist tocilizumab) does not rescue them.

This proposal imports a counterintuitive lesson from a distant field. In high-speed cluster interconnects, engineers found that guaranteeing timely delivery does not require ever more complex switches — a switch with 16 per-class queues can be replaced by a switch with only two queues, provided the network edge (the host interface) does the scheduling: it tags packets with deadlines, meters their release times, and refuses to admit more reserved traffic than the link can carry (Providing Full QoS Support in Clusters Using Only Two VCs at the Switches, Providing Full QoS with 2 VCs in High-Speed Switches). Simplification at the "core" works because control moved to the "edge." Applied here: progenitors' internal gene networks (the core) do not need to be re-engineered per cytokine; what needs controlling is the schedule and total flux of inflammatory signal arriving at the niche (the edge) — when it arrives relative to the cell's developmental stage, and how much of it arrives in aggregate.

Core Claim

The claim is that maturation arrest is governed by a single, low-dimensional quantity — a stage-weighted cumulative inflammatory dose — and that interventions which hold this quantity below a threshold will rescue maturation even when no individual cytokine is fully eliminated. Formally, with $\lambda(t)$ the aggregate inflammatory input rate (a receptor-occupancy-weighted sum of IFNγ, TNFα, IL-6, and so on), $s(t)$ the progenitor's stage-dependent susceptibility, and $\Theta$ the niche's maturation capacity,

$$D=\int_0^{T}\lambda(t)\,s(t)\,dt,\qquad \text{arrest predicted when } D>\Theta .$$

This is the direct analogue of Connection Admission Control in the 2-VC architecture, where deterministic guarantees hold only while the sum of reserved rates stays below link capacity (Providing Full QoS with 2 VCs in High-Speed Switches); tocilizumab's clinical failure is then exactly what the model predicts, since removing one contributor to $\lambda$ leaves the integral above $\Theta$ (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma). The susceptibility weighting is not decoration: IFNγ at day 36 of oligodendrocyte differentiation is devastating while day-60 exposure is largely tolerated (iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu), so the model's control variables are timing and aggregate flux, not cytokine identity.

Research Program

Three things get built. First, a susceptibility clock: rather than wall-clock days, we define cell stage through maturation regulons (the paired arrest and maturation programs already measured — GATA2/RUNX1/CEBPA up, ID2/MAFB down in arrested hematopoiesis; IDO1/CXCL9/TUJ1 redirection in arrested oligodendrocyte progenitors), fitting $s$ as a function of this internal state using the public datasets GSE147315 and GSE250444 from the two anchor studies. Second, the aggregate load variable $\lambda$: we use CINEMA-OT, a method that separates per-treatment effects from confounding cell state via independent component analysis and optimal transport (Causal identification of single-cell experimental perturbation effects with CINEMA-OT), to estimate individual cytokine contributions, then test whether a one-dimensional weighted sum predicts arrest across single- and mixed-cytokine conditions better than any single cytokine. We deliberately resist a high-dimensional gene-network controller: the CausalBench benchmark found that sophisticated causal-discovery algorithms fail to beat simple statistical baselines on real single-cell perturbation data because biological networks break their acyclicity assumptions (CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data) — the same empirical pattern that pushed interconnect design from 16 queues toward 2. Third, schedule synthesis: two intervention families derived from the edge-scheduling toolkit. The eligibility-time analogue is windowed anti-inflammatory cover — inflammation is made "ineligible" to reach progenitors only while $s(t)$ is high, directly generalizing the non-work-conserving shaper that holds packets until their eligibility time and which sustained an 85% video load where a conventional 4-queue switch collapsed at 45% (QoS Support for Video Transmission in High-Speed Interconnects). The admission-control analogue is flux capping — partial neutralization of all contributors, scaled so $D$ stays below $\Theta$. To map the boundary between rescue and arrest in the (dose, onset-day) plane we adapt the open-loop pulse-shaping machinery of Sootla et al., who showed that switching boundaries in bistable biological systems are monotone curves that random sampling resolves to about 4.5% relative error in under 90 seconds (Shaping pulses to control bistable biological systems); our extension makes the boundary explicitly time-varying through $s(t)$, which their static-bistable setting does not have.

Proposed Experiments

The experimental core is the 90-day human iPSC-to-oligodendrocyte protocol across the same donor structure as the original study (10 lines from healthy, stable MS, and progressive MS donors), with IFNγ at 10 ng/ml applied in pulses of fixed duration but varied onset day (approximately days 30, 36, 45, and 60), measuring the O4+ mature fraction, myelin basic protein, and the redirection markers IDO1, CXCL9, and TUJ1. Anchored against the reported homeostatic yield of 60–80% O4+ and the day-36 collapse to less than 20% (iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu), these data fit $s(t)$ and its interaction with dose and duration. The decisive mixture experiment replicates the ex vivo design of the CAR-T study (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma): healthy-donor hematopoietic stem and progenitor cells cultured with activated BCMA CAR-T supernatant (and, in parallel, a defined IFNγ/TNFα/IL-6 cocktail), scored on the same lineage markers (CD10+CD16+ neutrophilic, CD14+CD35+ monocytic, CD71+CD36+ erythroid). Arms compare single-cytokine neutralization, aggregate flux capping, and windowed cover, all at matched total antibody exposure. The model is then tested retrospectively: ferritin, the systemic marker whose peak correlated with persistent cytopenia in the 48-patient cohort, is a candidate clinical proxy for the integral $D$; we will test whether ferritin area-under-curve outperforms any single-cytokine peak at predicting the reported 90-day median platelet recovery. The clinical covariates are not stated as deposited alongside GSE250444, so this analysis depends on obtaining cohort access; the decisive test of the claim remains the in vitro mixture experiment.

Baselines

The do-nothing reference is unmodulated inflammation: day-36 IFNγ at 10 ng/ml, which the anchor study reports drives O4+ below 20% against a 60–80% homeostatic yield (iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu); the homeostatic yield itself serves as the ceiling any rescue arm is judged against, with "success" defined as recovering to within ten percentage points of it. The strongest existing biomedical competitor is single-cytokine neutralization — anti-IFNγ antibody in the pure-cytokine assay, and tocilizumab-style IL-6 receptor blockade in the mixture assay — which is exactly the intervention that failed clinically (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma). It will be run in the same donor lines, same batches, and same total antibody exposure as the shaping arms so that only the target and schedule differ; beating it means matching it in the single-cytokine assay while rescuing maturation in the mixture and supernatant assays where single blockade leaves O4+ and lineage yields suppressed. Computationally, the strongest competitors are CINEMA-OT, which estimates per-perturbation counterfactual effects but has no notion of timing or dosage scheduling (Causal identification of single-cell experimental perturbation effects with CINEMA-OT), and Fatecode, which predicts internal master regulators to perturb (it recovered 90% of known master regulators within its top 150 genes in simulation, including the OPC-maintenance factor Nr4a2) (Fatecode: Cell fate regulator prediction using classification autoencoder perturbation); both will be fit on the same training conditions and evaluated on held-out timing arms they were never shown, where a static method cannot in principle distinguish day-36 from day-60 exposure. The key ablation isolates the scheduling ingredient: windowed cover versus constant cover at identical cumulative drug exposure, and a susceptibility-weighted threshold ($D>\Theta$) versus a static dose threshold without $s(t)$. If constant cover or the static threshold performs as well on held-out arms, the proposal's distinctive claim — that when matters, not just how much — is falsified, and the simpler version wins.

Why not just…?

Why not just block the offending cytokine with an antibody? This is the simplest clinical move, and the literature already records its failure: tocilizumab did not prevent severe, persistent cytopenia in the CAR-T cohort, because IFNγ, TNFα, and IL-6 act in parallel (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma). I am honest that in the pure IFNγ assay, anti-IFNγ neutralization will almost certainly rescue as well as anything we propose; that is why the decisive comparison is staged in the multi-cytokine supernatant, where the model predicts single blockade fails and flux capping succeeds. If single blockade rescues the mixture too, the aggregate-load model is unnecessary and the proposal should die.

Why not just immunosuppress broadly? Broad suppression is the "add more queues" answer, and both fields document its cost: in CAR-T recipients it compromises the engineered cells' anti-tumor function, and in demyelinating disease systemic suppression eradicates needed phagocytic and efferocytotic clearance that tissue repair depends on (Molecular mechanisms promoting long-term cytopenia after BCMA CAR-T therapy in Multiple Myeloma, Animal models to investigate the effects of inflammation on remyelination in multiple sclerosis, Enucleated bone marrow-derived mesenchymal stromal cells regulate immune microenvironment and promote testosterone production through efferocytosis). This objection is actually the strongest argument for the proposal: windowed cover is the same drug, timed so that suppression ends once progenitors pass the susceptibility window. The matched-exposure ablation against constant cover is precisely the experiment that settles whether the timing, rather than the suppression, is doing the work.

Why not just transplant more progenitors? Over-provisioning fails because the constraint is environmental, not numeric: patient progenitors are intrinsically competent but arrested by the milieu (iPSCs from people with MS can differentiate into oligodendrocytes in a homeostatic but not an inflammatory milieu), and inflamed oligodendrocyte progenitors are additionally converted into antigen-presenting targets that CD8+ T cells kill (Oligodendrocyte precursor cells present antigen and are cytotoxic targets in inflammatory demyelination) — donor cells would meet the same fate. The interconnect literature independently warns that adding raw capacity without flow control does not fix blocking under hotspot load (Cost / Performance Trade-Offs and Fairness Evaluation of Queue Mapping Policies). A cheap decisive check: add surplus progenitors under day-36 inflammatory exposure and observe whether the mature fraction stays suppressed; if extra cells rescue yield, the overload framing is wrong.

Why not just model the full gene regulatory network and search for interventions in silico? Because the best current evidence says that does not work: on real large-scale single-cell perturbation data, elaborate causal and network-inference methods fail to beat simple statistical baselines (CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data), and even the pulse-control work that inspires our schedule design succeeded precisely by collapsing a high-dimensional control problem onto a two-dimensional monotone boundary (Shaping pulses to control bistable biological systems). We will still run Fatecode as a baseline; if its internal-regulator targets outperform edge scheduling on held-out conditions, the field should follow it instead.

Decisive Checks

The proposal stands or falls on three falsifiable outcomes. If single-cytokine blockade rescues the mixed-supernatant assay as well as flux capping, the aggregate-dose model is dead. If constant cover matches windowed cover at equal exposure, timing is not the active ingredient. If the susceptibility-weighted threshold fails to beat a static dose threshold on held-out onset days, the stage clock should be dropped and with it most of the claim. Conversely, rescue in the mixture assay by schedule and flux alone — with no individual cytokine fully eliminated — would establish that inflammatory microenvironment inhibition can be treated as a scheduling problem, exactly as two-queue switches established that quality of service is an edge property, not a core property, of a fabric.