Designing Off-the-Shelf Immune Cells on a Single GPU

"Off-the-shelf" immunotherapy means banking universal anti-cancer immune cells — typically CAR-T or CAR-NK cells, immune cells genetically rewired with a receptor that recognizes a tumor marker — so any patient can be treated immediately, without the weeks of patient-specific manufacturing that autologous therapies require. The price of universality is a stack of genetic edits: knock out the donor T-cell receptor so the graft does not attack the host, strip the donor's HLA molecules so the host's T cells cannot see the graft, then add "don't eat me" signals such as MIF or CD47 because stripping HLA provokes Natural Killer (NK) cells into destroying the now-invisible graft (Intrathecal CRISPR-edited allogeneic IL-13Rα2 CAR T Cells for recurrent high-grade Glioma: preclinical characterization and phase I trial; Macrophage Migration Inhibitory Factor Suppresses Natural Killer Cell Response and Promotes Hypoimmunogenic Stem Cell Engraftment Following Spinal Cord Injury). Each edit buys one kind of survival and costs another kind of risk — more cloaking means less tumor surveillance and higher toxicity risk. The combination space is far too large for wet labs to enumerate, and the two computational tools that touch this problem are mismatched to it: one optimizes a single receptor sequence with no allogeneic loci at all, and the other predicts static snapshots with heavy foundation models on cluster hardware.

This proposal builds the missing piece at the intersection of the two directions: a batch-native, GPU-resident simulator of the immune interaction itself, plus a compact reinforcement-learning (RL) policy that searches multiplex edit designs and dosing schedules under safety constraints — all on one consumer GPU, in the spirit of work that showed RL can reach top performance without clusters. The core mechanism is computational accessibility; the application regime is allogeneic cell engineering.

Core Claim

No published work formulates the multi-locus hypoimmune design problem — which of roughly eight loci to knock out, which inhibitory transgenes to insert and at what expression level, and how to schedule infusions — as a sequential decision problem solvable on desktop hardware. T-Cell Receptor Optimization with Reinforcement Learning and Mutation Policies for Precesion Immunotherapy applies RL to mutate a single CDR3 loop for affinity, against a static surrogate evaluator, with no allogeneic evasion loci and no dynamic interaction. Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport predicts post-perturbation single-cell distributions for CAR signaling domains with high accuracy, but as a static map between timepoints, conditioned on ESM2 protein embeddings trained at cluster scale, and restricted to autologous signaling-domain combinations. An open CAR-T single-cell atlas to enable in-depth characterization and rational engineering of CAR-T products supplies an observational catalog of 414,000 harmonized CAR-T cells but no environment or optimizer. Separately, the accessibility literature — Beyond The Rainbow: High Performance Deep Reinforcement Learning on a Desktop PC, Large Batch Simulation for Deep Reinforcement Learning, and Pgx: Hardware-Accelerated Parallel Game Simulators for Reinforcement Learning — has never been applied to any biological system. The claim of this proposal is that a co-designed pair (vectorized immune-interaction simulator + compact design policy) can make the off-the-shelf design problem tractable on one GPU, and that the simulator alone is a publishable artifact because it enables the first head-to-head comparison of the field's two contradictory cloaking strategies.

Research Program

The program has three components, each reusing a proven technique from the accessibility literature and adapting it to cellular biology.

First, the simulator. I will implement an agent-based model of a virtual tissue well: populations of allogeneic effector cells (CAR-T or iNK), tumor cells, host CD8+ and CD4+ T cells, NK cells, and macrophages, each cell carrying an editable state vector (surface HLA-I/II levels, TCR presence, MIF/CD47/HLA-E expression, CAR specificity). Cell-level rules encode the mechanisms documented in the file: NK "missing-self" lysis that switches on when MHC-I drops, macrophage clearance gated by CD47–SIRPα, fratricide when the CAR antigen is shared by the effector, antigen-loss escape by tumor subclones, and cytokine accumulation. The implementation follows the Pgx pattern — JAX code that auto-vectorizes across thousands of parallel wells on the GPU — with the Large Batch Simulation for Deep Reinforcement Learning insight that moving the environment entirely into GPU memory removes the CPU–GPU communication bottleneck. The known hard part, which that paper itself flags for biology, is heterogeneous branching logic: cell divisions and discrete deaths do not vectorize as cleanly as raycasting. The plan is to replace branching with dense masked updates — every cell computes both its "alive" and "dead" transitions each step and a mask selects one — trading arithmetic for branching, which is exactly what GPUs prefer.

Second, calibration against published numbers, because a simulator with free parameters can predict anything. The simulator will be fitted to reproduce, within tolerance, the specific reported results in the file: NK-92MI-mediated lysis at a 3:1 effector-to-target ratio of ~28% for HLA-KO cells versus ~17% for HLA-KO plus MIF, and the nearly 4× higher day-14 graft survival of the MIF group in nude rats from Macrophage Migration Inhibitory Factor Suppresses Natural Killer Cell Response and Promotes Hypoimmunogenic Stem Cell Engraftment Following Spinal Cord Injury; the in vivo median survival of 42–46 days for the CD33-knockout Loop CAR iNK cells versus 34–38 days for standard CAR constructs and 28–32 days for untreated controls in AML xenografts from CD33 <sup>KO</sup> -CD33-mesothelin Loop CAR design avoids fratricide and improves efficacy of iNK cells against acute myeloid leukemia; and the persistence-plus-clearance behavior of the MT026 cells in Intrathecal CRISPR-edited allogeneic IL-13Rα2 CAR T Cells for recurrent high-grade Glioma: preclinical characterization and phase I trial, where TRAC and HLA-I α-chain disruption prevented graft-versus-host disease (0 cases in 5 patients, 80% objective response rate) yet cells were still eventually cleared by host NK cells. A simulator that reproduces all three regimes from one parameter set is credible; one that fits each dataset only with incompatible parameters is not, and that failure would itself be reported.

Third, the design policy. A compact network in the Beyond The Rainbow mold — a few million parameters, spectral normalization, distributional value estimates, trained with parallel vectorized wells — selects a design and a schedule. The action space is the edit vector (per-locus knockout or knock-in level, up to roughly 8 loci, giving on the order of tens of thousands of static designs but an exponentially larger space once sequential dosing and second-round edits under antigen escape are included) plus infusion timing. The reward follows the dual-component pattern validated by T-Cell Receptor Optimization with Reinforcement Learning and Mutation Policies for Precesion Immunotherapy, where a recognition reward is balanced by a biological-validity penalty:

$$r = \underbrace{-\Delta T}_{\text{tumor reduction}} - \lambda_1 \underbrace{G_{\text{lost}}}_{\text{graft rejection}} - \lambda_2 \underbrace{f_{\text{IAC}}}_{\text{toxic phenotype}} - \lambda_3 \underbrace{\max(0,\, E - E_{\max})}_{\text{uncontrolled expansion}}$$

where $\Delta T$ is the change in tumor burden, $G_{\text{lost}}$ is effector loss to host immunity, $f_{\text{IAC}}$ is the fraction of the population matching the ICANS-Associated Cell phenotype that An open CAR-T single-cell atlas links to severe neurotoxicity with an odds ratio of 6.067, and the last term penalizes graft expansion beyond a safe ceiling, proxying the tumorigenesis risk that Navigating challenges in human pluripotent stem cell-derived islet therapy for type 1 diabetes identifies as the hypoimmune paradox. The $\lambda$ weights are swept explicitly to trace the efficacy–safety Pareto front, not hidden in a single scalar.

Baselines

The naive reference point is the fixed expert recipe: the standard-of-care edit set, specifically the TRAC plus HLA-I α-chain disruption used in MT026 (Intrathecal CRISPR-edited allogeneic IL-13Rα2 CAR T Cells...) and the B2M/CIITA knockout plus MIF recipe of Macrophage Migration Inhibitory Factor Suppresses Natural Killer Cell Response.... These are the strongest competitors in the sense that they are what the field actually builds, and the published evidence shows where they fail: MT026 cells were eventually cleared by NK cells and patients relapsed through IL-13Rα2-negative tumor outgrowth, while HLA-KO-alone grafts were largely rejected in humanized NSG mice by day 14. The policy beats this baseline if it finds designs or schedules with higher simulated day-14-plus persistence and lower tumor burden at equal or smaller edit count.

The strongest method-level competitor is random search plus a genetic algorithm run inside the same simulator under an identical evaluation and compute budget. This is the honest control, because on a static design space of tens of thousands of options, enumeration is a genuine threat (see below). In the nearest published head-to-head, RL beat a genetic algorithm on the McPAS TCR qualification benchmark 36.52% versus 25.18%, a 45% relative margin (T-Cell Receptor Optimization with Reinforcement Learning...) — but that was a sequence space where validity constraints bite hard, and I will not assume the same margin transfers to a smaller discrete design space. A second method-level baseline is a conCAROT-style static predictor (Modeling CAR Response at the Single-Cell Level Using Conditional Optimal Transport) retrained on simulator outputs to rank designs by predicted phenotype; it achieved mean $R^2 > 0.95$ on in-distribution CAR signaling variants, and beating it means matching its ranking quality while additionally producing sequential schedules it cannot represent. Both baselines get the same simulator, the same walltime budget on the same GPU, and the same tuning effort.

The key-ingredient ablation is to strip the validity and safety components of the reward ($\lambda_2 = \lambda_3 = 0$, TCRPPO's second reward term removed) and show that the unconstrained policy drifts toward designs that look excellent on tumor kill but sit on the hypoimmune-paradox cliff — high cloaking, high uncontrolled-expansion risk. A second ablation runs the same simulator and policy on a CPU-only, non-batched implementation to quantify what the batch co-design actually buys; the accessibility literature gives the expected shape of this gap, since Large Batch Simulation for Deep Reinforcement Learning reported a 110× speedup over a distributed CPU-to-GPU pipeline.

Why not just…?

Why not just enumerate all edit combinations on the GPU? For static designs this is a real threat and I will run it as the primary baseline rather than argue around it. The enumeration fails only when the search space becomes sequential: choosing infusion timing, second-round edits after observed antigen escape, and expression levels as continuous knobs grows the space beyond enumeration, and a policy amortizes across recipient immunotypes (varied host NK and T-cell profiles) rather than re-solving per instance. If enumeration matches the policy on the static subset, that result will be reported as the finding, and the contribution shifts to the calibrated simulator and the sequential extension — which is the more defensible position anyway.

Why not just rent a cloud cluster and brute-force it? The evidence says clusters do not buy proportional speed here: Large Batch Simulation for Deep Reinforcement Learning showed a single RTX 3090 reached 97% of a 64-GPU cluster's navigation accuracy in half the walltime (1.5 versus 3 days), because distributed setups bleed time into serialization and GPU starvation; Beyond The Rainbow reached a human-normalized IQM of 7.4 on Atari-60 with a 2.9-million-parameter model needing 0.9 A100 GPU-days versus 7.7 for Dreamer-v3. Beyond speed, the accessibility constraint is the point: cell-therapy labs do not have ML clusters, and a design tool they can run on the workstation next to the flow cytometer is one they will actually use.

Why not just extend conCAROT to the evasion loci? conCAROT predicts a static distribution of gene expression at fixed timepoints (Day 0, 6, 12); the design problem here is closed-loop kinetics — NK clearance rates that depend on current HLA levels, tumor antigen escape that depends on past killing, exhaustion that depends on cumulative stimulation. A static map cannot choose a dosing schedule or respond to escape, and adapting it would still leave the cluster-scale ESM2 backbone that this proposal deliberately avoids. The two are complementary: conCAROT could serve as an independent check on the top-ranked designs' predicted phenotypes.

Why not just use the established knockout recipe? Because the published recipes demonstrably saturate: MT026's dual knockout still ended in NK clearance and antigen-loss relapse, and the islet-review literature warns that uncalibrated cloaking trades rejection for unmonitorable tumorigenesis (Navigating challenges in human pluripotent stem cell-derived islet therapy for type 1 diabetes). More pointedly, the field has an unresolved internal contradiction: MT026's authors argue direct HLA-I heavy-chain disruption is superior because it softens NK missing-self attacks, while the MIF and islet lines insist complete B2M knockout plus inhibitory transgenes is the right route. No head-to-head comparison exists in the file — the simulator makes one possible.

Proposed Experiments and Decisive Checks

Experiment one is calibration: fit simulator parameters once, then test whether the single parameter set reproduces all three cited regimes (the ~28% versus ~17% NK lysis and 4× nude-rat survival from Li et al. 2025; the 42–46 versus 34–38 versus 28–32 day survival ordering from the Loop CAR iNK study; the MT026 persistence-then-clearance trajectory from Li et al. 2026). Calibration error per regime is the primary simulator metric.

Experiment two is the strategy head-to-head: run direct HLA-I α-chain disruption (Strategy A) against B2M/CIITA knockout plus MIF/CD47 (Strategy B) across a panel of simulated recipient immunotypes, varying NK and T-cell competence. The output is a predicted winner per immunotype, plus the boundary conditions where the ranking flips — a concrete, testable prediction for a collaborating wet lab using the NK-92MI assay and humanized NSG models already established in the cited protocols.

Experiment three is the design search: train the compact policy on vectorized wells and compare against random search, the genetic algorithm, and the conCAROT-style static ranker under matched compute on one RTX 4090-class GPU. Success criteria are stated in advance: the policy must match or beat the best baseline on simulated persistence and tumor burden, produce schedules no static method can, and trace a Pareto front in the $\lambda$-sweep that a domain expert judges biologically sensible. Throughput targets mirror the accessibility literature's envelope — following Sample Factory: Egocentric 3D Control from Pixels at 100000 FPS with Asynchronous Reinforcement Learning and Large Batch Simulation for Deep Reinforcement Learning as precedents — with a goal of tens of thousands of simulated cell-steps per second and full policy training in under two GPU-days, inside the 0.9 A100-GPU-day envelope that Beyond The Rainbow demonstrated is achievable for a compact agent. No reported result found on Lacuna for GPU throughput of any cellular or immune-interaction simulator, which is precisely the gap this fills.

The decisive check for the whole program is falsifiable: if the single calibrated parameter set cannot reproduce the cited regimes within a stated tolerance, or if enumeration matches the policy on the static design subset, the simulator and honest comparison remain the contribution, and the RL layer is reported as unnecessary — a negative result that would still settle whether off-the-shelf cell design is, or is not, a sequential decision problem.