Executable Simulations of Stem Cell Transplant Risk via Specialized Debugging Agents

Live stem cell transplants carry a built-in paradox: the same plasticity that lets cells repair tissue also lets them form tumors, and the tumor risk is not even cell-autonomous. Transplanted embryonic stem cells secrete Interleukin-34, which keeps host macrophages alive and reprograms them into vessel-growing, tumor-supporting cells; depleting macrophages shrinks teratomas from an average of 2502 mm³ to 83 mm³ in mice Embryonic Stem Cells Promoting Macrophage Survival and Function are Crucial for Teratoma Development. Cell-free alternatives exploit the same immune machinery in reverse: enucleated, non-replicating "Cargocytes" are deliberately engulfed by macrophages, flipping them from pro-inflammatory to anti-inflammatory and restoring testosterone to baseline in 14 days in a rat depletion model Enucleated bone marrow-derived mesenchymal stromal cells regulate immune microenvironment and promote testosterone production through efferocytosis. These competing dynamics — nonlinear, multi-scale, and full of feedback loops between graft and host immune system — are exactly the kind of system that hand-written mathematical models struggle to capture and that no current AI system simulates at all.

At the same time, a separate literature has learned how to make language models write correct code: close the loop with an execution environment, split the work across specialized agents, and roll back bad edits. Applied to hardware debugging, this recipe lifted a GPT-3.5 model above standalone GPT-4 MEIC: Re-thinking RTL Debug Automation using LLMs; applied to repository-scale software engineering, it improved issue resolution eight-fold over raw GPT-4 MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution. This proposal brings those two literatures together. The claim in one sentence: a compact (7–9B parameter) language model, post-trained with reinforcement learning against a numerical simulation oracle and embedded in a multi-agent write–run–score–rollback loop, can generate executable dynamical models of transplant-risk and cell-free-repair pathways that frontier models prompted in one shot cannot — and that this structure-plus-specialization route beats the scale route for biological simulation.

Core Claim

The closest existing work, VCWorld: A Biological World Model for Virtual Cell Simulation, predicts how cells respond to perturbations by retrieving paths from a biological knowledge graph and asking a large generalist model to reason over them in a single pass, reaching differential-expression accuracies of 0.70–0.76 on the GeneTAK benchmark. Its authors state that biological simulation quality scales with raw model intelligence, and they explicitly leave multi-agent extensions to future work. The software-engineering evidence points the opposite way: monolithic intelligence fails without execution feedback and role separation, and structured pipelines let small models rival large ones. This proposal treats that contradiction as its central empirical question, and predicts the structured route wins for executable, quantitative biological dynamics — not just qualitative gene-direction predictions — because the hard part is not knowledge retrieval but writing differential-equation code that runs, converges, and reproduces measured numbers.

Concretely, we will build two things. First, TransplantRiskBench, a benchmark of simulation-synthesis tasks assembled from the quantitative endpoints published in the transplant-risk literature: teratoma volume collapse under macrophage depletion and aortic-ring vessel sprouting of roughly 1.8 mm² Embryonic Stem Cells Promoting Macrophage Survival and Function are Crucial for Teratoma Development, the 72-hour viability window and 14-day testosterone recovery of Cargocytes Enucleated bone marrow-derived mesenchymal stromal cells regulate immune microenvironment and promote testosterone production through efferocytosis, and exosome-mediated wound-healing cascades Exosomal coactivator-associated arginine methyltransferase 1 derived from adipocytes accelerates diabetic wound healing by modulating inflammation and promoting angiogenesis, Exosomes from adipose-derived stem cells accelerate wound healing by increasing the release of IL-33 from macrophages. Each task gives the system a pathway description and literature extracts, and asks for runnable code — ordinary differential equation or stochastic kinetic models solved with standard numerical tools — whose simulated perturbation responses match the published direction and approximate magnitude of each endpoint. Second, GraftSim, a three-role agent pipeline: a Mechanism agent that proposes pathway structure (building on VCWorld's knowledge-graph grounding), a Coder agent that writes and edits the simulation script, and an independent Scorer agent that checks biological plausibility, dimensional consistency, and solver convergence, with a stateful rollback that reverts any edit that degrades the best stable state — the mechanism MEIC showed was essential to stop regression loops in hardware code.

The Coder and Scorer start from a compact open model and are post-trained with Group Relative Policy Optimization, the same family of rule-verified training that let a 7B molecular-reasoning model reach AVG4 scores of 64.7–67.9 on MoleculeNet, far above its base model's 45–49 and above 100B+ proprietary models AttriLens-Mol: Attribute Guided Reinforcement Learning for Molecular Property Prediction with Large Language Models, and that let a 9B medical agent score 79.6 on AutoMedBench-Lite, beating Claude Opus 4.6 with Claude Code at 77.5 BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics. The reward is computed entirely by a deterministic oracle — the simulator itself — so no human grading enters the training loop:

$$R = \mathbb{1}[\text{code executes}] + \mathbb{1}[\text{solver converges}] + \mathbb{1}[\text{sign of predicted change matches}] - \lambda\,\mathbb{1}[\text{edit regressed the best state}],$$

where $\mathbb{1}[\cdot]$ is 1 when the condition holds and 0 otherwise, and $\lambda$ penalizes regressive edits so the policy learns rollback-compatible behavior. Magnitude fidelity (whether the simulated teratoma shrinks by the right order, not just the right sign) is reserved for held-out evaluation, so the trained model is never rewarded on the test endpoints.

Baselines

The "do nothing" reference point is one-shot prompting of a frontier generalist model: give it the same task description and literature extracts and ask for the simulation script in a single pass. This is the strongest naive comparator because it is what a practitioner would do today, and the literature documents exactly how it fails under execution requirements — raw GPT-4 resolved only 1.74% of SWE-bench issues before structure was added MAGIS: LLM-Based Multi-Agent Framework for GitHub Issue Resolution, and standalone GPT-4 fixed only 62% of RTL syntax errors versus 93% inside the MEIC loop MEIC: Re-thinking RTL Debug Automation using LLMs. It will be run with the same literature extracts, the same solver, and the same scoring code as GraftSim, differing only in the absence of agents, iteration, and specialization.

The strongest existing method on this question is the VCWorld pipeline — knowledge-graph retrieval feeding single-pass chain-of-thought reasoning with a large generalist model — because it is the only published system that performs biological perturbation prediction with grounded LLM reasoning, reporting DE accuracies of 0.70–0.76 on GeneTAK VCWorld: A Biological World Model for Virtual Cell Simulation. We will run it verbatim on our benchmark's perturbation-transfer tasks (same inputs, same endpoints) and, where it produces qualitative predictions rather than code, score its directional accuracy and note that it produces no executable artifact. A second serious competitor is a single-agent execution loop — one model in a while-loop receiving simulator tracebacks, in the spirit of Gupta and Hollmann et al. — run with the same total token budget as GraftSim so the comparison is not confounded by compute. Beating the baselines means: a strictly higher task-pass rate (correct direction and magnitude on held-out endpoints) than both, at equal or lower token cost than the single-agent loop, with the gap large enough that it survives swapping the underlying base model.

The key ablation isolates specialization and rollback separately. GraftSim with a general instruct model instead of the GRPO-trained Coder isolates the contribution of simulator-grounded post-training; GraftSim with rollback disabled isolates the contribution of stateful reversion. The rollback ablation matters because VERINA: Benchmarking Verifiable Code Generation showed that 64 rounds of raw verifier feedback plateau near ~20% proof success with severe token thrashing, and Yan et al. showed single-model refinement plateaus after a few rounds — so the ablation should reproduce thrashing, and its absence in the full system is evidence the mechanism, not the model, is doing the work.

Why not just…?

Why not just prompt a frontier model like Gemini or GPT-4o in one shot? This is the VCWorld hypothesis, and it deserves a fair test rather than dismissal: VCWorld reports that jumping from an 8B open model to Gemini-2.5-Flash moved DE accuracy substantially, to 0.70–0.76. But that result covers qualitative gene-direction prediction, not executable quantitative dynamics, and the one-shot evidence on tasks requiring code that actually runs is uniformly bad — 1.74% on SWE-bench for raw GPT-4, and a 3.2% end-to-end success rate for OpenAI o3 on Lean 4 tasks requiring code, specification, and proof together VERINA: Benchmarking Verifiable Code Generation. Our main experiment runs the frontier one-shot baseline directly on TransplantRiskBench, so the scale-versus-structure question is settled empirically on the same tasks rather than by analogy. If the frontier one-shot passes most tasks, the proposal's premise is wrong and we will report that.

Why not just run a single agent in a loop with the simulator's error messages? It is the cheapest serious design, and it fixes the syntax-level failures that early systems like Gupta targeted. The evidence against it as a sufficient design is strong: single-model loops thrash — VERINA's ~20% plateau after 64 rounds is the cleanest number — and they can amplify latent flaws when the verifier checks only functional output Srivastava et al.. MEIC found that a lone debug agent falls into hallucination traps that an independent scorer plus rollback prevents MEIC: Re-thinking RTL Debug Automation using LLMs. We still treat this as a real threat, which is why it is a first-class baseline with an equal token budget: if the single-agent loop matches GraftSim's pass rate, the multi-agent machinery is not earning its cost.

Why not just do supervised fine-tuning or retrieval over a domain corpus? Fine-tuning and retrieval inject knowledge but not procedure. On the closest analogues in the file, retrieval/supervised baselines reached a 66% functional fix rate versus 78% for the agentic MEIC pipeline MEIC: Re-thinking RTL Debug Automation using LLMs, and base or SFT models scored roughly 45–49 on MoleculeNet versus 64.7–67.9 after oracle-grounded RL AttriLens-Mol: Attribute Guided Reinforcement Learning for Molecular Property Prediction with Large Language Models. The failure mode is specific: debugging a diverging ODE solver requires learning from execution outcomes, which static corpora do not contain. We will run a retrieval-augmented SFT variant as a third baseline to confirm this transfer holds for simulation code.

Why not just have systems biologists hand-write the ODE models? For the two best-characterized pathways — IL-34-driven macrophage reprogramming and Cargocyte efferocytosis — an expert could plausibly write a serviceable model by hand, and we do not claim otherwise; indeed, hand-built expert models would be the ideal fidelity reference, though no reported result found on Lacuna for expert-coded ODE reproduction of these specific endpoints, so we cannot calibrate against one. The honest answer is that the proposal's value is not replacing two hand-built models but scaling the process: the transplant-risk literature describes dozens of competing signaling cascades — exosomal Carm1 Exosomal coactivator-associated arginine methyltransferase 1 derived from adipocytes accelerates diabetic wound healing by modulating inflammation and promoting angiogenesis, macrophage IL-33 release Exosomes from adipose-derived stem cells accelerate wound healing by increasing the release of IL-33 from macrophages, lineage-commitment seesaws Fate decision of mesenchymal stem cells: adipocytes or osteoblasts? — and the bottleneck is the per-pathway human cost. Automation only matters if it generalizes, so the benchmark deliberately includes held-out pathways beyond the two flagship ones, and the decisive check is pass rate on those.

Proposed Experiments and Decisive Checks

The first experiment constructs and validates TransplantRiskBench. Roughly forty tasks are derived from published quantitative endpoints, split into a training-visible set (used only for the sign-match training reward, never for evaluation) and a held-out evaluation set. Two domain experts verify that each task's target signature is stated unambiguously and is actually implied by the source paper; a task passes when the generated model executes, converges, and reproduces the direction and order of magnitude of the held-out endpoint (for example, predicting a near-collapse of teratoma volume under macrophage depletion, consistent with the 2502 mm³ to 83 mm³ shift reported in Embryonic Stem Cells Promoting Macrophage Survival and Function are Crucial for Teratoma Development).

The second experiment is the main comparison: frontier one-shot, VCWorld pipeline, equal-budget single-agent loop, retrieval/SFT compact model, GraftSim with a general instruct Coder, and full GraftSim with the GRPO-trained Coder — all on identical tasks, identical solver, identical scoring code, and a matched total token budget. Metrics are task-pass rate, directional accuracy, token cost per pass, and wall-clock time. The third experiment transfers to the perturbation-setting where prior results exist: GeneTAK-style differential-expression prediction with models trained on Tahoe-100M: A Giga-Scale Single-Cell Perturbation Atlas data, scored against VCWorld's reported 0.70–0.76 DE accuracy, testing whether generating an explicit dynamical model helps even on the task the scale-first approach was built for.

The fourth experiment is the thrashing study: full GraftSim versus rollback-disabled and scorer-disabled variants across increasing refinement rounds, measuring convergence rate, regression frequency, and token bloat, directly testing whether rollback plus an independent scorer prevents the plateau documented in VERINA: Benchmarking Verifiable Code Generation (dataset at VERINA) and Yan et al.. The fifth check guards against over-specialization: following the degradation BaT observed when post-training a 4B model on non-domain benchmarks BaT: Towards Self-Evolving Medical Research Agent with Stage Rubrics, the specialized Coder is evaluated on general code benchmarks before and after training, and a size sweep over 4B, 7B, and 9B backbones maps where specialization stops paying for itself.

The decisive outcome, stated in advance: if the structured, specialized pipeline does not beat both the frontier one-shot baseline and the equal-budget single-agent loop on held-out pathway pass rate — or if it wins only on the two flagship pathways and not the held-out ones — then the claim that structure plus specialization beats scale for executable biological simulation is refuted, and the VCWorld scaling hypothesis stands for this domain. Either result resolves a genuine contradiction between two literatures that have never been run against each other on the same tasks.