What question did this study set out to answer?

To develop a multi-agent framework using large language models for simulating the pancreatic tumor microenvironment.

March 29, 20260 citationsOpen Access

OncoBiome Swarm: A Large Language Model-Based Multi-Agent Framework for Emergent Simulation of the KRAS G12D Pancreatic Tumor Microenvironment

Puntos clave

To develop a multi-agent framework using large language models for simulating the pancreatic tumor microenvironment.
Implemented a novel agent-based modeling (ABM) framework with independent LLMs for each agent.
Introduced NKCell and DendriticCell agents to study immune dynamics.
Conducted RMSE validation against published pancreatic cancer data.
Performed sensitivity analysis with rule engine runs and a reproducibility experiment.
LLM agents predicted immune collapse 43-74 days earlier than traditional deterministic models.
NK cell exhaustion was observed to consistently precede CD8+ T cell collapse across all runs.
Dendritic cells remained tolerogenic without IFN-γ priming under both modeling paradigms.
A non-linear NK cell dose response demonstrated significant impact on tumor reduction.

Resumen

OncoBiome Swarm is a novel agent-based modeling (ABM) framework in which every biological agent in the tumor microenvironment (TME) is driven by an independent large language model (LLM) with persistent episodic memory. Applied to KRAS G12D pancreatic ductal adenocarcinoma (PDAC), the framework reproducibly generates emergent immunosuppressive phenomena absent from deterministic rule-based systems. Sprint 4 introduces NKCell and DendriticCell agents, RMSE validation against published PDAC data, one-at-a-time and full factorial sensitivity analysis (34 rule engine runs, 0), and n=3 reproducibility for the innate-adaptive bridge experiment (8 CD8⁺ + 4 NK + 3 DC). Central findings: (1) LLM agents produce immune collapse 43–74 days earlier than deterministic rules (CV=8. 3%, n=3) ; (2) NK exhaustion consistently precedes CD8⁺ collapse by 4–9 cycles in all n=3 runs — unprogrammed emergent behavior; (3) DC remain tolerogenic without IFN-γ priming in both paradigms (Immunity 2023, PMID 37625410) ; (4) non-linear NK dose threshold: 4 NK → 0%, 8 NK → −37% tumor reduction. Nine emergent phenomena; none explicitly programmed. Total cost Sprints 0–4: 21. 76.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo