PO.MD01.01 · 分子诊断与数据
用于结直肠癌精准肿瘤学中临床、基因组及健康社会决定因素数据实时整合的多智能体对话式人工智能生态系统
A multi-agent conversational artificial intelligence ecosystem for real-time integration of clinical, genomic, and social determinants of health data in colorectal cancer precision oncology
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
引言：临床、基因组及健康社会决定因素（SDoH）数据的快速膨胀已超出传统分析方法的处理能力，迫切需要能够大规模整合和解读复杂肿瘤学数据集的智能系统。为应对这一挑战，我们团队开发了一套特定领域的对话式 AI 智能体，可实现对多组学结直肠癌（CRC）数据的实时、自然语言驱动的探索。这些智能体有助于发现特定人群的分子改变和治疗反应模式。
方法：每个智能体均利用经过微调的生物医学 LLaMA-3 模型、一个自然语言转代码解释器，以及一个连接到来自 TCGA、AACR Project GENIE 和 cBioPortal 的协调数据集的后端统计引擎。该平台自动化了队列创建、突变谱分析、生存分析、比值比检验，以及临床和 SDoH 变量的整合。专门的智能体包括 AI-HOPE-PI3K、TGFbeta、TP53、RTK-RAS、JAK-STAT、MAPK、WNT 和 AI-HOPE-PM，后者独特地整合了临床、基因组和 SDoH 特征。一个核心智能体维护整个生态系统的数据互操作性。所有分析均由通俗语言提示触发，并在数秒内返回可视化和叙述性输出。
结果：AI 智能体成功重现了已确立的临床基因组关联，并揭示了新颖的、具有临床意义的见解。AI-HOPE-PI3K 发现 INPP4B 突变在西班牙裔/拉丁裔早发性 CRC 中富集；AI-HOPE-TGFbeta 检测到 SMAD4 突变肿瘤中与 MSI 相关的生存获益；AI-HOPE-PM 揭示了经历经济压力的 TP53 突变型 CRC 病例生存更差，以及与食物不安全相关的化疗可及性差异。其他智能体则在 MAPK、RTK-RAS 和 WNT 通路内识别出跨分期、治疗暴露和人口统计学群体的预后差异。各应用案例的准确率均超过 90%，所有分析均实时完成，无需编程专业知识。
结论：多智能体对话式人工智能生态系统为精准肿瘤学提供了一个可扩展、可互操作且以人群信息为依据的多智能体架构。智能体到智能体（A2A）和模块化协作协议（MCPs）的开发将实现协调的跨领域分析和假设生成，推进面向癌症研究的协作式 AI 生态系统。通过在对话式界面下统一临床、分子和 SDoH 数据，AI 智能体生态系统引入了一种数据智能的新范式，加速生物标志物发现，并支持跨人群的、以人群信息为依据的精准肿瘤学。
查看英文原文 English abstract
Introduction: The rapid expansion of clinical, genomic, and social determinants of health (SDoH) data has outpaced traditional analytic approaches, creating an urgent need for intelligent systems capable of integrating and interpreting complex oncology datasets at scale. To address this challenge, our team developed a suite of domain-specific conversational AI agents that enable real-time, natural language-driven exploration of multi-omic colorectal cancer (CRC) data. These agents facilitate discovery of population-specific molecular alterations and treatment-response patterns.
Methods: Each agent leverages fine-tuned biomedical LLaMA-3 models, a natural language-to-code interpreter, and a backend statistical engine linked to harmonized datasets from TCGA, AACR Project GENIE, and cBioPortal. The platform automates cohort creation, mutation profiling, survival analysis, odds ratio testing, and integration of clinical and SDoH variables. Specialized agents include AI-HOPE-PI3K, TGFbeta, TP53, RTK-RAS, JAK-STAT, MAPK, WNT, and AI-HOPE-PM, the latter uniquely integrating clinical, genomic, and SDoH features. A core agent maintains data interoperability across the ecosystem. All analyses are triggered by plain-language prompts and return visual and narrative outputs within seconds.
Results:AI-agents successfully reproduced established clinical-genomic associations and uncovered novel, clinically meaningful insights. AI-HOPE-PI3K identified INPP4B mutations enriched in Hispanic/Latino early-onset CRC; AI-HOPE-TGFbeta detected MSI-associated survival benefits in SMAD4-mutant tumors; and AI-HOPE-PM revealed worse survival in TP53-mutant CRC cases experiencing financial strain, along with differences in chemotherapy access linked to food insecurity. Additional agents identified prognostic variation across stage, treatment exposure, and demographic groups within MAPK, RTK-RAS, and WNT pathways. Accuracy exceeded 90% across use cases, with all analyses completed in real time without requiring programming expertise.
Conclusions: Multi-agent conversational artificial intelligence ecosystem provides a scalable, interoperable, and population-informed multi-agent architecture for precision oncology. Development of Agent-to-Agent (A2A) and Modular Collaborative Protocols (MCPs) will enable coordinated, cross-domain analysis and hypothesis generation, advancing a collaborative AI ecosystem for cancer research. By unifying clinical, molecular, and SDoH data under a conversational interface, AI-agents ecosystem introduces a new paradigm for data intelligence, accelerates biomarker discovery, and supports population-informed precision oncology across populations.
利益披露 Disclosure
E. Velazquez-Villarreal, None..
B. Waldrup, None.