PO.BCS02.01 · 生物信息与计算
用于自主 CAR-T 开发的多智能体 AI 系统：面向癌症免疫治疗的集成化靶点发现、毒性预测与理性分子设计
Multi-agent AI system for autonomous CAR-T development: Integrated target discovery, toxicity prediction, and rational molecular design for cancer immunotherapy
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：嵌合抗原受体 T 细胞（CAR-T）疗法在血液系统恶性肿瘤中取得了显著成功，然而自 2017 年以来仅有 6 款产品获得 FDA 批准，靶点选择和毒性仍是关键瓶颈。传统研发流程需要 8-12 年，因验证不充分或安全性隐患导致的失败率高达 40-60%。若干显著的临床挫折凸显了在昂贵的临床试验之前预测安全风险的自主系统的迫切需求。
方法：我们开发了 Bio AI Agent，这是一套由大语言模型驱动、能够实现自主 CAR-T 开发的多智能体 AI 系统。该架构由六个智能体组成：(1) 靶点选择智能体，用于抗原优先级排序；(2) 毒性预测智能体，整合组织图谱和药物警戒数据库；(3) 分子设计智能体，用于模块化 CAR 工程；(4) 专利情报智能体，用于自由实施（freedom-to-operate）分析；(5) 临床转化智能体，用于监管指导；(6) 决策编排智能体，用于多智能体协调。各智能体通过配备向量数据库和自然语言接口的共享知识库进行通信。
结果：回顾性验证展示了跨关键阶段的自主能力。靶点评估将 3-4 个月的工作流程简化为快速处理。毒性预测通过表达谱分析和不良事件分析准确识别了有问题的靶点。专利情报标记了侵权风险，从而实现合规的设计策略。分子设计展示了具有实时预测的系统化优化。决策编排生成了涵盖验证、路径开发和临床转化的全面路线图。
结论：Bio AI Agent 通过在靶点发现、安全性预测和分子优化之间的智能协作，解决了 CAR-T 开发中的关键缺口。对存在隐患靶点的自主识别及缓解策略展示了减少研发损耗和改善安全性的潜力。多智能体架构实现了并行处理和专门推理，优于单体式系统。随着 CAR-T 向实体瘤扩展，自主平台对于加速精准肿瘤学将至关重要。
查看英文原文 English abstract
Background: Chimeric antigen receptor T-cell (CAR-T) therapy has achieved remarkable success in hematologic malignancies, yet only 6 products have gained FDA approval since 2017, with target selection and toxicity remaining critical bottlenecks. Conventional pipelines require 8-12 years with 40-60% failure rates due to inadequate validation or safety liabilities. Notable clinical setbacks underscore urgent need for autonomous systems predicting safety risks before costly trials.
Methods: We developed Bio AI Agent, a multi-agent AI system powered by large language models enabling autonomous CAR-T development. The architecture comprises six agents: (1) Target Selection for antigen prioritization, (2) Toxicity Prediction integrating tissue atlases and pharmacovigilance databases, (3) Molecular Design for modular CAR engineering, (4) Patent Intelligence for freedom-to-operate analysis, (5) Clinical Translation for regulatory guidance, and (6) Decision Orchestration for multi-agent coordination. Agents communicate through shared knowledge base with vector database and natural language interfaces.
Results: Retrospective validation demonstrated autonomous capabilities across key stages. Target assessment streamlined 3-4 month workflows to rapid processing. Toxicity prediction accurately identified problematic targets through expression profiling and adverse event analysis. Patent intelligence flagged infringement risks enabling compliant design strategies. Molecular design demonstrated systematic optimization with real-time prediction. Decision orchestration generated comprehensive roadmaps spanning validation, pathway development, and clinical translation.
Conclusions: Bio AI Agent addresses critical CAR-T development gaps through intelligent collaboration across target discovery, safety prediction, and molecular optimization. Autonomous identification of liability targets with mitigation strategies demonstrates potential for reducing attrition and improving safety. Multi-agent architecture enables parallel processing and specialized reasoning superior to monolithic systems. As CAR-T expands into solid tumors, autonomous platforms will be essential for accelerated precision oncology.
利益披露 Disclosure
Y. Ni, None..
L. Zhu, None.