TransMAgent: Dynamic Transcriptional Regulation Analysis Using Multi-Omics Aware Multi-Agent Systems

Transcriptional regulation research has long been hindered by critical challenges, including the difficulty of integrating multi-omics data, the complexity of analytical toolchains, and the steep technical expertise required for computational workflows. To address these limitations, we developed TransMAgent, an advanced multi-agent network system specifically engineered for transcriptional regulation analysis. By leveraging task decomposition with reflective mechanisms, multi-agent collaboration, dynamic context management, and a domain-specific knowledge base coupled with automated tool expansion, the system achieves end-to-end automation that encompasses the entire workflow from tool deployment and data acquisition to task decomposition and execution. TransMAgent enables efficient and precise analysis across diverse scenarios, extending from super enhancer identification to the inference of complex, multidimensional, single-cell regulatory networks, which is a substantial advancement from simple tool invocation to automated tool construction. Furthermore, through an efficient training strategy for domain-specific agent language models, we significantly improved the performance of the original base model on transcriptional regulation tasks. By transforming complex transcriptional regulation analysis into minimalist interactions driven purely by dialogue, this research provides the field with a highly scalable, comprehensive agent context engineering paradigm centered on large language models.转录调控（Transcriptional regulation）领域的研究长期以来受诸多关键挑战制约，包括多组学（multi-omics）数据整合难度大、分析工具链结构复杂，以及计算流程所需的专业技术门槛极高。为解决上述局限，本研究开发了TransMAgent——一款专为转录调控分析打造的先进多智能体（multi-agent）网络系统。该系统借助带反思机制的任务分解、多智能体协作、动态上下文管理，以及集成自动化工具扩展的领域专属知识库，实现了从工具部署、数据获取，到任务分解与执行的全流程端到端自动化。TransMAgent可在多元场景下实现高效精准的分析，覆盖从超级增强子（super enhancer）识别，到复杂多维单细胞调控网络推演的全谱系任务；这一成果实现了从简单工具调用到自动化工具构建的实质性跨越。此外，通过针对领域专属智能体语言模型的高效训练策略，本研究大幅提升了原始基础模型在转录调控任务上的性能表现。本研究将复杂的转录调控分析转化为纯粹由对话驱动的极简交互模式，为该领域提供了一种以大语言模型（Large Language Model）为核心的高可扩展、全功能智能体上下文工程范式。