PortrAIgent：用于端到端空间转录组学发现的协同科学家智能体

Background Spatial transcriptomics (ST) technology maps gene expression within tissue structures, offering unprecedented insights into tissue organization and cellular interactions. Analyzing these complex datasets, however, remains a significant bottleneck. Current workflows demand specialized, multi-domain expertise and rely on heavy manual intervention, an expertise barrier that hinders the rapid translation of data into biological insights. Method To address this challenge, we developed 'PortrAIgent,' a co-scientist AI agent that autonomously manages complex analysis workflows. The system operates by having (1) a multimodal LLM for reasoning, planning, and code generation; (2) a LangGraph-based workflow manager control the execution order of analysis steps; and (3) specialized computational tools, including a code execution environment and a real-time literature retrieval tool. A core feature of the agent is its Retrieval-Augmented Generation (RAG) system, which leverages codebases from scanpy, squidpy, and scvi-tools to integrate diverse informatics tools for specialized objectives like cell typing, spatial profiling, or data integration. Results Two workflows were successfully validated: 1) Hypothesis-Driven Exploration: When a user presents a biological hypothesis, the agent proposes alternative hypotheses, refines them via literature search, and then passes the plan to a ‘Reviewer Agent'. Following this validated plan, the agent proceeds to execution. 2) Analysis-Driven Workflow: For simple requests (e.g., "group comparison"), the agent's Planner checks the AnnData object's state to determine the optimal analysis functions before generating and executing the code. We demonstrated PortrAIgent by applying it to analyzing TME by ST data. Autonomous reasoning capability of this framework was validated by testing the system on multiple ST datasets with varying preprocessing states. The agent consistently detected missing steps (e.g., normalization, HVG selection, batch correction), revised its own plan, and regenerated the appropriate code without manual intervention. It interpreted the biological meaning of the discovered patterns using real-time literature search and generated a comprehensive research report summarizing the entire discovery process. Conclusion PortrAIgent offers a novel approach to automated scientific discovery in spatial biology for oncology. Beyond the primary case study, the system was tested across multiple tissue contexts, where it reliably adjusted analysis plans, corrected missing preprocessing steps, and generated coherent biological interpretations comparable to expert feedback. By integrating dynamic RAG-based code generation, autonomous execution, and comprehensive report generation, PortrAIgent streamlines complex spatial data analysis and lowers the expertise barrier for understanding complex biological systems.