PO.BCS02.01 · 生物信息与计算
用于加速 MERSCOPE 端到端空间生物学分析的 AI 分析智能体
AI analysis agent to accelerate end-to-end spatial biology analysis for MERSCOPE
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：MERSCOPE 平台上的空间生物学检测能够在肿瘤结构内实现 RNA 转录本和蛋白质的亚细胞定位。然而，这些检测会生成 TB 级规模的数据集，需要复杂的计算工作流程。采集后分析（包括细胞分割、细胞分型和空间域检测）仍然计算密集，并且需要专业的生物信息学专长。这可能限制许多癌症生物学实验室的可及性。为克服这些障碍，Vizgen 和 LatchBio 开发了一套由大语言模型（LLM）智能体驱动的 AI 工作流程，旨在简化并优化肿瘤学研究中 MERSCOPE 的端到端空间生物学分析。
方法：我们实现了一个针对 Vizgen 数据生态系统量身定制的 LLM 驱动智能体。该智能体：(1) 解析 MERSCOPE 输出（多通道 z-stack 图像、转录本坐标文件、元数据），并启动经过优化的 GPU 加速工作流程以进行图像分割和转录本分配；(2) 提供交互式 notebook 界面（Markdown、图表、组件），使用户能够用自然语言指定下游问题（例如“识别肿瘤边界区域内的细胞毒性 T 细胞”）；(3) 触发用于聚类、细胞类型注释、空间域检测以及差异表达/调控的生物信息学流程；(4) 与 LatchBio 可扩展的计算/存储基础设施集成，以进行大规模运行（每个数据集 >1 TB）。我们在 MERSCOPE 上处理的乳腺癌和结直肠癌样本上验证了性能。
结果：该智能体成功处理了大规模、多样本的 MERFISH 空间转录组学数据集。它在单次运行中完成了端到端分析（细胞重分割、无监督聚类、空间域检测以及选定细胞群体的差异表达）。输出以交互式、可复现的 notebook 形式交付，可快速查看细胞类型注释的聚类、空间域图谱和差异表达汇总。与传统的手动脚本化工作流程相比，该智能体在约 6 小时内完成了完整分析（传统方法约 72 小时）。在终端用户测试中，它将手动脚本编写和依赖管理工作量减少了约 60%，并降低了下游错误，提高了流程可靠性。
结论：Vizgen-LatchBio 的 AI 智能体工作流程为高级 MERSCOPE 数据分析提供了可扩展、易用的解决方案，使空间生物学实验室能够自助完成复杂的计算任务。通过抽象化计算复杂性并嵌入领域特定逻辑，该系统使生物学家能够专注于生物学洞见，而非技术工具链的串联。因此，它缩短了获得洞见的时间，并在药物发现、疾病研究和学术研究中促进了高通量工作流程。
查看英文原文 English abstract
Background: Spatial biology assays on the MERSCOPE platform enable subcellular mapping of RNA transcripts and proteins within the tumor architecture. However, these assays generate terabyte-scale datasets that require complex computational workflows. Post-acquisition analysis including cell segmentation, cell typing, and spatial domain detection remains computationally intensive and requires specialized bioinformatics expertise. This can limit accessibility for many cancer biology laboratories. To overcome these barriers, Vizgen and LatchBio have developed an AI-driven workflow powered by large language model (LLM) agents, designed to streamline and simplify end-to-end spatial biology analysis for MERSCOPE in oncology research.
Methods: We implemented an LLM-driven agent tailored to the Vizgen data ecosystem. The agent (1) parses MERSCOPE outputs (multi-channel z-stack images, transcript coordinate files, metadata) and launches optimized GPU-accelerated workflows for image segmentation and transcript assignment; (2) presents an interactive notebook interface (Markdown, plots, widgets) so users can specify downstream questions in natural language (e.g., “Identify Cytotoxic T-Cells within tumor border regions”); (3) triggers bioinformatics pipelines for clustering, cell-type annotation, spatial domain detection, and differential expression/regulation; and (4) integrates with LatchBio's scalable compute/storage infrastructure for large-scale runs (>1 TB per dataset). We validated performance on breast cancer and colorectal cancer samples processed on MERSCOPE.
Results: The agent successfully handled large-scale, multi-sample MERFISH spatial transcriptomics datasets. It performed end-to-end analysis (cell re-segmentation, unsupervised clustering, spatial domain detection, and differential expression in selected cell populations) within a single run. Outputs were delivered as interactive, reproducible notebooks for a rapid review of cell-type-annotated clusters, spatial domain maps, and differential expression summaries. Compared with a conventional manually scripted workflow, the agent completed full analyses in ~6 hours (vs ~72 h traditionally). In end-user testing, it reduced manual scripting and dependency management efforts by ~60 % and decreased downstream errors, improving pipeline reliability.
Conclusions: The Vizgen-LatchBio AI-agentic workflow provides a scalable, user-friendly solution for advanced MERSCOPE data analysis, enabling spatial biology labs to self-serve complex computational tasks. By abstracting computational complexity and embedding domain-specific logic, this system empowers biologists to focus on biological insights rather than technical tool-chaining. Consequently, it shortens time-to-insight and facilitates high-throughput workflows across drug discovery, disease research, and academic studies.
利益披露 Disclosure
H. Muralidharan, None..
C. Chen, None..
F. Preusser, None..
R. Cardenes, None..
K. Workman, None..
H. Le, None..
A. Andere, None..
L. Rognoni, None.