An agentic preparative thin-layer chromatography system for autonomous purification

英文原文与中文译文逐段对照 · English original with Chinese translation

Wendi Cai#, Boxuan Zhao#, Yansong Yue#, Changlin Liu, Chengchun Liu, Ying Cui, Zi-Hao Zhou, Xiaohui Tian, Zhong-Chao Zhang, Zhen Yang, Jie Zhu*, and Fanyang Mo*

蔡文迪#、赵博轩#、乐岩松#、刘昌霖、刘丞军、崔莹、周子豪、田晓辉、张仲超、杨震、朱杰* 和莫凡洋*

Cite This: J. Am. Chem. Soc. 2026, 148, 43269–43279. DOI: 10.1021/jacs.6c10800. Published October 14, 2026.

引用本文: J. Am. Chem. Soc. 2026, 148, 43269–43279. DOI: 10.1021/jacs.6c10800. 2026年10月14日正式刊出。

ABSTRACT: The rapid advancement of autonomous laboratories has significantly accelerated reaction discovery and optimization, yet downstream compound purification remains a manual and poorly integrated bottleneck. Preparative thin-layer chromatography (pTLC) is a core small-scale purification method, yet full automation has been hindered by the intrinsic coupling of multistep separation reasoning and millimeter-precision robotic recovery. Here, we present PLANAR (Preparative Layer Chromatography Agentic Network for Autonomous Recovery), a supervised agentic system that provides a gated, auditable route from target-directed purification requests to isolated products. PLANAR integrates a modular robotic workcell for end-to-end plate handling, development, and recovery; a two-stage predictive model (A2P-Cascade) that ranks candidate mobile phases based on planned band center and bandwidth priors; and an agentic controller that orchestrates these tools through explicit review gates. Trained on standardized TLC and pTLC records, A2P-Cascade achieves pooled R2 values of 0.835 and 0.673 for normalized band center and width on a compound-level formal-validation split. Across three proof-of-workflow cases spanning structural formula, reaction-image, and fixed-condition analytical TLC inputs, PLANAR delivered NMR-verified isolated products while grounding recovery decisions in evidence from preparative plates. This framework transforms purification into a traceable module for self-driving laboratories, supporting the generation of standardized run records for iterative model refinement and prospective integration with online process analytical technologies. By bridging the gap between synthesis design and verified isolated products, PLANAR moves autonomous chemical workflows toward complete end-to-end experimental discovery cycles.

摘要：自主实验室的快速发展极大加速了反应发现与优化，但下游的化合物纯化仍是人工操作、集成度低的瓶颈。制备型薄层色谱（preparative TLC, pTLC）是一种核心的小量纯化方法，然而其全自动化一直受制于多步分离推理与毫米级精度机器人回收之间的内在耦合。在此，我们提出 PLANAR（Preparative Layer Chromatography Agentic Network for Autonomous Recovery，制备层析自主回收智能体网络）——一个受监督的智能体系统，为目标导向的纯化请求到分离产物的获得提供了带门控、可审计的通路。PLANAR 集成了一个模块化机器人工作站，用于端到端的薄层板处理、展开与回收；一个两阶段预测模型（A2P-Cascade），依据计划的目标条带中心与条带宽度先验对候选流动相排序；以及一个通过显式复核关卡来编排上述工具的智能体控制器。在标准化的 TLC 与 pTLC 记录上训练后，A2P-Cascade 在化合物级别的形式化验证集上，对归一化条带中心与条带宽度分别取得 0.835 与 0.673 的合并 R²。在三个工作流验证案例中（分别以结构式、反应图像和固定条件分析型 TLC 作为输入），PLANAR 交付了经 NMR 验证的分离产物，并将回收决策建立在制备板的证据之上。该框架将纯化转变为自主实验室中可追溯的模块，支持生成标准化运行记录以用于模型的迭代精修，并为未来与在线过程分析技术的集成奠定基础。通过弥合合成设计与已验证分离产物之间的鸿沟，PLANAR 推动自主化学工作流走向完整的端到端实验发现闭环。

注：本文已在 Journal of the American Chemical Society 正式刊出（2026年10月14日，卷 148，期 40，第 43269–43279 页），全文双语正文待补充。期刊正式版本。