Researchers have introduced DianShi-RxnDB, a large-scale, fine-grained organic reaction data platform designed to support artificial intelligence for chemistry (AI4Chem). The work addresses a persistent bottleneck: high-quality structured reaction data remain scattered across patent text, images, and reaction schemes, making them difficult to use at scale.
According to the paper, the platform relies on a fully automated extraction and normalization pipeline that integrates patent text, images, and reaction schemes. Its corpus draws on organic synthesis patents from the USPTO and EPO published between 1976 and 2025, producing approximately 24 million reaction instances. Of these, roughly 14.8 million (61.7%) pass automated qualification checks.
Each qualified instance represents a specific single-step experiment and records participants, roles, quantities, temperatures, reaction times, yields, experimental procedures, and provenance links back to the source patents. In a manual evaluation of 1,300 sampled qualified instances, the authors report a micro-averaged field-level accuracy of 92.95%.
A matched comparison with Pistachio indicated advantages for DianShi-RxnDB in deduplicated record counts, representation granularity, and field-level exact agreement, according to the paper.
Beyond the dataset itself, the platform offers a Web research workbench for searching, filtering, comparing, and source-verifying records, plus a Model Context Protocol (MCP) service that provides AI agents with composable structured retrieval tools. The platform is available at https://dianshi.opendatalab.org.cn/.