Chemistry figures no longer need manual transcription: RxnMiner lifts stereochemistry-aware reaction F1 from 21.0% to 71.0%
Synopsis
Reaction figures in the chemical literature often pack several experiments into one shared template, substituent table and product grid, and recovering concrete reaction records from them has meant manual transcription. RxnMiner combines an expert-informed extraction procedure with a chemistry toolbox in one multimodal agent: it reads structures, binds evidence to each experimental entry, and instantiates undrawn reactants with a deterministic R-group engine. On ChemRxn-Bench, 603 reaction figures with 8,547 annotated records, it reaches 71.0% stereochemistry-aware reaction macro-F1 with a Qwen3.8-Flash-Next backbone, against 21.0% for direct prediction and 25.0% for the strongest existing extraction system.
Interpretation
A reaction figure can be turned automatically into individual concrete reaction records, each with fully instantiated reactant and product structures plus retained conditions, outcomes and provenance. Extracting reactions from journal figures previously required substantial manual effort; AbSynth turned nearly a century of total syntheses into a machine-readable resource only through undergraduate curation that assembled more than 3,000 total syntheses spanning close to 60,000 steps. On the 603 figures of ChemRxn-Bench with Qwen3.8-Flash-Next as backbone, RxnMiner reaches 71.0% stereochemistry-aware reaction macro-F1, versus 21.0% for the same model under direct prediction and 25.0% for the strongest existing extraction system, ChemEAGLE.
The agent organizes extraction by experimental entry: it reads molecules from the figure, binds evidence from shared templates, table rows and footnotes to each entry, then instantiates reactants that are not drawn separately. Recognizing molecules alone does not determine a reaction record, because shared templates and substituent tables draw information once while it applies to several experimental entries. Across five backbones, RxnMiner improves reaction macro-F1 over Bare Pi, the same base agent framework without the extraction procedure and chemistry toolbox, by 9.4 to 25.5 percentage points; Qwen rises from 59.6% to 71.0% and Kimi K3 from 57.7% to 69.0%.
The same extraction configuration scales to literature size without human intervention, producing a source-linked silver reaction corpus. Such corpora previously came mainly from patent text mining or contributed experimental records, while journal figures still needed manual transcription. Applied to the 10,112 figures of the RxnCaption-11k training split, it processed 10,107 figures and produced 88,202 source-linked reaction records, about 3.6 times the 24,716 reaction annotations associated with that source collection.
Perspective
Evaluation is limited to the 603 manually checked figure crops of ChemRxn-Bench, where the input contains the figure, caption, footnotes and associated table text but no article body or supporting information; the primary metric scores only molecular identity and reactant–product pairing, conditions and outcomes are assessed by a separate multimodal judge, and identifiers and reaction-graph links are not scored. The silver corpus comes from 10,107 figures of the RxnCaption-11k training split and has not undergone manual validation.
The silver corpus has not undergone manual validation; in a random sample of 1,000 figures, 25,223 of 25,225 molecular strings parsed with RDKit (99.99%), which indicates representational validity but not agreement with the source figures or correct pairing. Stereochemistry remains difficult: with the Qwen backbone the stereochemistry-aware and stereochemistry-blind scores are 71.0% and 78.0%, a seven-point gap, and case examples still show a single inverted stereocenter and stereochemistry assigned where the source leaves it unspecified.