High-quality structured organic reaction data are essential for developing artificial intelligence for chemistry (AI4Chem), yet much of this knowledge remains dispersed across patent text, images, and reaction schemes. We present DianShi-RxnDB, a large-scale, fine-grained organic reaction data platform built via a fully automated extraction and normalization pipeline integrating patent text, images, and reaction schemes. Its corpus covers organic synthesis patents from the USPTO and EPO published between 1976 and 2025, yielding approximately 24 million reaction instances, of which approximately 14.8 million (61.7%) pass automated qualification checks. Each instance represents a specific single-step experiment recording participants, roles, quantities, temperatures, reaction times, yields, experimental procedures, and provenance links to source patents. In a manual evaluation of 1,300 sampled qualified instances, the micro-averaged field-level accuracy was 92.95%. A matched comparison with Pistachio further indicated advantages in deduplicated record counts, representation granularity, and field-level exact agreement. The platform provides a Web research workbench for searching, filtering, comparing, and source-verifying records, and a Model Context Protocol (MCP) service offering AI agents composable structured retrieval tools. DianShi-RxnDB is available at https://dianshi.opendatalab.org.cn/ .
Figures & tables
Figure 1: System architecture and data-flow overview of DianShi-RxnDB. Through a fully automated information-extraction and normalization pipeline, DianShi-RxnDB transforms approximately 1.58 million patent documents from the USPTO and EPO, primarily covering 1976–2025, into approximately 24 million fine-grained, provenance-linked Reaction Instances , of which approximately 14.8 million (61.7%) pass automated qualification assessment. The instance-level data include reaction participants, experimental conditions, and procedures, with structured relationships among Substances , Reaction Instances , Reaction Groups , and References . These relationships support the organization and comparison of related single-step reaction records and provide a path back to source patents for verification. On this shared data foundation, the Web research workbench provides interactive search, comparison, and source verification for researchers, while the MCP service provides AI agents with structured, composable, multi-step retrieval capabilities.
Resource name
Reported scale
Literature coverage
Reaction and experimental fields
Provenance location
Access mode
USPTO-Lowe [ 21 , 27 ]
∼ 3.75 M
1976–2016
R⋅Q⋅C⋅P⋅Y
Document
Public download
CJHIF [ 19 , 18 ]
∼ 3.22 M
Not specified
R⋅Y
Not localized
Public repository
USPTO-LLM [ 37 , 36 , 13 ]
∼ 247,000
1976–2016
R⋅C⋅S
Document
Public download
Pistachio [ 26 , 23 , 24 ]
∼ 21 M
1971–present
R⋅Q⋅C⋅P⋅Y
Paragraph
Paid Web access
Reaxys [ 9 , 10 ]
∼ 73 M
1771–present
R⋅C⋅Y
Document
Paid Web; MCP/API license
CAS Reactions [ 1 , 2 ]
∼ 150 M
1840–present
R⋅Q⋅C⋅S⋅Y
Document
Paid Web access
Table 1: Representative organic reaction data resources and their system characteristics. The table compares the reported scale, original-literature coverage, reaction and experimental fields, provenance location, and access mode of each resource. Here, “reaction and experimental fields” refers to information captured at the reaction-record or reaction-instance level, including participant roles, quantities or scale, conditions, multistage support, structured procedures, yield, and other outcomes. The symbols indicate fields or capabilities documented in the cited materials and do not imply that every record contains every field. For the field notation, R denotes reaction participants and roles, Q quantities or scale, C reaction conditions, S support for multistage reactions, P a structured procedure mapped from the literature to predefined operations and parameters, Y yield, and O outcomes other than yield. For provenance location, “document” indicates an association with the original literature without confirmed paragraph-level localization, whereas “paragraph” indicates a paragraph-level location.
Processing stage
Input records
Removed records
Retained records
Source USPTO and EPO patent records
—
—
20,256,438
IPC-based filtering for organic chemistry relevance
20,256,438
17,743,513
2,512,925
Within-source consolidation and deduplication
2,512,925
544,695
1,968,230
Cross-source deduplication between USPTO and EPO
1,968,230
210,475
1,757,755
Removal of records without usable full text
1,757,755
176,816
1,580,939
Table 2: Processing stages from source patent records to the reaction-extraction corpus.
Available information such as quantities, equivalents, and product amounts
Reaction conditions
Available conditions such as temperature and time
Outcome information
Experimental results such as yield
Experimental process
Available experimental procedures and workup descriptions
Table 3: Principal information organized in a Reaction Instance .
Figure 2: DianShi-RxnDB transforms patent documents into provenance-linked Reaction Instances through a fully automated information-extraction and normalization pipeline. The pipeline parses patent text, images, and reaction schemes and, without manual extraction and curation of individual records, identifies organic synthesis experiments and organizes reaction participants, experimental conditions, yields, and experimental procedures into structured Reaction Instances ; chemical information in images and reaction schemes is primarily parsed by MinerU.Chem. Each instance retains a link to the relevant source-patent location to support source-text verification. Automated qualification then assigns each generated record a “qualified” or “not passed” status. Both categories are retained in the database with their extracted information and provenance links.
Figure 3: Conceptual relationships among Reaction Instances , Reaction Groups , and Reaction Templates . (A) A Reaction Instance represents a specific reported single-step reaction and independently retains its reaction conditions, yield, experimental procedure, and source information. A Reaction Group organizes multiple Reaction Instances using the same normalized reactant–product grouping key, without merging their experimental information. (B) A Reaction Template represents a local structural transformation pattern shared by different complete reaction instances, so instances with different specific reactant–product combinations may be associated with the same template. The local transformation shown in the figure is a simplified pattern for illustration and is not the SMARTS expression actually stored by the database.
Database object
Count
Counting convention
Source patent documents ( References )
608,309
USPTO or EPO patent documents linked to at least one retained Reaction Instance
Substances
6,261,797
Substance records organized by chemical identity in the database
Reaction Instances
23,999,236
Specific reported experiments extracted from patents and retained in the database
Reaction Groups
6,580,963
Groups formed from normalized reactant–product identity combinations
LocalRetro Reaction Templates
1,190,000
Deduplicated template collection generated using LocalRetro
RDChiral Reaction Templates
1,860,000
Deduplicated template collection generated using RDChiral
Table 4: Scale of core DianShi-RxnDB data objects.
Automated qualification status
Count
Share of all Reaction Instances
Qualified instances
14,808,205
61.7%
Not qualified instances
9,191,031
38.3%
Total
23,999,236
100%
Table 5: Automated qualification status of Reaction Instances .
Field
Correct judgments
Evaluated records
Field-level accuracy
Yield
1,233
1,300
94.85%
Reactant
1,226
1,300
94.31%
Reagent
1,100
1,300
84.62%
Catalyst
1,265
1,300
97.31%
Solvent
1,218
1,300
93.69%
Micro-average across five fields
6,042
6,500
92.95%
Table 6: Manual quality-evaluation results for five reaction fields in qualified Reaction Instances .
Dataset
Raw records
Records retained after deduplication
Duplicates removed
DianShi-RxnDB
4,222
4,093
129
Pistachio, text + image
3,630
2,992
638
Table 7: Reaction-record counts before and after deduplication in the matched sample of 100 US patents.
Selected differentiating feature
DianShi-RxnDB
Inspected Pistachio record
Separate Reagent and Catalyst roles
Yes
No ; broad Agent
Stage-level reaction object
Yes
No
Dedicated workup field
Yes
No
Explicit validation diagnostics
Yes
No
Table 8: Selected differentiating representation features observed for the shared reaction Intermediate 4 in patent US10975080. “Yes” indicates an explicitly represented feature; “No” indicates that no corresponding standalone field or object was exposed in the inspected record. This is not an exhaustive field inventory.
Field
Pistachio
DianShi-RxnDB
Gap
Reactant
623/660 (94.39%)
643/660 (97.42%)
+3.03%
Reagent
555/660 (84.09%)
576/660 (87.27%)
+3.18%
Catalyst
553/660 (83.79%)
569/660 (86.21%)
+2.42%
Solvent
615/660 (93.18%)
629/660 (95.30%)
+2.12%
Product
604/660 (91.52%)
609/660 (92.27%)
+0.75%
Yield
535/660 (81.06%)
656/660 (99.39%)
+18.33%
Table 9: Field-level exact agreement against source-grounded references in 660 source-paragraph-matched reaction pairs.
Figure 4: Structured Reaction Instance and source-patent comparison in the Web research workbench. The left panel shows experimental content from the source patent, while the right panel shows structured reaction information for the corresponding Reaction Instance , including reaction structures, participants, amounts, and experimental conditions. For records with available source content and location information, users can locate and compare the relevant source-patent paragraph or text region in the same interface to verify experimental procedures, participant roles, experimental conditions, and yields.
Instance
Reagent / catalyst
Solvent
Temp.
Time
Yield
Source patent
A
Sulfuric acid
Not recorded
50 ∘ C
20 min
70%
US10407376B2 [ 6 ]
B
SOMS
Dichloromethane
40 ∘ C
Not rec.
97%
US20220008887A1 [ 32 ]
C
Sulfated zirconia
Not recorded
120 ∘ C
30 min
92–95%
US20090082592A1 [ 17 ]
Table 10: Representative patent reaction instances for the formation of aspirin from salicylic acid and acetic anhydride.
Figure 5: Web use case from an aspirin target structure to patent reaction precedents and source verification. (A) The user enters or draws the aspirin structure, matches the corresponding Substance , restricts its participant role to Product, and enters the related reaction instances. (B) Among the 84 Reaction Instances associated with aspirin as a Product, the user locates the target Reaction Group in which salicylic acid and acetic anhydride form aspirin. This group contains 53 Reaction Instances from 10 source patents, with each instance’s experimental conditions, yield, and source information retained separately for comparison. (C) The user opens the source patent associated with Instance A and verifies the catalyst sulfuric acid, temperature of 50 ∘ C, and time of 20 minutes in the structured record.
Figure 6: From a natural-language task to multi-step MCP retrieval and example source inspection. (A) The AI agent calls the DianShi MCP according to the natural-language task: it retrieves 11 candidate Reaction Groups by the aspirin product structure, then locates 1 target Reaction Group from the salicylic-acid–acetic-anhydride reactant combination. The group contains 53 Reaction Instances , of which 20 are returned by the condition query in this call. (B) Based on the MCP results returned in this call, the agent organizes the target Reaction Group and selected representative Reaction Instances with their experimental conditions, yields, and source information. (C) Using Instance B as an example, the agent uses the reaction and source information returned by MCP to locate the corresponding record in the Web research workbench and inspect the source patent for SOMS, dichloromethane, 40 ∘ C, and 97% as reported experimental information.
Object or term
Definition
Primary use
Reference
A source patent document represented as a database object
Connects patent metadata and source content with Reaction Instances
Substance
A substance record organized by chemical identity in the database
Supports name, identifier, and structure retrieval; links to related reactions
Reaction Instance
Specific single-step reaction record extracted from a patent and retained in the database
Stores participant roles, conditions, yield, procedure, and provenance
Reaction Group
Groups Reaction Instances with the same normalized reactant–product identity
Organizes instances for one normalized reactant–product combination
Reaction Template
Abstract reaction representation using SMARTS
Provides an abstract representation beyond a specific reactant–product combination
Reaction Participant
A participation relationship between a Substance and a Reaction Instance
Records a Substance role as Reactant, Product, Reagent, Solvent, or Catalyst
Table A1 : Core DianShi-RxnDB objects and terminology.
Field
Description
Example
Document name
Patent name or title
Therapeutic agent for Parkinson disease
Document type
Type of source document
PATENT
Document number
Patent identifier
US054849209-19960116
Authors or organization
Patent inventors or organization
Suzuki; Fumio; Shimada; Junichi; Koike; Nobuaki et al.
Table B1: Principal Reference fields.
Field or group
Description
Example ( DianShi-RxnDB Web record )
In-document number
Subheading at the provenance location of the reaction in the patent
Table B3: Principal Substance fields and association statistics.
Operation
Definition
Add
Add a substance to the reactor
Centrifuge
Centrifuge the mixture to separate phases or collect solids
CollectLayer
Select aqueous or organic fraction(s)
Column
Execute column chromatography using adsorbent and eluent
Combine
Combine multiple batches, fractions, or extracts
Concentrate
Evaporate the solvent (e.g., using a rotary evaporator)
Table B4: Operational definitions for reaction-process details.
Raw JSON field
Text channel
Image channel
title
US07795244B2_0103
US07795244_C00043
data.paragraphText
In step 5, in a suitable reaction flask, sodium iodide (23.9 g, 159.45 mmol, 1.6 eq.) was dissolved in acetone (96 mL). 2-Ethoxy-isobutyric acid chloromethyl ester (18 g, 99.65 mmol, 1 eq.) was then added ...
Field absent
Deduplication result
Same reactant and product sets after removing atom-map numbers and sorting components
Same reactant and product sets after removing atom-map numbers and sorting components
Table D1 : Raw record identifiers, source text, and deduplication result for the text- and image-channel records of the same patent example in US07795244.
Field in this reaction
DianShi-RxnDB
Pistachio
Source pointer
Intermediate 4 ; source line 1306; full paragraph
US10975080B2_0290 ; paragraphText
Role labels
Reactant; Reagent; Solvent; Product
Reactant; broad Agent ; Solvent; Product
Catalyst field
Explicitly recorded and empty for this reaction
— No separate Catalyst field
Reaction stage
stage_1 groups substances, conditions, actions, and workup
— No stage-level object
Temperature
Stage field: 0 °C
Action parameter: Temperature = 0 °C
Time
Stage field: 2 h
Action parameter: Time = 2 h
Table D2 : Concrete field-level evidence for the shared reaction Intermediate 4 in patent US10975080.
Computer-assisted synthesis planning breaks target molecules into accessible precursors using large libraries of reaction rules that assign each transformation a deterministic, interpretable label. But chemistry is long-tailed, making manual encoding intractable, and existing tools rely on fixed rulesets that cannot adapt to new chemistries. Here we present a fully automated pipeline in which a multi-agent framework of large language models (LLMs) classifies reactions and writes the rules themselves across 665,901 US patent reactions, generating each rule under a verification loop that tests it against the corpus. It expands a standard taxonomy from 68 to 14,073 classes without human curation. With a lightweight fingerprint classifier, it classifies 97.7% of unseen reactions, matching a leading proprietary classifier while resolving chemistry more finely and extending on demand to chemistry outside its training distribution. The result is a living reactivity database and a general route to turning generative models into reliable, self-expanding symbolic systems.
Daniel Armstrong, Maarten Dobbelaere, Valentas Olikauskas +4
1École Polytechnique Fédérale de Lausanne (EPFL), Switzerland · 2Ghent University, Belgium · 3National Centre of Competence in Research (NCCR) Catalysis, Switzerland
Chemical reaction datasets such as USPTO suffer from substantial incompleteness, frequently missing byproducts, co-reactants, and stoichiometric coefficients. This limits their applicability and reliability in downstream applications. Here, we introduce CompleteRXN, a large-scale supervised benchmark for reaction completion under realistic missing-data conditions. We construct a dataset of aligned incomplete and atom-balanced reactions by mapping USPTO records to curated mechanistic reactions. We evaluate representative baselines, including a novel encoder-decoder reaction completion model with constrained decoding, the Constrained Reaction Balancer (CRB), and a recent algorithmic method, SynRBL. On our CompleteRXN benchmark, the CRB achieves high performance across splits of increasing difficulty, reaching 99.20% equivalence accuracy on the random split and 91.12% on the extreme out-of-distribution split. SynRBL produces many balanced and chemically plausible completions, but with lower accuracy on the benchmark test splits. Across all methods, performance degrades with increasing incompleteness. We observe a substantial drop when evaluating on reactions outside the benchmark (full uncurated USPTO), highlighting the gap between benchmark performance and practical robustness and motivating future work.
Gabriel Vogel, Minouk Noordsij, Evgeny Pidko +1
Department of Intelligent Systems Delft University of Technology Delft, 2629 HZ, The Netherlands · Department of Chemical Engineering Delft University of Technology Delft, 2629 HZ, The Netherlands
Chemistry literature synthesis often requires assembling specific findings scattered across many publications, yet existing literature-search systems primarily return ranked document lists. As a result, scientists and AI agents need to locate relevant information, verify their provenance, and assemble cross-paper answers manually. We present AskChem, a claim-centered infrastructure for cross-paper chemistry search. AskChem changes the unit of retrieval from the paper to the provenance-carrying claim: each paper is converted into atomic, typed claims, each grounded by a source DOI and a verbatim quote or an explicit evidence locator. Over this shared claim store, AskChem exposes complementary structures for search and synthesis: a stabilized faceted taxonomy for hierarchical retrieval and browsing, an evidence graph linking claims through relations, and an exploratory living taxonomy that situates indexed papers under scientific principles. AskChem currently indexes 2.4M claims from 147K papers and provides a web interface, as well as REST, SDK, and MCP access for AI agents. On AskChem-Bench, grounding a GPT-5.5 reader in AskChem yields 100% resolvable DOIs, compared with 88.3% without retrieval, and the highest citation density among five tested systems. AskChem is live at https://askchem.org.