concept · intermediate · partly written
How a reaction gets from a patent paragraph into a database record, and what goes wrong
A patent paragraph becomes a reaction record through a chain of guesses (find, tag, convert names, assign roles, map atoms), and only pieces of that chain have been measured.
Also called: patent text mining of reactions, Lowe patent extraction, USPTO extraction pipeline
How much can a patent-derived reaction record be trusted?
Treat these records as text-derived, not bench truth. The original pipeline's author checked 100 reactions and found 22% flawless. The atom mapper behind the public set got 38.47% of a hand-mapped benchmark right. Failed reactions are probably near absent. 12
Reported, not yet replicated
What does the extraction pipeline do, step by step? more
Daniel Lowe's 2012 thesis text-mined US patents because no open reaction database existed. The steps were to find the experimental section, tag chemical entities, convert names to structures with a parser, assign roles to each substance, and atom-map the result. The thesis set holds 424,621 reactions from 65,034 patents. 1
Reported, not yet replicated
What was measured about the original pipeline? more
On a sample of 100 reactions the author judged entity recall 96.4%, precision 88.9%, role assignment 91.8% correct, and quantities tied to the right product 64.9% of the time. Catalyst roles were right 41.7% of the time and 22% of reactions were flawless, though the product and main starting material were right in about 95%. These are the author's own judgements on a small sample, and whether they carry to the 3.3 million release is not shown. 1
Reported, not yet replicated
How good is the atom mapper behind the public set? more
On a hand-mapped benchmark of 1,851 reactions, Indigo, the mapper behind the public USPTO set, was right on 712 (38.47%), while RXNMapper was right on 83.74%. The paper also cites earlier work reporting erroneous maps in 89% of carboxyl-to-alcohol reductions and 31% of methyl ester formations on USPTO; it did not run those counts itself. 2
Reported, not yet replicated
What do the records leave out? more
In the Reaxys sample of 15.4 million filtered reactions, 54% had no yield, 53.9% no temperature and 98.4% no pressure. Patent data may include low-yield reactions, and failed reactions are probably near absent from the records, which is an inference from how patents are written and not a measured rate. 34
Our inference
Plain answer
Software finds the experimental paragraph in a patent, tags the chemicals, turns each name into a structure, guesses each one's role, and maps product atoms to reactant atoms. Each step can be wrong, and the only end-to-end test of the original pipeline was run by its author on 100 reactions. 1
Reported, not yet replicated
Still open here
- What is the end-to-end error rate of the public USPTO set as shipped? No audit was found.
- How often do the public records carry temperature and time? Not counted in any source opened.
Not yet written: Why it matters, Equation, Real numbers.
Linked from
Sources
- Daniel Lowe. Extraction of chemical structures and reactions from the literature. University of Cambridge (PhD thesis) (2012). paper, not peer reviewed; opened 2026-11-17 all records citing this
- Lin et al.. Atom-to-atom mapping benchmark on a hand-mapped set of 1,851 reactions [title not copied from the page]. Molecular Informatics (2022). paper; opened 2026-11-17 all records citing this
- Jacob et al.. Analysis of Reaxys reaction records for missing conditions [title not copied from the page]. Journal of Cheminformatics (2017). paper; opened 2026-11-17 all records citing this
- Connor W. Coley, et al.. Computer-assisted retrosynthesis based on molecular similarity. ACS Central Science (2017). paper; opened 2026-11-17 all records citing this
Where it sits on the map
History of this record (1)
This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.
- Status
- partly written
- Last updated
- 2026-11-17
- Record
- content/concepts/patent-reaction-extraction-and-its-errors.json