concept · intermediate · partly written
How a reaction is written for a computer: reaction SMILES, atom maps and templates
A reaction is stored as text, an atom map says which atom goes where, and a template generalises the reaction centre; each loses something.
Also called: reaction SMILES, atom-to-atom mapping, reaction templates, RXNMapper
What can each way of writing a reaction not say?
Reaction SMILES omits reagents and by-products that were not listed. Atom maps are imperfect even when made by good tools. Templates lose context beyond the reaction centre. Condition and yield models barely beat guessing the average. 1234
Reported, not yet replicated
What does reaction SMILES leave out? more
Reaction SMILES (SMILES is the Simplified Molecular-Input Line-Entry System) writes reactants, reagents and products as one line of text. The Molecular Transformer authors point to "an unlisted source of hydroxide" as the kind of species a patent record can omit. Patent data also lack reagent, solvent and temperature context and can include low-yield reactions. 15
Reported, not yet replicated
How accurate are atom maps? more
RXNMapper agreed with the reference maps on 96.8% of 49,000 reactions. Of 1,551 mismatches the authors judged 284 to be RXNMapper errors, 415 equivalent, 436 better than the reference, 369 questionable and 47 lacking key reagents, and after that review the agreement rose to 99.4%. The authors made the judgements by hand, so the reference itself was in doubt for about half of the mismatches. 2
Reported, not yet replicated
What does a template keep, and how much do sources share? more
A template is the generalised reaction centre of a reaction, extracted from examples and reapplied to new molecules. Only about 2% of templates were common to USPTO, Reaxys, Pistachio and an AstraZeneca notebook subset. RetroRules, a template-style rule set for enzymes, gives rules at several diameters; its abstract speaks of over 400,000 rules while its Table 1 never exceeds about 64,453 unique ones, and it reports no accuracy. 36
Reported, not yet replicated
Can models predict conditions and yield better than a simple guess? more
In one study, solvent prediction reached top-1 of about 35 to 46% (depending on the table) against a 29.2% popularity baseline. Yield prediction had a mean absolute error of 15.6% against 16.3% for always guessing the mean. Whether failed reactions are missing and whether adding them would help was not measured in this source. 4
Reported, not yet replicated
Plain answer
Why it matters
Still open here
- Does the reaction ontology RXNO cover the classes enzymologists use? Not covered by any source opened.
- What is the noise in a repeated literature yield? Only qualitative descriptions were found.
Not yet written: Equation, Real numbers.
Linked from
- How a reaction gets from a patent paragraph into a database record, and what goes wrong
- Reaction data: what is public, what is proprietary, and what each holds
- Reaction-prediction benchmarks: what a headline accuracy measures, and how splits inflate it
- How noisy is a reaction yield between repeats, and does adding failed reactions help prediction models?
Sources
- Philippe Schwaller, et al.. Molecular Transformer: A Model for Uncertainty-Calibrated Chemical Reaction Prediction. ACS Central Science (2019). paper; opened 2026-11-17 all records citing this
- Philippe Schwaller, et al.. Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Science Advances (2021). paper; opened 2026-11-17 all records citing this
- Amol Thakkar, et al.. Datasets and their influence on the development of computer assisted synthesis planning tools in the pharmaceutical domain. Chemical Science (2020). paper; opened 2026-11-17 all records citing this
- Beker et al.. Machine learning may sometimes simply capture literature popularity trends [title not copied from the page]. journal not copied (2022). 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
- Thomas Duigou, et al.. RetroRules: a database of reaction rules for engineering biology. Nucleic Acids Research (2019). 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/reaction-smiles-atom-mapping-and-templates.json