Reactions and retrosynthesis from 1985 to today

A reaction record lists reactants, products and the conditions between them. Its formats build on those for molecules, and pattern languages, datasets and retrosynthesis programs build on the formats. This post follows one reaction through all of them: the Fischer esterification of benzoic acid with ethanol.

A short history#

RXN file, 1985#

MDL's RXN file is a header, a line with the number of reactants and products, and one molfile per component, each after $MOL.¹ Atom-map numbers in the atom lines pair atoms across the arrow, and the RD file collects many reactions with data. Molfiles are described in Cheminformatics file formats.

$RXN

      RDKit

  2  2
$MOL

     RDKit          2D

  9  9  0  0  0  0  0  0  0  0999 V2000
    2.5833   -1.2990    0.0000 O   0  0  0  0  0  0  0  0  0  0  0  0

Reaction SMILES, 1988#

Reaction SMILES separates reactants, agents and products with >:

OC(=O)c1ccccc1.OCC>OS(=O)(=O)O>CCOC(=O)c1ccccc1.O

Sulfuric acid, the catalyst, is an agent, and water is a product. As with SMILES, one reaction has many valid strings and no canonical form shared between toolkits.

SMARTS, 1990s#

Daylight's SMARTS turns SMILES into a pattern language.² [#6] matches any carbon, [F,Cl,Br,I] any halogen and ~ any bond. The carboxylic acid pattern [CX3](=O)[OX2H1] matches benzoic acid but not ethyl benzoate. Substructure searches, structural alerts and reaction templates are written in SMARTS.

SMIRKS, 1990s#

SMIRKS writes a transformation with the same mapped atoms on both sides.³ This template covers the esterification of a carboxylic acid with an alcohol:

[C:1](=[O:2])[OH:3].[OH:4][C:5]>>[C:1](=[O:2])[O:4][C:5].[OH2:3]

Oxygen 3 leaves as water and oxygen 4 becomes the ester oxygen. Applied with RDKit to benzoic acid and ethanol, the template gives ethyl benzoate and water. Retrosynthesis programs run such templates in reverse.

USPTO reactions, 2012#

Daniel Lowe's PhD thesis described how to extract reactions from text.⁴ His data release covers US patents from 1976 to 2016 as CML and reaction SMILES, with atom maps that its own notes call wrong in many cases.⁵ Commercial databases cannot be redistributed, so most published retrosynthesis models are trained on this set.

RInChI, 2013#

The reaction InChI gives a whole reaction one canonical identifier, which reaction SMILES cannot, and the RInChIKey is its hash.⁶

ASKCOS, 2017#

MIT's open-source ASKCOS proposes routes from reaction templates, checks each step with a forward-prediction model and suggests conditions. Its planning informed a robotic flow-synthesis platform.⁷

IBM RXN, 2018#

IBM RXN's Molecular Transformer treats reaction prediction as translation from reactant SMILES to product SMILES, without templates. It was trained on USPTO reactions.⁸

AiZynthFinder, 2020#

AstraZeneca's AiZynthFinder runs a Monte Carlo tree search in which a neural network ranks the templates; a route is complete when every precursor is in stock. A company can replace the templates, the network and the stock list with its own.⁹

Open Reaction Database, 2021#

The Open Reaction Database defines a schema, in Protocol Buffers, for complete reaction reports: inputs with roles and amounts, conditions, and outcomes with yields and analytical data.¹⁰ That is what the USPTO set lacks and what models need to predict conditions.

Side by side#

On this site#

The reaction pages list named reactions and their participants, and Reaction scale turns a reaction's stoichiometry and one fixed amount into masses, volumes and moles. Notations, molfiles and spectra are covered in Chemistry line notations, Cheminformatics file formats and Spectroscopy file formats.

References#

1. Dalby A et al. Description of several chemical structure file formats used by computer programs developed at Molecular Design Limited. J. Chem. Inf. Comput. Sci. 32, 244–255 (1992). doi:10.1021/ci00007a012
2. Daylight Chemical Information Systems. SMARTS: a language for describing molecular patterns. daylight.com
3. Daylight Chemical Information Systems. SMIRKS: a reaction transform language. daylight.com
4. Lowe DM. Extraction of chemical structures and reactions from the literature. PhD thesis, University of Cambridge (2012). doi:10.17863/CAM.16293
5. Lowe DM. Chemical reactions from US patents (1976–Sep2016). figshare (2017). doi:10.6084/m9.figshare.5104873.v1
6. Grethe G, Goodman JM, Allen CHG. International chemical identifier for reactions (RInChI). J. Cheminform. 5, 45 (2013). doi:10.1186/1758-2946-5-45
7. Coley CW et al. A robotic platform for flow synthesis of organic compounds informed by AI planning. Science 365, eaax1566 (2019). doi:10.1126/science.aax1566
8. Schwaller P et al. Molecular Transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Cent. Sci. 5, 1572–1583 (2019). doi:10.1021/acscentsci.9b00576
9. Genheden S et al. AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. J. Cheminform. 12, 70 (2020). doi:10.1186/s13321-020-00472-1
10. Kearnes SM et al. The Open Reaction Database. J. Am. Chem. Soc. 143, 18820–18826 (2021). doi:10.1021/jacs.1c09820