Reactions and retrosynthesis — from 1985 to today

MDL's reaction file format. Header, then N reactant Mol blocks, then M product Mol blocks. ASCII; same V2000 / V3000 dialect as the molecule Mol files. Atom-mapping numbers in the reactant + product blocks pair atoms across the arrow (the same convention SMIRKS uses inline).

What it adds: reaction direction, atom mapping, and stoichiometry, with the full Mol-file detail per side. See Cheminformatics file formats § for the structural-format side of this overlap.

SMILES with a single new character: >. Reactants on the left, optional reagents in the middle, products on the right (reactants>reagents>products). Inherits everything from SMILES — same atom-and-bond grammar, same canonicalisation problem — and adds reaction direction. Atom-mapping numbers are optional in vanilla Reaction SMILES; SMIRKS makes them mandatory.

OC(=O)C1=CC=CC=C1.OCC>>O=C(OCC)C1=CC=CC=C1

What it adds: a single-line container for reactions that every SMILES-aware tool reads for free.

Daylight published the original specification as a companion to SMILES.¹ It is to SMILES what regular expressions are to strings: the parent grammar extended with wildcards, logical operators, and topological constraints. Lives in this article (rather than the line-notations one) because the primary use of SMARTS is filtering and substructure search across reaction templates, not describing a specific molecule.

c matches any aromatic carbon. [#6] matches any carbon. [F,Cl,Br,I] matches any halogen. [#6]=[O] matches a carbonyl. Bonds: ~ is any bond, @ is any ring bond, !@ excludes ring bonds. Recursion is supported via $().

[#16](=[#8])(=[#8])[#7]

What it adds: substructure search, filter rules, and the atom-half of every reaction template.

SMARTS plus a reaction arrow plus mandatory atom-mapping numbers.² Reads like a chemistry rule: "a carboxylic acid plus an alcohol gives an ester, and the carbonyl carbon is the same atom on both sides". Every reaction-template-driven tool — ASKCOS, RDKit's RunReactants, ChemAxon's reactor, RetroPath — speaks SMIRKS as input.

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

What it adds: a parseable, executable, vendor-neutral way to describe "this kind of reaction" — the unit retrosynthesis engines compose into multi-step plans.

Daniel Lowe's PhD thesis (Cambridge, 2012) extracted ~1.8M reactions from U.S. patent text using rule-based parsing. The corpus has been re-released, extended, and re-cleaned multiple times since;³ the most-cited public dump is the 1976–2016 grants + applications set (~3.7M reactions). This is the dataset most retrosynthesis models train on.

What it adds: a public, parseable, multi-million-reaction corpus that earlier reaction databases (Reaxys, CASRN reactions) did not permit redistributing. Most of modern ML retrosynthesis stands on it.

Reaction InChI. Grethe, Goodman, Allen, 2013.⁴ A canonicalised, layered identifier that hashes a complete reaction (reactants, products, agents) into a single string. The companion RInChIKey is the fixed-length URL-safe hash, the way InChIKey is for InChI. Built on the InChI Trust's reference C library.

What it adds: a deterministic identifier for "is this the same reaction?" queries across implementations. Reaction SMILES doesn't canonicalise; RInChI does.

MIT's open-source retrosynthesis platform.⁵ Combines a template-based retrosynthesis engine (templates extracted from USPTO via SMIRKS), a forward-prediction model (graph neural net on USPTO), and condition / solvent / catalyst recommendations. Web service at askcos.mit.edu; source on GitHub.

What it adds: the first end-to-end open-source pipeline from "target molecule" to "step-by-step synthetic route plus conditions".

IBM Research's transformer-based reaction prediction service.⁶ Treats reaction prediction as a sequence-to-sequence translation problem on Reaction SMILES — encoder reads reactants, decoder emits products. Trained originally on USPTO; the underlying RXN-Transformer architecture has been adapted for retrosynthesis, condition prediction, and yield estimation.

What it adds: end-to-end seq2seq prediction without explicit reaction templates — directly from atoms to atoms. The bench-mark every template-based system gets compared against.

AstraZeneca's open-source Monte Carlo Tree Search retrosynthesis tool.⁷ Template-based with a neural-network policy that ranks template applicability. Designed to be vendor-friendly: ships with USPTO-trained models, but the policy and stock-molecule lookup are pluggable so internal compound libraries can be substituted in.

What it adds: a production-shaped retrosynthesis engine intended to be deployed inside a pharma stack with private stocks and private templates.

Open Reaction Database (ORD).⁸ Cross-industry effort (initiated by Pfizer with Mike Burke, Connor Coley, and others). Defines a structured Protocol Buffer schema for reactions: every reagent has structure plus role plus quantity plus solvent context, every step has temperature/time/atmosphere, every outcome has yield + purity + analytics. Public data on GitHub; reading clients in Python, JavaScript, and Java.

What it adds: a structured schema for reaction reporting that captures the metadata Reaction SMILES and RXN files leave implicit. Most useful for ML: ORD records yield + condition data the USPTO corpus lacks.