Chemistry line notations — from 1949 to today

Wiswesser Line Notation. William Wiswesser was an industrial chemist who needed to write molecule structures on a typewriter. WLN encoded fused ring systems with letter-and-digit qualifiers (T6 for a six-membered alicyclic ring, L for carbocyclic, M for nitrogen) and a deterministic citation order with multiple priority rules for selecting the start atom and traversal direction.¹ A 1970s patent abstract may still contain a string like L66J BVQ DV1NN1. The most accurate modern parser² recovers about 75% of WLN strings in real archives; the residual failure rate is dominated by malformed source strings.

What it adds: typewriter-era graph encoding; the first compact line representation that survived as a real-world index for forty years.

CAS Registry Number, Chemical Abstracts Service. Three groups of digits separated by hyphens; the rightmost digit is a checksum.

58-08-2

What it adds: institutional memory. Reagent bottles, safety data sheets, regulatory filings, and IUPAC's own indexing all reference CAS numbers. They are not derivable from structure — every new molecule receives a number assigned by humans at CAS.

The IUPAC Nomenclature of Organic Chemistry is the most thorough molecular notation that exists. The current rulebook (the 2013 'Blue Book') runs over 1500 pages.

1,3,7-trimethyl-3,7-dihydro-1H-purine-2,6-dione

Generating a name from a structure is computationally hard and rule-dependent. ACD/Labs, ChemAxon, and OpenEye all sell commercial IUPAC name generators because the open-source ones do not compete. The widely-used free option is STOUT (Rajan et al., 2021)³ — originally an LSTM trained on PubChem name/structure pairs. The 2021 paper reports an average BLEU of about 90% and Tanimoto similarity above 0.9. STOUT V2.0 (2024)⁴ is the transformer version, trained on near-1B SMILES/IUPAC pairs.

What it adds: a name a human can read aloud and a referee can verify by hand.

Simplified Molecular-Input Line-Entry System. David Weininger, 1986–1988, at the U.S. EPA in Duluth.⁵ The brief was small: a notation a chemist could type in a single line of ASCII, that round-tripped through a parser, and that any modest computer of the day could read.

CN1C=NC2=C1C(=O)N(C(=O)N2C)C

SMILES describes a depth-first traversal of the molecular graph. Atoms are single letters (or two-letter elements in brackets); bonds are inferred or annotated (=, #, :); ring closures are matched digits; branches are parenthesised; stereo is annotated with @, /, \.

What it adds: a grammar simple enough to teach in an afternoon, and a graph-shaped semantics that maps directly onto the data structures of every modern toolkit.

SMILES is not unique. Caffeine can be written equally validly as CN1C(=O)N(C)c2ncn(C)c2C1=O — same molecule, different traversal start, totally different string. To compare two SMILES strings for chemical identity you have to canonicalise both first. The Morgan algorithm⁶ ranks atoms by extended connectivity and gives a deterministic traversal order. Every toolkit ships its own implementation. They mostly agree; they do not always agree.^7 8

RDKitCn1c(=O)c2c(ncn2C)n(C)c1=O
OpenEye / PubChemCN1C=NC2=C1C(=O)N(C(=O)N2C)C
Open BabelCn1cnc2c1c(=O)n(C)c(=O)n2C
IndigoCN1C(=O)C2=C(N=CN2C)N(C)C1=O

OpenSMILES⁹ sits alongside this work: a community effort started by Craig James and Andrew Dalke to document what every implementation actually does, so a parser written today matches a parser written in 1995.

What it adds: a deterministic dedup key inside a single toolkit; a path toward cross-toolkit agreement once all four canonicalisers converge.

IUPAC International Chemical Identifier. Standardised in 2005,¹² after roughly a decade of NIST/IUPAC work. The brief was the inverse of SMILES: a canonical identifier that two implementations anywhere in the world will compute identically for the same molecule.

InChI=1S/C8H10N4O2/c1-10-4-9-6-5(10)7(13)12(3)8(14)11(6)2/h4H,1-3H3

The layers are explicit: formula (C8H10N4O2), connectivity (c1-10-…), hydrogen positions (h4H,…), then optional stereo, isotope, charge, and tautomer layers. Two molecules are identical if and only if their standard InChIs are byte-equal.

InChI strings get long fast. The InChIKey is a 27-character SHA-256-based hash of the full InChI: fixed-length, URL-safe. The first 14 characters are a hash of the connectivity layer; the next 10 contain a hash of the remaining layers plus a version-and-protonation flag; the final character indicates protonation state.¹³ No real collision has been documented in the wild.¹⁴

RYYVLZVUVIJVGH-UHFFFAOYSA-N

What they add: a single agreed-on canonicalisation algorithm — implemented once, in C, by the InChI Trust — and a short hash that fits in a database column.

Noel O'Boyle and Andrew Dalke, 2018.¹⁷ The minimum viable fix for SMILES' machine-learning problems: drop matched parentheses (use a single ) per branch close), drop matched ring digits. Same length as SMILES, only one class of mutation breaks validity instead of three.

Cnc=O)ccncn5C))))nC)c6=O

What it adds: a near-zero-cost retrofit for pipelines that already speak SMILES.

Self-Referencing Embedded Strings. Krenn, Häse, Nigam, Friederich, Aspuru-Guzik, 2020.¹⁹ Born of a real ML problem: SMILES strings break under mutation. If a generative model trained on SMILES outputs c1ccccc1C(, that is a syntax error — unbalanced parens, no valid molecule. Roughly 99% of random SMILES mutations produce invalid output. SELFIES fix this by design: every syntactically valid SELFIES string maps to a valid molecule.

[C][N][C][=Branch1][C][=O][C][=C][Branch1][#Branch1][N][=C][N][Ring1][Branch1][C][N][Branch1][C][C][C][Ring1][N][=O]

What it adds: a closed grammar where the parser, not a downstream filter, enforces validity.

Balsa: A Compact Line Notation Based on SMILES. Richard L. Apodaca (Metamolecular, LLC), 2022.²⁰ Balsa keeps SMILES' character set and looks identical for most molecules — ethanol is CCO in both, aspirin is CC(=O)Oc1ccccc1C(=O)O. The break is in the grammar, not the strings: every place where SMILES is informal, Balsa is rigid.

Three concrete differences, lifted from the Rust reference grammar at github.com/metamolecular/balsa:

• No aromatic bond character. SMILES has six bond tokens — - = # : / \ — including : for aromatic. Balsa's BondKind enum has only five — Elided, Single, Double, Triple, Up, Down. Aromaticity is carried by the lowercase atom (Selection), never by the bond. SMILES allows benzene as either c1ccccc1 or C1=CC=CC=C1; Balsa specifies only the former.
• Organic-subset shortcuts are an explicit grammar rule. SMILES "if it's in the organic subset, you can drop the brackets" lives in prose. Balsa's AtomKind::Shortcut is a closed enum: B, C, N, O, F, P, S, Cl, Br, I. Anything else must use Bracket notation. A parser cannot accept [Mg] and Mg as equivalent input.
• Stereo is per-atom parity, not bond-direction soup. SMILES tetrahedral chirality uses @ / @@ on the atom plus implicit ordering of its neighbours; double-bond stereo uses / and \ on the surrounding bonds and depends on the parser's traversal order. Balsa's AtomParity is an explicit Clockwise / Counterclockwise token on the stereocentre's bracket atom; the canonical-order conventions are written into the grammar instead of the implementation.

What it adds: a parseable spec an implementer can target without reverse-engineering reference parsers. The strings look like SMILES; the difference is that two implementations targeting Balsa cannot disagree about what they decoded.

Cheng et al., 2023.²¹ Standard SELFIES tokenises by atom. Group SELFIES tokenises by chemically meaningful fragment — phenyl, amide, methyl ester, and so on. The vocabulary is larger, the strings shorter, and the model learns in chemistry-shaped chunks instead of one atom at a time.

What it adds: a shorter sequence and a more drug-like distribution out of the box, at the cost of fixing a fragment alphabet up front.

Sequential Attachment-based Fragment Embedding. Noutahi et al., 2024.²² SAFE rewrites a SMILES into a sequence of disconnected fragments joined by attachment-point indices. The fragments are themselves valid SMILES; the joins are explicit.

What it adds: fragment-conditional generation (keep this scaffold, regrow the side chain) without leaving the SMILES grammar — a SAFE string is also a valid SMILES once the attachment indices are resolved.