The word “SMILES” shows up in two completely unrelated corners of chemistry, and both are worth knowing about. One is a text-based language for writing molecular structures as short strings of characters, widely used in databases, drug discovery, and artificial intelligence. The other is a century-old organic chemistry reaction named after the British chemist Samuel Smiles. They share nothing except the name, yet both have become increasingly important in modern research for very different reasons.
SMILES Notation and Why Chemists Write Molecules as Text
SMILES stands for Simplified Molecular-Input Line-Entry System. Developed in the late 1980s by David Weininger, it gives chemists a way to represent a molecule’s structure as a compact line of ordinary keyboard characters. Water becomes O. Ethanol becomes CCO. Caffeine becomes Cn1cnc2c1c(=O)n(c(=O)n2C)C. The idea is that any molecule, no matter how complicated, can be turned into a string of letters, numbers, and a few punctuation marks that a computer can store, search, and compare far more efficiently than a drawn structural diagram.
The notation follows a few basic conventions. Atoms are represented by their element symbols, with uppercase letters for non-aromatic atoms and lowercase for atoms in aromatic rings. Carbon-carbon single bonds are implied between adjacent symbols, so you only need to mark double bonds (=) and triple bonds (#). Branches are shown in parentheses, and rings are indicated by pairing digits at the atoms where the ring opens and closes. A benzene ring, for instance, can be written as c1ccccc1. It reads like gibberish at first, but experienced chemists become fluent surprisingly quickly, and the real audience is software anyway.
What made SMILES take off was its simplicity compared to alternatives. Earlier molecular file formats like connection tables were bulky and hard to search. A SMILES string for a typical drug molecule fits in a tweet. That compactness turned out to matter enormously once chemical databases began scaling into the millions and then billions of structures. Today, virtually every major chemical database, from PubChem to ChEMBL, stores SMILES strings alongside their other representations.
The Canonicalization Problem
One wrinkle with SMILES is that the same molecule can be written in multiple valid ways depending on which atom you start from. Ethanol can be CCO or OCC. For a small molecule that is easy to sort out, but for a large database, you need every molecule to have exactly one “official” SMILES string so that searches reliably find duplicates. This is the canonicalization problem, and it has kept cheminformatics researchers busy for decades.
Weininger’s original CANON algorithm tackled this by assigning priorities to atoms using a connectivity-based method, then generating the unique string from a fixed starting point. But the algorithm had ambiguities, particularly with highly symmetrical molecules. Researchers have proposed fixes, including new atomic ring invariants that distinguish symmetry classes more reliably. One approach uses the product of prime numbers corresponding to each atom’s ring sizes as an unambiguous invariant, combined with modifications to the CANON algorithm that correct issues like arbitrary tie-breaking and incorrect handling of certain ring systems.1PubMed Central. Atomic ring invariant and Modified CANON extended connectivity algorithm for symmetry perception in molecular graphs and rigorous canonicalization of SMILES
Another approach, sometimes called Universal SMILES, uses the InChI (International Chemical Identifier) as a reference point for generating canonical SMILES. When tested against over a million compounds from the ChEMBL database, a shuffle test found that the InChI-based method had only 141 canonicalization failures, compared to roughly 2,400 for the standard Universal SMILES method. Results on a similar-sized subset of PubChem were nearly identical: 163 failures versus about 2,400.2PubMed Central. Towards a Universal SMILES representation – A standard method to generate canonical SMILES based on the InChI Those failure rates sound tiny, but when you are managing databases with hundreds of millions of entries, even a fraction of a percent can mean thousands of misidentified compounds.
SMARTS, SMIRKS, and BigSMILES
SMILES spawned a family of related languages. SMARTS (SMILES Arbitrary Target Specification) extends the notation to define patterns rather than specific molecules. Where a SMILES string says “this is ethanol,” a SMARTS pattern might say “find me anything with a hydroxyl group attached to a carbon chain.” That makes SMARTS essential for substructure searching, where a researcher wants every molecule in a database that contains a particular fragment.
SMIRKS takes the concept a step further by describing chemical reactions. It specifies a transformation: which bonds break, which form, and how atoms rearrange. Open-source tools have been built around SMIRKS to enable automated reaction searches and structure transformations in software.3PubMed Central. Ambit-SMIRKS: a software module for reaction representation, reaction search and structure transformation If SMILES is a dictionary entry, SMIRKS is a sentence describing a change from one state to another.
Standard SMILES was designed for small, precisely defined molecules. Polymers are a different beast: they consist of repeating units whose exact length, branching, and sequence vary from chain to chain. A single polymer sample is really a statistical mixture of slightly different molecules. BigSMILES was proposed as an extension that can capture this randomness. Rather than specifying one exact structure, BigSMILES describes the rules by which a polymer’s repeating units connect, making it possible to represent and index polymers in databases using a notation that respects their inherently stochastic nature. As a proof of concept, a standard dataset of polymer glass transition temperatures was converted into BigSMILES notation.4PubMed Central. BigSMILES: A Structurally-Based Line Notation for Describing Macromolecules
SMILES and Artificial Intelligence
The real explosion of interest in SMILES over the past decade has been driven by machine learning. Because SMILES strings are sequences of characters, they look a lot like sentences in a natural language. And the AI techniques that revolutionized language translation and text generation turn out to work remarkably well on molecular strings too.
The core idea is to treat SMILES like text and let a neural network learn molecular “grammar.” One influential approach, the SMILES Transformer, borrows directly from the Transformer architecture used in language models. It learns molecular fingerprints through unsupervised pretraining on large collections of SMILES strings, then applies those learned representations to predict molecular properties or biological activities with relatively little task-specific data.5PubMed Central. Application of Transformers in Cheminformatics Another strategy, SMILES-BERT, adapts the masked-language-model approach from text AI, treating each atom symbol as a “word” and each SMILES string as a “sentence,” then fine-tuning the pretrained model for specific prediction tasks.6PubMed Central. Application of Transformers in Cheminformatics A comprehensive review of these Transformer-based chemical language models found that the field is now large enough and varied enough to warrant its own dedicated survey literature.7PubMed Central. Transformer-based models for chemical SMILES representation: A comprehensive literature review
These approaches have practical payoff. In one example, a modeling framework that used SMILES-based molecular embeddings alongside protein representations achieved a mean correlation of about 0.92 when predicting how strongly compounds bind to viral proteins. The researchers applied the framework to identify a ranked list of 47 compounds targeting three key proteins of the SARS-CoV-2 virus, including antivirals, anticancer agents, and antibiotics already approved for human use.8Oxford Academic. A modeling framework for embedding-based predictions for compound–viral protein activity The ability to screen billions of SMILES strings computationally and prioritize a short list for laboratory testing has become central to how many pharmaceutical companies approach early-stage drug discovery.
When AI Struggles with SMILES
SMILES-based AI has a persistent headache: generative models frequently produce invalid strings. When a model is tasked with designing new molecules by outputting SMILES characters one at a time, it can easily generate something that looks plausible but violates the notation’s rules, like mismatched parentheses, unclosed rings, or chemically impossible valence states. An analysis of invalid SMILES produced by several generative models found that different architectures fail in different ways. Recurrent neural networks and generative adversarial networks mostly produced chemistry-related errors (atoms with impossible numbers of bonds, for example), while variational autoencoders tended to produce more grammar-level mistakes, like unbalanced parentheses and ring closure symbols.9PubMed Central. UnCorrupt SMILES: a novel approach to de novo design
This is not a minor inconvenience. If a generative model produces a batch of 10,000 candidate molecules and a significant fraction are nonsensical strings, the pipeline wastes computational effort filtering junk before any real chemistry can be evaluated. One response has been to develop “self-correcting” approaches that try to fix broken SMILES strings after generation. A more radical solution was the introduction in 2020 of SELFIES (SELF-referencing Embedded Strings), an alternative molecular language designed so that every possible string is a valid molecule. SELFIES achieves this by construction, making it impossible for a generative model to produce an invalid output no matter what characters it generates.10PubMed Central. SELFIES and the future of molecular string representations Whether SELFIES will replace SMILES or simply coexist alongside it remains an open question. SMILES has the advantage of decades of infrastructure, tooling, and human readability. SELFIES has the advantage of robustness. For now, most researchers use whichever fits their workflow.
Another practical challenge is converting SMILES strings back into three-dimensional molecular geometries. A SMILES string encodes connectivity (which atoms connect to which) but says nothing about the actual shape of the molecule in space. For simple, unstrained molecules, standard software can generate reasonable 3D structures from SMILES. But for molecules under strain, where bond angles are forced away from their preferred values, the default tools can produce distorted or incorrect geometries. A recently developed Python workflow called strainedSMILES2xyz addresses this by relaxing the constraints used during conformer generation, exploring stereoisomer variants, and refining the output with force-field calculations.11PubMed Central. strainedSMILES2xyz: a workflow for reliable 3D structures of strained molecules from SMILES
The Smiles Rearrangement, an Entirely Different Story
Switch gears completely. The Smiles rearrangement is a named reaction in organic chemistry that has nothing to do with text strings or databases. It is named after Samuel Smiles, a British chemist who reported the reaction in 1931. At its core, the Smiles rearrangement is an intramolecular nucleophilic aromatic substitution. A group within a molecule attacks its own aromatic ring at the position where another group is attached, breaking one bond and forming a new one in a single molecular reshuffling.12PubMed. Modern Aspects of the Smiles Rearrangement
The classic version requires an electron-poor aromatic ring (one bearing groups that pull electron density away from the ring, making it susceptible to attack by a nucleophile within the same molecule). The nucleophile, which might be a nitrogen, oxygen, or sulfur atom tethered to the ring through a short chain, swings over and displaces the original leaving group at the ipso position (the carbon directly bearing the leaving group). Computational studies on specific examples, like the intramolecular cyclization of certain pyridine-containing substrates, have confirmed that the attack at the ipso position proceeds through a two-step mechanism and is energetically favored over direct attack at an adjacent position on the ring.13Journal of Physical Organic Chemistry. Theoretical evidence on O–N type smiles rearrangement mechanism: a computational study on the intramolecular cyclization of N‐methyl‐2‐(2‐chloropyridin‐3‐yloxy)‐acetamide anion
For decades, the Smiles rearrangement was considered a somewhat specialized tool, limited by the requirement for an activated (electron-poor) aromatic ring. The closely related Truce-Smiles rearrangement, which involves migration of an aryl group to a carbon center rather than a heteroatom, expanded the scope, and a survey of the literature from the 1950s to the present documented a wide variety of substrate designs and tandem reaction sequences that exploit this variant.14Canadian Journal of Chemistry. The Truce–Smiles rearrangement and related reactions: a review
The Radical Renaissance
The transformation that took the Smiles rearrangement from a textbook curiosity to a hot research topic was the realization that it could proceed through radical intermediates, not just the traditional polar (nucleophilic) mechanism. In the radical version, instead of a lone pair of electrons attacking the ring, a carbon-centered radical does the job. This fundamentally changes the game because radical intermediates do not need an electron-poor aromatic ring to react. Unactivated arenes, including simple phenyl groups and heterocyclic rings, become viable substrates.
Visible-light photoredox catalysis has been the key enabler. By using a photocatalyst that absorbs visible light and generates radicals under mild conditions, chemists can trigger Smiles-Truce rearrangements at room temperature, in the presence of sensitive functional groups, and without harsh reagents.15Accounts of Chemical Research. Photoredox-Mediated Radical Smiles–Truce Rearrangement: From Aryl Migration to Molecular Complexity Generation The radical adds intramolecularly to the arene, forming a transient spirocyclic intermediate that then fragments to give the rearranged product. The net effect is that an aryl group migrates from one part of the molecule to another, and the whole process can be chained together with other radical transformations in cascade sequences that build molecular complexity quickly.
The applications have been diverse. One group reported the first visible-light-induced acyl radical Smiles rearrangement, converting biaryl ethers into hydroxybenzophenones under metal-free conditions using an organic photocatalyst and a hypervalent iodine reagent.16PubMed. Acyl Radical Smiles Rearrangement To Construct Hydroxybenzophenones by Photoredox Catalysis Another used neutral eosin Y, a common dye, as a direct hydrogen atom transfer photocatalyst to trigger a radical Smiles rearrangement that produced novel nitrogen-containing heterocycles as single diastereomers, some of which showed preliminary anticancer activity.17PubMed. A Radical Smiles Rearrangement Promoted by Neutral Eosin Y as a Direct Hydrogen Atom Transfer Photocatalyst These examples illustrate a broader trend: the radical Smiles rearrangement has become a general-purpose tool for moving aromatic groups around within a molecule, enabling bond disconnections that would be difficult or impossible by other methods.18PubMed Central. Radical Smiles Rearrangement: An Update
The reaction has also been used as a key step in complex natural-product synthesis. A synthesis of a ring-modified version of galanthamine, a drug used to treat Alzheimer’s disease symptoms, employed a radical-based Smiles rearrangement alongside an intramolecular Heck reaction to build the molecule’s challenging ring system.19PubMed. Synthesis of a D-ring isomer of galanthamine via a radical-based Smiles rearrangement reaction
Where the Radical Smiles Rearrangement Is Headed
Recent reviews have mapped out the current frontier and the gaps that remain. Photocatalytic and catalyst-free radical Smiles rearrangements now encompass alkene difunctionalizations (adding two groups across a double bond in one step), sulfur dioxide insertion cascades, decarboxylative arylations, and remote functionalization of bonds that are normally hard to reach. Emerging applications in polymer synthesis and biocatalysis have also been reported.20PubMed. Radical Smiles Rearrangements: Aryl Migration for Molecular Editing
The main limitations are familiar ones for radical chemistry. Controlling stereochemistry, the three-dimensional arrangement of atoms in the product, remains difficult because radical intermediates are often geometrically flat and unselective. Substrate scope, while dramatically expanded compared to the classical polar version, still has gaps. Some classes of arenes and heteroatom-containing tethers do not rearrange efficiently. And scaling these reactions for industrial use sometimes runs into problems with reproducibility under photoredox conditions, where light penetration and catalyst loading need careful optimization.
The most active areas of development include integrating Smiles rearrangements with electrochemical methods (using electricity instead of light to generate radicals), continuous-flow reactors (which solve many of the scaling problems of photochemistry), and data-driven catalyst design. The hope is that combining the rearrangement’s ability to reorganize molecular frameworks with these modern enabling technologies will open new routes to complex drug molecules and functional materials that are currently laborious or impossible to make by conventional methods.21PubMed. Radical Smiles Rearrangements: Aryl Migration for Molecular Editing There is a certain irony in the fact that SMILES notation, the text-string version, is being used in AI models that could eventually help design substrates for Smiles rearrangements, the reaction version. The two meanings of the word have never formally intersected, but the tools of computational chemistry are drawing them into the same workflows.