SMILES String

SMILES represents a molecule's graph structure as a linear sequence of characters, where atoms are represented by their elemental symbols, single bonds are implicit or denoted by -, double bonds by =, and triple bonds by #. Hydrogen atoms are typically omitted unless explicitly specified. Branches are enclosed in parentheses, and ring closures are indicated by digits placed after the atoms involved. For example, cyclohexane is C1CCCCC1.

A single molecule can have multiple valid SMILES strings depending on the starting atom and traversal path. Canonical SMILES is a deterministic algorithm that generates a unique, standardized representation for a given molecular graph. This is essential for:

• Database deduplication and indexing
• Ensuring consistent lookup in chemical registries
• Comparing molecular identity across datasets Canonicalization algorithms, such as the CANGEN method, assign unique ranks to atoms based on graph invariants.

Isomeric SMILES extends the basic specification to encode stereochemical and isotopic information. Tetrahedral chirality is indicated using @@ and @ symbols, while cis/trans isomerism around double bonds is specified with / and \ directional slashes. Isotopic mass is prepended to the atomic symbol (e.g., [13C]). This precision is critical for drug-target interaction prediction, where specific stereoisomers can exhibit vastly different binding affinities.

SMARTS is a superset of SMILES that allows for substructure pattern matching using logical operators and wildcard atoms. It enables:

• Defining pharmacophoric patterns for virtual screening
• Specifying reaction transforms in cheminformatics
• Filtering compound libraries by toxicophores or privileged scaffolds For example, [#6]~[#7] matches any carbon-nitrogen bond regardless of bond order, while [!C;R] matches any ring atom that is not carbon.

SMILES strings serve as the primary input representation for sequence-based molecular deep learning models. They are tokenized and fed into architectures such as:

• Recurrent Neural Networks (RNNs) and Transformers for de novo molecule generation
• Seq2Seq models for molecular translation tasks (e.g., retrosynthesis prediction)
• SMILES-BERT and similar pre-trained language models for molecular property prediction This string-based approach contrasts with graph neural networks, which operate directly on the molecular graph topology.

Standard SMILES has known limitations that led to successor formats:

• Poor 3D conformational encoding: SMILES is inherently a 2D topological representation and does not capture 3D coordinates. SDF and PDB formats are used for spatial data.
• Ambiguity in tautomeric forms: Different tautomers of the same molecule produce distinct SMILES strings.
• No explicit hydrogen bonding: Non-covalent interactions are not encoded. DeepSMILES and SELFIES are modern extensions that guarantee syntactic validity for every generated string, addressing a key challenge in generative chemistry models.

A fixed-length bit-string or count vector encoding the presence or absence of specific substructural features within a molecule. Unlike SMILES strings, fingerprints are not human-readable but are directly consumable by traditional machine learning models.

• Circular fingerprints (ECFP): Encode circular atom neighborhoods up to a specified diameter
• Path-based fingerprints (MACCS): Encode predefined structural keys
• Key use: Quantitative Structure-Activity Relationship (QSAR) modeling and similarity searching via Tanimoto similarity

The algorithmic process of generating a unique, deterministic SMILES string for a given molecular graph, regardless of the original atom numbering. This ensures that the same molecule is always represented identically.

• Morgan algorithm: The most widely used canonicalization method, iteratively assigning atom invariants based on neighboring atomic properties
• Isomeric SMILES: Preserves stereochemical and isotopic information
• Critical for deduplication in chemical databases and ensuring reproducible cheminformatics workflows

A metric for comparing the similarity of two sets, most commonly applied to binary molecular fingerprints derived from SMILES strings. It quantifies the structural overlap between two chemical compounds.

• Formula: Intersection size divided by union size of fingerprint bits
• Range: 0.0 (no shared features) to 1.0 (identical fingerprints)
• Typical threshold: 0.7–0.8 for identifying structurally similar analogs
• Used extensively in virtual screening and scaffold hopping campaigns

A large-scale deep learning model, often based on the Transformer architecture, pre-trained on massive protein sequence databases. These models learn the underlying grammar of protein structure and function, analogous to how SMILES encodes molecular syntax.

• ESM-2 (Meta): Trained on 250 million protein sequences
• ProtBERT: BERT-style model for protein sequence understanding
• Used for binding site prediction, variant effect scoring, and zero-shot structure prediction
• Complements SMILES-based ligand representations in drug-target interaction models

An umbrella term for neural network architectures designed to respect the symmetries and invariances of non-Euclidean data, such as the 3D rotational and translational symmetry of molecular structures represented by SMILES.

• Equivariant Neural Networks: Guarantee output transforms predictably under rotation
• SE(3)-Transformers: Extend attention mechanisms to 3D coordinates
• Key advantage: Predictions are independent of molecular orientation in space
• Applied to protein-ligand docking, molecular dynamics acceleration, and conformer generation