SELFIES

The defining feature of SELFIES is its guaranteed validity: every string generated from the SELFIES grammar corresponds to a valid molecular graph. This is achieved through a formal context-free grammar where derivation rules enforce valence constraints and bonding rules at the token level. Unlike SMILES, where a single misplaced character can produce an invalid molecule, SELFIES eliminates the need for post-hoc validity checks or rejection sampling during generative modeling.

While all SELFIES strings are syntactically valid, they can be semantically invalid. This occurs when a string represents a chemically unreasonable molecule, such as one with strained rings, charged species with missing counterions, or violations of the octet rule. The distinction is critical: syntactic validity guarantees a molecular graph exists, but semantic validity requires additional chemical filters or property predictors to assess drug-likeness or synthetic accessibility.

SELFIES uses a formal grammar with derivation rules that map each symbol to a specific graph-construction operation. Key rule types include:

• Atom symbols: Add an atom with a specified valence state (e.g., [C], [N], [O])
• Bond symbols: Specify the bond order between the most recently added atom and a previous atom (e.g., [=C], [#N])
• Branch symbols: Open and close branches for ring closures and side chains ([Branch1], [Ring1])
• State symbols: Modify atom states like charge or isotope without altering connectivity

SELFIES strings are typically 1.5-2x longer than equivalent SMILES, but the robustness trade-off is essential for variational autoencoders (VAEs), generative adversarial networks (GANs), and reinforcement learning pipelines where invalid outputs waste compute and degrade training.

The SELFIES pipeline operates in three stages:

1. Encoding: A SMILES string is parsed into a molecular graph, then converted to a SELFIES string using the derivation grammar
2. Manipulation: The SELFIES string is tokenized into a sequence of symbols for use in sequence-based deep learning models (LSTMs, Transformers, VAEs)
3. Decoding: The generated SELFIES string is deterministically converted back to a molecular graph, then optionally to SMILES for visualization

This bidirectional mapping ensures lossless conversion between representations while maintaining the robustness guarantee throughout the generative process.

SELFIES is particularly valuable in de novo drug design pipelines where generative models explore chemical space:

• Variational Autoencoders: The continuous latent space can be freely sampled without producing invalid molecules
• Reinforcement Learning: Reward functions can focus on property optimization rather than validity penalties
• Genetic Algorithms: Crossover and mutation operations always produce valid offspring
• Language Models: Autoregressive generation with SELFIES tokens guarantees valid outputs at every step

Libraries such as selfies (Python) and integrations with RDKit, DeepChem, and GuacaMol support adoption.

The formal mathematical structure underlying all molecular representations, where atoms are nodes and bonds are edges. SELFIES guarantees a bijective mapping to valid molecular graphs.

• Node features: Atom type, formal charge, hybridization state, chirality
• Edge features: Bond order (single, double, triple, aromatic), stereochemistry
• Graph constraints: Valence rules, octet rule compliance, kekulé structures for aromatic systems
• Representation: Adjacency matrices, edge lists, or feature-augmented graph objects used directly by Graph Neural Networks

An intermediate representation that modifies SMILES syntax to improve generative model validity by addressing ring closure and branching ambiguities. A stepping stone toward SELFIES-level robustness.

• Ring syntax: Uses explicit ring-opening and ring-closing tokens with paired indices
• Branching: Replaces parentheses with explicit branch-start and branch-end tokens
• Validity improvement: Reduces invalid outputs compared to standard SMILES but does not achieve 100% robustness
• Historical context: Demonstrated that syntax redesign could improve generation validity, paving the way for SELFIES

A class of generative methods that constrain outputs to a formal grammar, ensuring syntactic validity. SELFIES is a prime example where the grammar is designed to be semantically closed—every syntactically valid string maps to a valid molecule.

• Context-free grammars (CFG): Production rules define valid token sequences
• Attribute grammars: Extend CFGs with semantic rules to enforce chemical constraints
• SELFIES grammar: A deterministic finite-state automaton where each derivation step preserves chemical validity
• Alternative approach: Variational autoencoders with grammar-guided decoding for SMILES