SELFIES

The defining property of SELFIES is 100% robustness: any string generated by a model is guaranteed to decode into a chemically valid molecular graph. This is achieved through a formal grammar where each token either adds an atom, adds a bond, or branches, and the derivation rules inherently enforce valence constraints. Unlike SMILES, where a single misplaced character produces an invalid string, SELFIES eliminates the need for post-hoc validity filters or rejection sampling during generation.

SELFIES uses a context-free grammar with self-referencing constraints. The derivation state tracks the most recently derived atom, and bond tokens explicitly reference this atom to form connections. Key structural tokens include:

• [Ring] and [Branch] tokens for cyclic structures and branching points
• [=O] and [#N] for double and triple bonds to the last atom
• [C], [N], [O] for atom specification This design ensures that bond formation always respects the current valence state of the referenced atom.

SELFIES was explicitly created to solve the invalid output problem in deep generative models like variational autoencoders (VAEs), generative adversarial networks (GANs), and language models applied to molecular design. In SMILES-based generation, up to 90% of sampled strings can be invalid, wasting compute and complicating training. SELFIES constrains the output space to only valid graphs, allowing models to explore chemical space more efficiently and enabling gradient-based optimization directly in string space without validity penalties.

There exists a surjective mapping from SELFIES strings to molecular graphs: every SELFIES decodes to exactly one graph, and every graph has at least one SELFIES representation. This property is critical for machine learning because it means the representation space has no "dead zones" of invalid strings. The encoding algorithm converts a graph into a SELFIES string by traversing the graph in a deterministic order, while the decoding algorithm reconstructs the graph by sequentially applying derivation rules. This formal guarantee distinguishes SELFIES from all predecessor string representations.

SELFIES addresses fundamental limitations of prior representations:

• SMILES: Context-sensitive grammar with no validity guarantee; requires complex parsing and ring closure digit matching
• DeepSMILES: Simplifies ring and branch syntax but remains non-robust; invalid strings still possible
• SELFIES: Context-free grammar with formal validity guarantee; slightly longer strings than SMILES but zero invalid outputs
• InChI: Hierarchical but not designed for generative modeling; lacks sequential token-by-token construction The trade-off is string length—SELFIES strings are typically 10-20% longer than equivalent SMILES—but this is negligible compared to the elimination of invalid sampling.

SELFIES is supported by a lightweight, open-source Python library (selfies) that provides:

• selfies.encoder(): Convert SMILES to SELFIES
• selfies.decoder(): Convert SELFIES back to SMILES
• selfies.get_semantic_robustness(): Verify validity guarantees
• selfies.split_selfies(): Tokenize SELFIES into constituent symbols for model input The library integrates seamlessly with DeepChem, RDKit, and standard PyTorch/TensorFlow pipelines. The token vocabulary is finite and known in advance, simplifying one-hot encoding or learned embedding layers for sequence models.

A syntactic adaptation of SMILES designed to improve generative model performance by simplifying the grammar. DeepSMILES eliminates explicit ring closure digits and replaces them with a single ring closure symbol, reducing the complexity of the syntax tree. While this improves validity rates over standard SMILES, it still does not guarantee 100% robustness like SELFIES.

• Key innovation: Simplified ring and branching syntax
• Validity improvement: Reduces invalid outputs but not to zero
• Trade-off: Less human-readable than standard SMILES

The native mathematical representation of molecules as graphs where atoms are nodes and bonds are edges. Graph Neural Networks (GNNs) operate directly on this representation, bypassing string encodings entirely. While SELFIES converts graphs to strings for sequence-based models, GNNs preserve topological information natively and achieve state-of-the-art results in property prediction.

• Advantage: No information loss from string conversion
• Key architectures: Message Passing Neural Networks, Graph Attention Networks
• Complementary: SELFIES can be used alongside graph representations in hybrid models

Fixed-length binary or integer vectors encoding the presence or absence of specific substructural features. Unlike SELFIES, fingerprints are lossy representations optimized for similarity searching and quantitative structure-activity relationship (QSAR) modeling. Common types include ECFP4 (circular fingerprints based on Morgan algorithm) and MACCS keys (predefined structural keys).

• Dimensionality: Typically 1024–2048 bits
• Key use: Tanimoto similarity for virtual screening
• Contrast with SELFIES: Lossy vs. lossless; fixed-length vs. variable-length

The formal syntactic rules that define how molecular string representations are constructed and parsed. SELFIES uses a context-free grammar where every production rule guarantees chemical validity — a property known as semantic closure. This contrasts with SMILES, whose grammar requires complex semantic checks to validate the resulting molecular graph.

• Key concept: Formal language theory applied to chemistry
• SELFIES grammar: ~100 production rules covering organic chemistry
• Benefit: Enables formal verification of generative model outputs