SELFIES

SELFIES encodes molecules as self-referencing character strings that guarantee syntactic and semantic validity for use in machine learning and generative models.

Key Features:

• 100% Robustness: Every SELFIES string maps to a valid molecular structure, avoiding the syntactic and semantic invalidity issues sometimes encountered with SMILES.
• Implementation in the selfies library: The SELFIES representation is implemented in the selfies library, which provides the computational routines for encoding and decoding SELFIES strings.
• Generalization and Flexibility: The representation has been generalized to support a broader range of molecules and explicit semantic constraints.
• Streamlined Grammar: An underlying streamlined grammar enforces syntactic rules and supports the features provided by the selfies implementation.
• Integration with Machine Learning: SELFIES is designed to be used as direct input to machine learning and generative models to produce valid molecular graphs.

Scientific Applications:

• Cheminformatics: Representation and manipulation of molecular structures within cheminformatics workflows.
• Generative molecular design: Use in generative models to create new molecular structures that are syntactically and semantically valid.
• Chemical space exploration: Systematic exploration of chemical space while maintaining molecular validity constraints.
• Drug discovery and materials science: Application in computational discovery efforts where valid candidate molecules are required for downstream evaluation.

Methodology:

SELFIES employs a self-referencing string representation and a streamlined grammar, implemented in the selfies library, with updates to the representation and constraints to ensure generated strings correspond to valid molecular graphs.