SMILES Pair Encoding

SMILES Pair Encoding tokenizes SMILES strings into frequent, chemically meaningful substrings to improve deep learning for molecular generation and molecular property prediction.

Key Features:

• Vocabulary Learning: Learns a vocabulary of high-frequency SMILES substrings from large chemical datasets such as ChEMBL.
• Data-Driven Tokenization: Uses the learned vocabulary to tokenize SMILES into human-readable, chemically meaningful substrings that provide richer contextual information than atom-level tokenization.
• BPE Inspiration: Adopts a Byte-Pair Encoding (BPE)-inspired procedure that iteratively merges frequent token pairs to form new tokens.
• Performance Enhancement: Improves validity and novelty of generated SMILES and matches or surpasses atom-level tokenization performance across 24 benchmark datasets for molecular property prediction.

Scientific Applications:

• Molecular Generation: Enhances the quality, validity, and novelty of generated SMILES for designing novel molecules in drug discovery and materials science.
• Property Prediction: Boosts performance of molecular property prediction models across benchmark datasets, supporting prediction of chemical behavior and interactions.

Methodology:

SPE constructs a vocabulary from high-frequency SMILES substrings in large datasets (e.g., ChEMBL), applies a BPE-inspired iterative merging of frequent token pairs to learn tokens, and tokenizes new SMILES into those learned substrings.