SMILES Pair Encoding: A Data-Driven Substructure Tokenization Algorithm for Deep Learning.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Deep learning models utilizing Simplified Molecular Input Line Entry System (SMILES) are gaining traction in cheminformatics.
• Current atom-level tokenization methods for SMILES may limit model performance.
• A need exists for improved tokenization strategies to enhance chemical data representation.

Purpose of the Study:

• To introduce and evaluate SMILES Pair Encoding (SPE), a novel data-driven tokenization algorithm for SMILES.
• To compare the performance of SPE against traditional atom-level tokenization in deep learning tasks.
• To provide an open-source implementation of the SPE algorithm.

Main Methods:

• Developed SPE, a tokenization algorithm that learns a vocabulary of high-frequency SMILES substrings from large chemical datasets.
• Applied SPE to tokenize SMILES strings for training deep learning models.
• Evaluated SPE-based models on molecular generation and quantitative structure-activity relationship (QSAR) prediction tasks, comparing against atom-level tokenization.

Main Results:

• SPE-based generative models demonstrated superior novelty, diversity, and training set distribution resemblance compared to atom-level tokenization.
• SPE-based QSAR prediction models consistently matched or outperformed atom-level and k-mer tokenization across 24 benchmark datasets.
• The developed open-source Python package, SmilesPE, facilitates the implementation of this algorithm.

Conclusions:

• SMILES Pair Encoding (SPE) is a promising tokenization method for enhancing SMILES-based deep learning models in cheminformatics.
• SPE offers significant performance improvements in both molecular generation and QSAR prediction.
• The availability of the SmilesPE package promotes wider adoption and further research in this area.