Exploring data-driven chemical SMILES tokenization approaches to identify key protein-ligand binding moieties

Area of Science:

• Computational drug discovery
• cheminformatics
• bioinformatics

Background:

• Machine learning models frequently represent molecules as sequences for drug discovery tasks.
• Sequence-based models segment molecules into "chemical words" for natural language processing (NLP) applications.
• The chemical significance of these "chemical words" remains underexplored.

Purpose of the Study:

• To investigate the chemical characteristics and significance of "chemical words" derived from molecular sequences.
• To compare different data-driven tokenization techniques for identifying chemical words.
• To elucidate the role of these chemical words in protein-ligand binding.

Main Methods:

• Employed Byte Pair Encoding, WordPiece, and Unigram tokenization on molecular sequences (SMILES).
• Developed a language-inspired pipeline using tf-idf weighting to identify key chemical words in high-affinity ligands.
• Analyzed multiple protein-ligand affinity datasets and conducted target-specific case studies.

Main Results:

• Identified similar key chemical words across different subword tokenization algorithms, despite variations in vocabularies.
• Demonstrated that key chemical words are specific to protein targets.
• Correlated identified key chemical words with known pharmacophores and functional groups relevant to binding.

Conclusions:

• The study elucidates the chemical properties of machine learning-identified "chemical words".
• The identified key chemical words offer insights into molecular recognition and binding.
• This approach can aid in identifying significant chemical moieties for computational drug discovery.