Stereoisomers Are Not Machine Learning's Best Friends.

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Accurate stereoisomer identification is crucial for machine learning in cheminformatics.
• Existing methods like Isomeric SMILES are text-based and require conversion for machine learning.
• Current molecular embedding tools (e.g., Mol2vec) fail to distinguish stereoisomers due to limitations in capturing spatial configurations.

Purpose of the Study:

• To develop and evaluate new methods for stereoisomer discrimination in cheminformatics.
• To generate unique molecular vectors that accurately represent stereochemical information.
• To improve machine learning model performance in predicting host-guest binding constants.

Main Methods:

• Utilizing word embedding techniques adapted for molecular structures.
• Incorporating stereochemical information directly into molecular representations.
• Treating Isomeric SMILES as text for Natural Language Processing-based embeddings.
• Comparing proposed methods against a benchmark task of predicting cyclodextrin-guest association constants.

Main Results:

• The proposed methods successfully generate distinct molecular vectors, differentiating stereoisomers.
• Enhanced molecular representations improve the accuracy of machine learning predictions for association constants.
• The study demonstrates the feasibility of using NLP-inspired techniques for stereochemical data.

Conclusions:

• Novel word embedding approaches effectively address the challenge of stereoisomer identification in cheminformatics.
• Accurate stereochemical representation is vital for robust machine learning applications in molecular science.
• The developed methods offer a promising pathway for advancing molecular modeling and drug discovery.