Automatic Prediction of Molecular Properties Using Substructure Vector Embeddings within a Feature Selection...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials informatics

Background:

• Machine learning (ML) methods accelerate molecular property prediction by analyzing structure-property relationships.
• Unsupervised, self-supervised, and transformer models show promise but require extensive computational resources.
• Efficient screening of molecules is crucial for developing new pharmaceuticals and specialized chemical materials.

Purpose of the Study:

• To present a semisupervised strategy for predicting molecular and drug properties.
• To improve the balance between model accuracy and computational requirements in ML for chemistry.
• To enhance model interpretability through feature interaction analysis.

Main Methods:

• Developed a semisupervised strategy using substructure vector embeddings.
• Implemented an ML-based feature selection workflow.
• Evaluated the methodology on diverse regression and classification datasets.

Main Results:

• Achieved superior performance compared to many state-of-the-art algorithms.
• Demonstrated a favorable balance between predictive accuracy and computational demands.
• Provided insights into feature interactions, enhancing model interpretability.
• Successfully predicted chemical molecule lipophilicity in a case study.

Conclusions:

• The proposed semisupervised approach offers an efficient and interpretable alternative for molecular property prediction.
• Meticulous feature analysis and selection are vital for robust predictive modeling in cheminformatics.
• This strategy can accelerate drug discovery and materials design by optimizing computational workflows.