A merged molecular representation learning for molecular properties prediction with a web-based service

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Cheminformatics

Background:

• Deep learning significantly advances molecular property prediction for drug discovery.
• SMILES (Simplified Molecular Input Line Entry System) is widely used but struggles to capture chemical properties effectively.
• Existing methods often fail to integrate molecular structure and chemical context efficiently.

Purpose of the Study:

• To develop a novel self-supervised learning method for pre-training Transformer models.
• To simultaneously learn SMILES representations and the chemical contexts of molecules.
• To improve the accuracy and generalization of molecular property prediction.

Main Methods:

• Proposed a self-supervised Transformer pre-training approach.
• Incorporated adjacency matrix embedding for structural learning.
• Integrated Quantitative Estimation of Drug-likeness (QED) prediction for inferring molecular descriptors.
• Developed a web-based fine-tuning service for practical application.

Main Results:

• Achieved improved data generalization in molecular property prediction.
• Demonstrated superior average performance across various downstream tasks.
• Successfully benchmarked the model's effectiveness on diverse chemical datasets.

Conclusions:

• The proposed self-supervised method effectively learns both SMILES and chemical contexts.
• The approach enhances molecular property prediction accuracy and generalization.
• The developed service facilitates the model's application in drug discovery workflows.