Triview Molecular Representation Learning Combined with Multitask Optimization for Enhanced Molecular Property...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Current molecular property prediction methods often use single-view representations (e.g., SMILES strings).
• Limited research explores multi-view molecular representations beyond two views.
• Existing pretraining and fine-tuning approaches struggle with limited data and task correlations.

Purpose of the Study:

• To develop an improved molecular representation learning method.
• To enhance the accuracy and generalizability of molecular property prediction.
• To address limitations of single-view and dual-view molecular representations.

Main Methods:

• Integration of molecular sequences, graphs, and images into a unified representation.
• Utilizing three distinct encoders for extracting features from multiple molecular views.
• Employing contrastive learning to align the different molecular views.
• Implementing a multitask optimization strategy to leverage task correlations.
• Applying low-rank adaptation (LoRA) for efficient task-specific fine-tuning.

Main Results:

• Demonstrated enhanced accuracy and robustness in molecular property prediction.
• Validated performance across multiple benchmark datasets.
• Showcased the effectiveness of multi-view integration and multitask learning.

Conclusions:

• The proposed multi-view molecular representation learning method significantly improves prediction performance.
• The approach is effective even with limited data by leveraging cross-task information.
• This work offers a promising direction for advancing molecular property prediction in computational chemistry.