gSelformer-MV: Multiview, Subgraph-Augmented Group SELFIES Transformer for Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Data-driven methods are crucial for linking chemical structures to properties.
• Current molecular representation learning primarily uses atom-focused approaches (e.g., graph neural networks, chemical language models).
• Integrating functional group information into advanced models is an underexplored area.

Purpose of the Study:

• To introduce gSelformer-MV, a transformer model designed to represent molecules at both atomic and substructure levels.
• To leverage multiple views of Group SELFIES (a variant of SELFIES augmented with functional group tokens) for improved molecular property prediction.
• To address the gap in incorporating functional group data into advanced molecular representation learning.

Main Methods:

• Developed gSelformer-MV, a transformer architecture operating on multiple views of Group SELFIES.
• Constructed multiple subgraph-partitioned Group SELFIES views for joint training and inference.
• Compared gSelformer-MV against models trained solely on SELFIES strings.

Main Results:

• gSelformer-MV demonstrated superior accuracy and explainability over models using only SELFIES.
• Achieved state-of-the-art performance on multiple molecular regression benchmarks.
• Observed further performance improvements by focusing on high-confidence predictions.

Conclusions:

• Subgraph augmentation using Group SELFIES is an effective strategy for enhancing string-based molecular property prediction.
• gSelformer-MV offers a powerful new approach for molecular representation learning.
• The findings highlight the potential of incorporating substructure information for predictive modeling in chemistry.