Multi-Modal Fusion Frameworks of Subgraph-Optimized Graph Autoencoder for Molecular Property Prediction.

Area of Science:

• * Computational chemistry and cheminformatics.
• * Machine learning for drug discovery and materials science.

Background:

• * Molecular property prediction is crucial for drug design, but effective feature learning remains a challenge.
• * Existing graph models extract features but struggle with cross-task utilization.
• * Substructure features significantly impact molecular properties, necessitating advanced extraction methods.

Purpose of the Study:

• * To develop an improved method for molecular property prediction.
• * To enhance feature learning by effectively utilizing substructure information.
• * To leverage multimodal feature fusion for better cross-modal learning.

Main Methods:

• * Proposed a subgraph-optimized Graph Autoencoder (TurboGAE).
• * Introduced a subgraph-level graph tokenizer to capture substructure impacts.
• * Developed multimodal feature fusion strategies for aligning intermodal features during pretraining.

Main Results:

• * TurboGAE effectively captures the influence of substructure features on molecular properties.
• * Multimodal fusion strategies successfully align intermodal features, enhancing learning.
• * The proposed methods demonstrated excellent performance on downstream prediction tasks.

Conclusions:

• * TurboGAE offers a more effective approach to molecular feature representation.
• * Multimodal fusion strategies are vital for leveraging diverse molecular data.
• * The developed techniques show significant promise for advancing molecular property prediction in computational chemistry and drug design.