ProtoMol: enhancing molecular property prediction via prototype-guided multimodal learning.

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Multimodal molecular representation learning combines molecular graphs and text for improved predictions.
• Existing methods have limitations in hierarchical semantic dependency modeling and cross-modal alignment.

Purpose of the Study:

• To propose ProtoMol, a novel prototype-guided multimodal framework for fine-grained integration and semantic alignment of molecular graphs and text.
• To address limitations in current multimodal approaches for molecular representation learning.

Main Methods:

• Utilized dual-branch hierarchical encoders: Graph Neural Networks for molecular graphs and Transformers for text.
• Implemented a layer-wise bidirectional cross-modal attention mechanism for progressive semantic feature alignment.
• Introduced a shared prototype space with learnable anchors for coherent and discriminative representations.

Main Results:

• ProtoMol demonstrated consistent outperformance against state-of-the-art baselines across various molecular property prediction tasks.
• The framework achieved robust alignment and fine-grained integration between molecular graph and textual data.
• Layer-wise interaction and prototype space significantly improved predictive accuracy and interpretability.

Conclusions:

• ProtoMol offers a superior approach to multimodal molecular representation learning.
• The proposed framework enhances predictions of drug toxicity, bioactivity, and physicochemical properties.
• This work provides a foundation for more effective AI-driven drug discovery.