XMolCap: Advancing Molecular Captioning Through Multimodal Fusion and Explainable Graph Neural Networks

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence

Background:

• Large language models (LLMs) advance computational biology but current molecular captioning methods underutilize data modalities and lack interpretability.
• Integrating diverse molecular data (images, SMILES, graphs) is crucial for improving drug discovery.
• Existing methods often fail to provide clear explanations for their predictions.

Purpose of the Study:

• Introduce XMolCap, a novel explainable molecular captioning framework.
• Integrate multiple molecular data types for enhanced drug discovery.
• Provide interpretable and accurate molecular descriptions.

Main Methods:

• Developed XMolCap using a stacked multimodal fusion mechanism.
• Employed a BioT5-based encoder-decoder architecture with specialized models (SwinOCSR, SciBERT, GIN-MoMu).
• Integrated molecular images, SMILES strings, and graph-based structures for feature extraction.

Main Results:

• Achieved state-of-the-art performance on L+M-24 and ChEBI-20 benchmark datasets.
• Outperformed several strong baseline models in molecular captioning.
• Generated detailed, functional group-aware, and property-specific explanations via graph-based interpretation.

Conclusions:

• XMolCap offers accurate and interpretable molecular descriptions, advancing drug discovery.
• The framework enhances understanding of molecular properties and interactions.
• Publicly available for reproducibility and potential clinical/pharmaceutical applications.