Multimodal fusion with relational learning for molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials informatics

Background:

• Graph-based molecular representation is crucial for predicting properties in drug discovery and materials science.
• Current methods face challenges in capturing complex molecular relationships and often lack sufficient chemical knowledge.
• Multimodal fusion offers a promising avenue but is limited by narrow modality exploration and unexplored integration stages.

Purpose of the Study:

• To introduce MMFRL (Multimodal Fusion with Relational Learning), a novel framework to enhance molecular property prediction.
• To address limitations of existing multimodal fusion approaches, particularly the unavailability of auxiliary data in downstream tasks.
• To systematically investigate the impact of different modality fusion stages (early, intermediate, late) on predictive performance.

Main Methods:

• Leveraging relational learning to enrich embedding initialization during multimodal pre-training.
• Developing a framework that allows downstream models to utilize auxiliary modalities even when absent during inference.
• Conducting a systematic investigation of early, intermediate, and late-stage modality fusion.

Main Results:

• MMFRL significantly outperforms existing methods on MoleculeNet benchmarks, demonstrating superior accuracy and robustness.
• The framework successfully enables downstream models to benefit from auxiliary modalities, irrespective of their availability during inference.
• Early, intermediate, and late fusion stages exhibit distinct advantages and trade-offs, providing valuable insights into optimal integration strategies.

Conclusions:

• MMFRL represents a significant advancement in graph-based molecular representation learning and multimodal fusion.
• The framework enhances predictive performance and explainability, offering deeper insights into chemical properties.
• MMFRL has the potential to revolutionize applications in drug discovery and materials science by improving predictive modeling and understanding.