Unified Knowledge-Guided Molecular Graph Encoder with multimodal fusion and multi-task learning.

Area of Science:

• Computational chemistry and cheminformatics.
• Machine learning and artificial intelligence.
• Bioinformatics and computational biology.

Background:

• Graph Neural Networks (GNNs) excel at multimodal input assimilation, improving performance across domains.
• Current molecular modeling methods often fragment geometric and semantic data, limiting holistic integration.
• Harmonizing heterogeneous and sparse multimodal molecular datasets presents a significant challenge.

Purpose of the Study:

• To introduce the Unified Knowledge-Guided Molecular Graph Encoder (UKGE) for unified molecular representation.
• To reconcile geometric and semantic molecular features using knowledge graphs and meta-paths.
• To enhance the generalizability and efficacy of molecular models in downstream applications.

Main Methods:

• Constructing Unified Molecular Graphs by integrating elemental knowledge graphs (KGs) and meta-path definitions.
• Employing a Meta-Path Aware Message Passing mechanism for multimodal data integration.
• Utilizing a multi-task learning strategy to balance diverse data modalities.

Main Results:

• UKGE achieved 96.91% ACC and 99.14% AUC for drug-drug interaction (DDI) prediction in warm-start settings.
• Demonstrated 83.15% ACC in cold-start DDI prediction scenarios.
• Achieved 0.644 CI (Davis) and 0.659 CI (KIBA) for compound-protein interaction (CPI) prediction.
• In ligand-based drug design (LBDD), UKGE reached 99.3% validity, 98.4% uniqueness, and 98.9% novelty.

Conclusions:

• UKGE provides a comprehensive and unified molecular representation by reconciling heterogeneous data.
• The framework significantly advances multimodal molecular modeling, outperforming existing methods.
• UKGE establishes a new state-of-the-art in molecular modeling for various critical applications.