Multi-View Biomedical Foundation Models for Molecule-Target and Property Prediction

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Molecular modeling and simulation

Background:

• Foundation models offer powerful tools for bio-medical research, but often rely on single molecular representations.
• Developing robust molecular representations is crucial for accurate prediction across diverse downstream tasks.
• Existing models may have limitations due to a singular focus on one data modality or molecular view.

Purpose of the Study:

• To develop a novel multi-view molecular foundation model integrating diverse data types.
• To enhance prediction accuracy for various bio-medical tasks, including molecular properties and drug-target interactions.
• To apply the developed model for virtual screening of ligands targeting Alzheimer's disease-related G Protein-Coupled receptors (GPCRs).

Main Methods:

• Developed Multi-view Molecular Embedding with Late Fusion (MMELON) by integrating pre-trained graph, image, and text foundation models.
• Validated the MMELON model on over 120 diverse bio-medical tasks, including molecular solubility, ADME properties, and GPCR activity.
• Utilized the GPCR model array for virtual screening to identify potential ligands for Alzheimer's disease-related GPCRs.

Main Results:

• The MMELON model demonstrated robust performance across a wide range of bio-medical tasks.
• Successfully identified potential ligands targeting Alzheimer's disease-related GPCRs through virtual screening.
• Validated predictions using structure-based modeling and identified key binding motifs for selected ligands.

Conclusions:

• Multi-view molecular foundation models, like MMELON, significantly improve prediction accuracy and robustness in bio-medical research.
• MMELON facilitates efficient drug discovery by enabling accurate virtual screening and identification of potent drug candidates.
• The integration of multiple molecular views provides a more comprehensive understanding of molecular interactions and properties.