MAMMAL - Molecular Aligned Multi-Modal Architecture and Language for biomedical discovery.

Area of Science:

• Artificial Intelligence in Pharmacology
• Computational Drug Discovery
• Biomedical Data Integration

Background:

• Modern AI offers new avenues for pharmacology by improving disease mechanism and drug action modeling.
• Integrating diverse biomedical data modalities presents a significant challenge in developing effective AI models.

Purpose of the Study:

• Introduce MAMMAL (Molecular Aligned Multi Modal Architecture and Language), a foundation model for cross-modal learning.
• Address challenges in drug discovery tasks by enabling joint integration of disparate biomedical data.

Main Methods:

• Pre-trained MAMMAL on 2 billion samples across protein/antibody sequences, small molecules, and gene expression profiles.
• Developed MAMMAL to support classification, regression, and generative tasks on cross-modal inputs.
• Evaluated MAMMAL across eleven benchmarks covering multiple drug discovery pipeline stages.

Main Results:

• MAMMAL achieved state-of-the-art performance on nine out of eleven benchmarks.
• Demonstrated competitive results on the remaining two benchmarks.
• Fine-tuned MAMMAL significantly outperformed AlphaFold3 confidence scores in predicting antibody-antigen binding for five of seven targets.

Conclusions:

• MAMMAL represents a significant advancement in AI for drug discovery, capable of integrating diverse biological data.
• The model's strong performance across multiple benchmarks highlights its potential to accelerate the drug discovery pipeline.
• Public availability of the MAMMAL framework and models promotes open and collaborative research in the field.