Learning the language of protein-protein interactions

Area of Science:

• Computational Biology
• Biochemistry
• Genomics

Background:

• Protein Language Models (PLMs) are effective for individual protein properties.
• Existing PLMs struggle to model protein-protein interactions (PPIs) crucial for cellular processes.
• Understanding PPIs is vital for disease mechanisms and drug discovery.

Purpose of the Study:

• Introduce MINT, a novel PLM designed for contextual and scalable modeling of PPIs.
• Evaluate MINT's performance on various PPI-related tasks compared to existing models.
• Demonstrate MINT's utility in complex biological systems and therapeutic applications.

Main Methods:

• Unsupervised training of MINT on a large, curated PPI dataset from the STRING database.
• Benchmarking MINT against existing PLMs on binding affinity prediction and mutational effect estimation.
• Assessing MINT's performance in modeling protein assemblies, antibody-antigen interactions, and T cell receptor-epitope binding.

Main Results:

• MINT significantly outperforms existing PLMs in diverse PPI-related tasks.
• MINT excels at modeling complex protein assemblies and surpasses specialized models in antibody-antigen and TCR-epitope binding.
• MINT's predictions for oncogenic PPI mutations and antibody cross-neutralization align with experimental data.

Conclusions:

• MINT represents a significant advancement in modeling protein-protein interactions.
• The model offers a powerful, scalable tool for elucidating complex biological interactions.
• MINT has substantial implications for advancing biomedical research and accelerating therapeutic discovery.