Sliding Window Interaction Grammar (SWING): a generalized interaction language model for peptide and protein...

Area of Science:

• Computational biology
• Bioinformatics
• Protein science

Background:

• Protein language models are crucial for sequence embedding but struggle with interaction prediction.
• Understanding protein-protein interactions is vital for biological and disease research.

Purpose of the Study:

• To develop a novel interaction language model (iLM) for predicting protein interactions.
• To leverage amino acid properties for a specialized protein interaction vocabulary.
• To assess the model's performance on MHC class I and II interactions and variant effects.

Main Methods:

• Developed Sliding Window Interaction Grammar (SWING), an iLM architecture.
• Utilized differences in amino acid properties to create an interaction vocabulary.
• Applied SWING to predict peptide-major histocompatibility complex (pMHC) class I and II interactions.
• Evaluated SWING's ability to predict variant-disrupted interactions and cross-predict between MHC classes.

Main Results:

• SWING successfully predicted both pMHC class I and II interactions.
• The class I SWING model demonstrated unique cross-prediction capabilities for class II interactions.
• SWING accurately predicted murine pMHC class II interactions linked to autoimmune disease risk alleles.
• The model accurately predicted how sequence variants disrupt protein-protein interactions.

Conclusions:

• SWING is a generalizable, zero-shot iLM that effectively learns the language of protein-protein interactions.
• SWING outperforms passive protein language model embeddings for interaction prediction.
• The developed iLM architecture offers a valuable tool for predicting protein interaction disruptions from sequence data alone.