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

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• Protein language models (pLMs) are rapidly advancing due to increased sequence data.
• Current methods for predicting protein-protein or peptide-protein interactions (PPIs) often involve co-embedding or concatenating sequences.
• No existing method utilizes a language representation of the interaction itself.

Purpose of the Study:

• To develop a novel interaction language model (iLM) for predicting protein and peptide interactions.
• To introduce Sliding Window Interaction Grammar (SWING) for generating an interaction vocabulary from amino acid properties.
• To evaluate SWING's performance on peptide:MHC (pMHC) and mutation-induced PPI disruption prediction.

Main Methods:

• Developed an interaction LM (iLM) using a novel language to represent interactions.
• Utilized Sliding Window Interaction Grammar (SWING) to create an interaction vocabulary based on amino acid properties.
• Applied the LM with SWING representations as features for supervised prediction tasks.

Main Results:

• SWING achieved state-of-the-art prediction performance for Class I and Class II peptide:MHC (pMHC) interactions.
• A unique Mixed Class model jointly predicted both pMHC classes, and Class I training predicted Class II interactions.
• SWING accurately predicted Class II pMHC interactions in murine models and demonstrated generalizability by predicting mutation-specific disruptions in PPIs.

Conclusions:

• SWING is a first-in-class, generalizable zero-shot iLM that learns the language of PPIs.
• The model accurately predicts interaction-specific disruptions caused by missense mutations using only sequence information.
• SWING advances the field by offering a novel approach to understanding and predicting complex biological interactions.