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

Abstract

Protein language models embed protein sequences for different tasks. However, these are suboptimal at learning the language of protein interactions. We developed an interaction language model (iLM), Sliding Window Interaction Grammar (SWING) that leverages differences in amino-acid properties to generate an interaction vocabulary. SWING successfully predicted both class I and class II peptide–major histocompatibility complex interactions. Furthermore, the class I SWING model could uniquely cross-predict class II interactions, a complex prediction task not attempted by existing methods. Using human class I and II data, SWING accurately predicted murine class II peptide–major histocompatibility interactions involving risk alleles in systemic lupus erythematosus and type 1 diabetes. SWING accurately predicted how variants can disrupt specific protein–protein interactions, based on sequence information alone. SWING outperformed passive uses of protein language model embeddings, demonstrating the value of the unique iLM architecture. Overall, SWING is a generalizable zero-shot iLM that learns the language of protein–protein interactions. SWING is a versatile interaction language model that can learn the language of peptide and protein interactions.