Accurate prediction of the strength of interaction of polymorphic NK cell receptors with MHC class I ligands modulated by peptides using protein language models — ASN Events
  1. Schedule
  2. Poster Session One
  3. Accurate prediction of the strength of interaction of polymorphic NK cell receptors with MHC class I ligands modulated by peptides using protein language models

Accurate prediction of the strength of interaction of polymorphic NK cell receptors with MHC class I ligands modulated by peptides using protein language models (#102)

Abdallah AlShafey 1 , Madeline Nelson 1 , Mubashar Hassan 1 , Andrzej Kloczkowaski 1 , William Ray 1 , Salim I Khakoo 2 , Jayajit Das 1
  1. Nationwide Children's Hospital, Columbus, OHIO, United States
  2. University of Southampton, Southampton, England

Polymorphic killer-cell immunoglobulin-like receptors (KIRs) interact with polymorphic class I major histocompatibility complex (MHC-I) molecules which is modulated by short peptides (~8-mers). Experiments show peptide specificity regulating the interaction strength between KIR and MHC-I where a change in a few amino acid residues can turn a strong interaction between a particular KIR and MHC-I to below detection level. Peptides generated during viral infections (e.g., HCV and HIV) and recently in specific cancers (hepatocellular carcinoma) have been found to provide immune protection. Therefore, identifying peptides that can modulate the interaction between polymorphic KIR and MHC-I molecules are an attractive candidate for personalized immunotherapy against infection and cancer. However, a major challenge is to determine these peptides among the sheer number of possibilities. The state-of-the-art experiments currently explore a minuscule subset of all possible distinct interactions. To address this challenge, we leverage protein language foundation models (ProtBERT and ProtT5), protein structure prediction tools (AlphaFold2), deep learning approaches encoding protein structure (Foldseek), and available datasets for KIR binding with different peptide-HLA complexes. Our model takes the raw amino acid sequences of the KIR, peptide and HLA, and uses ProtBERT/ProtT5 and Foldseek to generate embeddings which are passed through a multilayer perceptron (MLP) model. The model was trained and tested on several datasets[1][2][3][4] and we report an area under receiver operator characteristic (AUROC) of >0.9.  Our analysis also indicates a higher dependence on the nature of the residues at the interaction site and lower emphasis on the overall structural information. Our model holds significant potential for advancing our understanding of immune regulation and the biophysical factors responsible for it, ultimately paving the way for novel therapeutic interventions.

 

[1] Science Immunology 8.87 (2023): eadh1781

[2] Proc. Natl. Acad. Sci. U.S.A. 107 (22) 10160-10165

[3] Cell 187.24 (2024): 7008-7024

[4] Nat Commun 14, 4809 (2023)