Improving generalizability for MHC-I binding peptide predictions through geometric deep learning

Abstract

Understanding the peptide presentation mechanism of Major Histocompatibility Complex (MHC) is crucial to study the recognition of pathogens, the treatment of autoimmune diseases and in developing cancer immunotherapies. To comprehend these mechanisms and design effective therapies at affordable costs, it is fundamental to be able to precisely predictin silicowhich peptides can be bound by each MHC. Many computational approaches have been developed to identify MHC-binding peptides based on their sequences, however, they present biases and their performance on less-studied MHC alleles with limited experimental binding data is severely restricted by their sequence-based (SeqB) nature. We reason that structure-based (StrB) methods could learn the physico-chemical and geometrical rules of peptide-MHC (pMHC) complexes conserved among the alleles, thus generalizing better. We designed three supervised StrB geometric deep learning (GDL) methods, showing superior generalization across all StrB methods over two SeqB methods when the test data is distant from the training data. To further explore data efficiency, we designed a novel self-supervised GDL approach trained only on ∼1K x-ray structures, 3D-SSL, which predicts binding affinities without seeing any during training. Finally, we showcase our StrB method on a case study from the HPV vaccine design to show how StrB methods are robust to biases in the binding data. These findings highlight the potentials of structure-based methods in efficiently utilizing limited data, enhancing generalizability. This study bears important implications on data hungry fields, for example, the long-standing T cell receptor (TCR) specificity challenge.

HLA class I peptidome analysis of synthetic long peptide-fed human dendritic cells for therapeutic vaccine design

“…Likely HLYSHPIILG is the product of endosomal cathepsins rather than of the (immune) proteasome and for this reason was not predicted as a binder for HLA-A2:01 66,67 . Structure-based modeling suggests that the binding of HLYSHPIILG to HLA-A*02:01 is, similar to its shorter counterpart, likely based on the leucine at P9 68 . This suggests that the T cell-exposed surface of the HLYSHPIIL-HLA-A2:01 and HLYSHPIILG-HLA-A2:01 are highly similar, although it remains to be determined whether they are indeed similarly recognized by the same TCRs.…”

SwiftTCR: Efficient Computational Docking protocol of TCRpMHC-I Complexes Using Restricted Rotation Matrices

“…With increasing advances in artificial intelligence (AI) approaches, structure-based TCR specificity predictions may hold the promise to decipher the long-elusive TCR recognition code. The integration of structural information into AI methods has been shown to enhance the development of generalizable predictive approaches for MHC-binding peptide predictions and drug design [11]. However, the available experimental TCRpMHC structures (∼350 experimental structures in PDB [12]/ IMGT [13]) are not enough to train a powerful AI predictor, necessitating the development of computational modeling methods that encompass the immense diversity of TCR, MHC, and peptide sequences.…”