Peptide-binding motif of HLA-A*6603

ACME: pan-specific peptide–MHC class I binding prediction through attention-based deep neural networks

Motivation

Prediction of peptide binding to the major histocompatibility complex (MHC) plays a vital role in the development of therapeutic vaccines for the treatment of cancer. Algorithms with improved correlations between predicted and actual binding affinities are needed to increase precision and reduce the number of false positive predictions.

Results

We present ACME (Attention-based Convolutional neural networks for MHC Epitope binding prediction), a new pan-specific algorithm to accurately predict the binding affinities between peptides and MHC class I molecules, even for those new alleles that are not seen in the training data. Extensive tests have demonstrated that ACME can significantly outperform other state-of-the-art prediction methods with an increase of the Pearson correlation coefficient between predicted and measured binding affinities by up to 23 percentage points. In addition, its ability to identify strong-binding peptides has been experimentally validated. Moreover, by integrating the convolutional neural network with attention mechanism, ACME is able to extract interpretable patterns that can provide useful and detailed insights into the binding preferences between peptides and their MHC partners. All these results have demonstrated that ACME can provide a powerful and practically useful tool for the studies of peptide–MHC class I interactions.

Availability and implementation

ACME is available as an open source software at https://github.com/HYsxe/ACME.

Supplementary information

Supplementary data are available at Bioinformatics online.

Pocketcheck: updating the HLA class I peptide specificity roadmap

The structural determination of peptide:HLA (human leucocyte antigen) class I complexes by X-ray crystallography has provided valuable information for understanding how peptides bind to individual HLA class I molecules and how this may influence the immune response. We compared 101 crystal structures of 9-mer peptide:HLA class I complexes available in the protein data bank (PDB) by performing a contact analysis using the Contact Map Analysis webserver http://ligin.weizmann.ac.il/cma. An InterSystems Caché 'post-relational' database containing residue position, amino acid (AA) and buried surface that contact a particular peptide position was then created allowing data comparison for all the structures (Pocketcheck). The analysis illustrates that the HLA class I residues 24, 45, 63 and 67 show high contact frequencies to both the p1 and/or p2 position of bound peptides, indicating that they might influence the nature of a peptide anchor. To determine the influence of these residues we utilized soluble HLA technology and mass spectrometry to analyze peptides derived from HLA-B*44:06 since it differs from the previously described allele B*44:02 by seven AA exchanges located in the alpha 1 domain (residues 24, 32, 41, 45, 63, 67 and 80). HLA-B*44:06 features an anchor motif of P or A at p2 and Y or W at the C-terminal. Additionally B*44:06-derived peptides feature an auxiliary anchor motif at p1, comprising D or E. Our results illustrate that structural analysis can provide valuable information to understand allogenicity and provides a further step towards intelligent HLA mismatching.

The impact of human leukocyte antigen (HLA) micropolymorphism on ligand specificity within the HLA-B*41 allotypic family

BACKGROUND

Polymorphic differences between human leukocyte antigen (HLA) molecules affect the specificity and conformation of their bound peptides and lead to differential selection of the T-cell repertoire. Mismatching during allogeneic transplantation can, therefore, lead to immunological reactions.

DESIGN AND METHODS

We investigated the structure-function relationships of six members of the HLA-B*41 allelic group that differ by six polymorphic amino acids, including positions 80, 95, 97 and 114 within the antigen-binding cleft. Peptide-binding motifs for B*41:01, *41:02, *41:03, *41:04, *41:05 and *41:06 were determined by sequencing self-peptides from recombinant B*41 molecules by electrospray ionization tandem mass spectrometry. The crystal structures of HLA-B*41:03 bound to a natural 16-mer self-ligand (AEMYGSVTEHPSPSPL) and HLA-B*41:04 bound to a natural 11-mer self-ligand (HEEAVSVDRVL) were solved.

RESULTS

Peptide analysis revealed that all B*41 alleles have an identical anchor motif at peptide position 2 (glutamic acid), but differ in their choice of C-terminal pΩ anchor (proline, valine, leucine). Additionally, B*41:04 displayed a greater preference for long peptides (>10 residues) when compared to the other B*41 allomorphs, while the longest peptide to be eluted from the allelic group (a 16mer) was obtained from B*41:03. The crystal structures of HLA-B*41:03 and HLA-B*41:04 revealed that both alleles interact in a highly conserved manner with the terminal regions of their respective ligands, while micropolymorphism-induced changes in the steric and electrostatic properties of the antigen-binding cleft account for differences in peptide repertoire and auxiliary anchoring.

CONCLUSIONS

Differences in peptide repertoire, and peptide length specificity reflect the significant functional evolution of these closely related allotypes and signal their importance in allogeneic transplantation, especially B*41:03 and B*41:04, which accommodate longer peptides, creating structurally distinct peptide-HLA complexes.

Implementing the modular MHC model for predicting peptide binding

The challenge of predicting which peptide sequences bind to which major histocompatibility complex (MHC) molecules has been met with various computational techniques. Scoring matrices, hidden Markov models, and artificial neural networks are examples of algorithms that have been successful in MHC-peptide-binding prediction. Because these algorithms are based on a limited amount of experimental peptide-binding data, prediction is only possible for a small fraction of the thousands of known MHC proteins. In the primary field of application for such algorithms--vaccine design--the ability to make predictions for the most frequent MHC alleles may be sufficient. However, emerging applications of leukemia-specific T cells require a patient-specific MHC-peptide-binding prediction. The modular model of MHC presented here is an attempt to maximize the number of predictable MHC alleles, based on a limited pool of experimentally determined peptide-binding data.

A modular concept of HLA for comprehensive peptide binding prediction

A variety of algorithms have been successful in predicting human leukocyte antigen (HLA)-peptide binding for HLA variants for which plentiful experimental binding data exist. Although predicting binding for only the most common HLA variants may provide sufficient population coverage for vaccine design, successful prediction for as many HLA variants as possible is necessary to understand the immune response in transplantation and immunotherapy. However, the high cost of obtaining peptide binding data limits the acquisition of binding data. Therefore, a prediction algorithm, which applies the binding information from well-studied HLA variants to HLA variants, for which no peptide data exist, is necessary. To this end, a modular concept of class I HLA-peptide binding prediction was developed. Accurate predictions were made for several alleles without using experimental peptide binding data specific to those alleles. We include a comparison of module-based prediction and supertype-based prediction. The modular concept increased the number of predictable alleles from 15 to 75 of HLA-A and 12 to 36 of HLA-B proteins. Under the modular concept, binding data of certain HLA alleles can make prediction possible for numerous additional alleles. We report here a ranking of HLA alleles, which have been identified to be the most informative. Modular peptide binding prediction is freely available to researchers on the web at http://www.peptidecheck.org .