1217 A novel transformer-based deep learning model for predicting binding interactions between HLA class I molecules and peptides

Le résumé fourni par la source

Background Predicting the binding affinities between human leukocyte antigen class I molecules (HLA-I) and peptides is crucial for understanding immune responses and developing immunotherapies. Although computational models have been developed to facilitate efficient and accurate screening of ligands with high binding probabilities, they often fail to generalize on peptide sequences of varying lengths. Deep learning models have been proposed, but they tend to bias toward common peptide lengths found in training data and often lack interpretability to reveal underlying mechanisms driving binding. Methods We introduce a novel deep learning model designed to predict the bindings between HLA-I molecules and peptides (pHLA-I). Our model, based on transformer architecture, encodes the amino acid sequences of HLA-I molecules and peptides using a co-attention mechanism that iteratively updates the representations of both sequences to capture their crucial interactions driving binding. The final representations of the peptides’ classification tokens (CLS) are then fed into a classification head to predict binding probabilities. Results Our model successfully competes against established pHLA-I prediction models, including TransPHLA,1 another transformer-based deep learning model using a self-attention mechanism, achieving an F1 score of 92.8% for independent test data and 87.8% for external data. Notably, our model is more robust on longer peptides with more than 12 residues, which constitute less than 5% of the training data. On the external data, our model showed F1-scores for binding predictions with peptides of lengths 12, 13, and 14 that were 7.2%, 6.9%, and 21.4% higher, respectively, compared to TransPHLA1 (F1-scores: 0.74, 0.62, and 0.51). Additionally, we facilitated the inference of binding motifs for each HLA-I allele with different ligand lengths from the attention scores at co-attention layers. We observed that the reconstructed binding motifs from attention scores aligned with binding motifs of well-studied HLA-I alleles. Conclusions Our model not only provides robust and accurate pHLA-I binding predictions with enhanced interpretability but also facilitates the inference of binding motifs for rare and under-examined HLA alleles. Furthermore, the updated representations of HLA and peptide sequences generated by our model could potentially extend to modeling other immunological tasks, such as T-cell receptor immunogenicity prediction, thereby broadening the model’s application in immunotherapy and vaccine development. Reference Chu Y, Zhang Y, Wang Q, et al. A transformer-based model to predict peptide–HLA class I binding and optimize mutated peptides for vaccine design. Nat Mach Intell. 2022;4:300–311. https://doi.org/10.1038/s42256-022-00459-7.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.