Improved Predictions of MHC-Peptide Binding using Protein Language Models

Nasser Hashemi, Boran Hao, Mikhail Ignatov, Ioannis Ch. Paschalidis, Pirooz Vakili, Sándor Vajda, Dima Kozakov

Abstract

Major histocompatibility complex (MHC) molecules bind to peptides from exogenous antigens and present them on the surface of cells, allowing the immune system (T cells) to detect them. Elucidating the process of this presentation is essential for regulation and potential manipulation of the cellular immune system. Predicting whether a given peptide will bind to the MHC is an important step in the above process, motivating the introduction of many computational approaches. NetMHCpan, a pan-specific model predicting binding of peptides to any MHC molecule, is one of the most widely used methods which focuses on solving this binary classification problem using a shallow neural network. The successful results of AI methods, especially Natural Language Processing (NLP-based) pretrained models in various applications including protein structure determination, motivated us to explore their use in this problem as well. Specifically, we considered fine-tuning these large deep learning models using as dataset the peptide-MHC sequences. Using standard metrics in this area, and the same training and test sets, we show that our model outperforms NetMHCpan 4.1 which has been shown to outperform all other earlier methods.

Introduction

The "Transformer revolution" that has transformed Natural Language Processing is now making significant waves in immunology. Researchers from Boston University and Stony Brook University have developed a new method for predicting MHC-peptide binding by fine-tuning large-scale protein language models (pLMs). This approach addresses the limitations of the long-standing industry standard, NetMHCpan 4.1, by leveraging the deep biological knowledge embedded in models like ESM1b.

The Challenge of MHC Binding

The Major Histocompatibility Complex (MHC) acts as the immune system's "display case." MHC molecules bind to small protein fragments (peptides) and present them on the cell surface for T-cell recognition. Accurate prediction of this binding is vital for vaccine design and cancer immunotherapy. However, the high polymorphism of MHC molecules and the vast sequence space of peptides make this a complex computational problem. Traditional models like NetMHCpan rely on shallow architectures and simplified representations of the binding site, which may miss intricate amino acid interactions.

Methodology: Fine-Tuning Protein Language Models

The proposed approach moves away from shallow networks toward Transfer Learning. The researchers adapted the ESM1b model using several innovative techniques:

  1. NLI-style Fine-tuning: Borrowing from Natural Language Inference, the MHC and peptide sequences are concatenated into a single string separated by a [SEP] token. This allows the Transformer's self-attention mechanism to explicitly model the interface between the two molecules.
  2. ESM1b-GAT: To incorporate structural information, a Graph Attention Network (GAT) layer was added on top of the ESM1b embeddings. This treats amino acids as nodes in a graph, using contact priors to refine the representation of the binding pocket.
  3. Multi-Instance Learning (MIL): To handle complex "Multi-Allele" data, a MIL framework was implemented to deconvolute signals where the specific MHC binder is unknown.
Figure 1: Model architecture overview showing the concatenation of MHC and peptide sequences into a pre-trained language model followed by GAT layers.

Key Results and Performance

The model was evaluated against NetMHCpan 4.1 using the same massive training dataset of ~13 million samples and an independent test set of 36 Eluted Ligand (EL) datasets. The results demonstrate a clear performance advantage:

  • Superior Accuracy: The ESM fine-tuning method outperformed NetMHCpan 4.1 in 35 out of 36 independent test sets when measured by Area Under the Curve (AUC).
  • Ranking Performance: Using Positive Predictive Value (PPV) across various hit-decoy ratios, the proposed model consistently ranked true binders higher than the baseline.
  • Robustness: The model showed particular strength in predicting binders for underrepresented MHC alleles and longer peptides (10-15 mers), suggesting that the structural priors and pre-training help the model generalize beyond common training examples.
Figure 2: Comparison of PPV scores between the proposed model and NetMHCpan 4.1 across various MHC groups.

Conclusion and Future Work

While this study focuses on MHC Class I, it sets a new benchmark for binding prediction. Future work will aim to address the increased complexity of MHC Class II molecules, which involve more variable peptide lengths. The researchers are currently integrating this high-accuracy model into the Cluspro server to provide the broader biological community with improved tools for immunological research.

Summary updated

Abstract

A bstract Major histocompatibility complex (MHC) molecules bind to peptides from exogenous antigens, and present them on the surface of cells, allowing the immune system (T cells) to detect them. Elucidating the process of this presentation is essential for regulation and potential manipulation of the cellular immune system [1]. Predicting whether a given peptide will bind to the MHC is an important step in the above process, motivating the introduction of many computational approaches. NetMHCPan [2], a pan-specific model predicting binding of peptides to any MHC molecule, is one of the most widely used methods which focuses on solving this binary classification problem using a shallow neural network. The successful results of AI methods, especially Natural Language Processing (NLP-based) pretrained models in various applications including protein structure determination, motivated us to explore their use in this problem as well. Specifically, we considered fine-tuning these large deep learning models using as dataset the peptide-MHC sequences. Using standard metrics in this area, and the same training and test sets, we show that our model outperforms NetMHCpan4.1 which has been shown to outperform all other earlier methods [2].
BibTeX
@misc{hashemi2022improved,
  author = {Nasser Hashemi and Boran Hao and Mikhail Ignatov and Ioannis Ch. Paschalidis and Pirooz Vakili and Sándor Vajda and Dima Kozakov},
  title = {{Improved Predictions of MHC-Peptide Binding using Protein Language Models}},
  year = {2022},
  url = {https://openreview.net/forum?id=iUeH4dWSuL},
}
Figure