Improving MHC Class I antigen processing predictions using representation learning and cleavage site-specific kernels

Patrick J. Lawrence, Xia NingbioRxiv 2021 · DOI

TL;DR

This paper introduces MHCrank, a novel deep learning model that significantly improves predictions of whether peptides will be processed for presentation by MHC Class I molecules. By incorporating learned amino acid representations and cleavage site-specific kernels, MHCrank outperforms existing state-of-the-art methods, paving the way for better drug and vaccine development.

What Problem Does This Paper Address?

The immune system relies on Major Histocompatibility Complex (MHC) Class I molecules to present fragments of intracellular proteins (peptides) to cytotoxic T cells, signaling for the elimination of infected or cancerous cells. For effective vaccine and drug design, it's crucial to predict which peptides will be processed and presented. While existing methods excel at predicting peptide binding affinity to MHC molecules, they often overlook the crucial upstream processing steps. This gap means that highly ranked peptides might not actually be presented, limiting their utility. This paper tackles this challenge by developing a model that specifically predicts the likelihood of a peptide being processed for MHC Class I presentation, aiming to bridge the gap between binding affinity and actual presentation.

Methods

The core of this work is the MHCrank model, a deep learning architecture designed to predict antigen processing. Key innovations include:

  • Cleavage Site-Specific Kernels: MHCrank utilizes a "cleavage site-specific kernel" (CSSK) that focuses on the C-terminal end of a peptide and the adjacent amino acids from the original protein. This is motivated by the observation that the C-terminus is critical for MHC binding and that specific motifs at this site are recognized by processing enzymes.
  • Learned Amino Acid Representations: Instead of relying on pre-defined matrices like BLOSUM, MHCrank learns problem-specific embeddings for each amino acid. This allows the model to capture nuanced relationships between amino acids that are relevant to the antigen processing task.
  • Sequence Pre-processing: MHCrank standardizes peptide sequences to a uniform length (10 amino acids) by padding or trimming, with a specific bias towards the N-terminal side for odd-length peptides. This is based on the understanding that C-terminal amino acids are more influential for processing.
  • Ensemble Methods: To further boost performance and robustness, the authors employ several ensemble strategies, combining predictions from multiple trained MHCrank models.

Datasets

The study utilizes datasets previously used to train and evaluate the MHCflurry-AP predictor.

  • Training Data: Comprised of 100 mass spectrometry (MS) experiments, including 8,537,960 peptides and 92 different HLA alleles. This data consists of "hits" (peptides confirmed to be processed for MHC Class I presentation) and randomly generated "decoys." The data was split into four training folds, with 10% of each fold reserved for validation.
  • Testing Data: Consists of 20 experiments from two studies, identifying 27,007 binding peptides. To create a fair benchmark, 99 decoys were generated for each hit, resulting in 2,700,700 peptides. Duplicates were removed, leaving 2,409,183 peptides. This dataset was specifically chosen because it was not available during the training of baseline methods, ensuring a rigorous evaluation.

Key Insights and Ideas

  • Processing Matters More Than Just Binding: The study strongly suggests that predicting peptide processing is a critical, often overlooked, step for accurate antigen presentation prediction, and that models focusing solely on binding affinity can be misleading.
  • C-terminal Focus is Key: The C-terminal cleavage site and its surrounding amino acids hold significant predictive power for antigen processing, justifying the development of specialized kernels for this region.
  • Learned Embeddings Outperform Static Matrices: The ability of MHCrank to learn amino acid representations from the data, rather than relying on fixed BLOSUM matrices, leads to a more nuanced understanding of amino acid interactions relevant to processing.
  • Central Amino Acids are Less Informative: The pre-processing step that trims central amino acids and the observed lack of enrichment in these regions for top-ranked peptides suggest they are less critical for MHC Class I presentation prediction.

Key Results

  • Significant Performance Improvement: MHCrank consistently outperformed two leading baseline methods, MHCflurry-AP and netMHCpan-EL, across various evaluation metrics (AUC, precision@k, NDCG@k).
  • AUC Gains: MHCrank ensembles achieved AUC values consistently above 0.9, indicating strong ability to distinguish between processed peptides (hits) and non-processed ones (decoys). The best performing ensemble, C-top2, reached an AUC of 0.9153.
  • Precision and NDCG Improvements: MHCrank ensembles showed significant improvements in precision@k and NDCG@k, particularly for higher values of k (e.g., k=500), demonstrating its ability to rank truly processed peptides higher. For instance, the Fw-top2 ensemble showed a percentage change of 10.23% in precision@k and 10.26% in NDCG@k over netMHCpan-EL at k=500.
  • Learned Embeddings Correlate with Physicochemical Properties: The cosine similarity matrix of learned amino acid embeddings showed strong correlations with known physicochemical properties, suggesting the model learned biologically relevant features.

Limitations and Future Work

  • Class Imbalance: The authors acknowledge that the class imbalance in the test data (few hits compared to decoys) might have optimistically inflated AUC scores.
  • Combinatorial Approach: Future work could integrate binding affinity predictions with processing predictions, similar to existing "presentation score predictors," to further enhance accuracy.
  • Protein Structure Integration: The current models rely solely on peptide sequences. Future research could explore incorporating protein structural data, which may offer a more complete understanding of antigen processing and presentation. Geometric deep learning models are suggested as a promising avenue for this.
  • Predicting Response Magnitude and Type: Beyond predicting presentation, future models could aim to predict the magnitude and type of immune response elicited by a presented peptide, potentially by analyzing interactions with T-cell receptors.

Summary updated

Abstract

Abstract In this work, we propose a new deep learning model, MHCrank , to predict the probability that a peptide will be processed for presentation within the MHC Class I pathway. We find that the performance of our model is significantly higher than two previously published baseline methods: MHCflurry and netMHCpan . Gains in performance result from the utilization of cleavage site-specific kernels and learned representations for amino acids. By visualizing the site-specific amino acid enrichment among top-ranked peptides, we find MHCrank’s top-ranked peptides are enriched at biologically relevant positions with amino acids that are consistent with previous work. Furthermore, the cosine similarity matrix derived from MHCrank’s learned embeddings for amino acids correlate highly with physiochemical properties that have been experimentally shown to be important in determining a peptide’s favorability to be processed. Altogether, the results reported in this work indicate that the proposed MHCrank demonstrates strong performance compared to existing methods and could have vast applicability to aid drug and vaccine development.
BibTeX
@misc{lawrence2021improving,
  author = {Patrick J. Lawrence and Xia Ning},
  title = {{Improving MHC Class I antigen processing predictions using representation learning and cleavage site-specific kernels}},
  year = {2021},
  doi = {10.1101/2021.10.17.464741},
  howpublished = {bioRxiv},
}
Figure