Peptide-MHC Binding Prediction

Each MHC allele possesses a unique binding groove characterized by specific anchor residue preferences. Prediction algorithms must model these polymorphic pockets.

• Anchor Residues: Specific positions (often P2 and P9 for MHC class I) that fit into allele-defined pockets.
• Polymorphism: Thousands of human leukocyte antigen (HLA) alleles exist, each requiring distinct predictive models.
• Pan-Specific Models: Modern neural networks learn a shared representation across alleles, enabling predictions for rare or uncharacterized HLA types.

Two distinct but related prediction tasks exist, often conflated in the literature.

• Binding Affinity (IC50): A regression task predicting the half-maximal inhibitory concentration in nanomolar units. Values below 50 nM typically indicate strong binders.
• Binary Ligand Classification: A classification task predicting whether a peptide is a naturally presented ligand, trained on mass spectrometry-eluted ligand data.
• Key Distinction: Not all binders are presented; ligand prediction incorporates antigen processing and transport steps implicitly.

Prediction methodologies fall into two primary architectural paradigms.

• Sequence-Based: Models like NetMHCpan use the peptide's linear amino acid sequence and the MHC pseudo-sequence as input, leveraging convolutional or transformer architectures.
• Structure-Based: Methods like Rosetta FlexPepDock or AlphaFold-based docking explicitly model the 3D geometry of the peptide-MHC complex.
• Trade-off: Sequence methods are fast and scalable for genome-wide screens; structure methods provide mechanistic insight into binding conformation but are computationally expensive.

MHC class I molecules predominantly bind peptides of 9 amino acids, but longer peptides are common in input data.

• Core Identification: Algorithms must first identify the optimal 9-mer binding core within a longer peptide sequence.
• Flanking Residues: Amino acids outside the core (PFRs) can influence binding stability and T-cell receptor recognition.
• Class II Complexity: MHC class II binding grooves are open-ended, accommodating peptides of highly variable lengths (typically 13-25 residues), making core registration significantly more challenging.

Model performance is fundamentally constrained by the quality and diversity of training data.

• Binding Affinity Databases: IEDB (Immune Epitope Database) contains curated quantitative affinity measurements, heavily biased toward common alleles and viral epitopes.
• Mass Spectrometry Data: Immunopeptidomics datasets provide in vivo presentation data but suffer from detection bias toward abundant peptides.
• Negative Set Definition: Defining true negatives (non-binders) is critical; random natural peptides are often used as a proxy, introducing noise.

Standardized metrics are essential for benchmarking predictor performance.

• AUC-ROC: Area Under the Receiver Operating Characteristic curve, measuring overall discriminative power.
• PPV at Fixed Sensitivity: Positive Predictive Value at a high sensitivity threshold (e.g., 80%) is clinically relevant for epitope discovery.
• Rank-Based Metrics: The percentile rank of a peptide's predicted affinity relative to a large background set of natural peptides is often more robust than raw IC50 values.
• External Validation: Performance on truly blind, prospective epitope discovery studies is the ultimate benchmark.

A set of cell surface proteins essential for the adaptive immune system. MHC Class I molecules present endogenous peptides (8-11 amino acids) to CD8+ cytotoxic T cells, while MHC Class II molecules present exogenous peptides (13-25 amino acids) to CD4+ helper T cells. The human MHC is called the Human Leukocyte Antigen (HLA) system, with thousands of allelic variants creating diverse peptide-binding specificities across populations. Each allele binds a distinct repertoire of peptides defined by anchor residue preferences at specific pocket positions.

A specific peptide fragment derived from a protein antigen that is recognized by a T-cell receptor (TCR) when presented on an MHC molecule. T-cell epitopes are the fundamental unit of adaptive immune recognition and the primary target of peptide-MHC binding prediction algorithms. Key characteristics include:

• Linear sequence: Unlike B-cell epitopes, T-cell epitopes are contiguous amino acid stretches
• MHC restriction: A given peptide is only immunogenic when bound to a specific MHC allele
• Proteasomal processing: Epitopes must be generated by cellular proteolytic machinery before MHC loading

The quantitative measure of peptide-MHC binding strength, typically expressed as the half-maximal inhibitory concentration (IC₅₀) in nanomolar (nM) units. Peptides with IC₅₀ values below 50 nM are generally classified as high-affinity binders, while those below 500 nM are considered weak binders. Prediction models are often trained on IC₅₀ data from competitive binding assays and output either:

• Continuous affinity predictions (regression)
• Binary binder/non-binder classifications (threshold-based)
• Percentile rank scores normalized against a background peptide distribution

A machine learning approach that trains a single model to predict binding across multiple MHC alleles simultaneously, rather than building separate models for each allele. Pan-specific methods leverage the pseudo-sequence concept—extracting only the amino acid residues forming the peptide-binding groove—to represent MHC molecules as feature vectors. This enables:

• Zero-shot prediction for alleles with no experimental binding data
• Generalization across the HLA supertype families
• Scalability to the thousands of known HLA variants Leading pan-specific tools include NetMHCpan and MHCflurry, which use artificial neural networks trained on mass spectrometry-eluted ligand data.

The large-scale experimental identification of peptides naturally presented on MHC molecules using liquid chromatography-tandem mass spectrometry (LC-MS/MS). Unlike affinity assays that measure binding in vitro, immunopeptidomics captures the end result of the entire antigen processing and presentation pathway, including:

• Proteasomal cleavage and transport via TAP
• MHC loading assisted by chaperone proteins
• Natural peptide abundance on the cell surface These datasets provide the gold-standard training data for modern deep learning predictors and have largely replaced synthetic peptide binding assays for model development.

The cellular machinery that generates peptide ligands for MHC presentation, a critical upstream step that determines which peptides are available for binding. Key components include:

• Proteasome: Degrades cytosolic proteins into peptide fragments; the immunoproteasome variant produces peptides with hydrophobic C-termini preferred by MHC Class I
• TAP (Transporter Associated with Antigen Processing): Translocates peptides into the endoplasmic reticulum with length and sequence selectivity
• ERAP (Endoplasmic Reticulum Aminopeptidase): Trims peptides to optimal MHC Class I binding length (8-10 residues) Integrated prediction pipelines combine proteasomal cleavage, TAP transport, and MHC binding scores for comprehensive epitope identification.