# Peptide-HLA binding and ligand prediction

Source: https://onco.cc/terms/peptide-hla-binding-prediction/  
OnCo record `peptide-hla-binding-prediction` (Term). Data CC BY-NC 4.0, attribute "Data from OnCo (onco.cc)"; commercial use needs a licence.

## TL;DR

A computer model can prioritize peptides expected to bind an immune display molecule or appear among its ligands. The prediction helps select candidates for testing.

## Summary

NetMHCpan-4.0 combined peptide-binding affinity data with eluted-ligand data, which also contain information about processing and natural presentation. Those training signals and prediction tasks should be distinguished. A consortium comparison assessed candidate tumor epitopes using measured immune interactions and developed improved prioritization features. A model score is a prediction with a validation setting; it is not an observation of peptide display, tumor recognition or patient benefit. The NetMHCpan-4.1 documentation expresses binder thresholds as percentile rank, the position of a predicted score against the scores of random natural peptides, and lets the user set the strong and weak binder thresholds, so a 0.4 per cent rank places a candidate within the top 0.4 per cent of that reference and is not a 99.6 per cent probability of presentation, recognition or benefit. The server predicts binding of a peptide to an MHC molecule of known sequence, and its model is trained on more than 850,000 binding-affinity and mass-spectrometry eluted-ligand peptides, two different training signals (Reynisson and colleagues, 2020). A model trained on more than 185,000 peptides eluted from 95 single-allele cell lines improved prediction of presented peptides (Sarkizova and colleagues, 2020). In a consortium study, 608 predicted epitopes were then tested for T-cell binding in patient-matched samples, a separate experiment from the prediction (Wells and colleagues, 2020), so expression, processing, display, T-cell receptor recognition and sparing of healthy tissue remain to be shown after a favourable score.

## Fields

- Kind: Term
- Last checked: 2026-11-10
- Also known as: peptide-MHC binding prediction; HLA ligand prediction

## Notes

- Assay distinction. A predicted binding IC50 estimates a competitive-binding assay endpoint. An IC50 depends on the assay conditions and is not automatically an equilibrium dissociation constant. Cer and colleagues' 2009 IC50-to-Ki tool calculates binding or inhibition constants using the interaction mechanism, concentrations and other assumptions; this is a biochemical distinction, not validation of an HLA model or a measure of cancer-cell growth inhibition. Reading guide (OnCo inference): Record the model and version, allele, peptide length, prediction task, units, reference distribution and threshold beside every score, and do not put a rank and an IC50 in one unlabelled column. A threshold is a selection rule for a stated workflow and cannot repair the wrong allele or an unsupported task. Inputs are peptide or source-protein sequences and an exact allele; scoring uses no tumour tissue; held-out validation, allele coverage and the endpoint (binding, recovery or T-cell recognition) decide what a good score means.
- Figure reading (fictional numbers, light and dark renders agree): a 1,000-score reference for HLA-A*02:01 9-mer peptides has bin counts 250, 240, 200, 130, 80, 50, 30, 16 and 4, which sum to 1,000. Four reference scores lie beyond the candidate boundary, so 4 / 1,000 x 100 = 0.4 per cent. A separate binding-affinity output of 40 nM is a predicted IC50. Neither value observes display.
- Fictional worked example. Candidate A has an eluted-ligand rank of 0.4 per cent and candidate B a predicted binding IC50 of 40 nM. The numbers come from different outputs with different units and tasks, so neither ranks the candidates against each other or shows tumour coverage; request matched outputs for the relevant allele and method, and test display and recognition separately.
- Output table in words. Predicted binding IC50 in nanomolar units: lower values mean stronger predicted competitive binding for that output, and it is not a measured concentration, an equilibrium dissociation constant or a functional T-cell response. Percentile rank: a lower rank places the candidate nearer the model's stronger-scoring reference peptides, and it is not a percent chance of tumour killing or the fraction of tumour cells covered.

## Sources

- NetMHCpan-4.0: Improved Peptide-MHC Class I Interaction Predictions Integrating Eluted Ligand and Peptide Binding Affinity Data.: https://doi.org/10.4049/jimmunol.1700893
- Key Parameters of Tumor Epitope Immunogenicity Revealed Through a Consortium Approach Improve Neoantigen Prediction.: https://doi.org/10.1016/j.cell.2020.09.015
- NetMHCpan-4.1 server documentation: https://services.healthtech.dtu.dk/services/NetMHCpan-4.1/
- NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data (Nucleic Acids Research 2020): https://doi.org/10.1093/nar/gkaa379
- A large peptidome dataset improves HLA class I epitope prediction across most of the human population (Nature Biotechnology 2020): https://doi.org/10.1038/s41587-019-0322-9
- Cer et al., IC50-to-Ki: conversion assumptions for enzyme inhibition and ligand binding (Nucleic Acids Research 2009): https://doi.org/10.1093/nar/gkp253

## Connected records

- terms: [Immune measurements and product evidence](https://onco.cc/terms/immune-product-evidence/), [Immunopeptidomics](https://onco.cc/terms/immunopeptidomics/), [Neoantigen](https://onco.cc/terms/neoantigen/), [Peptide binding, presentation and immune recognition](https://onco.cc/terms/binding-presentation-recognition/)

---
JSON: https://onco.cc/api/v1/entities/peptide-hla-binding-prediction.json