MHC Binding Prediction - Usage Guide

Overview

Predict peptide-MHC class I binding affinity and natural-presentation likelihood to nominate candidate CD8 T-cell epitopes, with the methodological judgment to read the right score (BA vs EL), threshold correctly (%Rank not nM), and know where the predictions fail (rare alleles, low-expression neoantigens). For CD4/HLA class II, see the mhc-class-ii-prediction skill.

Prerequisites

pip install mhcflurry
mhcflurry-downloads fetch
# NetMHCpan-4.1 and MixMHCpred are standalone academic downloads (not pip);
# the IEDB REST API wraps NetMHCpan if a local install is unavailable.

Quick Start

Tell your AI agent what you want to do: - "Predict class I presentation for these peptides with my patient's HLA genotype" - "Scan this protein for 9-mer epitopes and rank by %Rank" - "Which of these neoantigen peptides are presented, accounting for expression?" - "Should I read the binding-affinity or the eluted-ligand score for this question?"

Example Prompts

Presentation Prediction

"Score GILGFVFTL and NLVPMVATV against HLA-A02:01, A24:02, B*07:02 and report the best-presenting allele"

"Classify these peptides as strong/weak/non-binders using %Rank, not raw nM"

Protein Scanning

"Tile this spike protein into 8-11mers and return windows under 2% Rank for common HLA-A alleles"

"Find candidate CD8 epitopes in this antigen for a specific patient genotype"

Score Choice and Caveats

"Is this a binding-affinity question or a presentation question, and which tool/score fits?"

"These neoantigens are lowly expressed; how do I avoid the EL abundance bias misranking them?"

"This allele is rare and non-European; how much should I trust the prediction?"

What the Agent Will Do

1. Confirm the patient HLA class I genotype and whether the question is "can it bind" (BA) or "is it presented" (EL/presentation)
2. Choose a predictor (MHCflurry for scripting; NetMHCpan-4.1 for broadest coverage; MixMHCpred for MS-grounded motifs)
3. Score peptides in a batched call and report affinity (nM), %Rank, and presentation score
4. Threshold on %Rank (strong <= 0.5%, weak <= 2.0%) for cross-allele comparability
5. Flag rare/extrapolated alleles and integrate expression for neoantigen ranking
6. Caveat that a predicted binder is a candidate, not a validated epitope

Tips

• BA vs EL - read binding-affinity for "can it physically bind"; read eluted-ligand/presentation for "is it naturally presented" (the default question)
• %Rank, not nM - the 500 nM convention is allele-biased; always threshold on %Rank for multi-HLA work
• Abundance bias - EL/MS models over-rank peptides from abundant proteins; integrate expression when scoring low-expression neoantigens
• Rare alleles - pan-models extrapolate and never say "I don't know"; verify training support before trusting a number
• Length - 8-11mers, 9mers dominate; non-9mers rest on thinner training data
• Binding != epitope - presentation is one stage of a multi-stage funnel; do not report it as immunogenicity

Related Skills

• immunoinformatics/mhc-class-ii-prediction - CD4/HLA class II binding (open groove, register, DQ pairing)
• immunoinformatics/neoantigen-prediction - applies class I binding to tumor mutations
• immunoinformatics/immunogenicity-scoring - the separate, weaker prediction of T-cell response
• clinical-databases/hla-typing - determine the patient genotype