Bio Immunoinformatics Mhc Binding Prediction

Version Compatibility

Reference examples tested with: MHCflurry 2.1+, pandas 2.2+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Notes specific to this skill: MHCflurry 2.2.0+ switched its backend from TensorFlow to PyTorch (Python 3.10+) — confirm mhcflurry-downloads fetch succeeded and the backend imports before scoring. NetMHCpan-4.1 and MixMHCpred are standalone academic binaries, not pip-installable; the IEDB REST API wraps NetMHCpan if a local install is unavailable. Tool versions move fast — re-verify supported-allele lists and default %Rank thresholds against current docs.

MHC Binding Prediction

"Predict which peptides bind/are presented by MHC class I" -> Score peptide-HLA class I binding affinity and natural-presentation likelihood to nominate candidate CD8 epitopes.

• Python: mhcflurry.Class1PresentationPredictor.load().predict() (pip-installable, forgiving allele parser)
• CLI: netMHCpan (field default; EL score by default, -BA adds affinity) or MixMHCpred (MS-deconvolution, EL-only)

The Single Most Important Modern Insight -- a strong predicted binder is a candidate for the next experiment, not an epitope

Binding to MHC is necessary but nowhere near sufficient for immunogenicity. The real path is a funnel: expression -> proteasomal processing -> TAP transport and loading -> stable surface display -> a cognate T cell that survived thymic selection and activates. Binding prediction addresses essentially one stage. Each downstream stage discards a large fraction of binders, so the precision of "predicted binder -> validated epitope" is low even when the binding model itself is excellent. Two operational corollaries follow. First, never report a presentation score to a collaborator as an "immunogenicity" or "epitope" probability — that is a different, far weaker prediction (immunoinformatics/immunogenicity-scoring). Second, the modern EL/MS models that now define the field learned natural presentation from mass-spec immunopeptidomes, which over-represent peptides from highly expressed proteins; the model therefore partly learns "comes from an abundant protein" as a proxy for "is presented." That bias is exactly backwards for neoantigen discovery, where the targets are mutated and often lowly expressed, living in the under-detected tail the model systematically under-ranks.

Tool Taxonomy (Class I)

BA vs EL -- the conceptual axis that determines which score to read

BA (binding affinity) models train on in-vitro competitive-binding IC50 assays and measure only whether the groove can hold the peptide thermodynamically. EL (eluted-ligand / presentation) models train on mass-spec immunopeptidomics — peptides actually eluted from MHC on real cells — so the label implicitly folds in processing, transport, editing, and surface stability. Read BA for "could this bind the groove if delivered there"; read EL/presentation for "is this likely naturally presented," which is the default and recommended output of NetMHCpan-4.1, MHCflurry's presentation predictor, and MixMHCpred. IEDB codifies the split: netmhcpan_ba = recommended-binding, netmhcpan_el = recommended-epitope. Pick by intent.

Decision Tree by Scenario

Predict Presentation with MHCflurry

Goal: Score peptides against a patient genotype and report the best-presenting allele per peptide.

Approach: Load the presentation predictor; pass alleles as a sample->genotype dict so the model reports best_allele, affinity (nM), affinity_percentile (%Rank), and presentation_score (0-1, higher = more likely presented). Supply real n_flanks/c_flanks only if the genomic context is known.

from mhcflurry import Class1PresentationPredictor

predictor = Class1PresentationPredictor.load()
df = predictor.predict(
    peptides=['SIINFEKL', 'GILGFVFTL', 'NLVPMVATV'],
    alleles={'patient1': ['HLA-A*02:01', 'HLA-A*24:02', 'HLA-B*07:02']},
    include_affinity_percentile=True,   # required: %Rank column is off by default
    verbose=0,
)
# columns: peptide, sample_name, affinity, best_allele, processing_score,
#          presentation_score, and affinity_percentile (only with the flag above)
# affinity nM: LOWER is stronger. presentation_score: HIGHER is more likely presented.

Interpret with %Rank, Not Raw nM

Goal: Classify binding strength in a way that is comparable across alleles.

Approach: Threshold on %Rank (percentile of the score against random peptides for that same allele), not on absolute IC50. The 500 nM convention is allele-biased — it over-calls permissive alleles and under-calls restrictive ones, skewing a multi-HLA patient's epitope list toward a subset of the genotype.

def classify_by_percentile(affinity_percentile):
    '''Class I %Rank cutoffs (NetMHCpan convention). LOWER percentile = stronger.
    Strong binder <= 0.5%; weak binder <= 2.0%. Use %Rank for any cross-allele
    comparison; raw nM is only meaningful within a single allele.'''
    if affinity_percentile <= 0.5:
        return 'strong'
    elif affinity_percentile <= 2.0:
        return 'weak'
    return 'non-binder'

Scan a Protein for Class I Epitopes

Goal: Enumerate candidate epitopes across a protein for a patient genotype.

Approach: Tile 8-11mers (9mers dominate real ligands), score all windows in one batched call, keep windows under the 2% weak-binder cutoff. See examples/mhc_binding.py for the full tiling-and-rank script.

def scan_protein(protein_seq, genotype, lengths=(8, 9, 10, 11)):
    from mhcflurry import Class1PresentationPredictor
    predictor = Class1PresentationPredictor.load()
    peptides = [protein_seq[i:i + k] for k in lengths for i in range(len(protein_seq) - k + 1)]
    df = predictor.predict(peptides=peptides, alleles={'patient': list(genotype)},
                           include_affinity_percentile=True, verbose=0)
    return df[df['affinity_percentile'] <= 2.0].sort_values('affinity_percentile')

Per-Method Failure Modes

Pan-model extrapolation on rare alleles

Trigger: scoring an allele with little/no training support (much of HLA-C, many non-European alleles). Mechanism: pan-models emit a confident %Rank for any allele sequence — there is no built-in "I don't know." Symptom: a flat/mushy predicted motif; calls that don't validate. Fix: check the allele is in the trained/supported list and that close pseudosequence neighbors had real ligands; downgrade confidence when extrapolating.