{"architecture":{"backbone":"binding predictor (pan-allele) + processing predictor over flanking sequence; combined by logistic regression","family":"ensemble of feed-forward networks + antigen-processing network","input":"peptide encoded three times \u2014 left-aligned, centred, right-aligned \u2014 concatenated into 45 fixed positions, plus a 37-residue allele pseudosequence (34 contact residues + 3 to separate otherwise identical alleles); processing model additionally takes 15 aa of N- and C-terminal flanks","notes_en":"Where NetMHCpan searches for the right alignment, MHCflurry hands the network all three alignments at once and lets it decide. Multi-allelic data are handled by taking the best score among the alleles the sample carries, rather than by training-time deconvolution.","output":"affinity (nM), presentation score, %Rank"},"article":"/blog/modele-prezentacji-antygenu","cancer_slugs":["pan-cancer"],"category":"immunopeptidomics","confidence":"high","datasets":[{"name":"IEDB \u2014 Immune Epitope Database","note":"","role":"training","slug":"iedb"},{"name":"Mono-allelic HLA class I peptidome (Sarkizova / Abelin)","note":"","role":"training","slug":"monoallelic-peptidome"},{"name":"IEDB automated benchmark (MHC class I)","note":"","role":"benchmark","slug":"iedb-benchmark"}],"developer":"O'Donnell, Rubinsteyn and Laserson (Mount Sinai / OpenVax)","evaluation":[{"benchmark":"held-out mass-spec ligands (1:99 decoys from the same proteins)","external":true,"metric":"PPV","source":"https://www.cell.com/cell-systems/fulltext/S2405-4712(20)30331-8","value":"reported in paper; not comparable with PPV numbers computed at 1:999 by other tools"},{"benchmark":"effect of the antigen-processing model","external":false,"metric":"relative improvement","source":"https://www.cell.com/cell-systems/fulltext/S2405-4712(20)30331-8","value":"+3.3% on average"}],"hf":null,"kind":"task-model","license":"apache-2.0","limitations":"- Same assumed-negative problem and MS biases as every model in this family.\n- Decoys drawn from the same proteins make its PPV numbers stricter than some competitors' and not directly comparable.\n- One published benchmark had to exclude it because of train/test overlap \u2014 check your evaluation split.","links":{"demo":null,"docs":null,"doi":"10.1016/j.cels.2020.06.010","github":"https://github.com/openvax/mhcflurry","huggingface":null,"paper":"https://www.cell.com/cell-systems/fulltext/S2405-4712(20)30331-8","pmid":null},"modalities":["protein-sequence","immunopeptidomics"],"name":"MHCflurry 2.0","notable_uses":"The processing model recovers known biology: proline is depleted just before the peptide and enriched at position 2, matching the action of ERAP trimming. The measured gain from adding processing is real but modest \u2014 3.3% on average.","openness":"open-weights","regulatory":{"intended_use_en":"Research software.","intended_use_pl":"Oprogramowanie badawcze.","source_url":"https://github.com/openvax/mhcflurry","status":"research-only"},"regulatory_status":"research-only","release_date":"2020-07-08","run_snippet":"pip install mhcflurry\nmhcflurry-downloads fetch models_class1_presentation\n\nfrom mhcflurry import Class1PresentationPredictor\npredictor = Class1PresentationPredictor.load()\npredictor.predict(\n    peptides=['SIINFEKL', 'KLGGALQAK'],\n    alleles={'sample1': ['HLA-A*02:01', 'HLA-B*57:01']},\n    # n_flanks/c_flanks add the antigen-processing signal when you have the source protein\n)","settings":["basic-research","vaccine-design","immunotherapy"],"slug":"mhcflurry","sources":[{"label":"O'Donnell TJ, Rubinsteyn A, Laserson U. MHCflurry 2.0. Cell Systems 2020","url":"https://www.cell.com/cell-systems/fulltext/S2405-4712(20)30331-8"},{"label":"GitHub \u2014 openvax/mhcflurry","url":"https://github.com/openvax/mhcflurry"},{"label":"cancer3.ai \u2014 Trzydzie\u015bci cztery litery zamka","url":"https://cancer3.ai/blog/modele-prezentacji-antygenu"}],"summary":"Open-source pan-allele presentation predictor whose distinctive idea is a separate antigen-processing model: it reads the peptide together with fifteen amino acids of flanking sequence on each side, because the proteasome cut depends on what lies around the cut site.","tasks":["antigen-presentation","peptide-mhc-binding","neoantigen-prioritisation"],"training":{"institutions":"public IEDB and published immunopeptidomics datasets","summary_en":"IEDB binding affinities plus MS-eluted ligands from mono- and multi-allelic samples; negatives are decoys drawn from the same source proteins as the hits.","summary_pl":"Powinowactwa z IEDB oraz ligandy ze spektrometru z pr\u00f3bek jedno- i wieloallelicznych; negatywy to atrapy losowane z tych samych bia\u0142ek \u017ar\u00f3d\u0142owych co trafienia."},"updated_at":"2026-09-05T22:26:00.840541","url":"/ai-oncology/models/mhcflurry","usage":{"hardware":"CPU is fine; TensorFlow backend","library":"Python (pip install mhcflurry)","notes_en":"The presentation predictor needs the downloaded model bundle (mhcflurry-downloads fetch). Supplying flanking sequences is what activates the processing model \u2014 without them you are running the binding network only.","snippet":"pip install mhcflurry\nmhcflurry-downloads fetch models_class1_presentation\n\nfrom mhcflurry import Class1PresentationPredictor\npredictor = Class1PresentationPredictor.load()\npredictor.predict(\n    peptides=['SIINFEKL', 'KLGGALQAK'],\n    alleles={'sample1': ['HLA-A*02:01', 'HLA-B*57:01']},\n    # n_flanks/c_flanks add the antigen-processing signal when you have the source protein\n)"},"verified_at":"2026-09-05T22:26:00.839223","verified_by":"editorial","version":"2.0","what_it_does":"Predicts binding affinity and presentation; the final presentation score is a three-parameter logistic regression combining the binding network and the processing network. Installable with pip, which makes it the practical choice for reproducible pipelines."}