Choosing a predictor¶
Choose the biological question first, then check the model's inputs, installation requirements, and license. The predictor matrix lists every supported class and command-line name.
By question¶
| Question | Predictors |
|---|---|
| Class I binding affinity | NetMHCpan or MHCflurry |
| Class I presentation | NetMHCpan 4.1/4.2, MHCflurry, MixMHCpred, BigMHC (EL), CapHLA |
| Class II binding or presentation | NetMHCIIpan for affinity or presentation; MixMHC2pred for presentation |
| Peptide-MHC complex stability | NetMHCstabpan |
| Proteasomal cleavage | Pepsickle or NetChop |
| Class II cleavage | NetCleave (class II) |
| TAP transport | DeepTAP |
| ERAP1 trimming | ERAMER |
| T-cell immunogenicity | Calis, PRIME, BigMHC (IM), DeepImmuno; TLimmuno2 for class II |
| Recognition by a specific TCR | NetTCR, Tulip, MixTCRpred |
| Free-peptide half-life | PeptiVerse, PlifePred2 |
| Per-bond peptidase evidence | Peptidase activity |
The family guides explain each model's output and validation limits. Prediction kinds defines the corresponding result fields and units.
By constraint¶
License¶
MHCflurry, CapHLA, SMM/SMM-PMBEC, Pepsickle, DeepTAP, DeepImmuno, and Calis are open source or built in. The DTU tools require a license from DTU. The Gfeller lab tools, BigMHC, and NetTCR have academic, non-commercial terms. See licensing before installing a model.
Downloads¶
Calis and RandomBindingPredictor need no download. Other models need weights,
reference data, or an external tool; see getting models.
Runtime¶
MHCflurry, CapHLA, SMM, Calis, Pepsickle's neural models, and NetTCR run in the current Python environment. The DTU and Gfeller tools use external executables. DeepTAP, DeepImmuno, TLimmuno2, MixTCRpred, Tulip, PeptiVerse, and PlifePred2 use a separate interpreter; see optional backends and environment variables.
Inputs¶
Calis, DeepTAP, ERAMER, PeptiVerse, and PlifePred2 accept peptides without alleles. Cleavage predictors also use flanking residues. See input shapes for the other families.
Good habits¶
- Compare predictors of the same endpoint. The recipes show how to combine results in a table.
- Compare physical values only when the units and measurement context agree. Model scores use predictor-specific scales; see score, value, and percentile rank.
- Read the immunogenicity caveats before ranking neoepitopes. Performance falls toward chance on unseen tumor neoepitopes.
- Check the known limits of each model you select.