Track candidate model integrations (open weights, pip-installable, active)

[iskandr]

Tracking models worth integrating — criteria: open weights, installable via pip/conda, actively maintained.

Class I

Model	Install	Notes
MHCflurry	pip install mhcflurry	Already integrated. Affinity + presentation.
BigMHC	git clone	Already integrated. Presentation + immunogenicity.
MHCRoBERTa	GitHub	RoBERTa transfer learning for class I affinity. Published 2022. Check if pip-installable.

Class II

Model	Install	Notes
NetMHCIIpan 4.3	binary	Already integrated. DTU tool, affinity + EL.

Processing

Model	Install	Notes
Pepsickle	pip install pepsickle	Already integrated. Proteasomal cleavage (epitope model).

Not integrating

Model	Reason
HLApollo (Genentech)	Linux-only compiled binary, no Python API (#179)
BERTMHC	Unmaintained since 2020 (#162)

Open to suggestions — if you know of a model that's pip-installable with open weights, comment here.

[iskandr]

Surveyed recent models (April 2026):

• MHCRoBERTa — dead repo, 2 commits total, requires custom fairseq fork. Not integrating.
• RPEMHC (Bioinformatics 2024) — no pip package found
• ESM-MHC (2024) — no pip package found
• CapsNet-MHC (2023) — no pip package found
• HLApollo — Linux-only binary, already closed in Add HLApollo integration #179

The pattern is clear: most academic MHC prediction tools are published as one-off repos without proper packaging. The pip-installable tools we already integrate (MHCflurry, pepsickle) plus the DTU command-line tools (NetMHCpan, NetMHCIIpan) cover the practical use cases well.

Will check again when new tools appear on PyPI.

[iskandr]

2026-09-04 survey/runtime update:

• CapHLA is now tracked in Add CapHLA wrapper — modern PyTorch EL/BA for class I and II #270. It is the strongest near-term addition found: 2025 peer-reviewed EL+BA for both class I and II, MIT code/weights (~43 MB checkpoints), and its official 32-row fixture reproduced under Python 3.12 / PyTorch 2.7 CPU within 2.7e-7 without checkpoint conversion.
• ImmuneApp Evaluate ImmuneApp wrapper — class-I EL/BA + immunogenicity (legacy TF1.15) #238 successfully runs in a pinned x86 Linux Python 3.7 / TensorFlow 1.15 container, but the public repo exposes only EL, BA, and immunogenicity (not the advertised web-backend AP head). The environment is ~1.79 GB and the shallow checkout ~2.6 GB, so it is now correctly scoped and de-prioritized.
• PredIG Evaluate PredIG integration — aggregator, architectural mismatch with mhctools #214 is now peer-reviewed (Genome Medicine 2025) and has an official Docker image, but remains an aggregator over MHCflurry/NetCleave/NetCTLpan/NOAH rather than a thin base predictor. It still fits topiary/vaxrank better than mhctools.
• ImmunoStruct (Nature Machine Intelligence 2026) has released weights, but is not wrapper-ready: its own README still marks the end-to-end tool TODO, inference depends on precomputed or newly folded structures, and the code is non-commercially licensed.
• UniPMT (Nature Machine Intelligence 2025) is not a modern-runtime alternative despite its publication date: the official requirements pin PyTorch 1.2 and TensorFlow-GPU 1.13.1, and weights are distributed through Google Drive.
• HLAPepBinder is not self-contained (it requires nine upstream predictor outputs) and its repository has no license. TransHLA is allele-independent epitope propensity, so it does not map cleanly to mhctools' peptide/allele predictor contract.

For immediate class-I use, MHCflurry plus MixMHCpred 3.0 remain the practical pair. A 2025 independent benchmark reported stronger aggregate results for those two than ImmuneApp on its MHC-I dataset; CapHLA should initially be integrated as a complementary model, not declared the new default based only on its authors' benchmark.