Wrap ImmuneApp (and ImmuneApp-Neo) as an mhctools class-I predictor — presentation/binding (Kind.pMHC_presentation + Kind.pMHC_affinity) plus neoepitope immunogenicity (Kind.immunogenicity).
Why: it's the cleanest-licensed modern, peer-reviewed multi-task class-I package in the survey — one repo covering eluted-ligand + binding-affinity + antigen-processing heads, and (via ImmuneApp-Neo) CD8 immunogenicity.
Upstream: https://github.com/bsml320/ImmuneApp (Nat. Commun. 2024).
Pretrained weights: ✅ in-repo under models/ — separate EL, BA, AP, and immunogenicity heads. No retraining.
License: MIT — fully redistribution-clean (the most permissive of the modern class-I tools; contrast HLApollo/pep2vec, which forbid redistribution and benchmarking).
Framework / deps: ⚠️ legacy TensorFlow 1.15 / Python 3.7 — the main friction. Best isolated (pinned container or a sidecar env), like the isolation approach used for Tulip, rather than pulling TF1.15 into the main environment.
Inputs / output: peptide + class-I allele → EL score / BA / processing; ImmuneApp-Neo adds a neoepitope immunogenicity score. Map heads to pMHC_presentation, pMHC_affinity, antigen_processing, and immunogenicity kinds respectively.
Ease: Moderate — model loading is straightforward; the TF1.15/py3.7 pin is the real cost, so an isolated-env wrapper is the sane path.
Benchmark caveat: the presentation/affinity claims are self-reported (vs NetMHCpan-4.1/MHCflurry/MixMHCpred), and the immunogenicity head is subject to the field-wide ceiling — every independent immunogenicity benchmark (Buckley 2022, Nibeyro 2023, O'Brien 2023) puts neoepitope predictors near AUC 0.5–0.65. Worth it for reproducibility and the multi-task convenience, not as a validated immunogenicity oracle; state this in the docstring.
Suggested approach: lazy-import optional extra with an isolated TF1.15 runtime (Tulip-style), or subprocess.
Surfaced from a 2023–2026 survey of non-TCR immunology prediction tools/benchmarks. MIT-licensed, peer-reviewed, multi-task; legacy-TF friction.
https://claude.ai/code/session_01LZahFhBSCiehXTESCYQ7wG
Wrap ImmuneApp (and ImmuneApp-Neo) as an
mhctoolsclass-I predictor — presentation/binding (Kind.pMHC_presentation+Kind.pMHC_affinity) plus neoepitope immunogenicity (Kind.immunogenicity).Why: it's the cleanest-licensed modern, peer-reviewed multi-task class-I package in the survey — one repo covering eluted-ligand + binding-affinity + antigen-processing heads, and (via ImmuneApp-Neo) CD8 immunogenicity.
Upstream: https://github.com/bsml320/ImmuneApp (Nat. Commun. 2024).
Pretrained weights: ✅ in-repo under
models/— separate EL, BA, AP, and immunogenicity heads. No retraining.License: MIT — fully redistribution-clean (the most permissive of the modern class-I tools; contrast HLApollo/pep2vec, which forbid redistribution and benchmarking).
Framework / deps:⚠️ legacy TensorFlow 1.15 / Python 3.7 — the main friction. Best isolated (pinned container or a sidecar env), like the isolation approach used for Tulip, rather than pulling TF1.15 into the main environment.
Inputs / output: peptide + class-I allele → EL score / BA / processing; ImmuneApp-Neo adds a neoepitope immunogenicity score. Map heads to
pMHC_presentation,pMHC_affinity,antigen_processing, andimmunogenicitykinds respectively.Ease: Moderate — model loading is straightforward; the TF1.15/py3.7 pin is the real cost, so an isolated-env wrapper is the sane path.
Benchmark caveat: the presentation/affinity claims are self-reported (vs NetMHCpan-4.1/MHCflurry/MixMHCpred), and the immunogenicity head is subject to the field-wide ceiling — every independent immunogenicity benchmark (Buckley 2022, Nibeyro 2023, O'Brien 2023) puts neoepitope predictors near AUC 0.5–0.65. Worth it for reproducibility and the multi-task convenience, not as a validated immunogenicity oracle; state this in the docstring.
Suggested approach: lazy-import optional extra with an isolated TF1.15 runtime (Tulip-style), or subprocess.
Surfaced from a 2023–2026 survey of non-TCR immunology prediction tools/benchmarks. MIT-licensed, peer-reviewed, multi-task; legacy-TF friction.
https://claude.ai/code/session_01LZahFhBSCiehXTESCYQ7wG