Add DeepImmuno class-I immunogenicity predictor (#250)

[iskandr]

Closes #250.

Adds DeepImmuno (Li et al., Briefings in Bioinformatics 2021) — a small CNN that scores class-I CD8+ immunogenicity from a peptide and its HLA-A/B/C allele. It's the fourth immunogenicity predictor alongside Calis, PRIME, and BigMHC_IM, emitting one Kind.immunogenicity prediction per (peptide, allele).

Design

• Subprocess + sidecar interpreter (the DeepTAP pattern). DeepImmuno ships MIT-licensed weights in-repo but its deepimmuno-cnn.py rebuilds the net and loads them with an old Keras 2 / TensorFlow stack, so mhctools shells out to that CLI in a user-provided checkout (DEEPIMMUNO_HOME) via a user-provided interpreter (DEEPIMMUNO_PYTHON). On newer TensorFlow the interpreter only needs the tf-keras shim — the wrapper sets TF_USE_LEGACY_KERAS=1 for the subprocess so the Keras-2 checkpoint loads.
• 9- and 10-mers only (DeepImmuno's encoder handles no other length) — validated up front with a clear error.
• ~62 supported alleles; DeepImmuno snaps anything else to the nearest it knows, so the wrapper only normalizes punctuation (HLA-A*02:01 → HLA-A*0201).
• AsyncProcess/run_command gain cwd/env passthrough (backward-compatible) because DeepImmuno hardcodes ./data and ./models and must run in its own dir.

Verification

Confirmed end-to-end against a local DeepImmuno checkout on TF 2.17 + tf-keras: NLVPMVATV/HLA-A*02:01 = 0.9568, GILGFVFTL = 0.8872. Offline tests (parser, allele format, construction/validation) run in CI; the end-to-end tests are gated on DEEPIMMUNO_HOME. ruff clean.

⚠️ Same caveat as the other immunogenicity models: independent benchmarks put CD8 neoepitope immunogenicity near AUC 0.5–0.65 on unseen tumor neoepitopes. A prioritization aid, not ground truth.

Version bump 3.29.0 → 3.30.0.

https://claude.ai/code/session_01LZahFhBSCiehXTESCYQ7wG