Foreignness

WEIRDO exposes an allele-independent category score, using the verified model bundled with WEIRDO 2.2.0. Install it explicitly:

pip install 'mhctools[weirdo]'
mhctools predict-table --input peptides.csv --out scored.csv \
  --predictor weirdo:weirdo_foreignness:foreignness

peptides.csv must have a peptide column. The table command preserves input rows and adds the requested score column; use the Python prediction frame below when retaining full model provenance for serialization or cache replay.

from mhctools import WEIRDO

predictor = WEIRDO()
predictions = predictor.predict(["ACDEFGHI", "MTMDKSELV", "ACDEFGHI"])
frame = predictor.predict_dataframe(["ACDEFGHI"])
print(predictor.predictor_version)

The released human-self-complement-v1 policy returns one minus the model's human-category output. Higher scores mean lower predicted human-category membership under the training distribution. This is distinct from MHC presentation and immune recognition. No independent immune-response ranking benefit has been established; use in ranking is explicitly experimental. See WEIRDO's model and evaluation contract.

The model accepts canonical ASCII 8–11-mers, normalizes lowercase, rejects whitespace and ambiguous residues, preserves duplicate order and returns an empty result for an empty batch. Longer supported peptides average the category outputs of their 8-mer windows. Other lengths and reference hosts are unsupported. In particular, the model's rodent category is not a mouse-specific self reference.

Every result has kind="foreignness", an empty allele, policy semantics in its measurement context, and a predictor_version containing the adapter/package versions, immutable artifact, score policy and reference host. Keep the entire version in caches.

Prediction never downloads weights. artifact_status("weirdo") reports locally available files; integration_status("weirdo", check="reproduced") verifies the actual artifact and reference output. models_path accepts an identical explicit local copy.