Re-ranker Models

LAB

LightGBM training lives in the offline lab; this page registers the resulting booster so PROTEA can score predictions with it. Use Import booster to upload a fresh model.txt + spec.yaml + run.json, or Register by URI when the artefact already lives in MinIO.

⚠0 prediction set(s)⚠0 evaluation set(s)

Next: Evaluation →

LightGBM binary classifiers trained on dated GOA windows, LAFA-style: the model sees only annotations known at the earlier release date and is scored against those that appeared by the later one, never random splits. A re-ranker uses alignment, taxonomy, and aggregate features to re-score GO predictions with calibrated probabilities, replacing the raw embedding distance ranking.