MHCflurry 2.3.8
MHCflurry 2.3.8
This maintenance release preserves training provenance and makes evaluation
overlap claims more precise. Default weights remain 2.3.0; frozen-weight
predictions are unchanged.
- New affinity curation preserves study/sample/assay identities, source hashes
and raw row references, including every contributor to a deduplicated row. - New holdout policies exclude known whole source samples or studies in addition
to peptide–MHC pairs. Legacy policies retain their original scope. mhcflurry train release-holdout audit-samplesaudits all compared models'
components and pretraining/training/development/selection sources, resolves
explicit specimen aliases, and exports one shared cohort with hashes/counts.- Missing historical or external lineage remains unresolved. Passing the old
peptide/sample-ID checks is never presented as complete sample-disjoint proof.
This release does not retrain or republish weights and does not retrospectively
certify historical sample separation. See issue #444 for that remaining evidence
limitation and #469 for the planned processing-quality investigation.