openvax / topiary / 25532875627

90%

push

github

Plumb mhctools kind_support metadata through topiary predictors (#159)

* Plumb mhctools kind_support metadata through topiary predictors

Expose mhctools 3.13.7's per-(predictor, kind) MHC context metadata
on TopiaryPredictor and CachedPredictor so consumers can branch on
mhc_dependence (none / single_allele / haplotype) and mhc_class
without re-deriving it. The properties live on (predictor, kind), not
on Kind alone — e.g. MHCflurry presentation is single_allele or
haplotype depending on its configured mode.

- TopiaryPredictor.kind_support: {model_key: {kind: meta}}
- TopiaryPredictor.supported_kinds: tuple of unique kinds
- CachedPredictor.kind_support(): inferred from cached kinds; if a
  fallback exposes kind_support, prefer its entries (so haplotype
  semantics survive caching). supported_kinds property mirrors.
- Bump mhctools floor to >=3.13.7 (kind_support API).
- Bump topiary to 5.11.0 (additive minor).

* Add kind_support compatibility + processing-predictor tests

Lock in that TopiaryPredictor's per-(model_key, kind) inner dicts stay
byte-equal to what mhctools' own ``kind_support()`` returns, so consumers
can pass topiary's metadata anywhere mhctools' is accepted without
translation. Also cover the previously-missing ``mhc_dependence='none'``
case for processing predictors and ``proteasome_cleavage`` rows in
CachedPredictor (default and fallback paths).

- TopiaryPredictor: processing predictor (FakeProcessingPredictor) flows
  through with mhc_dependence='none' / mhc_class='none', and
  supported_kinds spans both binding and processing kinds.
- CachedPredictor: proteasome_cleavage rows default to single_allele/I
  (documented conservative behavior with no fallback); a processing
  fallback flows through 'none'/'none' faithfully.
- Compatibility: inner dicts match mhctools.predictor.kind_support()
  for the same model; all values stay within MHC_DEPENDENCE_VALUES /
  MHC_CLASS_VALUES vocabulary constants for both predictor types.

28 of 28 new or added lines in 3 files covered. (100.0%)

4115 of 4593 relevant lines covered (89.59%)

2.69 hits per line