openvax / topiary / 25672919678

90%

push

github

Add best-allele aggregation for haplotype-mode presentation (#158) (#160)

* Add best-allele aggregation accessors for haplotype-mode predictors

Issue #158 calls out that pMHC_presentation from haplotype-mode
predictors (MHCflurry's class1_presentation in haplotype mode) is
conceptually a multi-allele input → per-peptide answer, but mhctools
emits one row per (peptide, allele) so the DSL's per-allele grouping
returns per-allele scores by default. Add an explicit aggregator that
returns the best score across alleles plus the corresponding allele
attribution, gated behind new accessors so the default per-allele
behavior is preserved.

- BestAlleleField DSLNode: per-peptide max/min of (kind, field) across
  alleles, broadcast back to ctx.group_index. Direction by field:
  score=max, value=min, percentile_rank=min.
- KindAccessor properties: best_value, best_score, best_rank, plus
  *_allele variants returning the allele name string.
- DSL string parser: presentation['mhcflurry'].best_score and
  best_score_allele (and rank/value/percentile spellings) now parse to
  BestAlleleField.
- Refactor: factor _filter_kind_method_version helper out of Field.eval
  for reuse by BestAlleleField.
- Bump topiary to 5.12.0 (additive minor).

Use case (per #158): vaxrank's "presented anywhere" tier signal —
``Presentation['mhcflurry'].best_score`` returns the multi-allele
aggregate, ``best_score_allele`` returns the attribution.

* Address review: drop merge, per-kind directions, kind_support warnings

Four fixes against PR #160 review:

1. Replace the fragile ``__best__``/merge broadcast with
   ``ctx.group_index.droplevel('allele')`` + ``per_peptide.reindex()``.
   No intermediate `__best__` column, no positional-merge ordering
   assumption, fewer allocations. ``sub.copy()`` is gone too — work
   on the filter's already-fresh slice via index masks + ``.assign()``.

2. Replace the field-only ``_BEST_DIRECTIONS`` constant with a
   per-(kind, field) lookup ``_best_dir... (continued)

147 of 164 new or added lines in 4 files covered. (89.63%)

4214 of 4707 relevant lines covered (89.53%)

2.69 hits per line