openvax / topiary / 33427386457

93%

Build Type

push

github

Committed by web-flow

Commit Message

Per-peptide allele sets, and the candidates behind a resolved version (#222)

#219: EvalContext(alleles=) declared one set for the whole frame. A reader
emitting one row per (peptide, allele) that passed its own threshold --
LENS does -- gives peptides that were each reported against a different
subset of the genotype, so there is no single set to declare. Every LENS
fixture vaxrank has carries 2 to 8 distinct sets per file.

It now takes the same three forms from_predictions(allele_set=) does:
sequence for every peptide, mapping per peptide (keyed by the peptide
value or the full peptide-key tuple), or callable receiving the peptide
keys.

Declaring the union instead invents a group for every pairing that was
never scored. This hides well: an expression with an allele-scoped term
leaves the invented groups NaN, so output looks unchanged -- vaxrank
swapped in the union and every fixture came out byte-identical. An
expression reading only peptide-level evidence gives each invented group
a real number. Measured on a two-peptide frame with peptide_view(
proteasome_cleavage.score): union scores 6 groups, 2 of them pairings
never predicted; per-peptide scores 4, none invented.

A peptide declared nothing keeps only the groups its rows name rather
than inheriting another's genotype. A mapping key matching no peptide
raises, since a key that declares nothing looks exactly like a peptide
left undeclared on purpose. An empty set for one peptide is meaningful
even though an empty frame-wide sequence stays an error -- the same
absent-versus-empty distinction that has bitten this codebase twice now.

#220: describe_default_versions returns what resolve_default_versions
chose between, so a consumer can say "netMHCpan reports 4.1b and 4.2,
scoring with 4.2" without re-deriving the candidates -- a re-derivation
that re-implements "was a version named at all", the rule whose subtlety
caused the phantom-"nan" bug in both repos. Candidates are ordered by the
same PEP 4... (continued)

Coverage Stats

71 of 77 new or added lines in 2 files covered. (92.21%)

2 existing lines in 1 file now uncovered.

5828 of 6395 relevant lines covered (91.13%)

2.73 hits per line