openvax / topiary / 24415576156

88%

master: 90%

Pull #130

github

CachedPredictor — serve predictions from a pre-computed table (part 1 of #128)

Core class + four loaders, no sharding or version-match tests yet.
Implements enough of the mhctools predictor protocol
(predict_peptides_dataframe, predict_proteins_dataframe, alleles,
default_peptide_lengths) to drop into TopiaryPredictor(models=cache).

Core invariant: a single CachedPredictor holds predictions from
exactly one (prediction_method_name, predictor_version) pair — never
mixes versions.  Enforced at construction and on the first fallback
call (lazy version-check against the fallback's output).

Fallback semantics:
- fallback=None (default): miss raises KeyError with missed peptides.
- fallback=<predictor>: miss delegates to the fallback, result is
  merged back into the cache so subsequent queries serve locally.
  No populate-on-miss flag — caching fallback hits is always the
  right default for the expected (batch prediction) workload.

Loaders:
- from_dataframe: in-memory, with optional predictor_name /
  predictor_version backfill for DataFrames that lack those columns.
- from_topiary_output: Parquet / TSV round-trip of topiary's own
  predict_* output schema.
- from_tsv: generic tab- or comma-delimited third-party files with a
  caller-supplied column-name mapping.
- from_mhcflurry: mhcflurry-predict CSV — maps the mhcflurry_*
  columns onto canonical affinity / percentile_rank / score.

Defer to follow-up PRs:
- from_netmhcpan (.xls parsing is meatier; separate scope).
- concat / from_directory sharding.

23 tests covering construction, version invariant (mixed versions
rejected, name/version round-trips as string), predict_peptides +
predict_proteins sliding-window, fallback hit + miss + version
mismatch, all four loaders, and end-to-end integration with
TopiaryPredictor(models=cache).

170 of 187 new or added lines in 2 files covered. (90.91%)

3176 of 3598 relevant lines covered (88.27%)

2.65 hits per line