openvax / topiary / 24418527228

88%

master: 90%

Pull #130

github

Tighten CachedPredictor version handling + mhcflurry auto-compose

Six add-ons from the PR #128 design discussion, bundled:

1. Empty-cache + fallback mode — CachedPredictor(fallback=p) with no
   df is now valid; (name, version) are discovered from the fallback's
   output on the first query and locked in thereafter.  Pure read-
   through cache.

2. Reject null / NaN / empty-string on prediction_method_name and
   predictor_version at construction.  Silent "I don't know" agreement
   would mask the invariant; force the user to supply a real identity.

3. also_accept_versions opt-in equivalence set.  A fallback's version
   passes if it equals the cache's OR is in the caller's accepted set.
   Names stay strict — different predictors are different.  Default
   (empty set) keeps today's strict equality.

4. mhcflurry_composite_version() helper and auto-default in
   from_mhcflurry.  Discovers mhcflurry.__version__ plus
   mhcflurry.downloads.get_current_release() and composes e.g.
   "2.2.1+release-2.2.0".  Users never enumerate model bundles —
   CachedPredictor.from_mhcflurry("predictions.csv") just works.
   Explicit predictor_version= still overrides when a custom label is
   needed.

5. from_mhcflurry docstring now explains the mhcflurry-specific
   package-version + model-release composition (scoped to this loader,
   not the generic class — NetMHCpan et al. have their model data baked
   into the binary and don't need composite versions).

6. Lookup refactor: one pass to identify peptides with any miss, one
   fallback call per round, one pass to assemble rows from the updated
   cache.  Fixes a latent duplicate-row bug in the prior implementation
   (allele-level partial results + fallback union).

15 new tests across null-version rejection, empty-cache-plus-fallback
identity discovery, equivalence-set acceptance vs. rejection, the
helper's composition + error paths (stubbed mhcflurry module — avoids
tensorflow/libomp collisions in the test en... (continued)

208 of 229 new or added lines in 2 files covered. (90.83%)

3214 of 3640 relevant lines covered (88.3%)

2.65 hits per line