openvax / topiary / 25811341902

90%

push

github

Backfill value from score for [0,1]-score predictor kinds (#165) (#166)

* Backfill value from score for [0,1]-score predictor kinds (#165)

`_format_prediction_df` now fills `value` from `score` wherever `value`
is NaN. mhctools only populates `value` when the kind has a distinct
unit (IC50 nM for affinity, half-life for stability); for the other
kinds — most visibly `pMHC_presentation` — `value` came back as NaN
even though `score` carried the predictor's primary [0, 1] output.

NaN-in-value bit downstream consumers reading `value` uniformly across
kinds: strict `simplejson.dumps` rejects NaN/Inf, `sorted()` is
undefined on NaN keys, and arithmetic on `value` silently propagated.
Vaxrank just hit this on real data (vaxrank#289, fix in vaxrank#290).

Rows that already carry a unit-bearing `value` are untouched, so
affinity (IC50 nM) and stability (half-life) keep their semantics.
The `affinity` column continues to be populated only for
`pMHC_affinity` rows.

* Address review: kind-gated backfill, cover CachedPredictor path

- Extract `_backfill_value_from_score` helper that skips kinds in
  `mhctools.VALUE_BEST_DIRECTIONS` (pMHC_affinity, pMHC_stability).
  A NaN ``value`` on a unit-bearing kind means the unit is genuinely
  unknown; falling back to ``score`` there would falsely advertise a
  [0, 1] score as IC50 nM or hours.
- Apply the helper in `cached.py:_bindings_to_dataframe` and
  `_predictions_to_dataframe` so CachedPredictor's standalone path
  (mhcflurry class1_presentation, NetMHCpan ``-BA`` parsing, ...)
  gets the same fix, not just the TopiaryPredictor wrapper.
- Add end-to-end test through `predict_from_named_sequences` with a
  predictor that emits both affinity and presentation rows.
- Add regression tests for both cached.py helpers and for the
  unit-bearing-kind skip case.

14 of 16 new or added lines in 3 files covered. (87.5%)

4276 of 4771 relevant lines covered (89.62%)

2.69 hits per line