openvax / topiary / 24315494917

87%

master: 90%

Build Type

push

github

Committed by web-flow

Commit Message

DSL refactor: unify filter tree + vectorize eval (#111) (#113)

Consolidates the historical filter classes (EpitopeFilter, ColumnFilter,
ExprFilter) and the RankingStrategy composite into a single DSLNode tree
whose .eval(ctx) returns a pd.Series indexed by peptide-allele group
tuples. Every node — arithmetic, comparison, boolean — composes freely.

Breaking changes (warrants 5.0.0):
- Removed: EpitopeFilter, ColumnFilter, ExprFilter, RankingStrategy,
  SortSpec, affinity_filter, presentation_filter, apply_ranking_strategy.
- TopiaryPredictor now takes filter_by (DSLNode) + sort_by (list[DSLNode])
  + sort_direction. Legacy kwargs (ic50_cutoff, percentile_cutoff,
  ranking, ranking_strategy, filter, rank_by) are dropped.
- parse_expr, parse_filter, parse_ranking collapsed into a single parse().
- Parser now accepts mixed |/& with standard precedence (& tighter than |);
  previously this raised.
- Unqualified Affinity.value on multi-method DataFrames raises a clear
  ambiguity error pointing at Affinity["modelname"]. Old behavior
  silently passed if any row met the threshold.
- apply_filter errors on non-boolean input (accepts {True, False, 0, 1,
  NaN}). NaN still maps to False.

New:
- Field gains an optional version parameter; Affinity["netmhcpan", "4.1b"]
  and affinity['netmhcpan', '4.1b'] in string DSL filter on both
  prediction_method_name and predictor_version.
- apply_filter / apply_sort split as the public entry points. apply_sort
  preserves the NaN-fallthrough-to-next-tiebreaker semantic, now
  vectorized over a float matrix.
- AggExpr (mean/geomean/minimum/maximum/median) vectorized via pandas
  skipna ops; geomean via nanmean(log()) with non-positive masked.
- BoolOp honors NaN → False policy cleanly under all dtypes.
- DSLNode.to_expr_string() round-trips through parse() for every node.
- ranking.py split into a package: nodes.py (1376), parser.py (492),
  apply.py (194), __init__.py (103 — pure re-exports).

Tests: 843 passing. Added ... (continued)

Coverage Stats

1135 of 1299 new or added lines in 8 files covered. (87.37%)

1 existing line in 1 file now uncovered.

2595 of 2975 relevant lines covered (87.23%)

2.62 hits per line