Surface MHCflurry's antigen-processing (cleavage) score (#235) - #240

Closes #235.

MHCflurry's Class1PresentationPredictor.predict() already returns a processing_score column (its antigen-processing / cleavage head) right next to presentation_score — and the MHCflurry wrapper read presentation_score/presentation_percentile while discarding processing_score. We were computing the processing prediction on every run and throwing it away.

Change

Emit it as a Kind.antigen_processing prediction. It depends only on the peptide + flanks, not the allele (verified: SIINFEKL → 0.96719 for both HLA-A*02:01 and HLA-B*07:02), so it's emitted once per peptide, allele-less, carrying the n_flank/c_flank context.

from mhctools import MHCflurry, Kind

r = MHCflurry(alleles=["HLA-A*02:01", "HLA-B*07:02"]).predict(
    ["SIINFEKL"], n_flanks=["AAA"], c_flanks=["KKK"])[0]

r.processing            # -> Prediction(SIINFEKL | antigen_processing | score=0.9672 | mhcflurry)
r.processing.score      # 0.96719  (== MHCflurry's raw processing_score)
r.processing.allele     # ""  (allele-independent)

• No new dependency and no extra cost — the score was already computed.
• getattr(row, "processing_score", None) guards older/newer mhcflurry builds without the column (no-op if absent).
• kind_support() advertises antigen_processing as mhc_dependence/mhc_class = "none", matching the other allele-free processing predictors.

Small companions

• PeptideResult.processing accessor for Kind.antigen_processing. The existing ProcessingPredictor already emits this kind, but there was a cleavage accessor for proteasome_cleavage and none for antigen_processing — filled the gap.
• predict-table processing field token → (antigen_processing, score), so the Add predict-table command to annotate CSVs with predictor scores (#231) #234 command can pin the processing kind: --predictor mhcflurry:proc:processing (avoids the kind-agnostic score ambiguity when a predictor emits several kinds).

Tests

• test_mhcflurry_processing_score: asserts the emitted score matches MHCflurry's raw processing_score, is allele-less, carries the flanks, and that kind_support() advertises it as allele-independent.
• Updated test_mhcflurry_multiple_alleles for the extra prediction (2 affinity + 1 presentation + 1 processing).
• test_processing_field_token_pins_antigen_processing_kind in the (binary-free) annotate-table suite.

Compatibility

Additive by design. The new-API predict(), predict_dataframe(), and protein-scanning gain one extra antigen_processing prediction/row per peptide for MHCflurry (filterable by kind). The legacy predict_peptides and the CLI --output-csv path is untouched — it stays affinity-only, one BindingPrediction per (peptide, allele) — so existing long-format consumers see no change. Both are covered by regression tests (test_mhcflurry_legacy_predict_peptides_unchanged, test_mhcflurry_processing_score_without_flanks).

Version 3.21.0 → 3.22.0.

https://claude.ai/code/session_01LZahFhBSCiehXTESCYQ7wG