The existing fragment WT join matches the deconvolved allele as if it defined a haplotype score. When MT and WT select different presenters within the same configured genotype, the WT score is silently missing.
A deterministic presentation fixture with genotype HLA-A02:01/HLA-B07:02 gives MT SIINFEKL score 0.2 with presenter A02:01 and WT GIINFEKL score 0.8 with presenter B07:02. predict_from_fragments(..., predict_wt=True) produces wt_value=NaN, although both scores were returned. The explicit occurrence path in #451 correctly retains 0.8.
The join in _maybe_predict_wt_peptides includes allele unconditionally. Comparator matching must use declared MHC scope: the allele for per-allele scores, the full configured genotype for haplotype scores, and no allele key for allele-free scores. Keep the primary presenter label unchanged and do not invent WT context.
Reproducer on 5.90.0:
import pandas as pd
from topiary import TopiaryPredictor, ProteinFragment
class Haplotype:
alleles = ['HLA-A*02:01', 'HLA-B*07:02']
default_peptide_lengths = [8]
uses_flanking_sequences = True
def kind_support(self):
return {'pMHC_presentation': {'mhc_dependence': 'haplotype', 'mhc_class': 'I'}}
def predict_dataframe(self, peptides, n_flanks=None, c_flanks=None):
return pd.DataFrame([dict(peptide=p, allele=self.alleles[0 if p.startswith('S') else 1],
kind='pMHC_presentation', score=.2 if p.startswith('S') else .8,
value=.2 if p.startswith('S') else .8, percentile_rank=None,
n_flank=(n_flanks or [''] * len(peptides))[i], c_flank=(c_flanks or [''] * len(peptides))[i],
prediction_method_name='haplotype_fixture', predictor_version='1') for i, p in enumerate(peptides)])
def predict_proteins_dataframe(self, inputs):
return pd.concat([self.predict_dataframe([sequence], n_flanks=[''], c_flanks=['']).assign(
source_sequence_name=name, peptide_offset=0) for name, sequence in inputs.items()], ignore_index=True)
predictor = TopiaryPredictor(models=Haplotype(), predict_wt=True, only_novel_epitopes=False)
fragment = predictor.predict_from_fragments([ProteinFragment(
fragment_id='source', sequence='SIINFEKL', reference_sequence='GIINFEKL')])
print(fragment[['peptide', 'allele', 'allele_set', 'value', 'wt_value']])Acceptance: run fragment and exact-occurrence comparator paths through the same per-allele, haplotype and allele-free fixture battery. A changed haplotype presenter must retain the WT score, while distinct genotypes never match. Related #367 / PR #451.
The existing fragment WT join matches the deconvolved allele as if it defined a haplotype score. When MT and WT select different presenters within the same configured genotype, the WT score is silently missing.
A deterministic presentation fixture with genotype HLA-A02:01/HLA-B07:02 gives MT SIINFEKL score 0.2 with presenter A02:01 and WT GIINFEKL score 0.8 with presenter B07:02.
predict_from_fragments(..., predict_wt=True)produceswt_value=NaN, although both scores were returned. The explicit occurrence path in #451 correctly retains 0.8.The join in
_maybe_predict_wt_peptidesincludesalleleunconditionally. Comparator matching must use declared MHC scope: the allele for per-allele scores, the full configured genotype for haplotype scores, and no allele key for allele-free scores. Keep the primary presenter label unchanged and do not invent WT context.Reproducer on 5.90.0:
Acceptance: run fragment and exact-occurrence comparator paths through the same per-allele, haplotype and allele-free fixture battery. A changed haplotype presenter must retain the WT score, while distinct genotypes never match. Related #367 / PR #451.