Label/objective mismatch for residue-level prediction

[jeffreyHoelzel]

Current label generation expands peptide labels across every residue in a peptide alignment window. If a peptide is reactive, every residue in that peptide window becomes a positive residue.

Evidence:

• src/pepseqpred/core/labels/builder.py
  • _build_labels_for_protein marks def_mask[start:stop] = True for definite epitope peptides.
  • The FFNN then trains residue-level BCE on those expanded residue labels.

Why this can hurt:

• A reactive peptide means "the peptide contains an epitope signal", not necessarily "every residue in this peptide is epitope".
• This creates dense false-positive residue labels within positive peptides.
• The model is optimized for residue-wise correctness under noisy residue labels, while downstream use may care about peptide/protein regions or sparse true epitopes.
• The issue can compound in multi-pathogen data if peptide lengths, overlap density, or labeling criteria differ by source.

Planning direction:

• Add diagnostics comparing peptide-level labels to residue-level expansion density.
• Consider a multiple-instance learning objective, peptide-window objective, or soft/weak residue labels.
• If residue labels remain, consider down-weighting positive residues within reactive peptide spans or using boundary-aware smoothing.

Disclaimer: This may not be an issue we can resolve given the data we have, but this issue outlines some possible direction to consider.