Sharpening a Proteolytic Cleavage Oracle for Peptide Design

Designing a peptide construct means predicting where proteases will cut it, and a design loop can only use a cleavage predictor at the resolution that predictor is measured at. The model we build on, DeepPeptide, is scored by a criterion that accepts a predicted segment when both of its endpoints land within $\pm3$ residues of the annotation, a window comparable to the peptide itself at the short end of the 5--50 residue range, so the score rewards finding segments over placing them. We make placement the target of both the model and the measurement, and add two small blocks to that model. A zero-initialized boundary head gives its conditional random field the position-specific evidence its own state graph already expects, and a lightweight adapter re-projects the frozen protein language model embedding. Together they raise segment F1 from $0.576$ to $0.630$ on ESM-2 and from $0.588$ to $0.666$ on ESM-C 6B, under the $5\times4$ nested cross-validation DeepPeptide itself uses. Scored with no acceptance window at all, the head covers less of the annotation than the base while the adapter covers more, and no F1 change there is resolved, alone or together. Most of what the pair buys is placement: with detection held fixed, the share of boundaries landing on the annotated residue rises from $0.62$ to $0.70$ on all five outer folds. That is what a design loop needs and what a $\pm3$ window cannot report. We also quantify why this task is hard to measure on at all: homology-aware folds are taxonomically non-exchangeable, and their spread exceeds either addition measured on its own.