PepLab Bench: a source-model-selector benchmark for HLA-A*02:01 peptide selection

Retrospective HLA-A*02:01 peptide-selection benchmark: 160 archived nine-mers, 151 source/model/selector configurations, 200 shared starts and 30,200 episodes. At 32 reveals, the leading sequence-chemistry-contact GP/greedy pipeline recovered 3.265 of eight targets versus 2.920 for the strongest cheap chemistry/XGBoost/greedy pipeline. Matched contact gain: +0.450 targets; adding energy to that leader: -0.500. A clean public software installation reproduced 12 historical episodes (36 contexts). These are development comparisons, not independent biological validation or cost-aware acquisition. Author: Oguzcan Unver, Metastate Bio. Manuscript revision 6.4. Complete evidence: https://doi.org/10.5281/zenodo.24142178 . Code: https://github.com/otcan/peplab-bench .