BenchMHC: a reproducible framework for peptide-MHC presentation prediction

BenchMHC reproduces NetMHCpan-4.1 peptide–MHC presentation prediction with 0.794 per-allele Mean Top-K versus 0.790 on the official webserver, and 0.0025 Median FRANK versus 0.0022.

• Why it matters: Published peptide–MHC predictors are often inference-only and trained on different datasets, making it hard to separate method effects from training-data effects in neoantigen/epitope discovery.
• What they did: BenchMHC reproduces, retrains, and evaluates pMHC models on public data while recording every implementation decision; its NetMHCpan-4.1 reproduction scores 280,000 peptide/allele pairs in 27 s (~350× faster).
• The result: In public NetMHCpan-4.1 eluted-ligand sets, 86–99% of hits are in the human proteome versus ~12% of decoys, and redrawing decoys within source proteins (matching cysteine status) raises Mean Top-K from 0.734 to 0.755 and Median FRANK on CD8 epitopes from 0.0025 to 0.0053.