PR #318 can plot external predictors, including NetMHCpan 4.2 BA/EL, when those columns are present in the paper-figure artifacts. The current 2023 artifact bundle only contains NetMHCpan 4.0 BA/EL and MixMHCpred outputs, so NetMHCpan 4.2 panels are skipped by design.\n\nWe should add an optional release-evaluation stage that generates external predictor scores from the benchmark rows when the tools are available locally or on the remote evaluation machine. Suggested shape:\n\n- Detect or configure installed external predictors, especially NetMHCpan 4.2 BA and EL.\n- Prefer a thin wrapper around mhctools if it supports the installed tool/version; otherwise support explicit command-line templates for licensed binaries.\n- Keep this optional because NetMHCpan and MixMHCpred binaries/licenses are not redistributable with MHCflurry.\n- Write predictor columns using the artifact names already understood by paper-figures, e.g. netmhcpan4.2.ba, netmhcpan4.2.el, and *_best_allele.\n- Recompute accuracy_scores.multiallelic.csv percent-change columns against configurable baseline predictors.\n- Run on the remote GPU/CPU evaluation machine before artifact sync, not on the laptop.\n- Document exact tool versions, binary paths, and command provenance in the eval output.\n\nThis is distinct from the current MHCflurry-vs-MHCflurry compare-models path: rc14 does generate new MHCflurry predictions from the trained weights, but external predictors are currently artifact-driven rather than regenerated from scratch.
PR #318 can plot external predictors, including NetMHCpan 4.2 BA/EL, when those columns are present in the paper-figure artifacts. The current 2023 artifact bundle only contains NetMHCpan 4.0 BA/EL and MixMHCpred outputs, so NetMHCpan 4.2 panels are skipped by design.\n\nWe should add an optional release-evaluation stage that generates external predictor scores from the benchmark rows when the tools are available locally or on the remote evaluation machine. Suggested shape:\n\n- Detect or configure installed external predictors, especially NetMHCpan 4.2 BA and EL.\n- Prefer a thin wrapper around mhctools if it supports the installed tool/version; otherwise support explicit command-line templates for licensed binaries.\n- Keep this optional because NetMHCpan and MixMHCpred binaries/licenses are not redistributable with MHCflurry.\n- Write predictor columns using the artifact names already understood by paper-figures, e.g. netmhcpan4.2.ba, netmhcpan4.2.el, and *_best_allele.\n- Recompute accuracy_scores.multiallelic.csv percent-change columns against configurable baseline predictors.\n- Run on the remote GPU/CPU evaluation machine before artifact sync, not on the laptop.\n- Document exact tool versions, binary paths, and command provenance in the eval output.\n\nThis is distinct from the current MHCflurry-vs-MHCflurry compare-models path: rc14 does generate new MHCflurry predictions from the trained weights, but external predictors are currently artifact-driven rather than regenerated from scratch.