Two ways to train HLA predictors — on mass-spec peptides actually eluted from HLA (EL, capturing the whole presentation pathway) versus on lab-measured binding strengths (BA) — with EL generally the better signal for what gets presented.
Binding-affinity (BA) data comes from in-vitro assays that measure how tightly a synthetic peptide binds an HLA molecule. It models a single event — peptide-to-HLA binding — and ignores everything upstream. Eluted-ligand (EL) data comes from mass spectrometry of peptides physically pulled off HLA molecules on real cells (the immunopeptidome). EL therefore reflects not just binding but the whole antigen-presentation pathway: proteasomal cleavage, TAP transport, loading, and surface display. That makes EL a closer proxy for what is actually presented in vivo.
Modern predictors fuse both. NetMHCpan-4.1, trained on more than 850,000 BA and EL measurements, outputs both an EL score and a BA score, and its authors report that EL-trained models outperform BA-only models, with the combination best of all — because EL supplies presentation context while BA anchors the underlying binding chemistry and covers peptides that mass-spec hasn't sampled. In practice EL (often reported as a presentation %Rank) is the recommended default for ranking presentation likelihood, while BA remains useful as a complementary, physically interpretable affinity estimate.
So what for investors: knowing whether a pipeline ranks candidates by EL or BA tells you how realistically it models presentation. EL-driven selection should yield fewer false positives — peptides that bind in a tube but are never displayed on a cell. When assessing a neoantigen platform, ask which score drives candidate ranking; reliance on BA alone is a dated approach, while leveraging EL (and proprietary immunopeptidomics to train it) is a credible technical moat.
Binding-affinity (BA) data is in-vitro measurements of how strongly a peptide binds an isolated HLA molecule. Eluted-ligand (EL) data is mass-spec identification of peptides actually presented on HLA by living cells. BA captures only the binding step; EL captures the full presentation pathway (processing, transport, loading, display), so it better reflects what is presented in vivo.
EL is generally the stronger signal because it reflects real presentation, not just binding. The NetMHCpan-4.1 authors found EL-trained models outperform BA-only models, and combining both is best — BA contributes binding chemistry and broad peptide coverage, EL contributes presentation context.
Yes. NetMHCpan-4.1 was trained jointly on BA and EL data and outputs both an eluted-ligand (presentation) score and a binding-affinity score, usually reported as percentile ranks. The EL/presentation rank is typically the recommended metric for ranking candidate epitopes.