NetMHCpan
The canonical neural-net predictor of peptide–HLA binding — the tool you'll see cited as the baseline almost everywhere in this field.
NetMHCpan is a neural network that, given a peptide and an HLA allele, predicts how likely the peptide is to bind and be presented. Its "pan" design generalizes across alleles — including rare ones with little data — by learning from the alleles' sequences. MHCflurry is a popular open-source counterpart.
It matters as a benchmark for what's already solved: HLA-binding prediction is the mature, commoditized step, and beating NetMHCpan on binding alone is hard, so it's a weak basis for differentiation. That's exactly why the value — and the investment thesis — has moved up the stack to processing, presentation, and especially immunogenicity, where predictors are still weak and the headroom is large.
Is MHCflurry more accurate than NetMHCpan?
Neither is uniformly better. NetMHCpan-4.1 tends to lead on naturally-processed/presented peptides and rare-allele coverage thanks to its pan-allele design; MHCflurry is open-source, considerably faster, and competitive on non-9-mer peptides. Many pipelines run both and combine the scores rather than choosing one.
What is the difference between binding-affinity (BA) and eluted-ligand (EL) prediction?
Binding-affinity prediction is trained on lab-measured peptide–HLA binding strength. Eluted-ligand prediction is trained on peptides actually found presented on cells by mass spectrometry, so it captures the full antigen-processing pathway, not just binding. Modern predictors like NetMHCpan-4.1 output both; the EL score is usually the better presentation signal.
Why isn't beating NetMHCpan on binding enough to build a business on?
Because HLA-binding prediction is already a mature, commoditized step where NetMHCpan is a strong baseline. The unsolved, high-value problem is immunogenicity — predicting which presented peptides actually trigger a T-cell response — which is where current AI research and investment are concentrated.