Mono-allelic HLA class I peptidome (Sarkizova / Abelin)

At a glance

Labels and annotations

Every peptide carries an unambiguous allele label by construction: the B721.221 line makes no HLA class I of its own, so a single introduced allele is the only one that can present anything.

Details

Why it was needed

• A normal cell carries up to six class I molecules and the W6/32 antibody collects them all, so peptides come back without allele labels — nearly useless for a model whose whole point is 'this groove, these peptides'.

What it cost

• One engineered line per allele, roughly 100 million cells or a gram of tissue per pull-down, overnight antibody capture, acid elution, mass spectrometry with a 1–5% false-discovery rate.

Effect

• This one methodological change did more for prediction accuracy than any change of network architecture.

Models trained or evaluated on it

• training NetMHCpan-4.1 Health Tech, Technical University of Denmark (Nielsen lab) — mono-allelic MS data give unambiguous allele labels
• training MHCflurry 2.0 O'Donnell, Rubinsteyn and Laserson (Mount Sinai / OpenVax)
• training MixMHCpred 3.0 Gfeller lab, Ludwig Institute for Cancer Research / University of Lausanne
• training HLAthena Broad Institute / Dana-Farber Cancer Institute (Keskin, Wu, Carr labs) — this model and this dataset come from the same paper

Sources

1. Sarkizova S et al. Nat Biotechnol 2020
2. Abelin JG et al. Immunity 2017