Skip to main content

Advertisement

Table 3 MHC class I binding algorithm comparison

From: Best practices for bioinformatic characterization of neoantigens for clinical utility

Features/ software Algorithm type used Type of data used for training Number of HLA alleles used for training HLA alleles and peptide length that can be predicted Output information
Pickpocket (2009) Position-specific weight matrices In vitro quantitative binding data (> 150,000 data points) More than 150 different MHC molecules HLA-A, −B, −C, −E and -G alleles, also for non-human primates, mice, cattle and pigs. Peptides of 8–12 in length Prediction values are given in nM IC50 values
NetMHCcons (2012) Integration of NetMHC 3.4, NetMHCpan 2.8 and PickPocket 1.1 In vitro binding affinity data NetMHC 3.4 (94 MHC class I alleles), NetMHCpan 2.8 (> 120 different MHC molecules), PickPocket 1.1 (94 different MHC alleles) Can predict peptides to any MHC molecule of known sequence. Peptides of 8–15 amino acids in length Prediction values are given in nM IC50 values and as %rank to a set of 200,000 random natural peptides
NetMHC 4.0 (2016) Artificial neural networks In vitro binding affinity data 81 different human MHC alleles (HLA-A, −B, −C, and -E) and 41 animal alleles 81 different human MHC alleles (HLA-A, −B, −C, and -E) and 41 animal alleles. Any length but recommends 9 and discourages above 11 amino acids Core position for binding within the peptide, interaction core sequence, affinity in nM, rank of prediction compared with 400,000 random natural peptides (strong binders %rank < 0.5), and so on
NetMHCpan 4.0 (2017) Artificial neural networks Binding affinity (> 180,000 data points) and eluted ligand (MS) data 172 human and other animal MHC molecules Can predict peptides to any MHC molecule of known sequence Core position for binding within peptide, interaction core sequence, affinity in nM, rank of the predicted affinity compared to a set of random natural peptides (strong binders %rank < 0.5), and so on
MHCnuggets (2017) Gated recurrent neural networks IC50 values from immuno-fluorescent binding experiments for pMHC Class I pairs (137,654 data points) 106 unique MHC alleles Any MHC alleles, more reliable for alleles that are present in IEDB. Any peptide length is valid IC50 binding affinity prediction
MHCflurry (2018) Allele-specific feed forward neural networks Binding affinity and eluted ligand (MS) data (> 230,735 data points) Across 130 alleles from IEDB combined with benchmark dataset from Kim et al. [209] 112 alleles showed performance sufficient for their inclusion in predictor. Peptide lengths of 8–15 are supported Affinity given in nM, percentile predictions across the models, and quantile of affinity prediction among large number of random peptides tested
EDGE (2019) Deep neural network Peptide sequences from HLA immunoprecipitation followed by MS characterization Not explicitly specified 53 HLA alleles, 8–15-mer (inclusive) Not explicitly specified
  1. A direct comparison of a subset of popular MHC class I binding predictors showing their variability in algorithmic structure, training data, supported HLA alleles and valid peptide lengths