The predicted binding affinities and ranks that result from the IEDB prediction tools should be treated as ranking metrics as a way to prioritize peptides for experimental testing. There are many ways to rank the peptides. Here, we list only the latest recommendations.
MHC class I
For MHC class I T cell epitope predictions, predicted binders can be selected based on the percentile rank or MHC binding affinity.
The IEDB recommends making selections based on a percentile rank of <= 1% for each (MHC allele, length) combination to cover most of the immune responses.1, 2 Alternatively, an absolute binding affinity (IC50) threshold of 500 nM identifies strong binders.3 A paper from our group showed that the absolute binding affinity threshold correlates better with immunogenicity and also that, for even better correlation, MHC-specific thresholds should be used.4
The tables below show the allele-specific thresholds for the 38 most common HLA-A and HLA-B alleles, representing the nine major supertypes. The tables can also be downloaded as a PDF file (attached here: class_I_allele_specific_cutoff.pdf (37.7 KB)).
Table: Alleles sorted by name
| Allele | Population frequency of allele | Allele specific affinity cutoff (IC50 nM) |
|---|---|---|
| A*0101 | 16.2 | 884 |
| A*0201 | 25.2 | 255 |
| A*0203 | 3.3 | 92 |
| A*0206 | 4.9 | 60 |
| A*0301 | 15.4 | 602 |
| A*1101 | 12.9 | 382 |
| A*2301 | 6.4 | 740 |
| A*2402 | 16.8 | 849 |
| A*2501 | 2.5 | 795 |
| A*2601 | 4.7 | 815 |
| A*2902 | 2.9 | 641 |
| A*3001 | 5.1 | 109 |
| A*3002 | 5 | 674 |
| A*3101 | 4.7 | 329 |
| A*3201 | 5.7 | 131 |
| A*3301 | 3.2 | 606 |
| A*6801 | 4.6 | 197 |
| A*6802 | 3.3 | 259 |
| B*0702 | 13.3 | 687 |
| B*0801 | 11.5 | 663 |
| B*1402 | 2.8 | 700 |
| B*1501 | 5.2 | 528 |
| B*1801 | 4.4 | 732 |
| B*2705 | 2 | 584 |
| B*3501 | 6.5 | 348 |
| B*3503 | 1.2 | 888 |
| B*3801 | 2 | 944 |
| B*3901 | 2.9 | 542 |
| B*4001 | 10.3 | 639 |
| B*4002 | 3.5 | 590 |
| B*4402 | 9.2 | 904 |
| B*4403 | 7.6 | 780 |
| B*4601 | 4 | 926 |
| B*4801 | 1.8 | 887 |
| B*5101 | 5.5 | 939 |
| B*5301 | 5.4 | 538 |
| B*5701 | 3.2 | 716 |
| B*5801 | 3.6 | 446 |
Table: Alleles sorted by population frequency
| Allele | Population frequency of allele | Allele specific affinity cutoff (IC50 nM) |
|---|---|---|
| A*0201 | 25.2 | 255 |
| A*2402 | 16.8 | 849 |
| A*0101 | 16.2 | 884 |
| A*0301 | 15.4 | 602 |
| B*0702 | 13.3 | 687 |
| A*1101 | 12.9 | 382 |
| B*0801 | 11.5 | 663 |
| B*4001 | 10.3 | 639 |
| B*4402 | 9.2 | 904 |
| B*4403 | 7.6 | 780 |
| B*3501 | 6.5 | 348 |
| A*2301 | 6.4 | 740 |
| A*3201 | 5.7 | 131 |
| B*5101 | 5.5 | 939 |
| B*5301 | 5.4 | 538 |
| B*1501 | 5.2 | 528 |
| A*3001 | 5.1 | 109 |
| A*3002 | 5 | 674 |
| A*0206 | 4.9 | 60 |
| A*3101 | 4.7 | 329 |
| A*2601 | 4.7 | 815 |
| A*6801 | 4.6 | 197 |
| B*1801 | 4.4 | 732 |
| B*4601 | 4 | 926 |
| B*5801 | 3.6 | 446 |
| B*4002 | 3.5 | 590 |
| A*6802 | 3.3 | 259 |
| A*0203 | 3.3 | 92 |
| A*3301 | 3.2 | 606 |
| B*5701 | 3.2 | 716 |
| A*2902 | 2.9 | 641 |
| B*3901 | 2.9 | 542 |
| B*1402 | 2.8 | 700 |
| A*2501 | 2.5 | 795 |
| B*2705 | 2 | 584 |
| B*3801 | 2 | 944 |
| B*4801 | 1.8 | 887 |
| B*3503 | 1.2 | 888 |
NetMHCPan 4.0
The latest research from our group6 shows that setting a common threshold for eluted ligand (EL) percentile rank of ~1.1 in NetMHCPan 4.0 across all alleles results in 80% sensitivity for capturing immunogenic peptides. Allele-specific thresholds have also been established and are contained within the paper’s supplemental data. However, the increase in sensitivity and specificity is marginal unless there are relatively few alleles being considered with very divergent thresholds.
MHC class II
For MHC class II T cell epitope predictions, the selection of predicted binders can be made based on the percentile rank or MHC binding affinity. One recommended approach is to make selections based on a median percentile rank of 20% across a panel of alleles. When tested with a 7-allele panel, 50% of the immune response was recovered.7 Alternatively, selecting peptides predicted to bind at <=1,000nM or <=10 percentile rank is also supported by experimental data.5 Ranking the peptides by percentile rank and applying a fixed cutoff is another valid approach.
One additional thing to remember with class II instead of class I is the high degree of overlap for selected peptides due to common binding cores. Care should be taken to remove redundancy.
- Moutftsi et al., 2006, Nat. Biotech (PMID 16767078.pdf (164.2 KB))
- Kotturi et al., 2007, J. Virology (PMID 17329346.pdf (561.9 KB))
- Sette et al., 1994, J. Immunology (PMID 7527444.pdf (780.6 KB))
- Paul et al., 2013, J. Immunology (PMID 24190657.pdf (608.2 KB))
- Southwood et al., 1998, J. Immunology (PMID 9531296.pdf (112.2 KB))
- Reardon et al., 2021 Mol. Cell. Proteomics (PMID 34303001.pdf (2.6 MB))
- Paul et al., 2015 J. Imm. Methods (PMID 25862607.pdf (658.3 KB))