Table 2.
Performance evaluation on eluted ligand test NetMHCpan-4.1, [34] dataset including 77,053 peptides (lengths 8-14) across 36 distinct HLA-I molecules.
| RNN(no ambig.) | RNN | CNN(no ambig.) | CNN | NetMHCpan-4.1 | NetMHCpan-4.0 | MHCflurry | MixMHCpred | |
|---|---|---|---|---|---|---|---|---|
| Mean ROC AUC | 0.9657 | 0.9639 | 0.9615 | 0.9610 | 0.9498 | 0.9462 | 0.9335 | 0.9324 |
| Mean PPV @ 0.95 * n_pos |
0.8237 | 0.8038 | 0.8651 | 0.8446 | 0.8162 | 0.786 | 0.7255 | 0.7705 |
Our system identifies situations where predictions are ambiguous; columns with the “(no ambig.)” label do not include ambiguous data; this is in contrast to a forced decision in all cases for columns labeled only RNN or CNN. For consistency with referenced study methodology, PPV for each HLA was evaluated considering only a fraction of ranked predictions equal to 0.95 times the number of true positives. See methods for more detail. The mean of these metrics across HLAs is labeled “Mean PPV @ 0.95 * n_pos”.