Accurate MHC-Peptide Affinity Prediction

The server is trained on the largest number of quantitative peptide:MHC affinity measurements ever published using both affinity data from the Immune Epitope Database and Analysis Resource (IEDB) (6 (link)), eluted peptide data from the SYFPEITHI database (7 (link)) and proprietary affinity data. The predictions based on ANNs are trained essentially as described in (3 (link)) on data from 55 MHC alleles (43 Human and 12 non-human), and the predictions based on position specific scoring matrices (PSSMs) are trained as described in (2 (link)) for additional 67 HLA alleles. A large number of 9-mer MHC affinity data have become available from the IEDB database, since the training of the ANNs used at NetMHC-3.0, and all peptides not used in the training (6452 9-mer peptide affinity data points, covering 32 HLA alleles) were used for evaluation of the server performance. These data are available at the server. In this dataset, 3104 were measured to be binders (IC₅₀<500 nM), 76% of these were correctly predicted as such. 3030 peptides were predicted to bind to a given HLA, and 78% of these had a measured IC₅₀<500 nM. The average Pearson correlation coefficient (PCC) and area under a ROC curve (AUC) value using a 500 nM classification threshold were 0.71 and 0.86, respectively. For the full per allele results, see the Supplementary Material (Supplementary Table 1 and Supplementary Figure 1). NetMHC-3.0 uses a new approximation algorithm that reliably predicts the affinity of peptides of lengths 8, 10 and 11, for which affinity data for training are rare (8 ). The method uses predictors trained on peptides of length 9 to successfully extrapolate to other lengths. In short, the method approximates each peptide of any length to a number of 9-mers, by inserting X (for 8-mers) or deleting amino acid(s) (for 10- and 11-mers) and set the final prediction to an average of the 9-mer predictions. We had previously trained ANN predictors directly on 10-mer affinity data and since this training more than 2000 10-mer peptide:MHC affinities had become available from the IEDB database (6 (link)). Area under a ROC curve (AUC) values were calculated for each allele using either ANNs trained on 10-mers or the approximation method. For 12 of the 16 alleles, the approximation method performed better than the 10-mer trained ANNs (P < 0.01), see Supplementary Material Figure 2. However, for the four HLA-alleles, this evaluation showed better performance for ANNs trained on 10-mer peptides; these 10-mer trained ANNs are used for predictions by the server. For 8-mers, 2002 affinity data were extracted covering 35 MHC alleles. The overall PCC and AUC were 0.68 and 0.86, respectively. For 8-mer per allele performance, see the Supplementary Material Figure 4. For 8-mers, predictors trained on actual 8-mers seems to be better than the approximation method otherwise used, so for the alleles with available 8-mer affinity data, 8-mer trained ANNs are used for the predictions. In general, it is not possible to estimate how reliable a single prediction is. However, the stronger the affinity is predicted the higher are the chance that the actual affinity is stronger than the generally accepted binding threshold of 500 nM.