Fig. 3. Thermostability data of HLA-ligands improves immunogenicity prediction.

a The information content of the stability data for each of the two main alleles studied (HLA-A*02:01 and HLA-B*07:02) was investigated by predicting peptide binding affinity with netMHCpan-4.0 (BA)¹¹ of the 8-11mer eluted ligands for which stability measurements were achieved demonstrating no correlation between predicted affinity and measured stability. Thus, granularity through peptide stability for predicted high-affinity HLA ligands was observed for both alleles. b T_m values were calculated based on thermal melt curves for the identified peptides across 12 different temperature points ranging from 37 °C to 73 °C with n=3 biological replicates at each temperature point. Thermal melt curve data are presented as median values ± SD. We trained an ANN model using transformed T_m values as input for the identified 1094 peptides restricted by HLA-A*02:01. The resulting models were used to predict >8,000 allele-specific eluted ligands. Binding motifs were constructed using the 1500 best and poorest predictions in both models. c This demonstrated a separation in the motif identified for high and low stability binders which could not be achieved using netMHCpan-4.0 (EL). d We investigated whether the additional dimensionality of the thermostability data could improve the prediction of immunogenic cancer neoepitopes by training ANNs with T_m values rescaled to the interval [0.5;1] as positive training data and length-balanced, randomly sampled peptides from the human UniProt-Swissprot as negative training data. e The resulting Stability Predictor demonstrates superior performance to current prediction tools with 9 of the predicted peptides in the top 10 being true neoepitopes (Precision in Top10). BA: binding affinity. PCC: Pearson Correlation Coefficient. EL: eluted ligand. AUC: Area Under the Curve. ROC: Receiver Operating Characteristic. PPV: Positive Predictive Value.