Fig. 2: Predictive performance of TCR–antigen–HLA binding based on two external test sets using random-shuffle negative sampling.

a, AUROC, AUPR and PPVn scores for PISTE and competing models using two test sets. The red baseline represents a random classifier. b, The t-SNE embedding for the TCRs of the YLQ pHLA (top), TPR pHLA (middle) and KLG pHLA (bottom) learned by our model, in which PISTE facilitates the segregation of positive and negative TCRs, yielding distinct regions enriched with positive TCRs. All the testing triples whose antigen–HLA pairs were observed in the training data are removed from the test sets.