Fig. 3.

TCRconv evaluation. Results are obtained using stratified 10-fold cross-validation on VDJdbβ-large dataset. (A) Comparing TCRconv to other methods using average AUROC and AP scores. (B) The AUROC scores for TCRconv predictions correlate negatively with the diversity of the epitope specific TCRs (Pearson correlation −0.72). (C) Increasing the embedding context size increases the predictive AUROC score. The schematics on top show the approximate sections included in different context sizes. CDR3 + X refers to CDR3 embeddings with context size X and complete TCR to embeddings for complete TCRs without extracting only the CDR3 parts. TCRconv uses CDR3 + full (bolded)