Table 4.

AUC of TPP task with either known peptide and TCR (but unknown pairing TPP-I), known peptide unseen TCR (TPP-II), and unseen peptide and TCR (TPP-III).

Evaluation AUC	McPAS		VDJdb		McPAS+ VDJdb		Tumor
	AE	LSTM	AE	LSTM	AE	LSTM	AE	LSTM
TPP-I	0.860	0.859	0.840	0.842	0.776	0.761	0.805	0.813
TPP-II	0.810	0.798	0.792	0.764	0.770	0.745	0.805	0.813
TPP-III	0.601	0.562	0.669	0.522	0.636	0.674	0.570	0.646

The results are the test AUC using either AE or LSTM on McPAS (23) and VDJdb (27) separately or on the joined dataset. The final column is the prediction of tumor antigens TCR binding. Again, the AE consistently outperforms the LSTM, except for the TPP-III task, where increasing the training set size and the complexity of the encoders improves the performance. Bolded values are the best results.