Table 4.

CALIBER Conformational epitope prediction performance of BiLSTM, GCN, and boosting models on the Test dataset. Comparing the different embedding methods: Random initialization. Kidera, Kidera + biochemicals, ESM-2 and ESM-IF1.

Embedding	Model	AUC	BAC	MCC	PR-AUC
Random	BiLSTM	0.664	0.62	0.16	0.18
Kidera	BiLSTM	0.694	0.64	0.18	0.21
Kidera+bio	BiLSTM	0.718	0.66	0.21	0.25
ESM-2	BiLSTM	0.776	0.7	0.27	0.35
ESM-2	GCN	0.753	0.69	0.25	0.32
ESM-2	Boosting	0.777	0.68	0.23	0.37
ESM-IF1	BiLSTM	0.778	0.7	0.28	0.34
ESM-IF1	GCN	0.763	0.68	0.24	0.33
ESM-IF1	Boosting	0.788	0.69	0.24	0.38