Table 2.

Performance metrics of the transformer network trained and tested on the labeled transcription start sites obtained from RegulonDB [29], Etwiller . [9] (Cap(pable)-seq), Yan . [40] (SMRT-(Cappable-)seq), Ju . [18] (SEnd-seq) and the curated set : A model has been trained (rows) and evaluated (columns) on each set of annotations. Both the Area Under the Receiver Operating Characteristics Curve (ROC AUC) and Area Under the Precision Recall Curve (PR AUC) are given for each set-up.

Train set	Test set
	RegulonDB	Cap-seq	SMRT-seq	SEnd-seq	Custom
ROC AUC
RegulonDB [29]	0.882	0.815	0.923	0.882	0.885
Cap-seq [9]	0.790	0.961	0.938	0.945	0.945
SMRT-seq [40]	0.749	0.899	0.958	0.961	0.956
SEnd-seq [18]	0.669	0.835	0.944	0.978	0.964
Curated	0.740	0.920	0.976	0.981	0.976
PR AUC
RegulonDB [29]	0.030	0.026	0.053	0.057	0.064
Cap-seq [9]	0.014	0.132	0.029	0.039	0.044
SMRT-seq [40]	0.029	0.044	0.086	0.081	0.089
SEnd-seq [18]	0.035	0.052	0.098	0.128	0.137
Curated	0.039	0.057	0.141	0.128	0.141

The best performances on each test set are given in boldface.