TABLE 2.

Performance of models trained on prosite TP/TN on held‐out prosite FP/FN annotations

Site	prosite label	COLLAPSE	FEATURE	3DCNN	prosite total
ADH_SHORT	FN	11	8	7	14
	FP	30	33	33	33
EF_HAND_1	FN	40	28	34	48
	FP	120	106	125	128
EGF_1	FN	60	34	58	90
	FP	19	19	19	19
IG_MHC	FN	21	8	8	47
	FP	31	31	31	31
PROTEIN_KINASE_ST	FN	269	264	268	271
PROTEIN_KINASE_TYR	FN	3	3	3	3
	FP	14	20	20	20
TRYPSIN_HIS	FN	10	3	10	16
	FP	4	4	4	4
TRYPSIN_SER	FN	9	9	9	12
	FP	1	1	1	1

Note: Comparisons are made with FEATURE and 3DCNN numbers as reported in Torng et al. (2019). The number of proteins which are correctly reclassified (i.e., FPs predicted as negative, FNs predicted as positive) is reported for each method (higher is better). Numbers in bold indicate best performance on each site.