Table 4.

Remote homology and fold detection performance on the SCOP 1.67 benchmark dataset compared with literature results from GPkernel (Håndstad et al., 2007), LSTM_protein (Hochreiter et al., 2007) and ProDec-BLSTM (Li et al., 2017)

	Methods	Superfamily level		Fold level
		$\text{AUC}$	${\text{AUC}}_{50}$	$\text{AUC}$	${\text{AUC}}_{50}$
UDSMProt a	GPkernel	0.902	0.591	0.844	0.514
	LSTM_protein	0.942	0.773	0.821	0.571
	ProDec-BLSTM	0.969	0.849	—	—
	Fwd; from scratch	0.706	0.552	0.734	0.653
	Fwd; pretr.	0.957	0.880	0.834	0.734
	Bwd; pretr.	0.969	0.912	0.839	0.757
	Fwd+bwd; pretr.	0.972	0.914	0.862	0.776

Fwd/bwd, training in forward/backward direction; pretr., using language model pre-training. The best-performing classifiers are marked in bold face.

^aResults established in this work.