Table 3.

Comparing the performance of SPOT-1D-LM with single-sequence-based methods (SPIDER3-Single, ProteinUnet, and SPOT-1D-Single) and sequence-profile-based methods (SPOT-1D and NetSurfP-2.0) in the prediction of secondary structure in three (SS3) and eight (SS8) states, solvent accessibility (ASA), half-sphere-exposure-up (HSE-u), HSE-down (HSE-d), contact number (CN), backbone angles ($\psi$, $\varphi$, $\theta$ and $\tau$) for TEST2020. Performance measures are accuracy for SS3 and SS8, correlation coefficient for ASA, HSE-u, HSE-d, and CN, and mean absolute errors for the angles.

Model	SS3	SS8	ASA	HSE-u	HSE-d	CN	$\psi$	$\varphi$	$\theta$	$\tau$
SPIDER3-Single	71.31	57.57	0.596	0.358	0.417	0.434	45.64	23.48	11.52	46.04
ProteinUnet	72.20	58.71	0.555	0.366	0.426	0.441	44.87	23.19	10.49	44.95
SPOT-1D-Single	73.80	60.35	0.621	0.400	0.478	0.485	44.25	22.92	9.88	43.67
NetSurfP-2.0(profile)	79.42	66.36	0.702	–	–	–	35.07	20.70	–	–
SPOT-1D (profile)	80.52	67.76	0.691	0.516	0.594	0.600	34.46	20.33	8.50	33.64
SPOT-1D-LM (This work)	79.82	66.68	0.731	0.522	0.597	0.704	36.57	20.67	8.67	35.80