. Author manuscript; available in PMC: 2022 Mar 22.

Published in final edited form as: Nat Comput Sci. 2021 Jul 15;1(7):462–469. doi: 10.1038/s43588-021-00098-9

Table 3.

GNNRefine’s performance with different features and training data on the CASP13 data. For the AtomEmb (with local frame), we use Cα, N, and C to define the reference frame of atom coordinates for each residue.

Features	Training data	GDT-HA	GDT-TS	lDDT	Degradation
Features	Training data	GDT-HA	GDT-TS	lDDT	0	−1	−2
All features	In-house	+3.15	+1.96	+2.88	1	0	0
All features	DeepAccNet data	+3.19	+1.75	+2.74	3	1	1
All features	CASP models only	+1.42	+0.92	+1.35	8	6	3
no Orientation	In-house	+2.21	+1.28	+2.26	4	2	0
no Dihedral&SS&RSA	In-house	+2.53	+1.67	+2.31	2	0	0
no AtomEmb	In-house	+3.25	+2.03	+2.57	2	0	0
AtomEmb (with local frame)	In-house	+3.05	+1.82	+2.50	3	1	1

GDT-HA: Global Distance Test High Accuracy, GDT-TS: Global Distance Test Total Score, lDDT: Local Distance Difference Test, Degradation: the number of refined models has quality worse than their initial models by a given threshold based on GDT-HA.