Skip to main content
. Author manuscript; available in PMC: 2025 Jan 4.
Published in final edited form as: Structure. 2023 Nov 23;32(1):97–111.e6. doi: 10.1016/j.str.2023.10.013

Table 2. Comparison of all-atom reconstruction accuracies with different methods.

See also Figures S3, S4, and S5.

Input Method RMSD1 χ-angle accuracy1,2 MolProbity1

Backbone [Å] Heavy atom [Å] χ1 [%] χ1+2 [%] Score Clash score Rama favor [%] Rotamer outlier [%]

cg2all 0.18 (0.05) 0.96 (0.12) 86.2 (3.0) 71.4 (4.8) 2.07 (0.21) 31.2 (9.5) 97.9 (1.1) 0.8 (0.7)
w/SCRWL3 1.06 (0.14) 83.2 (3.6) 70.1 (5.2) 2.00 (0.18) 28.6 (8.1) 97.9 (1.1) 0.0 (0.1)
cg2all (fixed)4 0.17 (0.05) 0.93 (0.12) 86.5 (3.0) 71.8 (4.7) 2.13 (0.21) 34.2 (10.3) 98.0 (1.1) 1.0 (0.7)

PULCHRA5 0.47 (0.11) 1.57 (0.14) 59.2 (4.0) 39.6 (4.5) 3.79 (0.23) 164.4 (28.5) 86.4 (3.9) 4.9 (1.7)
w/SCWRL3 1.36 (0.14) 73.0 (4.3) 58.4 (5.6) 2.90 (0.20) 67.4 (17.8) 86.4 (3.9) 0.1 (0.2)

REMO5,6 0.81 (0.42) 2.09 (0.41) 43.6 (4.3) 28.5 (5.4) 4.37 (0.21) 200.2 (36.8) 78.4 (7.2) 14.8 (3.8)
w/SCWRL3 1.74 (0.52) 68.5 (6.5) 53.8 (7.5) 3.15 (0.25) 95.7 (43.6) 78.4 (7.2) 0.2 (0.3)

ModRefiner5,6 0.66 (0.20) 1.51 (0.21) 71.8 (4.1) 55.1 (5.6) 2.38 (0.24) 56.5 (19.6) 97.0 (1.7) 0.6 (0.5)
MODELLER 0.97 (0.88) 2.12 (0.76) 42.8 (4.0) 25.4 (4.0) 3.63 (0.19) 93.3 (15.7) 86.4 (3.4) 6.0 (1.8)

CM cg2all 0.22 (0.05) 0.46 (0.06) 95.4 (1.9) 85.9 (3.9) 2.00 (0.23) 20.6 (6.2) 97.3 (1.4) 1.1 (0.7)
w/SCWRL3 1.00 (0.14) 83.9 (3.6) 70.9 (5.2) 2.01 (0.20) 25.5 (7.3) 97.3 (1.4) 0.0 (0.1)

PULCHRA5 1.08 (0.08) 1.91 (0.12) 46.6 (4.0) 29.0 (4.1) 4.49 (0.13) 291.6 (36.2) 72.9 (4.4) 10.5 (2.3)
w/SCWRL3 1.08 (0.08) 1.84 (0.12) 55.3 (4.0) 36.3 (4.5) 3.51 (0.11) 176.2 (27.6) 72.9 (4.4) 0.2 (0.3)
MODELLER5 1.63 (1.03) 2.38 (0.94) 43.7 (3.9) 25.5 (3.8) 3.80 (0.18) 123.6 (19.9) 81.8 (3.4) 5.5 (1.8)

N, Cα, C, O cg2all 0.04 (0.01) 0.82 (0.11) 89.6 (2.8) 75.6 (4.7) 2.08 (0.22) 26.9 (8.3) 97.4 (1.3) 1.0 (0.8)
cg2all (fixed)4 - 0.82 (0.11) 89.7 (2.8) 75.7 (4.6) 2.05 (0.22) 27.3 (8.3) 97.9 (1.2) 1.1 (0.8)

SCWRL - 0.97 (0.13) 85.6 (3.4) 73.0 (5.0) 1.97 (0.19) 26.7 (7.8) 97.9 (1.2) 0.0 (0.1)

AttnPacker - 0.61 (0.11) 90.9 (2.9) 73.3 (5.4) 2.32 (0.25) 22.8 (7.7) 97.9 (1.2) 3.8 (1.5)
+local min. - 0.67 (0.11) 92.4 (2.8) 80.6 (4.9) 1.86 (0.28) 11.9 (3.9) 97.9 (1.2) 2.0 (1.1)

MARTINI cg2all 0.08 (0.02) 0.31 (0.05) 98.8 (1.0) 93.1 (2.8) 1.88 (0.21) 17.2 (5.4) 97.5 (1.2) 0.9 (0.7)
w/SCWRL3 0.98 (0.13) 85.2 (3.6) 72.6 (5.2) 2.00 (0.19) 26.1 (7.5) 97.5 (1.2) 0.0 (0.1)

Backward7 0.84 (0.07) 1.06 (0.08) 60.9 (5.8) 46.1 (7.2) 2.71 (0.29) 4.5 (1.7) 86.7 (4.5) 15.9 (3.8)
w/SCWRL3 1.64 (016. 64.5 (55) 50.4 (66. 2.93 (0.22) 74.4 (2L4) 86.7 (4.5) 0.1 (02.
1

The average reconstruction accuracy measures for the test set protein structures (n=720) are given with their standard deviations in the parentheses.

2

Side chain χ-angles were considered accurate when their deviations from experimental values were less than 30 degrees.

3

Side chains were reconstructed using SCWRL4 after building a backbone structure using other methods (e.g., cg2all, PULCHRA, and REMO).

4

The atomic coordinates in the input files were preserved, while the original method does not. For instance, cg2all model for “Cα (fixed)” generates output structures with the exact same Cα coordinates of the input structures. On the other hand, the original cg2all model generates slightly altered Cα coordinates.

5

Chains in multi-chain targets were separately converted to all-atom structures and superposed onto the original Cα-trace.

6

Conversions of several structures failed because they cannot handle short peptides or were not completed within a reasonable time frame (12 hours). Successful conversions for REMO: 684/720; for ModRefiner: 714/720.

7

Conversion of one structure failed because short peptides (<3 residues) cannot be handled.