Table V.

Illustration of the effect of each step in the wfFUIK method on 13 CASP10 refinement targets.

TARGET	Start GDT_TS	Best in Unique/ Filtered Set	Lowest E dDFIRE	Lowest E Rosetta	Best in ICON	Best after ICON+ KoBaMIN	Best Submitted Model	Best Model 1 in CASP10	Best MODEL in CASP
TR661	80.68	78.51	59.46	55.27	69.69	71.49	65.81	81.35	81.35
TR663	69.24	69.74	41.94	53.62	54.44	54.77	54.93	74.84	77.3
TR679	71.73	74.25	68.22	67.71	71.48	71.61	69.85	73.74	74.37
TR688	78.24	78.24	68.78	67.43	72.7	72.43	71.76	79.73	80.14
TR696	71.5	73.5	61.75	51.25	63.5	63.5	63.5	75.5	76
TR698	65.55	72.27	64.71	59.66	64.71	65.13	64.71	67.02	67.65
TR705	64.84	73.7	65.36	63.02	70.57	70.83	64.84	70.05	71.35
TR710	75.13	83.76	71.78	70.62	80.41	79.51	78.09	77.83	80.28
TR722	57.09	69.69	42.32	58.27	61.02	65.75	65.75	65.75	65.75
TR723	85.11	89.31	72.52	86.45	86.45	86.83	86.83	88.17	91.41
TR747	83.61	89.44	86.39	78.33	86.39	87.78	83.89	85.28	86.11
TR752	90.37	91.22	86.32	83.28	88.01	88.51	88.51	90.71	90.88
TR754	78.31	80.15	48.16	56.25	72.79	71.32	71.32	79.04	81.98
ΣGDT_TS	971.4	1023.78	837.71	851.16	942.16	949.46	929.79	1009.01	1024.57
ΔΣGDT_TS from Start	0	52.38	−133.69	−120.24	−29.24	−21.94	−41.61	37.61	53.17

The first column indicates the CASP10 target. The second column indicates the starting structure’s GDT_TS value as calculated using the TMScore program⁴⁸. The third and fourth columns indicate the GDT_TS of the lowest energy dDFIRE²⁴ and Rosetta²³ structures, respectively. The fifth column indicates the best structure contained in the set of candidates after ICON clustering, in addition to the lowest energy dDFIRE and Rosetta structures. The seventh column indicates the GDT_TS of the best structure after refinement by KoBaMIN²⁷. The eighth column is the best structure submitted in blind predictions during CASP10 by the wfFUIK method. The ninth and tenth columns indicate the best Model 1 and best of the five Models submitted in CASP10 by any method. text indicates that the structure was a refinement of the start. text indicates that the best contained model is more accurate than any submitted during CASP10. This demonstrates the selection challenge the structure prediction community faces. That is, even though there are sampling methods that can generate high-accuracy structures, the best forcefields and selection methods still cannot pick these “best” structures among the ensembles. If one were able to select the best produced Foldit structure, generated by the game players, this would be comparable to the best structure submitted in CASP by all groups.