QMEANclust: estimation of protein model quality by combining a composite scoring function with structural density information

Table 2 Comparison between QMEAN, various QMEANclust implementations and selfQMEAN on all CASP7 server models.

QMEAN implementation	Pearson	Spearman	Sum(GDT)
*QMEAN:*
QMEAN3	0.645	0.551	50.17
QMEAN3 * fraction modelled	0.663	0.605	51.92
QMEAN4	0.647	0.540	49.57
QMEAN4 * fraction modelled	0.671	0.609	52.65
QMEAN5	0.729	0.630	54.87
QMEAN5 * fraction modelled	0.740	0.676	55.32
QMEAN6	0.741	0.638	56.36
QMEAN6 * fraction modelled	0.752	0.684	56.70
*QMEANclust: no preselection*
Median	0.872	0.812	56.64
Mean (~3D-jury based on GDT_TS)	0.889	0.821	57.16
Weighted mean	0.883	0.824	57.63
*QMEANclust: QMEAN Z-score > x*
Median: Z-score > -1	0.877	0.815	57.05
Mean: Z-score > -1	0.876	0.817	57.30
Weighted mean: Z-score > -1	0.882	0.823	57.60
Median: Z-score > 0	0.884	0.824	57.52
Mean: Z-score > 0	0.879	0.822	57.35
Weighted mean: Z-score > 0	0.882	0.826	57.31
Median: Z-score > 0.5	0.885	0.828	57.33
Mean: Z-score > 0.5	0.880	0.830	56.96
Weighted mean: Z-score > 0.5	0.883	0.832	57.18
*QMEANclust: top × percent models*
Median: 20% TBM, 20% FM	0.888	0.842	57.37
Median: 10% TBM, 10% FM	0.890	0.844	57.83
Median: 5% TBM, 5% FM	0.873	0.826	56.98
Median: 10% TBM, 20% FM	0.886	0.844	57.23
Median: 20% TBM, 10% FM	0.892	0.842	57.97
*QMEANclust: ΔQMEAN-score from max*
Median: Δ < 0.05 Å TBM, Δ < 0.05 Å FM	0.867	0.826	57.65
Median: Δ < 0.1 Å TBM, Δ < 0.1 Å FM	0.892	0.837	57.69
Median: Δ < 0.05 Å TBM, Δ < 0.1 Å FM	0.892	0.841	58.11
Median: Δ < 0.1 Å TBM, Δ < 0.05 Å FM	0.868	0.822	57.23
*selfQMEAN:*
Linear combination of 5 terms (w/o all-atom)	0.811	0.755	55.53
Sum of Z-scores (5 terms)	0.830	0.749	56.60
Sum of Z-scores (6 terms)	0.832	0.753	55.60

Average correlation coefficient and total maximum GDT_TS score of the selected models of different QMEAN versions obtained on the test set containing all CASP7 server models. A description of all QMEAN versions is given in Table 1. For the QMEANclust consensus score, a multitude of strategies for pre-selecting reference models based on QMEAN score is investigated. The models of the reference set are defined based on a certain Z-score cut-off, by using only a percentage of top scoring models or by including only models being close to the highest scoring model. The different cut-offs used for template-based modelling targets (TBM) of free modelling targets (FM) are indicated. Underlined values are used in Table 3 for comparison to other methods. The selfQMEAN scoring function is based on ensemble-specific statistical potentials.

ISSN: 1472-6807