Optimization of protein models

Stages of integrative structure determination and analogous structure prediction. Structure determination by integrating varied data from experiments and modeling can be divided into four steps12: (1) generation of structural data by experiment, (2) data translation into spatial restraints, (3) optimization, and (4) ensemble analysis. These four steps are also characteristic for structure prediction, shown here for easy modeling cases (unambiguous data derived from homologous proteins) and difficult ones (ambiguous or sparse data) embedding fold recognition methods, together with de novo modeling.

Distribution of Δcrmsd (coordinate root mean square deviation) for each method. Successful results are below the red dashed line. Only for eight of 34 methods have the sets of the resulting structures mean Δcrmsd lower than zero.

Distribution of coordinate root mean square deviation (crmsd) differences between structures before and after refinement, obtained during the CASP9 experiment. Green color (negative Δcrmsd) means improvements and red (positive Δcrmsd) means worsening of the original models.

Distribution of coordinate root mean square deviation (crmsd) of the starting structures for the refinement in CASP8 and CASP9. The majority of models are close to the native structure.

Results of TR592 refinement from the CASP9 refinement competition. Top: Three superimposed structures are shown—native in blue, starting model in gray [coordinate root mean square deviation (crmsd): 1.26], and the best model in magenta (crmsd: 0.96). Significant improvement from the starting model can be observed in the loop region marked with a circle. It is noteworthy that this region was pointed out by the competition assessors to the participating groups as one of the main areas for refinement. Bottom: In the same size and orientation as mentioned above; all the predicted models (designated by participating groups as the best) are shown in magenta, together with the native in blue (a few heavily mismatched models were removed for clarity).

Approximate dependence of sequence alignment accuracy and expected models accuracy on the percentage of sequence identity. The classification of model accuracy (see Table 1) practically agrees with alignment accuracy, which falls into one of the following three zones of sequence similarity (defined by Rost15): safe, twilight, and midnight zone. The twilight zone denotes a huge drop in alignment accuracy (roughly in the range of 25%–30% of sequence identity) from the safe zone (high level of sequence similarity—proteins also have similar structures and functions) to the midnight zone (low level of sequence similarity—protein similarity cannot be detected from sequence comparisons alone).