All of the largest fragments were again statistically significant. Remodeling proteins onto templates with different bind ing preferences generally did not enhance similarity between the thenthereby binding cavities of modeled proteins and the cavity in the template. In fact remodeling appears to frequently accentuate structural differences. These results suggest that, in the context of a general applica tion, remodeling a set of proteins with both similar and different binding preferences may contribute to a more sensitive classification of proteins with different binding preferences. Binding site variations in structure predictions We performed medial remodeling on all dataset struc tures onto all template structures.
For each of the 100 models generated between each template sequence pair, Inhibitors,Modulators,Libraries we measured the volume of the largest fragment between the binding site of the Inhibitors,Modulators,Libraries model and that of the Inhibitors,Modulators,Libraries template. These volumes varied considerably between maximum and minimum, even when range between the 25th percentile volume and the 75th volume was very narrow. For example, among the enolases in our dataset, with the 1e9i template, the largest fragments were larger than 750 3, even though more than half of the largest fragments were approximately 75 3. With the 1ebh template, the largest fragments were larger than 850 3, though most of the largest fragments were approximately 75 3. Among the kinases, with the 1qcf template, the largest fragment was 1387 3, though most of the largest fragments were approximately 92 3. With the 2hz4 template, the largest fragment was 1438 3, and most of the largest fragments were approxi mately 342 3.
In isolation, the generation of a modeled structure with an unusual binding cavity is not very high. How ever, in the context of a structural classification effort, where many structures must be modeled, the probability of generating at least one unusual model increases with the size of the dataset, through multiple testing. By eliminating extrema, Inhibitors,Modulators,Libraries we hypothesize that medial remodeling can maintain accurate classification despite the nondeterministic nature of structure prediction. Evaluating medial remodeling We performed medial remodeling on all dataset struc tures onto all six template structures, computing the median of the volumes of the largest fragments in all 100 models.
For dataset structures Inhibitors,Modulators,Libraries remodeled onto the eno lase templates, 1e9i and 1ebh with similar binding preferences, the median volume of the largest fragment was never statistically significant, except in the case of 2pa6. The largest fragment computed with an unmodeled structure was larger, sometimes considerably larger, than the median volume, selleck products and larger still in cases of conformation change. In the case of 2XSX and most of the sequentially redundant enolase structures, the lar gest fragment from an unmodeled structure is statistically significant, and thus indistinguishable from proteins with different binding preferences.