The authors of the STRuster ( 9) method explore the calculation of root mean square deviations (RMSD) and use their algorithm to cluster alternative structural models from the PDB (i.e. models that correspond to different structure determination experiments). In addition to the traditional RMSD measure, the STRuster method uses two filters to define the final scoring metric called dissimilarity measure M ( 9). These two filters are introduced in order to identify both large and small (but significant) backbone conformational changes by reducing the influence in local large distances (only distances below 14.0 Å are considered) and also to restrict the analysis to significant structural differences (the distances above 1.0 Å). An approach for structural comparisons, fundamentally different from those using RMSD, was proposed by Rogen and Fain ( 10). They introduced the SGM (Scaled Gauss Metric), which is a metric derived from knot theoretical ideas to cluster proteins according to their structural topologies. They applied their method to predicting membership of proteins in CATH and achieved 95% accuracy at all levels of the classification hierarchy.
In order to achieve a high level of agreement with other clustering schemes, some algorithms that use a multi-criterion approach (weighted combination of different scoring schemes), are initially trained on labeled data from an existing structural hierarchy (SCOP or CATH) and use cross-validation (or similar methods) to select the best parameters for their classifiers. For example, ProtClass ( 11) uses a nearest-neighbor-based classification scheme and several structural features to classify proteins at the fold level of the SCOP hierarchy.
Their features include secondary structure elements predicted by the Stride program ( 12), the sequence length, and the percentage of observed helices. SCOPmap ( 13) is an approach that achieves roughly 95% accuracy when classifying proteins into the superfamily level of the SCOP hierarchy.
0 kommentar(er)
