Protein Structure and Function. Application of Bioinformatics Methods - John Rigden 2014

Fold Recognition
Alignment Accuracy, Model Quality, and Statistical Significance

Structure/29.html">Fold Recognition can be divided into two main challenges: 1) identifying a suitable template, and 2) aligning the target sequence with that template. Obviously, any useful method for Cell/13.html">Protein Structure Prediction based on template modeling must at least be capable of identifying appropriate templates. Nevertheless, the quality of the resulting model is not constrained by template detection alone, but is entirely determined by the quality of the alignment. Errors in alignment will inevitably lead to a low-quality model, even when a high-quality template is used.

Up to this point, we have assumed that a system capable of accurately identifying templates will also produce accurate alignments. Exceptions include certain support vector machine-based classifiers discussed in Section 3.3. Although this assumption generally holds true, there are numerous instances where it fails. First, in most of the described systems, templates are essentially ranked using some alignment score. In other words, such Methods always yield a model with the highest score. However, the fact that a particular alignment scores higher than others does not mean it is error-free. Second, most of these methods attempt to detect features of extremely distant Homology. Consequently, they may simply identify a few conserved motifs, or patches of similarity, interrupted by long stretches of sequence where no similarity can be established. This ultimately yields an alignment that is essentially "noise." In turn, this leads to significant inaccuracies in the resulting 3D structural model.

For these reasons, many research groups are investigating methods to improve alignment accuracy and evaluate the quality of the resulting models. There are three general approaches to addressing the alignment accuracy problem: 1) directly improving alignment algorithms; 2) generating a large ensemble of alternative alignments combined with a Selection system to choose the best one; and 3) building Spatial Models from multiple alternative alignments and assessing the resulting models.