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

Prediction of protein function based on their theoretical models
Protein models as a publicly available resource
Model quality

The quality of a Cell/13.html">Protein Structure determines its suitability for functional prediction. While the majority of experimentally solved structures possess sufficient resolution to elucidate their chemical mechanisms, the same cannot always be said for protein models. At the same time, the reliability of any functional insights derived from a structure or model directly depends on its quality. For instance, a model of average quality should be used with extreme caution for Applications such as molecular docking of drug candidates and their analogs. As mentioned above, the accuracy of Homology-based models decreases as sequence identity between the target and template Proteins diminishes (see also Chapter 3). This leads to an empirical rule of thumb: models built on closely related templates tend to be “good,” whereas those derived from distant templates are typically “poor.” However, this rule assumes that optimal alignments were used during model construction and that no refinement of the initial conformation toward the true structure was performed. Furthermore, the degree of sequence identity to a template cannot serve as a measure of model quality for de novo models (Chapter 1), as they do not rely on templates at all.

Recently, several software tools have been developed to assess the quality of protein models without requiring information about their native structures. Some of these approaches (e.g., PROQ (Wallner and Elofsson 2003)) are designed to distinguish between native, near-native, and non-native structures, whereas others (e.g., PROQres (Wallner and Elofsson 2006), ModFold (McGuffin 2008), or MetaMQAP (Pawłowski et al. 2008)) focus on predicting how severely various regions of a model may deviate from their counterparts in the native structure. Currently, there is no universal scoring function capable of reliably estimating both global and local accuracy of structural models. Nevertheless, functional inferences drawn from theoretical models must take into account the predicted global model quality when considering overall properties (such as charge or surface topology), and local quality when focusing on active sites or other localized regions. In the latter case, Methods that predict the deviation of specific model fragments from the unknown native structure prove particularly valuable.

Incorporating such quality estimates can at least help prevent overinterpretation of a model. For example, it is clearly pointless to analyze the interaction geometry between side chains within a potential catalytic center if the backbone coordinates of those residues are predicted with an accuracy of only 5 Å.