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

Prediction of protein function based on their theoretical models
Accuracy and added value of model-based predictions

Although an ideal algorithm for assessing overall model accuracy has yet to be developed, it is already possible to determine the average prediction accuracy of specific properties derived from structural models. Most Proteins rely on intermolecular recognition to carry out their Functions, engaging ligands that range from small molecules to multi-subunit Structure/178.html">Protein Complexes. Consequently, Surface Properties—particularly residue exposure, accessible surface area, clefts, and electrostatic potential—are directly related to protein function.

The accuracy of model-based property predictions was evaluated through a large-scale analysis of simple comparative models (Chakravarty et al. 2005). The findings demonstrated that the overall accuracy of all structural properties included in the analysis decreases as the sequence identity between the target and template proteins declines; however, this degradation impacts the accuracy of individual properties to varying degrees (Table 12.2).

Class="center">Table 12.2. Accuracy and added value of structural properties in single-template comparative modeling (Chakravarty and Sanchez 2004; Chakravarty et al. 2005)

For instance, alignment errors have a negligible effect on the prediction accuracy of the accessible surface area (ASA), whereas the accuracy of electrostatic potential predictions begins to deteriorate when sequence identity drops below 50% (Table 12.2). It is worth emphasizing that this analysis was performed using a set of single-template comparative models without loop modeling. Therefore, these results are representative of high-throughput automated modeling approaches. More rigorous modeling protocols—incorporating multiple templates and model optimization—can yield superior results and enhance the prediction accuracy of specific properties at a given sequence identity.

Once the reliability of various model properties is established, it is of interest to determine what additional information (added value) they provide relative to the template structure used for model construction. As before, a systematic analysis of added value was conducted on a large set of single-template models (Chakravarty and Sanchez 2004), yielding valuable insights into which specific model properties are truly informative (Table 12.2).

In general, the greater the divergence between the template and target sequences, the more significant the added value becomes. This follows from the fact that at low sequence similarity, the template contains less information regarding the size and PHYSICOCHEMICAL PROPERTIES OF specific residues in the target. However, not all structural properties gain additional information compared to the template. For those properties governed primarily by residue position—such as exposure, buried residue neighborhood, and surface pocket count—models offer no added value. This is likely because buried residues form the structural core responsible for protein integrity and are consequently more conserved than exposed residues.

For other structural properties, such as the exposed residue neighborhood or total accessible surface area, models do provide a degree of added value. This is particularly important when solvent-accessible residues mediate interactions with other molecules, thereby determining the biological function of the protein.

Finally, for protein properties heavily dependent on the PHYSICOCHEMICAL CHARACTERISTICS OF the Amino Acid Sequence, such as pocket composition and electrostatic potential, models deliver substantial added value. Indeed, identifying charged regions is critical, as these areas frequently comprise binding sites or active centers (see Chapter 7).

In summary, the work by Chakravarty et al. demonstrated that, with the exception of pocket detection, the majority of structural model properties possess measurable added value. Furthermore, the more strongly a property depends on The amino acid sequence, the more useful the model proves in estimating that property. Encouragingly, depending on the specific property evaluated, a target-template sequence identity of 25–40% yields an accuracy comparable to that obtainable from an NMR structure.