From Protein Structure to Function with Bioinformatics.pdf

32 L.A. Kelleyalways unique) structure driven by the energetically favourable interactionsbetween the amino acids within the structure, and between the amino acids and thesurrounding solvent.If one were able to understand what spatial and solvent interactions stabilise agiven structure, then one could both detect compatible sequences given a structureand design sequences that fit that structure. This is the concept of threading. Givena sequence whose structure we wish to predict, one aligns or ‘drapes’ this sequenceover each of the known structures in our database. In each case one calculates ascore to represent how favourable our sequence is with each structure. A structurewith a highly favourable score will be our prediction. But what are these favourableinteractions and how do we calculate their magnitude? Fortunately, thanks to thediligent work of many experimentalists around the world, we have a database ofnative protein structures; a database of favourable interactions.By careful statistical analysis of the distribution of the different amino acid typesthroughout known protein structures, powerful sequence-structure relationships can beinferred, and used to tackle prediction problems. These empirically-derived or ‘knowledge-based’force fields are widely used across the entire spectrum of protein structureprediction techniques and their key role in ab initio modelling means many of thedetails may be found in that chapter. Nevertheless, a brief summary will be useful.2.2.1 Knowledge-Based PotentialsTo empirically derive rules relating protein sequence to three-dimensional structurerequires (1) a large number of examples of sequences and their correspondingstructures and (2) a structural feature of proteins one wishes to analyse. A simpleillustration of the technique is the generation of a solvation potential. Any globularprotein in its folded native state has some residues buried in the (largely hydrophobic)interior and some residues (largely hydrophilic) on the surface exposed to the surroundingsolvent. It is straightforward to calculate to what extent a given residue R isexposed or buried in a protein of known structure. One method, albeit crude, issimply to measure how many other residues are within a certain distance of theresidue R (more sophisticated methods are usually used; Richmond 1984; Kabschand Sander 1983). So it is possible to compile a list of every residue in every knownprotein structure together with its associated level of solvent accessibility (in termsof neighbours). With these data it is now possible to use a variety of statisticaltechniques to attempt to discover any relationship between amino acid type and itspropensity to be on the interior or exterior of the protein. One of the common methodsused is based on statistical mechanics or Bayesian statistics (for a comparison withother methods see Xia and Levitt 2000). First proposed by Tanaka and Scheraga(1976) and later refined by Sippl (1990) and Myazawa and Jernigan (1996), thesemethods all rely on Boltzmann statistics.First one assumes that protein structures in the database constitute a kind ofensemble and that the levels of solvent exposure of a residue type within proteins