From Protein Structure to Function with Bioinformatics.pdf

4 Membrane Protein Structure Prediction 105process is continued until all subsets have been validated. Two types are commonin TM topology prediction. In K-fold cross-validation, the data set is partitionedinto K subsets. Of the K subsets, a single subset containing a number of sequencesis retained as validation data for testing the model, while the remaining K-1 subsetsare used as training data. This process is then repeated K times (folds), with eachof the K subsets being used exactly once as the validation data. The K results fromthe folds can then either be combined or averaged to produce a single estimation.A more stringent, although computationally more intensive form of cross-validationis leave-one-out cross-validation (LOOCV), also referred to as a jack knife test.Jack knifing involves testing a single sequence from the data set against the remainingsequences which make up the training set, then repeating the test such that everysequence is validated once. This is the same as a K-fold cross-validation with Kbeing equal to the number of sequences in the data set.While some studies have attempted to compare TM topology prediction accuracybetween different methods (e.g. Melén et al. 2003), significant progress hasbeen made since then. Currently, the best TM topology predictors claim to predictcorrect topologies for 80–93% of proteins, though in the absence of independentcross-validation using a common test set it is difficult to accurately compare methods.Those which perform well when tested on a particular data set, e.g. one containingfew signal peptides, may perform poorly when tested on a data set whichcontains many signal peptides. Methods optimised on a data set containing manyweakly hydrophobic TM helices may tend to over predict TM helices in other datasets. Current gold-standard TM protein data sets with topologies derived solelyfrom structural data contain no more than 150 sequences when homology reduced(Lomize et al. 2006b), but a lack of consensus amongst these combined with thescarcity of necessary cross-validation files means that differences in accuracybetween methods may thus be a result of differences in training and validation datasets rather than significant differences in performance.4.7 3D Structure PredictionAs with globular proteins, 3D structure prediction of TM proteins can be dealt withvia two approaches, homology modelling and ab initio modelling, covered inChapters 1 and 3 of this book.Homology modelling, also known as comparative modelling, involves the use of arelated template structure in order to build a 3D model of a target protein. The methodis based on the observation that protein structure is conserved more highly than aminoacid sequence, hence even proteins that have diverged significantly in sequence butstill share detectable similarity (>30% sequence identity) may also share commonstructural properties, particularly the overall fold. Due to the difficulties involved inobtaining high-resolution crystal structures, particularly with regard to TM proteins,homology modelling can provide useful structural models for generating hypothesesabout a protein’s function and directing further experimental work. The process can be