13.07.2015 Views

computer modeling in molecular biology.pdf

computer modeling in molecular biology.pdf

computer modeling in molecular biology.pdf

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

32 Tim J I? Hubbard and Arthur M. LeskNote Added <strong>in</strong> ProofThere is a serious ‘catch 22’ like problem <strong>in</strong> evaluat<strong>in</strong>g the effectiveness of prote<strong>in</strong>modell<strong>in</strong>g: if you model a structure that is known, you cannot be sure how biasedyou were by that prior knowledge (s<strong>in</strong>ce almost no modell<strong>in</strong>g system is entirely ablack box) whereas if you model a structure that is unknown you cannot assess theaccuracy of your models. One way around this is to build models ‘just <strong>in</strong> time’, i. e.immediately before publication of an experimentally-determ<strong>in</strong>ed structure, so it ispossible to evaluate the accuracy of your model with the confidence that it was abl<strong>in</strong>d prediction. When this chapter was written there had been isolated examples ofthis sort of arrangement between theoreticians and experimentalists but they werequite rare.In the last month there has been a meet<strong>in</strong>g to evaluate the first ever large scaleprote<strong>in</strong> structure prediction competition, which ran for most of 1994 [94]. -35groups made - 150 predictions about - 25 target prote<strong>in</strong>s. The predictions were considered<strong>in</strong> three categories : homology modell<strong>in</strong>g, fold recognition and ab <strong>in</strong>itioprediction. The results were <strong>in</strong>structive:Homology modell<strong>in</strong>g naturally gives the most reliable predictions, but despite theefforts made to automate the modell<strong>in</strong>g process, it is clear that where the templatestructure used to build the model differs substantially from the experimental structurethe model is generally wrong: we are unable to model the variations that commonlyoccur between homologous prote<strong>in</strong>s (loops, man cha<strong>in</strong> shift and the associateddifferent side cha<strong>in</strong> pack<strong>in</strong>g) with much greater accuracy than was possible byhand 10-15 years ago.Although the accuracy of homology modell<strong>in</strong>g was disappo<strong>in</strong>t<strong>in</strong>g, the number oftargets that could potentially be modelled based on a template structure is go<strong>in</strong>g to<strong>in</strong>crease, s<strong>in</strong>ce the meet<strong>in</strong>g demonstrated that fold recognition techniques (us<strong>in</strong>g newmethods such as ‘thread<strong>in</strong>g’) can already <strong>in</strong>dentify the most similar fold <strong>in</strong> the structuredatabase <strong>in</strong> a substantial number of cases. Thread<strong>in</strong>g is still a very young techniqueand it is clear that many improvements can be made, so the accuracy and sensitivitycan only <strong>in</strong>crease.F<strong>in</strong>ally, it does appear that useful ab <strong>in</strong>itio structure predictions can be made fortargets where there are many homologous sequences. Secondary structure predictionby the PHD method [13, 141 <strong>in</strong> such cases is sufficiently reliable for predictors to considerhow these secondary structural elements might be assembled (i. e. to attempta full tertiary prediction) and new techniques are emerg<strong>in</strong>g to predict such long range<strong>in</strong>teractions based on specialized potentials [95] and correlation <strong>in</strong>formation 1961.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!