Predictio no fCoordinatio nNumbe ran dRelativ eSolvent Accessibilit yi nProteins
Pierr eBaldi · 2002
Knowin gth ecoordinatio nnumber an drelativ esolven taccessibilit yo fal lth eresidues i na protei ni scrucia lfo rderivin gconstraint suseful i nmodelin gprotei nfoldin gan dprotei nstructure an di nscorin gremot ehomolog ysearches .W ede- velo pensemble so fbidirectiona lrecurren tneural networ karchitecture st oimprov eth estat eo fthe ar ti nbot hcontac tan daccessibilit yprediction, leveragin ga larg ecorpu so fcurate ddat atogether wit hevolutionar yinformation .Th eensemble sare use dt odiscriminat ebetwee ntw odifferen tstate sof residu econtact so rrelativ esolven taccessibility, highe ro rlowe rtha na threshol ddetermine db ythe averag evalu eo fth eresidu edistributio no rthe accessibilit ycutoff .Fo rcoordinatio nnumbers ,the ensembl eachieve sperformance srangin gwithin 70. 6-73.9 %dependin go nth eradiu sadopte dt odis- criminat econtact s(6A-12A) .Thes eperformances represen tgain so f1 6-20 %ove rth ebaselin estatisti- ca lpredictor ,alway sassignin ga namin oaci dt othe larges tclass ,an dar e 4-7 %bette rtha nan yprevious method .A combinatio no fdifferen tradiu spredic- tor sfurthe rimprove sperformance .Fo raccessibil- it ythreshold si nth erelevan t15-30 %range ,the ensembl econsistentl yachieve sa performanc eabove 77% ,whic hi s1 0-16 %abov eth ebaselin eprediction an dbette rtha nothe rexistin gpredictors ,b yu pto severa lpercentag epoints .Fo rbot hproblems ,we quantif yth eimprovemen tdu et oevolutionar yinfor- matio ni nth efor mo fPSI-BLAST-generate dprofiles ove rBLAS Tprofiles .Th epredictio nprogram sare implemente di nth efor mo ftw owe bservers ,CON- pr oan dACCpro ,availabl ea thttp://promoter.ics. uci.edu/BRNN-PRED/ .Protein s2002;47:142-153.