Using PubChem’s database with data mining and machine learning algorithms for the prediction of EGFR inhibitors: a comparative study

i" "" " " " " " " " " " " " " " " " " " " " " " " " " " " " " " Using&PubChem’s&da abase&wi h&Da a&Mining&and& Machine&Lea ning&Algo i hms& o & he&p edic ion& o &EGFR&inhibi o s&! Liliana"Mon ei o"Rosa" A"compa a i e"s udy! Disse a ion"p esen ed"as"pa ial" equi emen " o "ob aining" he"Mas e ’s"deg ee"in"In o ma ion"Managemen " " " i" " ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! Ti le:"Using"PubChem’s"da abase"and"Machine"Lea ning"Algo i hms" o " he"p edic ion"o "EGFR"inhibi o s" Sub i le:"A"compa a i e"s udy" S uden "" ull"name:"Liliana"Mon ei o"Rosa " MGI" 2017" 2017" Ti le:"Using"PubChem’s"da abase"and"Machine"Lea ning"Algo i hms" o " he"p edic ion"o "EGFR"inhibi o s" Sub i le:"A"compa a i e"s udy" " S uden " ull"name:"Liliana"Mon ei o"Rosa " MGI" i" " ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ii" " NOVA!In o ma ion!Managemen !School! Ins i u o!Supe io !de!Es a ís ica!e!Ges ão!de!In o mação" Uni e sidade"No a"de"Lisboa! ! USING!PUBCHEM’S!DATABASE!WITH!DATA!MINING!AND!MACHINE! LEARNING!ALGORITHMS!FOR!THE!PREDICTION!OF!EGFR! INHIBITORS:!A!COMPARATIVE!STUDY! ! ! by" Liliana"Mon ei o"Rosa" " " " " " " Disse a ion"p esen ed"as"pa ial" equi emen " o "ob aining" he"Mas e ’s"deg ee"in"In o ma ion" Managemen ,"wi h"a"specializa ion"in"Knowledge"Managemen "and"Business"In elligence." " " " Ad iso :"Mau o"Cas elli" " " " Augus "2017" " iii" " ACKNOWLEDGEMENTS! Fi s ly,"I"would"like" o" hank"P o esso "Mau o"Cas elli" o "accep ing"being"my"Ad iso "in" his" hesis"and" o "con ibu ing" o"my"in e es "in" his" opic"wi h"his"classes." Thank"you" o"all"my" amily"bu "mos "o "all" o"my"mo he ,"who"made"me"who"I"am"and" o"whom"I"owe" e e y hing" ha "I"ha e"accomplished,"academic"o "o he wise."" I"also"need" o" hank"my"aun "Cla isse"who"was"a"second"mo he " o"me"and"p o ided"me"wi h"inc edible" suppo " h oughou "my"academic"pa h."" To"Ana," o "being"my"pe son,"bo h"in"solida i y"in" he"pa h" o "comple ing"a"Mas e ’s"deg ee"and"in" e e y hing"else"in"li e." To"Rúben,"whose"suppo "was"e e "p esen "bo h"in" he"decision" o"en oll"in" his"Mas e ’s"deg ee"and" du ing"i s"comple ion"and"in"so"many"o he " imes,"e en" o"his"own"de imen ."" Thank"you"all." " " ! ! ! ! ! i " " ABSTRACT! Da a"Mining"and"Machine"Lea ning"algo i hms"and"me hods"ha e"become"inc easingly"impo an " o " se e al"indus ies"due" o" he"amoun "o "a ailable"da a" ha "has"g own"exponen ially"in" ecen "yea s" and"led" o" he"need"o "e ec i e"ways"o "gaining"insigh s" om" ha "da a." In" his" s udy," hese" me hods" a e" applied" o" he" p edic ion" o " Epide mal" G ow h" Fac o " Recep o " inhibi o s"using"da a"ex ac ed" om"PubChem’s"da abase."PubChem"is"a" eely"accessible"chemical" eposi o y" ha "con ains"in o ma ion"submi ed" om"se e al"di e en "sou ces,"and" ha "comp ises" h ee"da abases,"one"o "which"p o ides"in o ma ion"abou "BioAssays," ha "is,"assays"wi h" he"pu pose" o " sc eening" nume ous"compounds" o " ac i i y" on" a" pa icula " biological" a ge ." In" his" wo k," he" da ase "used" o" ain"and"e alua e" he"de eloped"models" esul ed" om" he"in o ma ion"ga he ed" om" he"assays"pe o med" o"iden i y"inhibi o s"o "EGFR"and" he"sou ce" o " he" ea u es"used" o"cha ac e ize" he"compounds"was"PubChem’s"own"chemical"desc ip o ," he"Subs uc u e"Finge p in ." The"wo k"comp ises"a"li e a u e" e iew"on" his"subjec "and" he"implemen a ion"o "a"me hodology" ha " es s" he"pe o mance"o "di e en " ypes"o "classi ie s" o " he"p oblem"a "hand,"namely"Naï e"Bayes," Decision" T ee," Logis ic" Reg ession," !-Nea es " Neighbo s," Suppo " Vec o " Machine," Mul ilaye " Pe cep on,"Random"Fo es ,"Ex emely"Randomized"T ees,"Bagging,"Boos ing"and"Vo ing." Conside ing"bo h" he"e alua ed"quali y"me ics"and" he"model’s"compu a ional"bu den," he"Mul ilaye " Pe cep on" was" conside ed" he" bes " model," al hough" some" o " he" o he "models" had" close" pe o mances."" I " was" concluded" ha " he" used" me hodology" and" de eloped" models" had" good" quali y," as" did" PubChem’s"Subs uc u e"Finge p in "as"a"desc ip o ,"bu " ha " he e"was"s ill" oom" o "imp o emen " ha "could"be"achie ed"wi h" u he "expe imen a ion"on"di e en "aspec s"o " he"me hodology." " " " KEYWORDS! Da a"Mining;"Machine"Lea ning;"Epide mal"G ow h"Fac o "Recep o ;"PubChem" " " " "" INDEX! 1."In oduc ion"................................................................................................................."1" 1.1."Objec i es"and"S udy"Rele ance"..........................................................................."1" 1.2."Documen "S uc u e"............................................................................................."2" 2."Li e a u e" e iew"........................................................................................................."3" 2.1."PubChem".............................................................................................................."3" 2.2."Re iew"o "wo ks"using" i ual"sc eening"on"PubChem"da a."................................"5" 2.3."Epide mal"G ow h"Fac o "Recep o "(EGFR)".........................................................."9" 2.4."Da a"Mining"and"Machine"Lea ning"Algo i hms"Theo e ical"F amewo k"..........."10" 2.4.1."Types"o "Da a"Mining"P oblems"..................................................................."11" 2.4.2."Theo e ical"backg ound"o " he"used"algo i hms".........................................."12" 3."Me hodology"............................................................................................................."24" 3.1."Used" ools"..........................................................................................................."24" 3.2."Da a"Ga he ing"and"T ea men ".........................................................................."24" 3.3."Choice"o "Algo i hms".........................................................................................."25" 3.4."Fea u e"Selec ion"................................................................................................"25" 3.5."T ea men "o "Imbalanced"da a"..........................................................................."26" 3.6."Model"E alua ion"................................................................................................"27" 3.6.1."The"Con usion"Ma ix".................................................................................."27" 3.6.2."Accu acy"Measu es"......................................................................................"28" 3.7."O e i ing".........................................................................................................."29" 3.8."Models’"Pa ame e "Op imiza ion"......................................................................."30" 3.9."S udy"wo k low"..................................................................................................."30" 4."Resul s"and"discussion"..............................................................................................."32" 4.1."Resul s"................................................................................................................"32" 4.1.1."Resul s" om"C oss"Valida ed"G id"Sea ch" o "Bes "Pa ame e s".................."32" 4.1.2."Resul s" om" en- old"C oss"Valida ion"........................................................"32" 4.1.3."Resul " om"Tes "Se "...................................................................................."33" 4.2."Discussion"..........................................................................................................."34" 4.2.1."Indi idual"Classi ie "Discussion"...................................................................."34" 4.2.2."Model"Compa ison"......................................................................................"36" 4.2.3."Gene al"Discussion"......................................................................................"39" 5."Conclusions"................................................................................................................"40" 6."Limi a ions"and" ecommenda ions" o " u u e"wo ks"................................................."41" i" " 7."Bibliog aphy"..............................................................................................................."42" 8."Annexes"(op ional)"...................................................................................................."47" 8.1."Code"...................................................................................................................."47" 8.1.1."Gaussian"Naï e"Bays"...................................................................................."47" 8.1.2."Be noulli"Naï e"Bayes".................................................................................."49" 8.1.3."Decision"T ee"..............................................................................................."50" 8.1.4."Logis ic"Reg ession"......................................................................................"55" 8.1.5.""-Nea es "Neighbo s"..................................................................................."52" 8.1.6."SVM"wi h"Linea "Ke nel"..............................................................................."57" 8.1.7."SVM"wi h"RBF"Ke nel"..................................................................................."59" 8.1.8."Mul ilaye "Pe cep on"................................................................................."61" 8.1.9."Random"Fo es "............................................................................................"63" 8.1.10."Ex emely"Randomized"T ee"................................................................."66" 8.1.11."Bagging"wi h"DT"base"classi ie "............................................................."70" 8.1.12."Bagging"wi h"MLP"base"classi ie ".........................................................."72" 8.1.13."AdaBoos "..............................................................................................."68" 8.1.14."Vo ing"...................................................................................................."74" 8.2."PubChem"Finge p in ".........................................................................................."76" " " ! " ! " ! " " ! ii" " LIST!OF!FIGURES! Figu e"1"-"Pubchem"S uc u e"...................................................................................................."4" Figu e"2"-"Decision"T ee"S uc u e"..........................................................................................."12" Figu e"3"-"Rep esen a ion"o "Maximum"Ma ginal"Hype plane"................................................"17" Figu e"4"-"Simpli ied" ep esen a ion"o "Mul ilaye "Pe cep on,"whe e"Wij"is" he"weigh "be ween" uni s"i"and"j"and"qj"is" he"bias"o "uni "j".............................................................................."19" Figu e"5"–"Con usion"Ma ix"...................................................................................................."27" Figu e"6"–"Con usion"Ma ix" o "Bina y"Classi ica ion"............................................................."28" Figu e"7"-"!"- old"C oss"Valida ion"..........................................................................................."29" Figu e"8"-"S udy"Wo k low"......................................................................................................."31" Figu e"9"-"A e age"Accu acy" o "10- old"C oss"Valida ion"......................................................."36" Figu e"10"-"A e age"Sensi i i y" o "10- old"C oss"Valida ion"..................................................."36" Figu e"11"-"A e age"P ecision" o "10- old"C oss"Valida ion"....................................................."37" Figu e"12"-"Accu acy"on"Tes "Se "............................................................................................."37" Figu e"13"-"Sensi i i y"on"Tes "Se "..........................................................................................."37" Figu e"14"-"P ecision"on"Tes "Se "............................................................................................."38" Figu e"15"-"MCC"on"Tes "Se "...................................................................................................."38" ! ! ! ! ! ! ! 5" "" • Easy"iden i ica ion"o "exis en "pa en s" o "compounds."Gi en" he"impo ance"o "being"able" o" pa en "po en ial"d ugs"in" esea ch"p og ams,"PubChem"o e s"links"be ween"pa en "documen s" om"U.S.,"Eu ope"and"Wo d"In ellec ual"P ope y"O ganiza ion"(WIPO)"and"unique"chemical" s uc u es."" • Possibili y"o "in e ing," o "less"s udied"compounds," hei "chemical"cha ac e is ics" h ough" he" compa ison" wi h" s uc u ally" simila " compounds." PubChem" makes" his" p ocess" easie " by" p o iding" a" p ecompu ed" lis " o " such" molecules," called" “neighbo s”." This" simila i y" can" be" compu ed"in" ela ion" o"2-D"o "3-D"cha ac e is ics,"gi ing" ise" o"2-D"o "3-D"neighbo s."" • P og amma ic"access" o"PubChem."The e"a e"se e al"p og amma ic"ways"by"which"i "is"possible" o"access"PubChem’s"da a."These"include"En ez"U ili ies,"Powe "Use "Ga eway"(PUG),"PUG- SOAP,"and"PUG-REST."I "is"also"possible" o"download" he"da abases"in"di e en " ile" o ma s" including"XML"and"SDF"(Spa ial"Da a"File"–"a" ype"o "chemical"da a" ile" o ma )" h ough" he"File" T ans e "P o ocol." " 2.2. !REVIEW!OF!WORKS!USING!VIRTUAL!SCREENING!! The e"ha e"been"se e al"published"wo ks"whe e" i ual"sc eening"o "PubChem"was"pe o med"wi h" di e en "app oaches."He e"will"be"men ioned" he"ones" ound"mos " ele an "in" e ms"o "simila i y"o " objec i es"wi h" he"s udy"a "hand." Chen"and"Wild"(2010)"ha e"conduc ed"a"s udy"whe e"a"Naï e"Bayes"model"was"c ea ed" o"p edic " ac i i y"using"1133"bioassays" om" he"PubChem"da abase."A"pa icula " ype"o "molecula " inge p in s" (FCFP_6"ci cula )"and" hei "encoded"s uc u al" ea u es"we e"used" o " he"model"c ea ion."The"au ho s" concluded" ha "Bayesian"models"gene a ed" om"PubChem"Da ase s"a e" easonably"accu a e"bu " ha " he" a iabili y"in" hei "accu acy"is"s ill"qui e"highe " han" ha "obse ed"in" he"mo e" adi ional"QSAR" (Quan i a i e"S uc u e-Ac i i y"Rela ionship)"modeling"used" o "d ug" esea ch"(Che kaso "e "al.,"2014)." This" ac "makes" hem"sui able" o " i ual"sc eening"whe e" he"main"objec i e"is" o"ob ain"a"smalle " numbe "o "po en ially"ac i e"compounds" o"be" es ed,"bu "no " o "p edic ion"whe e"high"le els"o " accu acy"a e"needed."The"au ho s"poin " ha "a"way" o"imp o e" he"accu acy"would"be" o"include"mo e" in o ma ion"abou "inac i e"compounds."(Chen"&"Wild,"2010)"" In"2008,"Weis,"Visco"and"Faulon"(2008)"de eloped"a"Suppo "Vec o "Machine"(SVM)"classi ie " o"iden i y" compounds" ha " ac " as" Fac o " XIa" inhibi o s." The" classi ie " was" ained" on" one" o " he" bioassays" conduc ed" o " he"iden i ica ion"o "Fac o "XIa"inhibi o s,"using" he"Signa u e"molecula "desc ip o ."The" bioassay"used"was"chosen"conside ing" wo" ac o s:" he" ac " ha "i "was"a"con i ma o y"assay"which" diminishes" he"numbe "o " alse"posi i es"p esen "and" ha "i "had"a"balanced"numbe "o "posi i e"and" nega i e"compounds"which"elimina es"a"p oblem" ha "is"common"in"HTS"da a."This"p oblem" esides"in" he" ac " ha " he"se s"a e"usually"highly"imbalanced"wi h"a"much"smalle "numbe "o "ac i e" han"inac i e" compounds,"leading" o"classi ie s" ha "may"ha e"high"accu acy"bu "a e"unable" o"iden i y" he"posi i e" compounds,"a"p oblem" ha "will"also"be" u he "men ioned"in" his"s udy."A e "implemen a ion"o " he" au ho s’" e sion"o " ecu si e"clus e "elimina ion" o " ea u e"selec ion," he" inal"model’s"10- old"c oss- alida ion"accu acy"was"imp o ed" om" ha "o "a" andom"classi ie " o"89%,"p o ing"i s"adequacy" o " he" 6" "" usage"in" he"case"a "hand."I "was"also"used" o"sc een" he"12"million"compounds" ha "we e"p esen "a " he" ime"in"PubChem’s"da abase."(Weis,"Visco,"&"Faulon,"2008)" Han"e "al."(2012),"e alua ed"Suppo "Vec o "Machines" o " he"iden i ica ion"o "S c"inhibi o s"in"la ge" compound"lib a ies"by" aining"and" es ing" he"models"on"1703"inhibi o s"and"63318"pu a i e"non- inhibi o s" epo ed"be o e"2011,"wi h" he" esul "o "co ec ly"iden i ying"93.53%~"95.01%"inhibi o s"and" 99.81%~" 99.90%" non-inhibi o s," in" 5- old" c oss" alida ion" s udies." The" model" co ec ly" iden i ied" 70.45%"o " he"44"inhibi o s" epo ed"since"2011,"wi h" he"model"being"applied"bo h" o" he"comple e" PubChem"da abase"and" he"MDDR"da abase"(A"bioac i i y"da abase"p oduced"by"BIOVIA"and"Thomson" Reu e s"wi h"in o ma ion"ga he ed" om"pa en "li e a u e,"jou nals,"mee ings,"and"cong esses"(“BIOVIA" Da abases"|"Bioac i i y"Da abases:"MDDR,”"n.d.))."(B."Han"e "al.,"2012)" A"s udy" o"p edic "ac i i y"agains "pa asi ic"nema odes"was"conduc ed"by"Khanna"and"Rangana han" (2011)"whe e"a"Suppo "Vec o "Machine"model"was" ained"using"da a" om" a ious"sou ces"including" PubChem" o" ga he " he" ac i e" compounds," while" he" inac i e" compound" se " was" de i ed" om" D ugBank"(Law"e "al.,"2014)."The" alida ion"o "each"model"was"done"using" en- old"s a i ied"c oss- alida ion"and" he"bes " esul s,"wi h"an"accu acy"o "81.79%"on"an"independen " es "se ,"we e"ob ained" using" he" adial"basis" unc ion"ke nel."The"au ho s"concluded" ha " he"de eloped"model"would"be"able" o"iden i y"new"po en ially"an helmin ic"ac i e"compounds."(Khanna"&"Rangana han,"2011)" P edic ion" o " ac i i y" on" human" e he -a-go-go" ela ed" gene" (hERG)" po assium" ion" channel" was" pe o med"by"Shen,"Su,"Esposi o,"Hop inge "and"Tseng"(2011)"using"an"SVM"model"and"di e en "hERG" Bioassay"da ase s" o " aining"and" alida ion,"while"a" es "se "was"de i ed" om"li e a u e"da a."The" e alua ion" was" pe o med" wi h" 10- old" c oss- alida ion," and" he" bes " model" had" an" accu acy," sensi i i y,"and"speci ici y"o "95%,"90%"and"96%," espec i ely"and"an"o e all"accu acy" o " he" es ing" se " o " 87%." The" conclusions" d awn" by" he" au ho s" we e" ha " he" model" was" able" o" p edic " “p edisposi ion”" o"block"hERG"ion"channels"and" ha "i "was" obus "ac oss" he"s uc u al"di e si y"o " he" aining"se ."(Shen,"Su,"Esposi o,"Hop inge ,"&"Tseng,"2011)" Cheng"e "al."(2011)"conduc ed"a"s udy" o"iden i y"inhibi o y"ac i i y"o "compounds"on"cy och ome"P450" (CYP)"since"i s"inhibi ion"is"an"impo an " ac o "in"d ug-d ug"in e ac ions."Fo " his"pu pose," he"au ho s" used"a"da a"se "composed"o "24700"compounds"ex ac ed" om"PubChem"and"a"combined"classi ie " algo i hm."This"algo i hm"is"an"ensemble"o "di e en "independen "machine"lea ning"classi ie s:"Suppo " Vec o " Machine," C4.5" Decision" T ee," !-Nea es " Neighbo s" And" Naï e" Bayes," used" by" a" back- p opaga ion" Neu al" Ne wo k." The" models" we e" alida ed" by" 5- old" c oss- alida ion" and" sepa a e" alida ion"se ."The" esul s"ob ained"led" o" he"conclusion" ha " hese"models"a e"applicable" o" he" i ual" sc eening"o "inhibi o s"o " he"di e en "iso o ms"o "CYP."(Cheng"e "al.,"2011)" Ano he "s udy" o"p edic "inhibi o y"ac i i y"on"CYP"was"conduc ed"by"Su"e "al."(2015)."The"au ho s"used" a" ule-based" C5.0" algo i hm" and" di e en " desc ip o s," including," amongs " o he s," PubChem’s" Subs uc u e" inge p in s."An"algo i hm"o " a ional"sampling"was"also"de eloped" o"selec "compounds" om" he" aining"se "in"o de " o"enhance" he"pe o mance"o " he"models."The"op imized"model"showed" imp o emen s"in" ela ion" o"p e iously"exis ing"models,"being"use ul" o " he"sc eening"o "la ge"da a" se s"o "compounds"and"also"p o iding" he"mos "impo an " ules" o"iden i y"p obable"inhibi o s"which" can"gi e"new"insigh s"abou " he"s uc u al" ea u es" ha "a e"impo an " o " his"ac i i y."(Su"e "al.,"2015)" 7" "" Han," Wang" and" B yan " (2008)"de eloped" Decision" T ee" (DT)" models" o" p edic " ac i i y" o " 5HT1a" agonis s,"an agonis s,"and"HIV-1"RT-RNase"H"inhibi o s"using"PubChem"Bioassay"da a"and"compound" inge p in s."The"models"we e"e alua ed"by"10- old"c oss- alida ion"and"ob ained"sensi i i y,"speci ici y" and" Ma hews" Co ela ion" Coe icien " (MCC)" in" he" anges" 57.2%-80.5%," 97.3%-99.0%" and" 0.4-0.5" espec i ely,"wi h" he"conclusion" ha " he"DT"models"de eloped"can"be"used" o " i ual"sc eening"as" well"as"a"complemen " o"o he "mo e" adi ional"app oaches" o"ac i i y"p edic ion."(L."Han,"Wang,"&" B yan ,"2008)" Schie z"(2009)"used"Weka’s"(Hall"e "al.,"2009)"cos -sensi i e"implemen a ion"o " ou "classi ie s"(Suppo " Vec o "Machines,"C4.5"Decision"T ee,"Naï e"Bayes"And"Random"Fo es )" ha "we e"applied" o"da a" om" se e al"Bioassays"on"di e en " a ge s."The"au ho s"concluded" ha "Weka's"implemen a ions"o " he" Suppo "Vec o "Machine"and"C4.5"Decision"T ee"lea ne "pe o med" ela i ely"well"and" ha "ca e"should" be" aken"wi h" he"use"o "p ima y"sc eenings" o " hei "high"numbe "o " alse"posi i es"as"well"as"wi h" he" use"o "Weka’s"cos -sensi i e"classi ie s"as"“ac oss" he"boa d"misclassi ica ion"cos s"based"on"class" a ios" should"no "be"used"when"compa ing"di e ing"classi ie s" o " he"same"da ase .”"(Schie z,"2009,"p."21)" A" consensus" model" using" he" !-Nea es "Neighbo "algo i hm"was"de eloped"by"Cha an,"Abdelaziz," Wiklande "and"Nicholls"(2016)" o " he"classi ica ion"o "hERG"po assium"channel"blocke s."The"au ho s" i s "cons uc ed"8" models"based"on" 8" di e en "kinds" o "signa u es,"one" o "which"was"PubChem’s" Subs uc u e"Signa u e," ha "we e"ob ained" o "a"da a"se "o "172"channel"blocke s" ha "was"c ea ed" based"on"in o ma ion" e ie ed" om"OCHEM"(“Online"Chemical"Modeling"En i onmen ,”"n.d.)"and" Fenichel"(“Recep o "Binding,”"n.d.)."The"au ho s" hen"c ea ed"consensus"models"based"on"majo i y" ule" and"on" he"sensi i i y"o " he"indi idual"models"(as"in" his"case"i "was"mo e"impo an " he"abili y" o" iden i y"ac i e"compounds)"using"3,"5"and"7"di e en "signa u es."The" inal"consensus"model"showed" sensi i i y"and"speci ici y"o "0.78"and"0.61" o " he"in e nal"da ase "compounds"and"0.63"and"0.54" o " an"ex e nal" alida ion"se "o "PubChem"da a."(Cha an,"Abdelaziz,"Wiklande ,"&"Nicholls,"2016)" Wang," Xie," Wang," Zhu" and"Niu" (2016)" applied" ou " di e en " Machine" Lea ning" algo i hms" o" he" p edic ion"o "selec i e"es ogen" ecep o "be a"(ER-b)"agonis "ac i i y:"Naï e"Bays,"!-Nea es "Neighbo ," Random" Fo es " and" Suppo " Vec o " Machine." The" da a" abou " he" ac i e" chemical" s uc u es" was" e ie ed" om"public"chemogenomics"da abases"and" i e" ypes"o "chemical"desc ip o s"we e"used," including"PubChem’s" inge p in ."The"models"we e"e alua ed"by"5- old"c oss- alida ion"wi h"a" epo ed" ange"o "classi ica ion"accu acies"be ween"77.10%"and"88.34%,"and"a" ange"o "a ea"unde " he"ROC" ( ecei e " ope a ing" cha ac e is ic)" cu e" be ween" 0.8151" and" 0.9475." The" s udy" sugges s" ha " he" Random"Fo es "and" he"Suppo "Vec o "Machine"classi ie s"a e"mo e"sui ed" o " he"classi ica ion"o " selec i e"ER-β"agonis s" han" he"o he "e alua ed"classi ie s."(S."Wang,"Xie,"Wang,"Zhu,"&"Niu,"2016)" No o a skyi,"Sushko,"Kö ne ,"Pandey"and"Te ko"(2011)"conduc ed"a"s udy" ha "compa ed"di e en " models" o " hei "e icacy"in"p edic ing"bioac i i y"on"CYP1A2."These"models"we e"buil "using" a ious" combina ions"o "Machine"Lea ning"me hods"and"chemical"desc ip o s."The"ML"algo i hms"used"we e" Associa i e"Neu al"Ne wo ks"(“a"combina ion"o "an"ensemble"o " eed- o wa d"neu al"ne wo ks"and" KNN”"(No o a skyi,"Sushko,"Kö ne ,"Pandey,"&"Te ko,"2011)),"k-Nea es "Neighbo s,"Random"T ee,"C4.5" Decision"T ee"and"Suppo "Vec o "Machine"wi h"all"o " he"models"being"also"used"in"combina ion"wi h" he" Bagging" echnique." The" di e en " combina ions" o " hese" me hods" wi h" he" di e en " chemical" desc ip o s"and" he"usage"o "ei he " he" ull"se "o "desc ip o s"o "a"subse "o "selec ed"ones" esul ed"in" 80"e alua ed"models."The"au ho s"concluded" ha "desc ip o "selec ion"did"no "imp o e" he"quali y"o " 8" "" he"models"and" ha " he"bes "pe o ming"model"was"ASNN"wi h" he" ull"desc ip o "se "wi h"83%"and" 68%"o "accu acy"in" he"in e nal"and"ex e nal" es "se s," espec i ely."(No o a skyi"e "al.,"2011)" In"a"wo k"conduc ed"by"Poulio ,"Chiang"and"Bu e"(2011)" he"au ho s"buil "Logis ic"Reg ession"models" wi h" he" goal" o " co ela ing" pos ma ke ing" ad e se" eac ions" (ADRs)" wi h" sc eening" da a" om" PubChem’s"Bioassay"da abase."The"de eloped"pipeline"used"508"BioAssays"o " he"PubChem"da abase" wi h"485"di e en "d ug"componen s."The"ADRs"we e"g ouped"in"di e en "sys em"o gan"classes"and" models"we e"buil " o "each"o " hese."The"models"we e"e alua ed"using"Lea e"One"Ou "C oss-Valida ion" (LOOCV)"and" he"au ho s" epo "a"be e "pe o mance" han"expec ed"gi en" he"simplici y"o " he"Logis ic" Reg ession"models"wi h"hal "o " he"models"ha ing"an"AUC"o "³"0.7"and"all"o " he"models"ha ing"AUC"³" 0.6."(Poulio ,"Chiang,"&"Bu e,"2011)" Recen ly,"Yu,"Shi,"Tian,"Gao"and"Li"(2017)"p esen ed"a"s udy"wi h" he"objec i e"o "de eloping"a"model" o " he" classi ica ion" o " CYP450" 1A2" inhibi o s" and" non-inhibi o s" using" a" mul i- ie ed" deep" belie " ne wo k" (DBN)" on" a" la ge" da ase ." The" au ho s" used" a" da ase " o " o e " 13000" compounds" om" PubChem"and"245"molecula "desc ip o s"including"bo h"2D"and"3D"desc ip o s" ha "we e"calcula ed"by" molecula "compu a ional"so wa e."Wi h" he"objec i e"o "imp o ing" he"classi ie ’s"pe o mance"and" dec ease" he"compu a ional" ime"a"desc ip o "selec ion"was"pe o med"by"implemen a ion"o " h ee" ules:" emo al"o "desc ip o s"wi h" oo"many"ze os,"wi h"small"s anda d"de ia ion" alues"(<"0.5%)"and" wi h" co ela ion" coe icien s" highe " han" 0.9." Fo " compa ison" pu poses," shallow" machine" lea ning" models"we e"also" ained,"namely"Suppo "Vec o "Machine"and"A i icial"Neu al"Ne wo k."All"models" we e" un" se e al" imes" o" de e mine" he" bes " pa ame e s" o" use" and" e alua ed" by" 5- old" c oss" alida ion"and"by"an"ex e nal"da ase ."The"bes " esul s"we e"ob ained"by" he"DBN"model"using"bo h"2D" and"3D"desc ip o s"wi h"an"in e nal"o e all"accu acy"o "83.6%"and"an"ex e nal"accu acy"o "77.0%."(Yu," Shi,"Tian,"Gao,"&"Li,"2017)" In" he"s udy"by"Bilsland"e "al."(2015),"A i icial"Neu al"Ne wo ks"we e"used" o " i ual"sc eening"o " Selec i e"G1-Phase"Benzimidazolone"inhibi o s."The"da ase "used"was"de i ed" om"PubChem"Bioassay" da a"wi h" he"au ho s"op ing" o" educe" he"numbe "o "inac i e"compounds"by"applying"simila i y" il e s" using"chemical"so wa e" o"ob ain"a"balanced"da ase ."The" inal" aining"se "con ained"3924"compounds" o " which" 1859" we e" ac i e" and" 2065" we e" inac i e." The" desc ip o s" used" included" PubChem’s" inge p in s"amongs "o he s," esul ing"in"a"g oup"o "2780" ea u es."Successi e" uns"we e"pe o med" o" iden i y" he"bes "pa ame e "selec ion," o"elimina e"compounds" ha "we e"consis en ly"misclassi ied"and" o"de e mine" he"op imal"subse "o "desc ip o s."The"au ho s"op ed" o"use"as"a" inal"model"an"ensemble" o "10"ne wo ks" ained"using" he"op imal"combina ion"o "pa ame e s,"compounds"and"desc ip o "se ." The"o e all"sensi i i y,"speci ici y"and"accu acy"e alua ed"by"10- old"c oss" alida ion"we e," espec i ely," 83.1%,"82.4%"and"82.7%"wi h" he"au ho s"conside ing" hese" esul s" o"indica e" e y"good"p edic i e" pe o mance."(Bilsland"e "al.,"2015)" While"mos "o " he"QSAR"s udies" o " his" a ge "ha e"used" eg ession" echniques," he"wo ks"desc ibed" below"ha e"used"classi ica ion"models" o"p edic " he"inhibi o y"ac i i y"o "compounds"on"EGFR:" In" he"wo k"published"by"Kong,"Qu,"Chen,"Gong"and"Yan"(2016)" he"au ho s"ha e"de eloped"models" o " he"classi ica ion"o "compounds"as"inhibi o s"o "non-inhibi o s"o "EGFR"by"using"Kohonen’s"Sel - O ganizing"Map"(SOM)"and"Suppo "Vec o "Machine"(SVM)"algo i hms."The"used"da ase "was"compiled" om"CHEMBL"(Ben o"e "al.,"2014)"keeping"only" he"compounds"wi h"inhibi o y"concen a ion"(IC50)" unde "10"µM," esul ing"in"1248"inhibi o s."Fo " he"inac i e"compounds"3093"decoys"whe e"ga he ed" 9" "" om" he"DUD"da abase"(Huang,"Shoiche ,"&"I win,"2006)."A"PCA"analysis"was"pe o med"on"some" p ope ies" o"de e mine" ha " he e"was"o e lapping"in" he"ac i e"and"inac i e"compounds"as" o"make" he" iden i ica ion" challenging." The" inal" da ase " was" di ided" in o" aining" and" es " se " o" pe o m" e alua ion"o " he"model."Fo " he"molecula "desc ip o s,"ADRIANA.Code""(Gas eige †,"2006)"desc ip o s" we e"calcula ed"and" hen"a"subse "selec ed"based"on"co ela ion"wi h"ac i i y."The"au ho s" epo ed" ha " he" inal"models"had"p edic ion"accu acies"on" aining"and" es ing"se ," espec i ely,"o "98.5%"and" 96.3%" o " he"SOM"model"and"99.0%"and"97.0%" o " he"SVM"model,"and"sensi i i y,"speci ici y"and" MCC," espec i ely,"o "94.0%,"97,3%"and"0.91" o " he"SOM"model"and"94.2%,"98.2%"and"0.93" o " he" SVM" model," concluding" ha " bo h" models" had" good" pe o mance" when" dis inguishing" be ween" inhibi o s"and"decoys"o "EGFR."(Kong,"Qu,"Chen,"Gong,"&"Yan,"2016)" Zhao"e "al."(2017)"conduc ed"a" ecen ly"published"s udy"whe e" he"au ho s"cons uc ed"2D"and"3D- QSAR" models" wi h" he" 2D" model" being" buil " using" a" Suppo " Vec o " Machine"classi ie ." The" used" da ase "was"cons uc ed"using"100"inhibi o s" e ie ed" om" he"li e a u e"and"185"inhibi o s" om" he" DUD"da abase."Fo " he"2D"s udy" he"da ase "was"di ided"in o" h ee" aining"se s"which"accoun ed" o " 75%," 70%" and" 50%" o " he" whole" da ase ." Fo y- i e" molecula " desc ip o s" whe e" calcula ed" using" ChemO ice"(I win*,"2005)"and"a"subse "o "9"desc ip o s"was"selec ed"using"Co ela ion-Based"Fea u e" Selec ion"combined"wi h"Gene ic"Sea ch"algo i hms."The" aining"o " he"model"was"conduc ed"wi h" he" h ee"di e en " aining"se s"wi h" he"au ho s"op ing" o"use" he"da ase "accoun ing" o "70%"o " he"da a" which"led" o" he"highe "accu acy."The" inal"model"p esen ed"sensi i i y,"speci ici y,"accu acy"and"MCC" o "98.55%,"99.23%,"98,99%"and"0.978," espec i ely"on" he" aining"se "e alua ed"by" en- old"c oss- alida ion"and"96.77%,"98.18%,"97.67%"and"0.950"on" he" es "se ,"indica ing"good"pe o mance"o " he" model."(Zhao"e "al.,"2017)" " Singh"e "al."(2015)"de eloped"a"model" o " he"classi ica ion"o "compounds"as"inhibi o s"o "non-inhibi o s" o " EGFR." Fo " his"pu pose," he" au ho s" ob ained" 3528" an i-EGFR" compounds" and" hei " inhibi o y" concen a ion" (IC50)" om" a" da abase" ha " was" also" de eloped" by" he" au ho s" and" ha " con ains" in o ma ion"ga he ed" om"a ound"350" esea ch"a icles."Wi h" hese"compounds," he"au ho s"buil " h ee"di e en "da a"se s"wi h"di e en "p opo ions"o "ac i e"and"inac i e"compounds"acco ding" o" he" chosen"IC50"le el" h eshold:"EGFR10,"EGFR100,"EGFR1000."As" he"chemical"desc ip o ," he"au ho s"used" PubChem’s"Subs uc u e" inge p in s"calcula ed" o " he"compounds"p esen "in" he"da ase s"buil "by" he" au ho s."Each"o " he"da ase s"was"di ided"in o" aining"and" alida ion"se s" o " he"pu pose"o "model" e alua ion" and" di e en " ML" algo i hms" implemen ed" in" Weka" we e" applied," including" IBK" (an" implemen a ion"o "!-Nea es "Neighbo s),"Naï e"Bayes,"Suppo "Vec o "Machine"and"Random"Fo es ." The" au ho s" ound" ha " he" bes " pe o ming" model"was" he" Random" Fo es " wi h" an" accu acy," sensi i i y,"speci ici y,"and"MCC"o "83.66%,"69.89%,"86.03%"and"0.49," espec i ely,"when"e alua ed"on" he"EGFR10"da ase "which"had" he"lowes " h eshold"o "IC50"and" hus"a"smalle "p opo ion"o "ac i e" compounds."(Singh"e "al.,"2015)"" 2.3. !EPIDERMAL!GROWTH!FACTOR!RECEPTOR!(EGFR)! The" Epide mal" G ow h" Fac o " Recep o " (EGFR)" o " e bB1/HER1" is" a" ansmemb ane" glycop o ein" (He bs ,"2004)."I "is"a"membe "o " he"e bB/human"epide mal"g ow h" ac o " ecep o " amily"o " y osine" kinases,"which"also"includes"e bB2/HER2,"e bB3/HER3"and"e bB4/HER4"(T oiani"e "al.,"2012).""EGFR"is" ound" no " only" in" he" plasma" memb ane," bu " i s" exp ession" le els" a e" also" high" in" he" nucleus," endosomes,"lysosomes,"and"mi ochond ia"(H."Li,"You,"Xie,"Pan,"&"Han,"2017)." 10" " The"EGFR"signaling"pa hway"is"o "ex eme"impo ance"in"mammalian"cells,"ha ing" oles"in"g ow h," su i al,"p oli e a ion,"and"di e en ia ion"o "cells"(Oda,"Ma suoka,"Funahashi,"&"Ki ano,"2005)."I "is"also" o e exp essed"in"a" a ie y"o "cance s:""EGFR"is"o e exp essed"in"50–80%"o "non-small"cell"lung"cance s" and"E bB2"and"E bB3"a e"o e exp essed"in"25–30%"and"63%"o "b eas "cance s," espec i ely"(Scha adin" e "al.,"2017)."I "is"also" ela ed" o"inc easing" esis ance" o"chemo he apy"and" adia ion" he apy"o " umo " cells"(He bs ,"2004),"and"is"o e exp essed"in"cance s"o " e y"poo "p ognosis"such"as"panc ea ic"cance " (T oiani"e "al.,"2012)"making"i "a"c i ical"d ug" a ge ,"being"i s"inhibi ion"o "pa icula "in e es ."" Cu en ly," he apies"di ec ed"a "EGFR"a e"included"in" wo"gene al"ca ego ies:"monoclonal"an ibodies" ha " a ge " he" ex acellula " domain" and" small" molecule" y osine" kinase" inhibi o s" ha " show" e ec i eness"bu "e en ually"lead" o" esis ance"(Scha adin"e "al.,"2017)"which"jus i ies" ha " he"sea ch" o "new"po en ial"inhibi o s"is"con inually"necessa y."" Because"o "i s"high"p e alence"in"a"numbe "o "pa hways"bo h"heal hy"and"pa hogenic" ha "make"i "a" po en ial" d ug" a ge " o " such" impo ance" and" o " he" de elopmen " o " esis ance," he" esea ch" o " compounds"wi h"bioac i i y"on" his" a ge "has"been"g ea "and"con inuous"since" he"disco e y"o "i s" ole" in"cance "in" he"1980’s"(Vas ag,"2005)."This"has"led" o" he"exis ence"o " as "li e a u e"on" he"subjec " which"makes"da a"on"known"inhibi o s" eely"a ailable"(Singh"e "al.,"2015)."" All" he" p e iously"men ioned"cha ac e is ics," mainly" he" impo ance" as" a" d ug" a ge ," he" e e - inc easing"need" o "new"inhibi o s"d i en"by" esis ance,"and" he"a ailabili y"o "da a"makes"EGFR"a"g ea " a ge " o " i ual"sc eening." 2.4. !DATA!MINING!AND!MACHINE!LEARNING!ALGORITHMS!THEORETICAL!FRAMEWORK! The" e m"Da a"Mining"(DM)"was"ini ially"a"de oga o y" e m" ha "mean " he"ac "o "sea ching" o "an" insigh " ha "was"no "suppo ed"by" he"da a"(Lesko ec,"Raja aman,"&"Ullman,"2011)."Wi h" he"inc ease" o " eadily"a ailable"da a"bo h"in"quan i y"and"size,"i "has" aken"a"posi i e"meaning"and"can"be"desc ibed" as"a"p ocess" o"disco e "pa e ns"and" ela ionships"in"da a" ha "can"b ing"p e iously"unknown"insigh s" and"allow" he"making"o " alid"p edic ions"(Edels ein,"1999)."I " esul s" om" he"c ossing"o "se e al" ields" including"Da abase"Managemen ,"A i icial"In elligence,"Machine"Lea ning,"Pa e n"Recogni ion,"and" Da a"Visualiza ion"(F iedman"&"F iedman,"1997)."Da a"Mining"has"applica ions"in"any"indus y."These" applica ions"include"cus ome "segmen a ion"and" a ge ing,"c edi "sco ing," aud"de ec ion"and"d ug" e ec "iden i ica ion"in"d ug" ials"(“Da a"Mining"F om"A" o"Z,”"n.d.)."" Some"au ho s"conside "Da a"Mining"and"Machine"Lea ning"(ML)"as"synonyms"and"Da a"Mining"does" use"algo i hms" om"ML"in"i s"p ocesses"o "Knowledge"disco e y"(Lesko ec"e "al.,"2011)."Bu "Machine" Lea ning"can"be"said" o"be" he"discipline" ha "aims"a "making"compu e s"modi y"o "adap " hei "ac ions" (which"can"be"making"p edic ions"o "o he s)"so" ha " hese"ac ions"become"mo e"accu a e"in" e ms"o " he"goal"(Ma sland,"2015)"using"me ics" ha "e alua e" his"adap a ion" o"guide" he"p ocess."" When"i "comes" o" he"kind"o " asks" ela ed" o"Da a"Science,"we"can"say" ha "Machine"Lea ning" e e s" o" he"c ea ion"and"use"o "models" ha "a e"lea ned" om"Da a,"which"will" ypically"ha e"as"a"goal" he" p edic ion"o "a"ce ain"ou come"(G us,"2015)." In" he"las "decade," he"mul idisciplina y"na u e"o "Machine"Lea ning"has"become"appa en ."Concep s" om" Neu oscience" and" Biology," S a is ics," Ma hema ics" and"Physics"ha e" all" con ibu ed" o" he" de elopmen "o "p ocesses" o"make"compu e s"lea n"(Ma sland,"2015),"which"is"appa en "in"algo i hms" 11" " such"as"A i icial"Neu al"Ne wo ks"(Rumelha ,"Wid ow,"&"Leh ,"1994),"and"Gene ic"algo i hms"(Koza," 1992)." 2.4.1. Types!o !Da a!Mining!P oblems! 2.4.1.1. Reg ession/Classi ica ion!P oblems! Da a"mining"p oblems"can"be"classi ied"as"ei he "Reg ession"o "Classi ica ion"p oblems."Bo h" he"inpu " and"p edic ed" a iables"used" o"de elop"a"model"can"be"ei he "quan i a i e"o "ca ego ical."In"a"gene al" sense,"we"can"say" ha "a"p oblem" ha "aims" o"p edic "a"quan i a i e" alue"is"a"Reg ession"p oblem" while"p oblems" ha "aim" o"assign"a" eco d" o"a"ce ain"ca ego y"a e"Classi ica ion"p oblems"(James," Wi en,"Has ie,"&"Tibshi ani,"2013)."" 2.4.1.2. Types!o !Lea ning! Da a"Mining"p oblems"can"also"be"classi ied"in" e ms"o " he"amoun "and" ype"o "supe ision" hey"ge " du ing" aining."These"include:" • Supe ised!Lea ning:"A" aining"se "o " eco ds"wi h" hei "co ec " esponses"( a ge s"o "labels)" is"p o ided"(Ma sland,"2015)."In" his"case," he"aim"is" o" i "a"model" ha " ela es" he"p edic o s" wi h" he" esponse," in" o de " o" be" able" o" p edic " he" esponse" o " new" obse a ions" (p edic ion)" and/o " o" be e " unde s and" he" ela ionship" be ween" he" esponse" and" he" p edic o s"(in e ence)"(James"e "al.,"2013)."" " • Unsupe ised!Lea ning:!On" he"o he "hand,"in"unsupe ised"lea ning," he" aining"da ase "does" no "include"labels" o "a" a ge " a iable,"and"in" his"case," he"aim"is" o" ind" ela ionships"be ween" he" a iables"o "be ween" he"obse a ions"(James"e "al.,"2013).! ! • Semi-supe ised!Lea ning:!In" his"kind"o "lea ning" he e"is"a"la ge"amoun "o "unlabeled"da a" and"a"smalle "amoun "o "labeled"da a."Mos "semi-supe ised"sys ems"consis "o "combina ions" o "supe ised"and"unsupe ised"algo i hms"(Gé on,"2017).! ! • E olu iona y!Lea ning:!Lea ning"sys ems!inspi ed"by!biological"e olu ion"using" he"concep "o " i ness"as"a"measu e"o "how"good"is" he"cu en "solu ion"(Ma sland,"2015)." ! • Rein o cemen !Lea ning:!In" ein o cemen "lea ning" he"algo i hm"is" old"when" he"answe "is" w ong"bu "no "how" o"ge " he" igh "one."In" his"case," he"algo i hm"o "agen "can"obse e" he" en i onmen "and"expe imen "wi h"di e en "solu ions" ha "ge " ewa ded,"wi h" he"agen "ha ing" o" igu e"ou "which"is" he"bes "policy" o"inc ease" hese" ewa ds"(Gé on,"2017;"Ma sland,"2015).! In" his" s udy," he" p oblem" a "hand" is" a" Classi ica ion" p oblem" wi h" supe ised" lea ning" since" he" objec i e"is" o"classi y" he"compounds"in"PubChem’s"da abase"o "o he s"as"ei he "ac i e"o "inac i e"and" he"algo i hms"used"will"be"p o ided"wi h" he"co ec "ca ego y" o " he" aining"examples." 12" " 2.4.2. Theo e ical!backg ound!o ! he!used!algo i hms!! 2.4.2.1. Decision!T ees! A"simple"way" o"desc ibe"Decision"T ees"is" ha " hey"a e"a"way"o " ep esen ing"a"se "o " ules" ha "when" applied" o"an"obse a ion"can"lead" o"a"class"o " alue"(Edels ein,"1999)."They"a e"composed"o "a" oo " node,"decision"nodes"and"lea "nodes" ha "a e"connec ed"by"b anches:" " Figu e"2"-"Decision"T ee"S uc u e" A " he" oo "( he" i s "spli )"and"a " he"decision"nodes,"each"a ibu e"is"e alua ed"acco ding" o"a" ule" which"is"applied"in"each"b anch"un il"a"lea "node"is" eached"whe e" he e"a e"no"mo e"a ibu es" o" e alua e"on"(Daniel"T."La ose,"2015)."The"g ea e " he"pu i y"o " he"lea "nodes"and" he"dis ance"be ween" hem" he"be e "(Edels ein,"1999)."" Decision"T ees"ha e" he"ad an ages"o "being"simple"and"allowing"easy"in e p e a ion,"bu "usually"don’ " pe o m"as"well"as"o he "mo e"complex"algo i hms"in" e ms"o "accu acy"(James"e "al.,"2013)"and"o e i " o" he" aining"se " e y"easily"which"means"less"gene aliza ion"capaci y"(G us,"2015)." In" he"p ocess"o "building"decision" ees,"i "is"necessa y" o"de e mine"which"ques ions"a e"being"asked" in"each"decision"node"and"in"wha "o de "(G us,"2015)."To"achie e" his"goal,"i "is"necessa y" o"ha e" measu es" o " pu i y" o " he" nodes" be o e" and" a e " he" pa i ion" is" applied" and" o" calcula e" ha " di e ence"which"will"be" he"measu e"o "how"much"in o ma ion"will"be"gained"by"applying"a"ce ain" pa i ion."The" wo"mos "used"measu es" o " his"pu pose"a e"En opy"and"Gini"Impu i y"(Ma sland," 2015):" • En opy"o "a"se "o "p obabili ies"#$:" %&'()#* = , − #$log1#$ $ 1" • Gini"Impu i y" o "a"pa icula " ea u e"k:"" " " 34= 1 − 5 6 1 7 $89 , 2 " Roo $Node Decision$ Node Decision$ Node Lea $Node Lea $Node Lea $NodeLea $Node B anches 13" " whe e"c"is" he" o al"numbe "o "classes"and"N(i),is" he" ac ion"o " eco ds" ha "belong" o"class"i." T ees" ha "a e"allowed" o"g ow"inde ini ely"will"o e i " o" he" aining"da a."In"o de " o"a oid" his," s opping" ules"mus "be"applied."Common"s opping" ules"a e"simply"limi ing" he"maximum"dep h" ha " he" ee"can" each"o "es ablishing"a"lowe "limi " o" he"numbe "o " eco ds"in"a"node."Al e na i ely,"i "is" possible" o"p une" he" ee,"whe e" he" ee"is"allowed" o"g ow" o" ull"size"and" hen"is"p uned"back" o" he" smalles "size" ha "does"no "comp omise"accu acy."(Edels ein,"1999)" The e"a e"di e en "algo i hms" ha "can"be"used" o"build"a" ee" ha "include"ID3,"C4.5,"C5.0,"and"CART." In" his"wo k," he"algo i hm"used"is"CART." The"CART"algo i hm"p oduces"bina y" ees," ha "is," o "each" ea u e,"i "spli s" he" aining"se "in o" wo" subse s"based"on"a"ce ain" h eshold" alue" o " ha " ea u e."The" ea u e"and" h eshold"used"a e"chosen" so" ha " he"pu es "subse s"a e"ob ained."I " ollows" hese"s eps" ecu si ely"un il" he"s opping"condi ion" is"me ."CART"is"a"g eedy"algo i hm"which"means" ha " he"op imum"spli ing"choice"is"made"a "each"le el" wi hou "checking"i "i "will"lead" o" he"pu es "subse s"in" he"le els"below."This"usually"leads" o"a"good" solu ion"bu "doesn’ "gua an y"an"op imal"one"(Gé on,"2017)."" 2.4.2.2. Naï e!Bayes! S udies"ha e" ound" he"Naï e"Bayes"classi ie " o"ha e"compa able"pe o mance" o"Decision"T ees"and" some" Neu al"Ne wo k" classi ie s"(J."Han,"Kambe ," &" Pei,"2012)." I " is"called" Naï e"Bayes"because"i " assumes" ha " he"obse a ion" a iables"a e"independen "o "each"o he "which"will"mos "o " he" imes" no "be" ue."This"classi ie "is"based"on"Bayes’" heo em" ha "s a es" ha "(Ma sland,"2015):" @ A B = , @ B A @ A @ B , 3 " whe e"@ A B "is" he"condi ional"p obabili y"o "H"gi en"X,"@ B A "is" he"condi ional"p obabili y"o "X" gi en"H"and"@ A "and"@ B "a e" he"a*p io i"p obabili ies"o "H"and"X," espec i ely."Fo " he"pu pose"o " classi ie s,"we"can"conside "@ A " o"be" he"p obabili y"o " he"hypo hesis" ha " he"obse a ion"X"belongs" o" a" ce ain" class," and"@ B " he" p obabili y" ha " an" obse a ion" X" is" equal" o" a" ce ain" ec o " o " a iables.""" Wi h" he" use" o " he" assump ion" o " independence" o " he" a iables" in" he" da ase ," we" come" o" a" simpli ied"equa ion" ha "s a es" ha " he"p obabili y"o "an"obse a ion"Xj"being"equal" o"a"ce ain" ec o " o " a iables"gi en" ha "i "belongs" o"class"D$"(@ BED$= @(BE 9, BE 1, … BE G|D$),"whe e" he"supe sc ip s" o " X" ep esen " he" index" o " he" a iables" o " he" ec o )," is" equal" o" he" p oduc " o " he" indi idual" p obabili ies:" " @ BE 4= , I4D$J = ,@ BE 9= , I9D$,×,@ BE 1= , I1D$,× …,×,@ BE G= , IGD$, 4 " and" he"classi ie "will"selec " he"class"D$" o "which" he" ollowing"compu a ion"is" he"maximum:" @(D$) @(BE 4= , I4|D$ 4 ). 5 " Al hough" his"compu a ion"is" he" esul "o "an"ob iously"inco ec "assump ion,"se e al"empi ical"s udies" 14" " show" ha "Bayesian"classi ie s"pe o m"well"and"a e"compa able" o"o he "mo e"complex"algo i hms"(J." Han" e " al.," 2012)"as" he" assump ion" made" ends" o" no " hu " classi ica ion" pe o mance" (Fos e " &" Fawce ,"2013)." This"classi ie "is" e y"e icien "in" e ms"o "used"s o age"space"and"compu a ional" ime"and"i "is"also"a" na u al"“inc emen al"lea ne ”"as"i "can"upda e"i s"model"one"example"a "a" ime"wi hou "ha ing" o" ep ocess"all"pas " aining"examples"(Fos e "&"Fawce ,"2013)."" In" heo y,"Bayesian"classi ie s"will"ha e" he"minimum"e o " a e"in"compa ison" o"o he "models"which" no "always"happens"in"p ac ice"due" o" he"inaccu acies" esul ing" om" he"assump ions"made"(J."Han" e "al.,"2012)."None heless,"i "is"a" e y"commonly"used"classi ie " o"se e"as"a"baseline" o"which"o he " models"a e"compa ed"(Fos e "&"Fawce ,"2013)." 2.4.2.3. Logis ic!Reg ession! Linea "Reg ession"aims" o"app oxima e" he" ela ionship" ha "exis s"be ween"a"se "o " a iables"and"a" con inuous" esponse," bu " when" he" esponse" is" ca ego ical" Linea " Reg ession" is" no " applicable." Howe e ,"an"analogous"me hod"can"be"used,"Logis ic"Reg ession"(Daniel"T."La ose,"2015)."I "is"mos ly" used" o" p edic " bina y" a iables" bu " can" also" be" applied" o" he" p edic ion" o " mul i-class" a iables" (Edels ein,"1999)." In"a"classi ica ion"p oblem,"we"wan " he"examples" ha "a e" u he "away" om" he"bounda y"be ween" classes" o"ha e"a"highe "p obabili y"o "belonging" o" ha "class."The"p oblem"o "using"Linea "Reg ession" is" ha " he"dis ance" om" he"bo de "can" ange" om"−∞" o"+∞"while" he"p obabili ies"should"be"in" he" ange"ze o" o"one"(Fos e "&"Fawce ,"2013)."" Since" he" a ge " a iable"is"disc e e"and"i "is"no "possible" o"di ec ly"model"using"linea " eg ession," ins ead"o "p edic ing"i " he"e en "i sel "will"happen" he"logis ic"model"p edic s" he"loga i hm"o " he"odds" o "i s"occu ence"(Edels ein,"1999)."" To"mee " he"objec i e"o "ge ing"p obabili ies"be ween"ze o"and"one"we"can"use" he"logis ic" unc ion" (James"e "al.,"2013):"" # B = , ℯRSTRUV 1 +,ℯRSTRUV, 6 " whe e" he"XY"and"X9," ep esen " he"coe icien s" o "a"single"p edic o "X." Wi h"some"manipula ion"o " he" o mula"we"a i e"a :" log # B 1 − # B = , XY+,X9B , 7 " wi h" he"le "side"being" he"log-odds*o "logi " ha "is"linea " o"X."" Fo "a" eg ession"wi h"mul iple"p edic o s" he"p e ious"equa ion"can"be"gene alized" o:" log # B 1 − # B = , XY+,X9B9+ ⋯ +,X B . 8 " 21" " ha e"a"long" aining" ime,"al hough" he"p edic ions"a e"p o ided"quickly"(Edels ein,"1999)." 2.4.2.7. Ensemble!me hods! Ensemble"me hods"a e"based"on"a"p inciple"simila " o" ha "o " he"wisdom*o * he*c owd"which"is"a" phenomenon"whe e" he"a e age"answe "o "a"la ge"numbe "o "people" o"a"ce ain"ques ion"is"o en" be e " han" he"single"answe "o "an"expe "(Gé on,"2017)." An" ensemble" model" o " classi ica ion" is" a" composi e" model" ha " esul s" om" combining" di e en " classi ie s."Ensemble"classi ie s" end" o"pe o m"be e " han" hei "composing"models"indi idually"(J." Han"e "al.,"2012)." Vo ing)Classi ie s) The"simples " o m"o "Ensemble"me hods"is" he"Vo ing"me hod"whe e" he"p edic ions" o " he"indi idual" classi ie s"a e"agg ega ed,"and" he"p edic ion"wi h" he"majo i y"o " o es" om" he"composing"classi ie s" is" he"one"chosen." This"s a egy,"albei "i s"simplici y,"will"gene ally"ou pe o m" he"bes "classi ie "in" he"ensemble"and" ends" o"achie e"good"pe o mance"e en"i "all" he"composing"classi ie s"a e"weak"lea ne s," ha "is,"only" sligh ly"be e " han" andom"guessing."This"happens"because"o " he*law*o *la ge*numbe s," ha "s a es" ha "o e "a"la ge"numbe "o " ials" he"a e age" esul "will"be"close" o" he"expec ed" alue."In" he"case"o " classi ie s," his" ansla es" o" he" ac " ha "i "all"classi ie s"ha e"an"accu acy" o " abo e"50%,"as" hei " numbe "inc eases," he"accu acy"o " he"combined"models"will"also"inc ease."(Gé on,"2017)" Bagging) Bagging"is"a"me hod" ha "also"uses"majo i y" o e" o"ge " he" inal"p edic ion"o " he"ensemble,"bu "i " in oduces" he" a ie y"o "classi ie s"in"a"di e en "way" han"desc ibed"p e iously."The"se e al"indi idual" classi ie s"a e"buil "using" he"same"algo i hm"bu "on"di e en "samples" aken" om" he"o iginal"da ase " wi h" eplacemen ."This"implies" ha "a"ce ain"sample"will"likely"exclude"some"examples"and"duplica e" o he s" om" he"o iginal"da ase ."The"aim"o " his"p ocess"is" o"dec ease"a"sou ce"o "e o " o "indi idual" classi ie s" ha "a ises" om" he"use"o "a"pa icula " aining"se " ha "is"ine i ably" ini e"and"no "comple ely" ep esen a i e"o " he"whole"popula ion."Tha "is,"i "in ends" o"diminish" he" a iance"componen "o " he" e o "o " he"classi ie "in" he"bias- a iance" ade-o ."Because"o " his,"Bagging"is"usually"mos "use ul" when"used"wi h"algo i hms" ha "a e"by" hemsel es"uns able,"such"as"Decision"T ees."(Wi en"&"F ank," 2005)" Boos ing) In"Boos ing,"weigh s"a e"assigned" o"each" eco d"and,"a e "a"classi ie "is" ained," hese"weigh s"a e" upda ed" so" ha " he" ollowing" classi ie " will" gi e" mo e" impo ance" o" he" eco ds" ha " we e" misclassi ied."In" he" end," he"ensemble"chooses" he"co ec "p edic ion" based" on" o ing" wi h" each" classi ie ’s" o e"weigh "being"a" unc ion"o "i s"accu acy."(J."Han"e "al.,"2012)" AdaBoos ) " The"mos "commonly"used"algo i hm" o"pe o m"boos ing"is"AdaBoos "( om"adap i e"boos ing)."I "can" be"desc ibed"by" he" ollowing"(J."Han"e "al.,"2012):" 22" " Conside ing"we"ha e"a"da ase "ä,"o "e"labeled" uples"(B9, *9),"(B1, *1),"…,"(Bã, *ã)"whe e"*$"is" he" class"label"o "B$,"AdaBoos "will" i s ly"assign"a"weigh "o "1 e" o"each"ins ance."Following" his" i s "s ep," !" ounds"a e"pe o med" o "!"gene a ed"classi ie s."On" ound"6"a"sample"ä$,"o "size"e"is"gene a ed"wi h" eplacemen " o" o m" he" aining"se "so"each" uple"may"appea "mo e" han"once"and" hei "p obabili y" o "selec ion"is"dependen "on"i s"assigned"weigh ."A"classi ie ,"å$,"is"gene a ed"and"i s"e o "is"calcula ed" using"ä$"as"a" es "se ."This"e o "is"calcula ed"using:" ~(()( å$= hE×~(( BE ã E89 ,27 " whe e" ~(((BE)"is" he" misclassi ica ion" e o " o " ins ance" BE"and" is" equal" o" 1" i " he" ins ance" was" misclassi ied" and" 0" o he wise." The" weigh s" o " each" co ec ly" classi ied" uple" a e" hen" upda ed" mul iplying"by:" ~(()( å$ 1 − ~(()( å$ .28 " A e " his"s ep,"all"weigh s"a e"no malized,"including" hose"o " he"misclassi ied" uples,"which" esul s"in" an"inc ease"o " he"weigh s"o " he"misclassi ied" uples"and"a"dec ease"o " he"weigh s"o " he"co ec ly" classi ied"ones."" A " he"end"o " he"!" ounds," he" inal"p edic ion"is"made"by" o ing,"as"p e iously"men ioned,"and" he" weigh "gi en" o"each"classi ie ’s" o e"is"gi en"by:"" log 1 − ~(()( å$ ~(()( å$ .29 " Fo "each"class," he"weigh s"o "each"classi ie " ha " o ed" o "i "as" he"co ec "one"a e"summed,"and" he" class"wi h" he"highes "sum"is" he"p edic ion"made"by" he"ensemble." In" compa ison" wi h" Bagging," Boos ing" ends" o" achie e" highe " accu acy" bu " has" a" highe " isk" o " o e i ing"(J."Han"e "al.,"2012)." Random)Fo es ) Random"Fo es s"a e"an"ensemble"o "Decision"T ees"gene ally" ained"using"Bagging"(Gé on,"2017)."I " so," hey" a e" buil " using" Bagging" in" combina ion" wi h" andom" a ibu e" selec ion" a " each" node" o" de e mine" he"spli ."Wi h"a" aining"se "ä,"o "size"e," o"gene a e"!"Decision"T ees," o "each"i e a ion" 6(6 = 1,2, … , !),"a"sample"o "size"e"is"gene a ed"wi h" eplacemen ."A "each"node,"an"ç"numbe "o " a ibu es"is" andomly"selec ed"as"candida es" o " he"spli "wi h"ç"being"much"smalle " han" he"numbe " o " o al"a ibu es."The"used"algo i hm" o"g ow" he" ees"is"CART,"and" he" ees"a e"no "p uned."Random" Fo es "is"compa able" o"AdaBoos "in" e ms"o "accu acy"bu "is"mo e" obus " o"e o s"and"ou lie s""(J."Han" e "al.,"2012)." Ex emely)Randomized)T ees) Ex emely"Randomized"T ees"a e"simila " o"Random"Fo es s,"bu "besides"choosing"a" andom"numbe " o " ea u es" o"be"conside ed" o "spli ing," he" h eshold"used" o "each" ea u e"is"also" andom"ins ead" o "sea ching" o " he"bes "possible"one."This"will" ade"bias" o " a iance," ha "is,"will"inc ease" he"e o " 23" " ha " a ises" om" di e ences" in" he" aining" da ase " and" dec ease" he" e o " ha " a ises" om" he" assump ions"made"by" he"model."Since" he"choice"o " h eshold"is" andom,"Ex emely"Randomized" T ees" ain"much" as e " han"no mal"Random"Fo es "bu "is"no "possible" o" ell"be o ehand"which"will" pe o m"be e "and"bo h"need" o"be"applied"and"compa ed" o"de e mine" he"bes "one."(Gé on,"2017)" 24" " 3. METHODOLOGY! The"p ocess"o "de ining" he"me hodology" o " his"wo k"was"an"i e a i e"and" ecu si e"one"as"i "is" o " mos " Da a" Mining/Machine" Lea ning"p ojec s." I s" majo " pa s" in ol ed" he" expe imen a ion" and" selec ion"o " he"bes " ools," he"ga he ing"and" ea men "o " he"da a,"and" he"choice"o "algo i hms" o" use,"o " hei "pa ame e s’" alues"and"o "how" o"e alua e" hem,"each"s ep"consis ing"o "a"p ocess"o " ial" and"e o " o"achie e" he"bes " esul s." 3.1. !USED!TOOLS! The" ool" o " choice" o " his" wo k" was" he" Py hon" (“Welcome" o" Py hon.o g,”" n.d.)"p og amming" language"and"i s"da a"analysis"and"da a"science"lib a ies,"mainly"Numpy"(“NumPy"—"NumPy,”"n.d.)," Pandas"(“Py hon"Da a"Analysis"Lib a y"—"pandas:"Py hon"Da a"Analysis"Lib a y,”"n.d.)"and"Sciki -lea n" (“sciki -lea n:" machine" lea ning" in" Py hon" —" sciki -lea n" 0.19.0" documen a ion,”" n.d.)." A e " conside ing"o he " ools," he"choice"o "using"Py hon"was"due" o"se e al" easons."This"language"and"i s" lib a ies"a e"one" he"mos "popula " ools" o "da a"science"because" hey"ha e"se e al"ad an ages."They" a e"sui ed" o" he" as "and"easy"handling"o "la ge"amoun s"o "da a," hey"a e"in ui i e" o"use"and"easy" o" lea n." Pa icula ly" in" he" case" o " Sciki -lea n," he" compu a ions" a e" as ," he" models" a e" easy" o" implemen "while"s ill"o e ing" lexibili y"and"use "cus omiza ion,"and"mos "p ocesses"necessa y"in"a" da a"science"p ojec "a e"co e ed." 3.2. DATA!GATHERING!AND!TREATMENT! As"p e iously"men ioned," he"da a" ha "was"used"in" his"s udy"came" om" he"PubChem"da abase," namely" om" he" Bioassay" da abase" whe e" he" in o ma ion" abou "all" he" assays" ha " ha e" been" uploaded" is" ga he ed," and" da ase s" a e" a ailable" o " each" bioassay." These" da ase s" con ain" he" in o ma ion"abou " he" ype"o "assay"o "sou ce,"which"could"be"a"p ima y"sc eening,"a"con i ma o y" assay"o "a"li e a u e"based"da ase ," he" es ed"compounds"and" hei "classi ica ion"as"ac i e"o "inac i e."" In" he"case"o "EGFR," he"da a"a ailable"was"qui e"ex ensi e"in"compa ison" o"some"o he "compounds" and"a"choice"was"made" o"use" he"da ase " ha "ga he ed"all" he"a ailable"in o ma ion" o " his" a ge "as" almos "all"o " he"assays"we e"con i ma o y"o " he"in o ma ion"was"collec ed" om" he"li e a u e."This" gi es" con idence" in" he" quali y" o " he" esul s" and" ha " he" numbe " o " alse" posi i es" will" be" p opo ionally"low"compa ed" o" he"case"o "p ima y"sc een"assays"whe e"a"la ge"numbe "o "compounds" is" es ed" and" he" IC50" h eshold" o " he" compound" o" be" conside ed" ac i e" is" highe ," while" o " con i ma o y"assays" he" h eshold"is" igh e " o"con i m" he" indings"o " he"p ima y"sc eens."The" o al" da ase "was"composed"o "13116"compounds"o "which"4692"a e"labeled"as"ac i e." The"da ase s"a e"a ailable" o "download"in"di e en " o ma s"and"con ain"in o ma ion"abou " he"assays" and" he"compounds" ha "we e" es ed," oge he "wi h" hei "classi ica ion"as"ac i e"o "inac i e."Fo " his" s udy," he"only"in o ma ion" ha "was"kep "was" he"CIDs" o " he"compounds"( hei "PubChem"iden i ie s" ha "allow" he"ga he ing"o "in o ma ion"abou " hem" om" he"da abase)"and" hei "classi ica ion." In" his"wo k," he"chemical"desc ip o s" ha "we e"chosen" o"cha ac e ize" he"compounds"and"be" he" ea u es"used"in" he"model"building"we e"PubChem’s"own"pa icula "desc ip o s"named"Subs uc u e" Finge p in s" ha "consis "o "a"bina y"s ing"o "881"bi s"whe e"each"bi "codes" he"p esence"o "absence"o " 25" " a"chemical"s uc u e.""The"coding" o "each"bi "can"be" ound"in" he"annexes"sec ion"o " his"wo k."This" inge p in "is"a ailable"as"a"base64"encoded"s ing."" To"ex ac " hese" inge p in s," he"CIDs"we e"used"and"we e"ga he ed" om"PubChem"using"a"Py hon" lib a y,"PubChemPy"(“PubChemPy"documen a ion"—"PubChemPy"1.0.4"documen a ion,”"n.d.)," ha " uses"PubChem’s"API" o"allow" he"use " o"ge "only" he"in o ma ion"needed"p og amma ically."" As" such," he" p ocess" o" ob ain" he" inal" da ase " consis ed" o " he" download" and" ea men " o " he" da ase ," he" impo " o " he" inge p in s" in" base64," he" con e sion" o " hose" inge p in s" o" bina y," emo al"o " he"padding,"inpu ing"each"o " he" esul ing"881"bi s" o"a"sepa a e" ea u e,"and"assigning" he" alue"one" o" he"ac i e"compounds"and"ze o" o" he"inac i e"ones,"so" ha " he" esul ing"da ase " consis ed"o "881"bina y"inpu " a iables,"one"bina y" a ge " a iable,"and"13116" eco ds." This" esul ing"da ase "was" hen"di ided"in o" he" aining"and" es ing"se s"in"a"p opo ion"o "70"and"30%" espec i ely,"as"will"be"explained"in" he"sec ion"abou "model"e alua ion." 3.3. CHOICE!OF!ALGORITHMS! Since" one" o " he" objec i es" o " his" s udy" was" he" compa ison" o " se e al" algo i hms" o " hei " pe o mance"on" he"p oblem"a "hand," he"op ion" ollowed"was" o"use"a"wide"di e si y"o "classi ie s" known" o" be" able" o" gi e" good" esul s," ha " could" be" applied" and" e alua ed" using" he" same" me hodology"and" ool,"which"mean " he"implemen a ions"o "Sciki -lea n."As"such," he"classi ie s" ha " we e"applied"in" his"wo k"we e:" • Decision"T ee"wi h"CART"algo i hm," • Gaussian"Naï e"Bayes,"whe e" he"likelihood"o " he" ea u es"is"assumed" o"be"Gaussian"and" Be noulli"Naï e"Bayes" ha "assumes" he"da a"is"dis ibu ed"acco ding" o"mul i a ia e"Be noulli" dis ibu ions"and"as"such," equi es"samples" o"be" ep esen ed"as"bina y- alued" ea u e" ec o s" which"is" he"case" o " he"used"da ase "(“sciki -lea n:"machine"lea ning"in"Py hon"—"sciki -lea n" 0.19.0"documen a ion,”"n.d.)," • Logis ic"Reg ession," • K-Nea es "Neighbo s," • Suppo "Vec o "Machine," • Neu al"Ne wo k"–"Mul ilaye "Pe cep on," • Ensemble"Me hods:" o Random"Fo es ," o Ex emely"Randomized"T ee," o Bagging," o Boos ing"wi h"AdaBoos "algo i hm," o Vo ing." 3.4. !FEATURE!SELECTION! I " is" known" ha " ea u e" selec ion" is" an" impo an " pa " o " mos " Da a" Mining" o " Machine" Lea ning" p ojec s,"and"i s"impo ance"is"inc easing"in" ecen " imes"as"la ge"amoun s"o "da a"become"easily" a ailable," and" da a" scien is s" a e" aced" wi h" da ase s" ha " o en" con ain" a iables" ha " a e" ei he " 26" " i ele an " o"model" he"p oblem"being"s udied"o "a e" edundan ," ha "is,"con ey"in o ma ion" ha "is" al eady"encoded"in"o he " a iables"(Fe nandez-Lozano"e "al.,"2013)." The e"a e"se e al" echniques" ha "can"be"used" o"pe o m" ea u e" educ ion" ha "ha e" he"pu pose"o " inding"an"op imal"subse "o " ea u es" ha "a e"needed" o" ind" he"solu ion" o "a"p oblem."The"bene i s" o "applying" hese" echniques"include" he"need" o "a"smalle "numbe "o "samples" o"ob ain"an"op imal" esul ,"less" unning" ime"(Fe nandez-Lozano"e "al.,"2013)"and"be e "pe o mance"o " he"model"by" diminishing" he" isk"o "o e i ing"(G us,"2015)." Conside ing" hese"ad an ages,"di e en " echniques"we e" ied"and"applied" o" he"da a" ha "was"used" o"build" he"p edic i e"models."The"expe imen ed" echniques"we e" he"Sciki -lea n"implemen a ions" o :" • Fea u e" emo al"based"on"low" a iance"–" his" echnique"elimina es"all" he" ea u es" ha "don’ " mee "a"ce ain" a iance" h eshold," ha "is," ha "ha e" he"same" alue" o "a"ce ain"p opo ion"o " he"examples;" • P incipal"Componen "Analysis"–"PCA"is"an"unsupe ised"app oach" ha "uses" o a ion"me hods" o"g oup" he" a iables"in"such"a"way" ha " he" o al" a iance"explained"is"maximum," esul ing"in" a" educed" se " o " a iables" ( he" P incipal" Componen s)" ha " a e" linea " combina ions" o " he" o iginal"ones,"o de ed"by" hei " a iance"(Das,"Cha opadhyay,"&"Gup a,"2016)." • Uni a ia e" ea u e"selec ion"–"This" echnique"uses"s a is ic" es ing" o"selec " he"!" ea u es"wi h" he"s onges " ela ionship"wi h" he" a ge " a iable."In" his"case," he" es "used"was" he"é1" es ." • Linea "models" wi h" egula iza ion"–" These"me hods" use" coe icien s" om"linea "models" o" selec " he"bes " ea u es"since,"i " he" a iables"a e"on" he"same"scale," he"mos "impo an "ones" o " he"model"will"ha e" he"highe "coe icien s,"and" he"ones"unco ela ed" o" he" a ge " a iable" will"ha e"coe icien s"close" o"ze o."The"use"o " egula iza ion"adds"a"penal y" o" he"loss" unc ion" o"a oid"o e i ing."" • T ee-based" ea u e"selec ion"–" hese"models"use" he"measu e"o "pu i y"inc ease" esul ing" om" he" ee"models’"pa i ions" o"selec " he"bes " ea u es." A e " he"expe imen a ion"wi h" hese"di e en " echniques,"i "was" e i ied" ha " he"ob ained" esul s," depending"on" he"model,"we e"ei he "simila "o "wo se" han"was" he"case"wi h" he"use"o " he" o al" numbe "o " ea u es."As"such,"and"because" he"di e ence"in"compu a ion" imes"was"no " e y"signi ican ," he"op ion" aken"was" o"use" he" ull"se "o " ea u es." These" esul s"may"be" ela ed" o" he"na u e"o " he"da ase "whe e" he" ea u es"all"encode"chemical" s uc u es"which"a e" he"base"o "molecula " a ge "ac i i y"which"is"wha "is"being"modeled." 3.5. !TREATMENT!OF!IMBALANCED!DATA! HTS"da a"is"usually"cha ac e ized"by"highly" imbalanced"da a," ha "is,"da a" ha "has"a"much"highe " p opo ion"o "one"class" han"ano he ,"as" o "each" es " he e"will"usually"be"a"much"smalle "numbe "o " ac i e" compounds" han" inac i e" compounds." This" ype"o " da ase s" may" lead" o" a" weakened" pe o mance"o "models"as"i "may"skew" he"accu acy"(Q."Li,"Wang,"&"B yan ,"2009)."I " he" eco ds"o "a" da ase "a e"composed"almos "only"o "one"o " he"classes," he"model"can"ha e"a"high"accu acy"by"igno ing" one"o " he"classes"which"migh "de ea " he"pu pose"o " he"p ojec ." 27" " In" his"wo k," he"da ase "used"was" no "highly"imbalanced"as"i "was"mos ly"based"on" con i ma o y" assays."Ne e heless,"a"mo e"balanced"da ase "was"ob ained."Bo h"unde sampling"and"o e sampling" echniques"we e" ied" o"achie e" his"objec i e." To"unde sample" he"da ase ,"a"simple"me hod"o " andomly" emo ing"a"p opo ion"o " he"non-ac i e" eco ds"was"used,"bu " he"bes " esul s"we e"achie ed"by"o e sampling" he"mino i y"class," he"ac i e" compounds,"using" he"SMOTE"(Syn he ic"Mino i y"O e "Sampling)"me hod." The"SMOTE"me hod"c ea es"syn he ic" eco ds"o " he"mino i y"class"by"selec ing" o "each"ins ance"o " he"mino i y"class"!"ins ances"closes "in"Euclidean"dis ance,"calcula ing" he"di e ence"be ween" he" ins ance"and"i s"neighbo s,"mul iplying" ha " alue"by"a" andom"numbe "and"adding" he" esul " o" he" a iables"o " he"mino i y" eco d."The" alue"o "!"is"dependen "on" he"in ended" a io"be ween"classes." (Ramezankhani"e "al.,"2016)" Fo " his"s udy,"di e en " a ios"we e" ied,"and"al hough" he"di e ences"in" esul s"wi h"models"using" he"da ase "be o e"and"a e "cons uc ing"a"mo e"balanced"da ase "we e"no " e y"la ge," he"bes " esul s" we e" achie ed" using" he" SMOTE" me hod" o " o e sampling" o" c ea e" a" da ase " wi h" a" a io" o " 0.7" be ween" he"ac i e"and"inac i e"classes."The"SMOTE"me hod"was"applied"a e " he"sepa a ion"o " he" da ase "in" aining"and" es ing"se "so"as" o"keep"a"se "whe e" he" a io"was" he"ini ial"one"and"whe e" he" gene aliza ion"capaci y"o " he"model"could"be"assessed"wi hou "bias." 3.6. !MODEL!EVALUATION! To" choose" he" bes " model" o" sol e" a" p edic ion" p oblem" i " is" necessa y" o" ha e" measu es" o " i s" pe o mance" ha " e lec " he"di e en "aspec s"o " he"quali y"o " he"models." 3.6.1. The!Con usion!Ma ix! A " he"basis"o " he"measu es"o "quali y"o "classi ica ion"p oblems"is" he"con usion"ma ix."I "is"simply"a" ma ix" wi h" he" labels" o " all" he" possible" classes" lis ed" bo h" ho izon ally" and" e ically," wi h" he" p edic ed"classes"lis ed"on"one"o ien a ion"and" he"ac ual"classes"lis ed"pe pendicula ly."So,"i "we"ha e" wo"classes" o " he" a ge " a iable,"D9"and"D1,"we"would"ha e:" " Figu e"5"–"Con usion"Ma ix" Whe e,"we"would"ha e"on" he" op"le "co ne " he"numbe "o "ins ances"co ec ly"classi ied"as"belonging" o"D9,"on" he"bo om"le "co ne " he"numbe "o "ins ances"classi ied"as"D9" ha "ac ually"belong" o"class" D1,"on" he" op" igh " he"numbe "o "ins ances"classi ied"as"D1" ha "ac ually"belong" o"D9"and"on" he" bo om" igh " he"numbe "o "ins ances"co ec ly"classi ied"as"D1." 28" " So,"i "we"would"ha e"D9"as" he"“nega i e”"class"and"D1"as" he"“posi i e”"class"we"would"ha e:" " Figu e"6"–"Con usion"Ma ix" o "Bina y"Classi ica ion" Whe e"TN,"FP,"FN,"and"TP"mean" ue"nega i es," alse"posi i es," alse"nega i es"and" ue"posi i es," espec i ely." 3.6.2. Accu acy!Measu es! The"mos "gene ally"used"measu e" o"e alua e" he"quali y"o "a"classi ie "is"i s"accu acy."I "p o ides"a"way" o"measu e" he"gene al"p edic i e"capabili y"o " he"classi ie ,"and"i s" o mula"is:" á@ +á5 á@ +á5 +ç@ +ç5 30 " Al hough"accu acy"is"a" aluable"indica o "o " he"model’s"quali y,"i "does"no "p o ide"a"comp ehensi e" iew"o " he"model’s"pe o mance,"and"o he "me ics"a e"necessa y" o"make" ha "in e p e a ion."The" ones" ha "we e"used"in" his"s udy"a e"desc ibed"below:" è~& 6'6•6'* =á@ á@ +ç5 31 " è#~ê6Å6ê6'* =á5 á5 +ç@ 32 " @(~ê6 6)& =á@ á@ +ç@ 33 " ç9= 2 ∙ #(~ê6 6)&.(~êI^^ #(~ê6 6)& +(~êI^^ 34 " Sensi i i y" (o " ecall" as" i " is" named" in" Sciki -lea n)" gi es" a" me ic" o" e alua e" how" well" he" model" “cap u es”" he"posi i e"ins ances"in" ha "i "is" he"p opo ion"o " he"numbe "o " co ec ly"p edic ed" posi i es" o" he" o al"numbe "o "posi i es"while"speci ici y"is" he"same"me ic"applied" o"nega i es." P ecision,"on" he"o he "hand,"allows" he"e alua ion"o "wha "p opo ion"o " he" eco ds"classi ied"as" posi i e"a e"ac ually"posi i e."ç9"gi es"a"measu e"o " he"balance"be ween"p ecision"and" ecall." A"measu e" ha "can"be"used"e en"in" he"case"o "highly"imbalanced"da ase s"and"gi es"a"good"o e all" measu e"o " he"quali y"o " he"model"is" he"Ma hew’s"Co ela ion"Coe icien :" åDD =á@×á5 −ç@×ç5 á@ +ç@ á@ +ç5 á5 +ç@ á5 +ç5 35 " 29" " 3.7. OVERFITTING! O e i ing"co esponds" o" he"e o "o "no " ollowing" he"Occam’s"Razo "p inciple"which"is"a"pa simony" p inciple" ha "s a es" ha " he"used"model"should"only"ha e" he"complexi y"necessa y" o"model" he" s udied"p oblem"and"no"mo e."This"addi ional"complexi y"can"a ise" om"di e en " ac o s"such"as" he" inclusion"o "i ele an "componen s"as," o "example,"using"a"polynomial"o "excessi e"deg ee," he"use"o " i ele an "p edic o s"o " unning" he"lea ning"phase" o " oo"long."(Hawkins*,"2003)" This"excessi e"complexi y"o " he"models"leads" o"a"g ea "p oblem" ha "is" he"incapaci y"o " he"model" o" gene alize" o"unseen"da a"which"means" ha " he"model"migh "ha e"a" e y"high"pe o mance"on" aining" da a"and"no "pe o m"any"be e " han"a" andom"classi ie "on"unseen"da a." To" p e en " his" p oblem" and" co ec ly" assess" he" gene aliza ion" capaci y" o " models," di e en " echniques"may"be"applied,"and"in" his"s udy," wo"app oaches"we e"used"as" o"be"able" o"con iden ly" make" his" assessmen ." He e," he" used" echniques" we e" 10- old" C oss" Valida ion" and" keeping" an" independen " es "se " ha "was"unal e ed"and"ne e "seen"by" he"model"du ing" aining." !-Fold" c oss" alida ion" is" used" o" assess" he" models’" gene aliza ion" capaci y" and" p o ide" a" mo e" accu a e"me ic"o " he"models’"pe o mance."I "wo ks"by"di iding" he"da ase "in"!" olds"and" ain"and" es " he"model"!" imes,"each" ime" es ing"on"one"o " he" olds"while" he" aining"is"done"on" he" emaining" ! − 1" olds"all" oge he ."The"assessmen "me ics"a e" hen"ob ained"by"calcula ing" he"a e age"o " he" indi idual" esul s"on"each"o " he" olds"when"used"as"a" es ing"se ."In" his"case,"10- old"c oss" alida ion" was"chosen"since"i "is" he"one"mos "commonly"used"because"o "i s"empi ically"demons a ed"good" esul s"(J.-H."Kim,"2009)." A"scheme" o " his"p ocess"is:" " Figu e"7"-"!"- old"C oss"Valida ion" To"p o ide"an"added"le el"o "con idence"on" he"assessmen "o " he"pe o mance"o " he"models"on" unseen"da a,"a" es "se "wi h"30%"o " he"da a"was"ini ially"spli " om" he"o iginal"da ase ," ha "was"ne e " shown" o" he"model"du ing" aining,"and" ha "was"no "o e sampled"and,"as"such,"p o ides"a"good" es ima e"o " he"quali y"o " he"models"in" e ms"o " hei "gene aliza ion"capabili y." !" !# !$ !% T aining'Se T aining'Folds Tes ing' Fold &i e a ion 3 d i e a ion 2nd i e a ion 1s i e a ion ! = 1 &)!* " *+% … 30" " 3.8. !MODELS’!PARAMETER!OPTIMIZATION! Mos "classi ie s"will"ha e"pa ame e s" ha "in luence" he"pe o mance"o " he"model."These"pa ame e s" can"de e mine" he"con igu a ion"and"complexi y"o " he"model"and,"as"such,"de e mine"i s"lea ning"and" whe he "i "unde "o "o e i s" he"da a." To" op imize" he" choice" o " hese" pa ame e s" as" o" ge " he" bes " model" possible" om" he" applied" algo i hm," he" simples " app oach" would" be" o" i e a e" he" aining"o " he" model" using" di e en " pa ame e s"and" hen"compa e" he" esul s,"bu " he e"a e"mo e"au oma ic"me hods"o "making" ha " selec ion." In" his"wo k,"bo h"app oaches"we e"used."Fi s ly,"an"i e a i e"p ocess"o " unning" he"lea ning"o " he" models"se e al" imes"wi h"pa ame e s"wi h"a"wide " ange"was"pe o med," o" hen"ob ain"a"smalle " in e al"o " alues" ha "a e" hen"subjec " o" he"p ocess"o "au oma ic"sea ch"o " he"op imal" alue." The"au oma ed"pa "o " he"sea ch"was"conduc ed"using"an"implemen a ion" om"Sciki -lea n"o "a"G id" Sea ch" e alua ed" by" c oss- alida ion." This" me hod" es s" all" possible" combina ions" o " pa ame e s," e alua ing" each" combina ion" h ough" c oss- alida ion," which" a oids" he" choice" o " models" o " unnecessa y" high" complexi y" ha " could" lead" o" o e i ing," and" using" a" sco ing" unc ion" ha " is" maximized."The"use "de e mines" he"pa ame e "g id," he"numbe "o " olds" o"use"in" he"c oss" alida ion" and" he"sco ing" unc ion" ha "by"de aul "is"accu acy."In" his"case," he"op ion"was" o"use"10- old"c oss" alida ion" o " he" easons"p e iously"men ioned,"and" ecall"as" he"sco ing" unc ion"because"o " he" na u e"o " he"p oblem,"conside ing" ha " he"ul ima e"goal"is" o"sc een"la ge"da abases" o " he"ac i e" compounds" ha "usually"will"be"in"much"smalle "numbe " han" he"non-ac i e"ones"and,"as"such," he" mos "impo an " ea u e"o " he"model"will"be"i s"abili y" o"iden i y" he"la ges "p opo ion"possible"ou " o " he"ac ual"posi i e"ins ances" ha "exis "in" he"da abase." 3.9. STUDY!WORKFLOW!! Al hough" abo e" he" di e en " pa s" o " he" s udy" we e" p esen ed" sepa a ely," he" wo k" was" done" con inuously"and"i e a i ely."Bellow,"a"wo k low"o " he"s udy"is"p esen ed,"al hough" he"o de "o " he" wo k"was"a " imes"no "as"linea "and"mo e" ecu si e" han"is"p esen ed"in" he"scheme."" 37" " " Figu e"11"-"A e age"P ecision" o "10- old"C oss"Valida ion" " Figu e"12"-"Accu acy"on"Tes "Se " " Figu e"13"-"Sensi i i y"on"Tes "Se " 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 A e age!P ecision 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Accu acy 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 Sensi i i y 38" " " Figu e"14"-"P ecision"on"Tes "Se " " Figu e"15"-"MCC"on"Tes "Se " The" cha s" used" we e" buil " o " he" me ics" ha " we e" conside ed" mos " impo an ."In" pa icula " Sensi i i y"as" he" measu e" o "abili y" o"‘cap u e’" he"ac i e"compounds"and" Accu acy" and"MCC" as" measu es"o "o e all"quali y"o " he"models." We" can" see" om" all" he" measu es" ha " he" wo s " pe o ming" classi ie s" a e" he" Naï e" Bayes," he" AdaBoos ," he"Linea "SVM,"Logis ic"Reg ession"and"Decision"T ee"while"ERT,"Bagging"wi h"DT"base" classi ie "and"!-NN"a e"in" he"mid- ange." As"such,"in"gene al" e ms,"we"can"say" ha " he"models" ha "pe o med"be e "we e"SVM"wi h"RBF"ke nel" and"MLP" om" he"indi idual"models"and"Random"Fo es ,"Bagging"wi h"MLP"as" he"base"classi ie "and" Vo ing" om" he"ensemble"models." Some"o he "conside a ions"can" be"made,"such" as" he" ac " ha "al hough" he"Gaussian" NB" had" he" highes " Sensi i i y," he" measu e" conside ed" o " mos " impo ance," i " was"no " conside ed" a" good" pe o me "as" ha "came"a " he"cos "o "a" e y"high"numbe "o " alse"posi i es"which"would"de ea " he" pu pose"o " he"model."On" he"o he "hand,"we"can"see" ha " he"models" ha "ha e"high"sensi i i y"ha e" 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 P ecision 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 MCC 39" " a"lowe " alue"o "p ecision"and" ice- e sa"which"is"expec ed,"and"in" his" ade-o " he"ones"wi h"highe " sensi i i y" will" be" p e e ed" o " his" s udy," as" he" ul ima e" objec i e" is" he" abili y" o" sc een" la ge" da abases"whe e" he"numbe "o "posi i es"is"expec ed" o"be"qui e"lowe " han" he"numbe "o "nega i es." Also," he" esul s"ob ained"show" ha " he"di e ences"in"pe o mance" om" he"bes "pe o ming"models" a e"no " e y" ele an ,"wi h" e y"simila " alues"bo h"in" he"a e age"measu es" om" he"c oss" alida ion" and"in" he" es "se ."As"such,"conside ing" ha " he"ensemble"and"SVM"models" ake"longe " o" un"and" ha " he"MLP"classi ie "showed" he"bes " esul s"in" e ms"o " he"sensi i i y"in" he" es "se "and"has"a" smalle " ime"o "p edic ion," his"could"be"conside ed" he"bes "model" o"be"used" o"sc een" he"en i e" PubChem"da abase,"subsec ions"o "i "o "o he "da abases."" 4.2.3. Gene al!Discussion! Some"conside a ions"can"be"made" aking"in o"accoun " he" esul s"o " his"s udy."In"a"gene al"way,"we" can"say" ha " he"bes "models"ob ained"good"pe o mances"as"classi ie s"o "EGFR"inhibi o s"wi h" he" MLP"classi ie ,"conside ed" he"bes ,"ha ing"high" alues" o " he"me ics"conside ed"mos "impo an ," namely"accu acy"and"sensi i i y."These"we e"89%" o "bo h" he"a e age"accu acy"and"sensi i i y" o "10- old"CV"and"87%"and"90%," espec i ely"on" he" es "se "and"also"an"MCC"o "0.74." P e ious"s udies"using"Machine"Lea ning"and"Da a"Mining"me hods" o"build"p edic i e"models"o "EGFR" inhibi o s"ha e"had" esul s"bo h"below"(Singh"e "al.,"2015)"and"abo e"(Kong"e "al.,"2016;"Zhao"e "al.," 2017)" he"ones" epo ed"in" his"s udy."This"could"be" ela ed"wi h"se e al" ac o s"in" he"me hodology" such"as" he"chemical"desc ip o s"used."" Ne e heless,"i "is"di icul " o"make"a"di ec "compa ison" o" he"p e ious"s udies"as" he"used"da ase s" we e" e y"di e en "and" he"models"c ea ed"migh "ha e"a"di e en "balance"be ween" he"e o s"de i ed" om"bias"and" a iance"and" he"la e " e e s" o" he"sensi i i y" o" he"used"da ase ."On" he"o he "hand," he" ac " ha " he"da ase "used"in" his"p ojec "comp ises"a"much"la ge "numbe "o "compounds"and," p esumably,"mo e"di e si y"o " ep esen ed"chemical"s uc u es,"will"likely"ha e" esul ed"in"models"wi h" be e "gene aliza ion"capabili y"and"less" a iance"de i ed"e o ."Also,"making"a"gene al"compa ison" wi h" he"s udies" ha "had"a"mo e"simila "me hodology" o" his"one"in" e ms"o " he"sou ce"and"size"o " he" used"da ase ,"we"see" ha " he" esul s"ob ained"we e"wo se" han" he"ones" o " he"s udies"aimed"a " p edic ing"inhibi o y"ac i i y"on"EGFR" ha "used"small" aining"se s,"which"co obo a es" he"in ui ion" ha " he"use"o "di e en "da ase s"wi h" e y"dispa a e"sizes"will"lead" o"di e en " esul s." Addi ionally,"since" he" alues" o " he"quali y"me ics"o " his"s udy"a e"calcula ed"based"on"10- old"c oss" alida ion"and"on" es ing"on"unseen"da a,"i "is"possible" o"be"con iden " ha " he"applica ion"o " hese" models"in"o he "da ase s"would"pe o m"simila ly"wi hou " he" isk"o "ha ing"a"g ea "dec ease"in" he" quali y"o " he"p edic ions"because"o "o e i ing." Finally,"i "is"no ewo hy" ha "because" he" esul ing"bes "models"a e"wha "is"conside ed"‘black"boxes’," ha "is,"do"no "p oduce"in e p e able" ela ionships"be ween" he"inpu "and"ou pu " a iables,"and" he" expe imen ed"me hods"o " ea u e"selec ion"did"no "p oduce"be e " esul s," he"de e mina ion"o " he" mos "impo an " ea u es" o " his"classi ica ion"p oblem"was"no "possible,"which"could"be"in e es ing" in o ma ion,"al hough" o " hese"kind"o "p ojec s,"wi h" he"objec i e"o "sc eening"la ge"da abases,"wha " is" mos " impo an " is" o " he" model" o" be" able" o" make" accu a e" p edic ions" and" no " so" much" o" unde s and" he"unde lying" ela ionships"in" he"da a." 40" " 5. CONCLUSIONS! A " he"end"o " his"s udy,"i "is"possible" o"say" ha " he"objec i es"we e"me ."" The"p oposed" e iew"o "simila "s udies,"al hough"no "exhaus i e,"is"bo h"comp ehensi e"and" ele an ," se ing"as" he"g ounds"o " he"cu en "s a e"o " he"a "bo h"in"s udies"aimed"a " he"p edic ion"o "EGFR" inhibi o s"and"in"gene al"p edic i e"s udies" ha "used"PubChem’s"da abase"as"a"da a"sou ce." Ano he "one"o " he"objec i es"was" o"es ablish"a"me hodology" ha "could"easily"be"implemen ed" o" simila "s udies"which"was"achie ed,"as"all" he" ools"used"a e" eadily"a ailable"and"allow" he"use " o" easily"gain" he"necessa y"knowledge" o"use" hem." I "was" e i ied" ha "PubChem"can"indeed"be"used"bo h"as"a"da a"sou ce" o " he"ac i i y"o "compounds" and" o " he"necessa y"in o ma ion" o"build"chemical"desc ip o s."Also,"i "was"shown" ha "PubChem’s" Subs uc u e"Finge p in s"a e"good"chemical"desc ip o s" o"be"used"as"inpu " ea u es"conside ing" he" quali y"o " he"models"de eloped."" On" he"o he "hand," he"bes "models"de eloped"could"easily"be"used" o "making"p edic ions"bo h"on" PubChem’s"da abase"o "o he s,"as"using" he"chemical" o mula"o "ano he "iden i ie "o "a"compound,"i " is"possible" o"impo " he"Subs uc u e"Finge p in s" om"PubChem"and" hus"c ea e"a"da ase "whe e" he" model"could"make"p edic ions."" Rega ding" he"o e all"quali y"o " he"models"p oduced,"as"s a ed"ea lie ," he e"ha e"been"s udies" o " he" same" pu pose" bo h" wi h" be e " and" wo se" esul s" bu " a" di ec " compa ison" o " he" models’" pe o mance" is" no " possible" o " he" easons" men ioned," in" pa icula ," because" o " he" use" o " such" di e en " aining"da ase s." As"such,"i "is"conside ed" ha " he"bes "models"de eloped"ha e"good"pe o mance"al hough"wi h" oom" o "imp o emen ," ha "could"be"achie ed"wi h" he"ex ension"o " his"wo k"bo h"in" he"da a,"desc ip o s" and"algo i hms"used." " 41" " 6. LIMITATIONS!AND!RECOMMENDATIONS!FOR!FUTURE!WORKS! The"main"limi a ions"o " his"s udy"we e" ime"and"compu a ional"powe "cons ain s."As"such,"in" u u e" wo ks" ha "would"ex end"on" his"one," he e"would"be"app oaches" ha "could"be" ied" o"imp o e" he" esul s"ob ained."" In" his"s udy,"one"o " he"objec i es"was"de e mining" he"quali y"o "PubChem’s"Subs uc u e" inge p in s" as"desc ip o s" o "p edic i e"modelling,"bu "wi h" he"possibili y" o"ex end" he"s udy,"i "would"be"o " in e es " o" also" use" o he " chemical" desc ip o s," specially" combining" 2D" and" 3D" desc ip o s," o" in es iga e"i " hose"would"p oduce"a"be e "model,"since"we"ha e"seen"o he "au ho s"use"o he "kinds" o "chemical"desc ip o s"wi h"good" esul s."" Also," since" he" Mul ilaye " Pe cep on" was" he" model" ha " was" conside ed" he" bes " o " u u e" p edic ions"on"unseen"da a," aking"in o"accoun "bo h"i s"me ics"and" unning" ime,"i "would"be"o " in e es " o" y"bo h" he"use"o "o he " ools" ha "allow"mo e" ine" uning"o " he"model" o"see"i "i "would" be" possible" o" imp o e" on" i " and" also" o he " a ia ions" o " Neu al" Ne wo ks," such" as" Deep" Belie " Ne wo ks."O he "algo i hms"could"also"be" ied,"such"as"Gene ic"Algo i hms." In"a"s udy"no "only" ocused"on"PubChem’s"da abase," he"use"o "addi ional"da a"abou "inhibi o s"no " p esen "in" he"used"da abase,"i "possible" o"ga he ,"would"p obably"bene i " he"model." On" he" o he " hand," i " would" also" be" use ul" o" es " he" model" on" a" di e en " da ase " and" make" p edic ions"on"PubChem"i sel "o "o he "da abases"and"analyze" he" ype"o "compounds"p edic ed"as" posi i e" o " he"inhibi ion"o "EGFR." " " 42" " 7. BIBLIOGRAPHY! Aus in,"C."P.,"B ady,"L."S.,"Insel,"T."R.,"&"Collins,"F."S."(2004)."NIH"Molecula "Lib a ies"Ini ia i e."Science* (New*Yo k,*N.Y.),"306(5699),"1138–9."h ps://doi.o g/10.1126/science.1105511" Ben o,"A."P.,"Gaul on,"A.,"He sey,"A.,"Bellis,"L."J.,"Chambe s,"J.,"Da ies,"M.,"…"O e ing on,"J."P."(2014)." The"ChEMBL"bioac i i y"da abase:"an"upda e."Nucleic*Acids*Resea ch,"42(D1),"D1083–D1090." h ps://doi.o g/10.1093/na /gk 1031" Be nd ,"E."R.,"Nass,"D.,"Klein ock,"M.,"&"Ai ken,"M."(2015)."Decline"in"economic" e u ns" om"new" d ugs" aises"ques ions"abou "sus aining"inno a ions."Heal h*A ai s*(P ojec *Hope),"34(2),"245– 52."h ps://doi.o g/10.1377/hl ha .2014.1029" Bilsland,"A."E.,"Pugliese,"A.,"Liu,"Y.,"Re ie,"J.,"Bu ns,"S.,"McCo mick,"C.,"…"Kei h,"W."N."(2015)." Iden i ica ion"o "a"Selec i e"G1-Phase"Benzimidazolone"Inhibi o "by"a"Senescence-Ta ge ed" Vi ual"Sc een"Using"A i icial"Neu al"Ne wo ks."Neoplasia,"17(9),"704–715." h ps://doi.o g/10.1016/j.neo.2015.08.009" BIOVIA"Da abases"|"Bioac i i y"Da abases:"MDDR."(n.d.)."Re ie ed"July"24,"2017," om" h p://accel ys.com/p oduc s/collabo a i e-science/da abases/bioac i i y- da abases/mdd .h ml" Cha an,"S.,"Abdelaziz,"A.,"Wiklande ,"J."G.,"&"Nicholls,"I."A."(2016)."A"k-nea es "neighbo "classi ica ion" o "hERG"K+"channel"blocke s."Jou nal*o *Compu e -Aided*Molecula *Design,"30(3),"229–236." h ps://doi.o g/10.1007/s10822-016-9898-z" Chen,"B.,"&"Wild,"D."J."(2010)."PubChem"BioAssays"as"a"da a"sou ce" o "p edic i e"models."Jou nal*o * Molecula *G aphics*and*Modelling,"28(5),"420–426." h ps://doi.o g/10.1016/j.jmgm.2009.10.001" Cheng,"F.,"Yu,"Y.,"Shen,"J.,"Yang,"L.,"Li,"W.,"Liu,"G.,"…"Tang,"Y."(2011)."Classi ica ion"o "Cy och ome"P450" Inhibi o s"and"Noninhibi o s"Using"Combined"Classi ie s."Jou nal*o *Chemical*In o ma ion*and* Modeling,"51(5),"996–1011."h ps://doi.o g/10.1021/ci200028n" Che kaso ,"A.,"Mu a o ,"E."N.,"Fou ches,"D.,"Va nek,"A.,"Baskin,"I."I.,"C onin,"M.,"…"T opsha,"A."(2014)." QSAR"Modeling:"Whe e"Ha e"You"Been?"Whe e"A e"You"Going"To?"Jou nal*o *Medicinal* Chemis y,"57(12),"4977–5010."h ps://doi.o g/10.1021/jm4004285" Daniel"T."La ose,"C."D."L."(2015)."Da a*Mining*and*P edic i e*Analy ics*(Wiley*Se ies*on*Me hods*and* Applica ions*in*Da a*Mining)"(Second"Edi,"Vol."25)."John"Wiley"&"Sons,"Inc." h ps://doi.o g/10.1007/s13398-014-0173-7.2" Das,"G.,"Cha opadhyay,"M.,"&"Gup a,"S."(2016)."A"compa ison"o "sel -o ganising"maps"and"p incipal" componen s"analysis."In e na ional*Jou nal*o *Ma ke *Resea ch,"58(6),"815." h ps://doi.o g/10.2501/IJMR-2016-039" Da a"Mining"F om"A" o"Z."(n.d.)."Re ie ed"July"29,"2017," om" h ps://www.sas.com/en_us/whi epape s/da a-mining- om-a-z-104937.h ml" DiMasi,"J."A.,"G abowski,"H."G.,"&"Hansen,"R."W."(2016)."Inno a ion"in" he"pha maceu ical"indus y:" New"es ima es"o "R&D"cos s."Jou nal*o *Heal h*Economics,"47,"20–33." h ps://doi.o g/10.1016/j.jhealeco.2016.01.012" Edels ein,"H."A."(1999)."In oduc ion* o*Da a*Mining*and*Knowledge*Disco e y"(3 d"ed.)."Two"C ows" 43" " Co po a ion." Fe nandez-Lozano,"C.,"Can o,"C.,"Ges al,"M.,"And ade-Ga da,"J."M.,"Rabuñal,"J."R.,"Do ado,"J.,"&"Pazos," A."(2013)."Hyb id"model"based"on"Gene ic"Algo i hms"and"SVM"applied" o" a iable"selec ion" wi hin" ui "juice"classi ica ion."TheScien i icWo ldJou nal,"2013,"982438." h ps://doi.o g/10.1155/2013/982438" Fonge ,"G."C.,"Hakkinen,"P.,"Jo dan,"S.,"&"Publicke ,"S."(2014)."The"Na ional"Lib a y"o "Medicine’s" (NLM)"Haza dous"Subs ances"Da a"Bank"(HSDB):"backg ound," ecen "enhancemen s"and" u u e" plans."Toxicology,"325,"209–16."h ps://doi.o g/10.1016/j. ox.2014.09.003" Fos e ,"P.,"&"Fawce ,"T."(2013)."Da a*Science* o *Business"(Fi s "Edi )."O’Reilly"Media,"Inc." F iedman,"J."H.,"&"F iedman,"J."H."(1997)."Da a"mining"and"s a is ics:"Wha ’s" he"connec ion." PROCEEDINGS*OF*THE*29TH*SYMPOSIUM*ON*THE*INTERFACE*BETWEEN*COMPUTER*SCIENCE* AND*STATISTICS."Re ie ed" om" h p://ci esee x.is .psu.edu/ iewdoc/summa y?doi=10.1.1.374.7489" Gas eige †,"J."(2006)."O "Molecules"and"Humans."h ps://doi.o g/10.1021/JM0608964" Gé on,"A."(2017)."Hands-On*Machine*Lea ning*wi h*Sciki -Lea n*and*Tenso Flow"(Fi s "Edi )."O’Reilly" Media,"Inc." G us,"J."(2015)."Da a*Science* om*Sc a ch"(Fi s "Edi )."O’Reilly"Media,"Inc." Hall,"M.,"F ank,"E.,"Holmes,"G.,"P ah inge ,"B.,"Reu emann,"P.,"&"Wi en,"I."H."(2009)."The"WEKA"da a" mining"so wa e."ACM*SIGKDD*Explo a ions*Newsle e ,"11(1),"10." h ps://doi.o g/10.1145/1656274.1656278" Han,"B.,"Ma,"X.,"Zhao,"R.,"Zhang,"J.,"Wei,"X.,"Liu,"X.,"…"Chen,"Y."(2012)."De elopmen "and"expe imen al" es "o "suppo " ec o "machines" i ual"sc eening"me hod" o "sea ching"S c"inhibi o s" om"la ge" compound"lib a ies."Chemis y*Cen al*Jou nal,"6(1),"139."h ps://doi.o g/10.1186/1752-153X-6- 139" Han,"J.,"Kambe ,"M.,"&"Pei,"J."(2012)."Da a*Mining*Concep s*and*Techniques"(Thi d"Edi )."Else ie ." Han,"L.,"Wang,"Y.,"&"B yan ,"S."H."(2008)."De eloping"and" alida ing"p edic i e"decision" ee"models" om"mining"chemical"s uc u al" inge p in s"and"high– h oughpu "sc eening"da a"in"PubChem." BMC*Bioin o ma ics,"9(1),"401."h ps://doi.o g/10.1186/1471-2105-9-401" Hawkins*,"D."M."(2003)."The"P oblem"o "O e i ing."h ps://doi.o g/10.1021/CI0342472" He bs ,"R."S."(2004)."Re iew"o "epide mal"g ow h" ac o " ecep o "biology."In e na ional*Jou nal*o * Radia ion*Oncology*Biology*Physics,"59(2),"S21–S26." h ps://doi.o g/10.1016/j.ij obp.2003.11.041" Huang,"N.,"Shoiche ,"B."K.,"&"I win,"J."J."(2006)."Benchma king"Se s" o "Molecula "Docking."Jou nal*o * Medicinal*Chemis y,"49(23),"6789–6801."h ps://doi.o g/10.1021/jm0608356" I win*,"J."J."(2005)."So wa e"Re iew: "ChemO ice"2005"P o"by"Camb idgeSo ." h ps://doi.o g/10.1021/CI050286X" James,"G.,"Wi en,"D.,"Has ie,"T.,"&"Tibshi ani,"R."(2013)."An*In oduc ion* o*S a is ical*Lea ning." Sp inge "Science+Business"Media." Khanna,"V.,"&"Rangana han,"S."(2011)."In"silico"app oach" o"sc een"compounds"ac i e"agains "pa asi ic" 44" " nema odes"o "majo "socio-economic"impo ance."BMC*Bioin o ma ics,"12(Suppl"13),"S25." h ps://doi.o g/10.1186/1471-2105-12-S13-S25" Kim,"J.-H."(2009)."Es ima ing"classi ica ion"e o " a e:"Repea ed"c oss- alida ion," epea ed"hold-ou " and"boo s ap."Compu a ional*S a is ics*&*Da a*Analysis,"53(11),"3735–3745." h ps://doi.o g/10.1016/j.csda.2009.04.009" Kim,"S."(2016)."Ge ing" he"mos "ou "o "PubChem" o " i ual"sc eening."Expe *Opinion*on*D ug* Disco e y,"11(9),"843–855."h ps://doi.o g/10.1080/17460441.2016.1216967" Kim,"S.,"Thiessen,"P."A.,"Bol on,"E."E.,"Chen,"J.,"Fu,"G.,"Ginduly e,"A.,"…"B yan ,"S."H."(2016)."PubChem" Subs ance"and"Compound"da abases."Nucleic*Acids*Resea ch,"44(D1),"D1202–D1213." h ps://doi.o g/10.1093/na /gk 951" Kong,"Y.,"Qu,"D.,"Chen,"X.,"Gong,"Y.-N.,"&"Yan,"A."(2016)."Sel -O ganizing"Map"(SOM)"and"Suppo " Vec o "Machine"(SVM)"Models" o " he"P edic ion"o "Human"Epide mal"G ow h"Fac o "Recep o " (EGFR/"E bB-1)"Inhibi o s."Combina o ial*Chemis y*&*High*Th oughpu *Sc eening,"19(5),"400– 411."h ps://doi.o g/10.2174/1386207319666160414105044" Koza,"J."R."(1992)."Gene ic*p og amming :*on* he*p og amming*o *compu e s*by*means*o *na u al* selec ion."MIT"P ess." Law,"V.,"Knox,"C.,"Djoumbou,"Y.,"Jewison,"T.,"Guo,"A."C.,"Liu,"Y.,"…"Wisha ,"D."S."(2014)."D ugBank"4.0:" shedding"new"ligh "on"d ug"me abolism."Nucleic*Acids*Resea ch,"42(Da abase"issue),"D1091-7." h ps://doi.o g/10.1093/na /gk 1068" Lesko ec,"J.,"Raja aman,"A.,"&"Ullman,"J."D."(2011)."Mining*o *Massi e*Da ase s."Lec u e*No es* o * S an o d*CS345A*Web*Mining"(Fi s "Edi ,"Vol."67)."Camb idge"Uni e si y"P ess." h ps://doi.o g/10.1017/CBO9781139058452" Li,"H.,"You,"L.,"Xie,"J.,"Pan,"H.,"&"Han,"W."(2017)."The" oles"o "subcellula ly"loca ed"EGFR"in"au ophagy." Cellula *Signalling,"35,"223–230."h ps://doi.o g/10.1016/j.cellsig.2017.04.012" Li,"Q.,"Wang,"Y.,"&"B yan ,"S."H."(2009)."A"no el"me hod" o "mining"highly"imbalanced"high- h oughpu " sc eening"da a"in"PubChem."Bioin o ma ics,"25(24),"3310–3316." h ps://doi.o g/10.1093/bioin o ma ics/b p589" Madej,"T.,"Lanczycki,"C."J.,"Zhang,"D.,"Thiessen,"P."A.,"Gee ,"R."C.,"Ma chle -Baue ,"A.,"&"B yan ,"S."H." (2014)."MMDB"and"VAST+:" acking"s uc u al"simila i ies"be ween"mac omolecula "complexes." Nucleic*Acids*Resea ch,"42(Da abase"issue),"D297-303."h ps://doi.o g/10.1093/na /gk 1208" Ma sland,"S."(2015)."Machine*Lea ning :*An*ALgo i hmic*Pe pec i e"(Second)."CRC"P ess." McInnes,"C."(2007)."Vi ual"sc eening"s a egies"in"d ug"disco e y."Cu en *Opinion*in*Chemical* Biology,"11(5),"494–502."h ps://doi.o g/10.1016/j.cbpa.2007.08.033" No o a skyi,"S.,"Sushko,"I.,"Kö ne ,"R.,"Pandey,"A."K.,"&"Te ko,"I."V."(2011)."A"compa ison"o "di e en " QSAR"app oaches" o"modeling"CYP450"1A2"inhibi ion."Jou nal*o *Chemical*In o ma ion*and* Modeling,"51(6),"1271–1280."h ps://doi.o g/10.1021/ci200091h" NumPy"—"NumPy."(n.d.)."Re ie ed"Augus "23,"2017," om"h p://www.numpy.o g/" Oda,"K.,"Ma suoka,"Y.,"Funahashi,"A.,"&"Ki ano,"H."(2005)."A"comp ehensi e"pa hway"map"o " epide mal"g ow h" ac o " ecep o "signaling."Molecula *Sys ems*Biology,"1,"2005.0010." h ps://doi.o g/10.1038/msb4100014" 45" " Online"Chemical"Modeling"En i onmen ."(n.d.)."Re ie ed"Augus "14,"2017," om" h ps://ochem.eu/home/show.do" Poulio ,"Y.,"Chiang,"A."P.,"&"Bu e,"A."J."(2011)."P edic ing"Ad e se"D ug"Reac ions"Using"Publicly" A ailable"PubChem"BioAssay"Da a."Clinical*Pha macology*&*The apeu ics,"90(1),"90–99." h ps://doi.o g/10.1038/clp .2011.81" PubChem"Da a"Sou ces."(n.d.)."Re ie ed"July"9,"2017," om" h ps://pubchem.ncbi.nlm.nih.go /sou ces/" PubChemPy"documen a ion"—"PubChemPy"1.0.4"documen a ion."(n.d.)."Re ie ed"Augus "30,"2017," om"h ps://pubchempy. ead hedocs.io/en/la es /" Py hon"Da a"Analysis"Lib a y"—"pandas:"Py hon"Da a"Analysis"Lib a y."(n.d.)."Re ie ed"Augus "23," 2017," om"h p://pandas.pyda a.o g/" Ramezankhani,"A.,"Pou nik,"O.,"Shah abi,"J.,"Azizi,"F.,"Hadaegh,"F.,"&"Khalili,"D."(2016)."The"Impac "o " O e sampling"wi h"SMOTE"on" he"Pe o mance"o "3"Classi ie s"in"P edic ion"o "Type"2"Diabe es." Medical*Decision*Making,"36(1),"137–144."h ps://doi.o g/10.1177/0272989X14560647" Recep o "Binding."(n.d.)."Re ie ed"July"25,"2017," om" h p://www. enichel.ne /pages/P o essional/subpages/QT/Tables/pbyd ug.h m" Rumelha ,"D."E.,"Wid ow,"B.,"&"Leh ,"M."A."(1994)."The"basic"ideas"in"neu al"ne wo ks." Communica ions*o * he*ACM,"37(3),"87–92."h ps://doi.o g/10.1145/175247.175256" Scha adin,"T."M.,"He,"W.,"Yiannakou,"Y.,"Tomilo ,"A."A.,"Saldana,"M.,"Co opassi,"G."A.,"…"Hende son," P."T."(2017)."Syn hesis"and"biochemical"cha ac e iza ion"o "EGF" ecep o "in"a"wa e -soluble" memb ane"model"sys em."PLOS*ONE,"12(6),"e0177761." h ps://doi.o g/10.1371/jou nal.pone.0177761" Schie z,"A."C."(2009)."Vi ual"sc eening"o "bioassay"da a."Jou nal*o *Chemin o ma ics,"1(1),"21." h ps://doi.o g/10.1186/1758-2946-1-21" sciki -lea n:"machine"lea ning"in"Py hon"—"sciki -lea n"0.19.0"documen a ion."(n.d.)."Re ie ed"Augus " 23,"2017," om"h p://sciki -lea n.o g/s able/" Shen,"M.,"Su,"B.-H.,"Esposi o,"E."X.,"Hop inge ,"A."J.,"&"Tseng,"Y."J."(2011)."A"Comp ehensi e"Suppo " Vec o "Machine"Bina y"hERG"Classi ica ion"Model"Based"on"Ex ensi e"bu "Biased"End"Poin " hERG"Da a"Se s."Chemical*Resea ch*in*Toxicology,"24(6),"934–949." h ps://doi.o g/10.1021/ x200099j" Singh,"H.,"Singh,"S.,"Singla,"D.,"Aga wal,"S."M.,"Ragha a,"G."P."S.,"&"Aga wal,"S."(2015)."QSAR"based" model" o "disc imina ing"EGFR"inhibi o s"and"non-inhibi o s"using"Random" o es ."Biology*Di ec * 2015*10:1,"8(1),"28."h ps://doi.o g/10.1186/1745-6150-8-28" Su,"B.-H.,"Tu,"Y.,"Lin,"C.,"Shao,"C.-Y.,"Lin,"O."A.,"&"Tseng,"Y."J."(2015)."Rule-Based"P edic ion"Models"o " Cy och ome"P450"Inhibi ion."Jou nal*o *Chemical*In o ma ion*and*Modeling,"55(7),"1426–1434." h ps://doi.o g/10.1021/acs.jcim.5b00130" Szymański,"P.,"Ma kowicz,"M.,"&"Mikiciuk-Olasik,"E."(2012)."Adap a ion"o "high- h oughpu "sc eening"""""" ………in"d ug"disco ey- oxicological"sc eening" es s."In e na ional*Jou nal*o *Molecula *Sciences,"13(1)," ………427–452."h ps://doi.o g/10.3390/ijms13010427" " The"PubChem"P ojec ."(n.d.)."Re ie ed"July"9,"2017," om"h ps://pubchem.ncbi.nlm.nih.go /" 46" " T oiani,"T.,"Ma inelli,"E.,"Capasso,"A.,"Mo gillo,"F.,"O di u a,"M.,"De"Vi a,"F.,"&"Cia diello,"F."(2012)." Ta ge ing"EGFR"in"panc ea ic"cance " ea men ."Cu en *D ug*Ta ge s,"13(6),"802–10."Re ie ed" om"h p://www.ncbi.nlm.nih.go /pubmed/22458527" Vas ag,"B."(2005)."Fo "EGFR"Resea ch,"New"Ta ge ed"D ugs"Mean"New"Ques ions."JNCI*Jou nal*o * he* Na ional*Cance *Ins i u e,"97(9),"628–630."h ps://doi.o g/10.1093/jnci/97.9.628" Wang,"S.,"Xie,"L.,"Wang,"H.,"Zhu,"B.,"&"Niu,"A."(2016)."P edic ion"o "selec i e"es ogen" ecep o "be a" agonis "using"open"da a"and"machine"lea ning"app oach."D ug*Design,*De elopmen *and* The apy,"Volume*10,"2323–2331."h ps://doi.o g/10.2147/DDDT.S110603" Wang,"Y.,"Suzek,"T.,"Zhang,"J.,"Wang,"J.,"He,"S.,"Cheng,"T.,"…"B yan ,"S."H."(2014)."PubChem"BioAssay:" 2014"upda e."Nucleic*Acids*Resea ch,"42(D1),"D1075–D1082." h ps://doi.o g/10.1093/na /gk 978" Wang,"Y.,"Xiao,"J.,"Suzek,"T."O.,"Zhang,"J.,"Wang,"J.,"&"B yan ,"S."H."(2009)."PubChem:"a"public" in o ma ion"sys em" o "analyzing"bioac i i ies"o "small"molecules."Nucleic*Acids*Resea ch," 37(Web"Se e "issue),"W623-33."h ps://doi.o g/10.1093/na /gkp456" Weis,"D."C.,"Visco,"D."P.,"&"Faulon,"J."L."(2008)."Da a"mining"PubChem"using"a"suppo " ec o "machine" wi h" he"Signa u e"molecula "desc ip o :"Classi ica ion"o " ac o "XIa"inhibi o s."Jou nal*o * Molecula *G aphics*and*Modelling,"27(4),"466–475." h ps://doi.o g/10.1016/j.jmgm.2008.08.004" Welcome" o"Py hon.o g."(n.d.)."Re ie ed"Augus "23,"2017," om"h ps://www.py hon.o g/" Wi en,"I."H.,"&"F ank,"E."(2005)."Da a*Mining:*P ac ical*Machine*Lea ning*Tools*and*Techniques." Complemen a y*li e a u e*None"(Second"Edi)."Else ie ."h ps://doi.o g/0120884070," 9780120884070" Ya den,"Y.,"&"Sliwkowski,"M."X."(2001)."Un angling" he"E bB"signalling"ne wo k."Na u e*Re iews* Molecula *Cell*Biology,"2(2),"127–137."h ps://doi.o g/10.1038/35052073" Yu,"L.,"Shi,"X.,"Tian,"S.,"Gao,"S.,"&"Li,"L."(2017)."Classi ica ion"o "Cy och ome"P450"1A2"Inhibi o s"and" Noninhibi o s"Based"on"Deep"Belie "Ne wo k."In e na ional*Jou nal*o *Compu a ional* In elligence*and*Applica ions,"16(1),"1750002."h ps://doi.o g/10.1142/S146902681750002X" Zhao,"M.,"Wang,"L.,"Zheng,"L.,"Zhang,"M.,"Qiu,"C.,"Zhang,"Y.,"…"Niu,"B."(2017)."2D-QSAR"and"3D-QSAR" Analyses" o "EGFR"Inhibi o s."BioMed*Resea ch*In e na ional,"2017,"1–11." h ps://doi.o g/10.1155/2017/4649191" " " " " " " " 53" " C: Use s Liliana Rosa Anaconda3 lib si e-packages sklea n c oss_ alida ion.py:41: D ep eca ionWa ning: This module was dep eca ed in e sion 0.18 in a o o he model _selec ion module in o which all he e ac o ed classes and unc ions a e mo ed. Al so no e ha he in e ace o he new CV i e a o s a e di e en om ha o his module. This module will be emo ed in 0.20. "This module will be emo ed in 0.20.", Dep eca ionWa ning) In [2]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [3]: om sklea n.model_selec ion impo G idSea chCV In [4]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [5]: cls = KNeighbo sClassi ie () In [6]: K = lis ( ange(5,15)) pa am_g id = {'n_neighbo s': K} sco ing = ['accu acy', ' ecall', 'p ecision'] In [7]: gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing=' ecall', n_jobs=-1, c =10) In [8]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'n_neighbo s': 5} In [10]: cls_bes = KNeighbo sClassi ie (n_neighbo s=5) In [11]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) 54" " In [12]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.87 +/- 0.02 A e age Recall 0.88 +/- 0.03 A e age P ecision 0.83 +/- 0.02 In [13]: model = cls_bes . i (X_ ain_ es, y_ ain_ es) In [14]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.91 T aining ecall: 0.92 T aining p ecision: 0.87 Tes accu acy: 0.87 Tes ecall: 0.86 Tes p ecision: 0.80 p ecision ecall 1-sco e suppo 0 0.92 0.87 0.90 2497 1 0.80 0.86 0.83 1438 a g / o al 0.87 0.87 0.87 3935 [[2184 313] [ 198 1240]] 55" " 8.1.5. Logis ic!Reg ession! In [1]: impo numpy as np impo pandas as pd om sklea n.linea _model impo Logis icReg ession om sklea n.model_selec ion impo c oss_ alida e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om imblea n.o e _sampling impo SMOTE C: Use s Liliana Rosa Anaconda3 lib si e-packages sklea n c oss_ alida ion.py:41: D ep eca ionWa ning: This module was dep eca ed in e sion 0.18 in a o o he model _selec ion module in o which all he e ac o ed classes and unc ions a e mo ed. Al so no e ha he in e ace o he new CV i e a o s a e di e en om ha o his module. This module will be emo ed in 0.20. "This module will be emo ed in 0.20.", Dep eca ionWa ning) In [2]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [3]: om sklea n.model_selec ion impo G idSea chCV In [4]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [5]: cls = Logis icReg ession(sol e ='sag', max_i e =4000) In [6]: pa am_g id ={'C': [ 0.1, 1, 10]} sco ing = ['accu acy', ' ecall', 'p ecision'] In [7]: gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing=' ecall', n_jobs=-1, c =10) 56" " In [8]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'C': 10} In [9]: cls_bes = Logis icReg ession(C = 10, sol e ='sag', max_i e =4000) In [10]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [11]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.84 +/- 0.01 A e age Recall 0.82 +/- 0.02 A e age P ecision 0.81 +/- 0.01 In [12]: model = cls_bes . i (X_ ain_ es, y_ ain_ es) In [13]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.87 T aining ecall: 0.85 T aining p ecision: 0.83 Tes accu acy: 0.84 57" " Tes ecall: 0.79 Tes p ecision: 0.78 p ecision ecall 1-sco e suppo 0 0.88 0.87 0.87 2497 1 0.78 0.79 0.79 1438 a g / o al 0.84 0.84 0.84 3935 [[2167 330] [ 295 1143]] 8.1.6. SVM!wi h!Linea !Ke nel! In [28]: impo numpy as np impo pandas as pd om sklea n.s m impo SVC om sklea n.s m impo Linea SVC om sklea n.model_selec ion impo c oss_ alida e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om imblea n.o e _sampling impo SMOTE In [29]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [30]: om sklea n.model_selec ion impo G idSea chCV In [31]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [43]: cls = Linea SVC( andom_s a e=24) In [44]: pa am_g id = {'C': [1, 10, 100]}, sco ing = ['accu acy', ' ecall', 'p ecision'] In [45]: 58" " gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing=' ecall', n_jobs=-1, c =10) In [46]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'C': 1} In [47]: cls_bes = Linea SVC(C=1, andom_s a e= 24) In [48]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [49]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.84 +/- 0.01 A e age Recall 0.82 +/- 0.02 A e age P ecision 0.81 +/- 0.02 In [50]: model = cls. i (X_ ain_ es, y_ ain_ es) In [51]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) 59" " p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.87 T aining ecall: 0.85 T aining p ecision: 0.83 Tes accu acy: 0.84 Tes ecall: 0.80 Tes p ecision: 0.77 p ecision ecall 1-sco e suppo 0 0.88 0.86 0.87 2497 1 0.77 0.80 0.78 1438 a g / o al 0.84 0.84 0.84 3935 [[2147 350] [ 287 1151]] 8.1.7. SVM!wi h!RBF!Ke nel! In [22]: impo numpy as np impo pandas as pd om sklea n.s m impo SVC om sklea n.s m impo Linea SVC om sklea n.model_selec ion impo c oss_ alida e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om imblea n.o e _sampling impo SMOTE In [23]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [24]: om sklea n.model_selec ion impo G idSea chCV In [25]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [30]: 60" " cls = SVC(ke nel=' b ', andom_s a e=24) In [31]: pa am_g id ={'C': [1, 10, 100]} sco ing = ['accu acy', ' ecall', 'p ecision'] In [32]: gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing=' ecall', n_jobs=-1, c =10) In [33]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'C': 100} In [35]: cls_bes = SVC(ke nel=' b ', C=100, andom_s a e=24) In [36]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [37]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.89 +/- 0.01 A e age Recall 0.89 +/- 0.03 A e age P ecision 0.84 +/- 0.01 In [39]: model = cls_bes . i (X_ ain_ es, y_ ain_ es) In [40]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) 61" " es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.91 T aining ecall: 0.91 T aining p ecision: 0.88 Tes accu acy: 0.88 Tes ecall: 0.86 Tes p ecision: 0.82 p ecision ecall 1-sco e suppo 0 0.92 0.89 0.90 2497 1 0.82 0.86 0.84 1438 a g / o al 0.88 0.88 0.88 3935 [[2230 267] [ 207 1231]] 8.1.8. Mul ilaye !Pe cep on! In [1]: impo numpy as np impo pandas as pd om sklea n.neu al_ne wo k impo MLPClassi ie om sklea n.c oss_ alida ion impo c oss_ al_sco e om sklea n.model_selec ion impo c oss_ alida e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om imblea n.o e _sampling impo SMOTE /Lib a y/F amewo ks/Py hon. amewo k/Ve sions/3.6/lib/py hon3.6/si e-packages/sklea n/c oss_ alida ion.py:41: Dep eca ionWa ning: This module was dep eca ed in e sio n 0.18 in a o o he model_selec ion module in o which all he e ac o ed classes and unc ions a e mo ed. Also no e ha he in e ace o he new CV i e a o s a e d i e en om ha o his module. This module will be emo ed in 0.20. "This module will be emo ed in 0.20.", Dep eca ionWa ning) In [2]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] 62" " X = X.d opna(axis=1) In [3]: om sklea n.model_selec ion impo G idSea chCV In [4]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [5]: cls = MLPClassi ie ( andom_s a e=6409) In [6]: laye _sizes = [(100,100),(300),(400)] lea ning_ a e_ini = [0.0001,0.001] pa am_g id = {'hidden_laye _sizes': laye _sizes, 'lea ning_ a e_ini ' : lea ning_ a e_ini } sco ing = [' ecall','accu acy', 'p ecision'] In [7]: gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing=' ecall', n_jobs=-1, c =10) In [8]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'hidden_laye _sizes': 400, 'lea ning_ a e_ini ': 0.001} In [9]: cls_bes = MLPClassi ie (hidden_laye _sizes=(400), lea ning_ a e_ini =0.001, andom _s a e=6409) In [10]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [11]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) 69" " In [11]: es ima o s = lis ( ange(700,1700,100)) pa am_g id = {'n_es ima o s': es ima o s} sco ing = [' ecall','accu acy', 'p ecision'] In [7]: gs = G idSea chCV(es ima o =cls, pa am_g id=pa am_g id, sco ing='accu acy', n_jobs=-1, c =10) In [8]: gs_ i = gs. i (X_ ain_ es, y_ ain_ es) p in ('Bes pa ame e s %s' % gs_ i .bes _pa ams_) Bes pa ame e s {'n_es ima o s': 1600} In [9]: cls_bes = AdaBoos Classi ie (n_es ima o s=1600) In [12]: sco es = c oss_ alida e(cls_bes , X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [13]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.84 +/- 0.03 A e age Recall 0.79 +/- 0.08 A e age P ecision 0.81 +/- 0.02 In [14]: model = cls_bes . i (X_ ain_ es, y_ ain_ es) In [15]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) 70" " p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.85 T aining ecall: 0.81 T aining p ecision: 0.83 Tes accu acy: 0.82 Tes ecall: 0.73 Tes p ecision: 0.77 p ecision ecall 1-sco e suppo 0 0.85 0.87 0.86 2497 1 0.77 0.73 0.75 1438 a g / o al 0.82 0.82 0.82 3935 [[2174 323] [ 386 1052]] 8.1.12. Bagging!wi h!DT!base!classi ie ! In [1]: impo numpy as np impo pandas as pd In [2]: om sklea n.ensemble impo BaggingClassi ie , RandomFo es Classi ie om sklea n.neighbo s impo KNeighbo sClassi ie om sklea n. ee impo DecisionT eeClassi ie om sklea n.neu al_ne wo k impo MLPClassi ie om sklea n.s m impo SVC om sklea n.c oss_ alida ion impo c oss_ al_sco e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om sklea n.model_selec ion impo c oss_ alida e /Lib a y/F amewo ks/Py hon. amewo k/Ve sions/3.6/lib/py hon3.6/si e-packages/sklea n/c oss_ alida ion.py:41: Dep eca ionWa ning: This module was dep eca ed in e sio n 0.18 in a o o he model_selec ion module in o which all he e ac o ed classes and unc ions a e mo ed. Also no e ha he in e ace o he new CV i e a o s a e d i e en om ha o his module. This module will be emo ed in 0.20. "This module will be emo ed in 0.20.", Dep eca ionWa ning) In [3]: 71" " om imblea n.o e _sampling impo SMOTE In [4]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [5]: om sklea n.model_selec ion impo G idSea chCV In [6]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [7]: cls = BaggingClassi ie () sco ing = [' ecall','accu acy', 'p ecision'] In [8]: sco es = c oss_ alida e(cls, X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [9]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.88 +/- 0.02 A e age Recall 0.84 +/- 0.05 A e age P ecision 0.87 +/- 0.01 In [10]: model = cls. i (X_ ain_ es, y_ ain_ es) In [11]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) 72" " p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.95 T aining ecall: 0.94 T aining p ecision: 0.94 Tes accu acy: 0.87 Tes ecall: 0.82 Tes p ecision: 0.83 p ecision ecall 1-sco e suppo 0 0.89 0.90 0.90 2497 1 0.83 0.82 0.82 1438 a g / o al 0.87 0.87 0.87 3935 [[2254 243] [ 265 1173]] 8.1.13. Bagging!wi h!MLP!base!classi ie ! In [1]: impo numpy as np impo pandas as pd In [9]: om sklea n.ensemble impo BaggingClassi ie , RandomFo es Classi ie om sklea n.neighbo s impo KNeighbo sClassi ie om sklea n. ee impo DecisionT eeClassi ie om sklea n.neu al_ne wo k impo MLPClassi ie om sklea n.s m impo SVC om sklea n.c oss_ alida ion impo c oss_ al_sco e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om sklea n.model_selec ion impo c oss_ alida e In [6]: om imblea n.o e _sampling impo SMOTE In [3]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) 73" " X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [4]: om sklea n.model_selec ion impo G idSea chCV In [7]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [20]: cls = BaggingClassi ie (base_es ima o =(MLPClassi ie (hidden_laye _sizes=(400)))) sco ing = [' ecall','accu acy', 'p ecision'] In [21]: sco es = c oss_ alida e(cls, X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [22]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.89 +/- 0.02 A e age Recall 0.89 +/- 0.04 A e age P ecision 0.86 +/- 0.01 In [23]: model = cls. i (X_ ain_ es, y_ ain_ es) In [24]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) 74" " p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.94 T aining ecall: 0.93 T aining p ecision: 0.92 Tes accu acy: 0.88 Tes ecall: 0.83 Tes p ecision: 0.84 p ecision ecall 1-sco e suppo 0 0.90 0.91 0.91 2497 1 0.84 0.83 0.83 1438 a g / o al 0.88 0.88 0.88 3935 [[2269 228] [ 247 1191]] 8.1.14. Vo ing!! In [4]: impo numpy as np impo pandas as pd om sklea n.ensemble impo AdaBoos Classi ie om sklea n.c oss_ alida ion impo c oss_ al_sco e om sklea n.model_selec ion impo c oss_ alida e om sklea n.c oss_ alida ion impo ain_ es _spli om sklea n.me ics impo accu acy_sco e, ecall_sco e, p ecision_sco e om sklea n.me ics impo classi ica ion_ epo ,con usion_ma ix om imblea n.o e _sampling impo SMOTE om sklea n.neu al_ne wo k impo MLPClassi ie om sklea n.ensemble impo RandomFo es Classi ie om sklea n.neighbo s impo KNeighbo sClassi ie om sklea n. ee impo DecisionT eeClassi ie om sklea n.neu al_ne wo k impo MLPClassi ie om sklea n.s m impo SVC om sklea n.linea _model impo Logis icReg ession om sklea n.ensemble impo Vo ingClassi ie om sklea n.linea _model impo Logis icReg ession om sklea n.nai e_bayes impo GaussianNB In [5]: d = pd. ead_cs ('Da ase _EGFR.cs ', index_col = False) X = d .d op(['PUBCHEM_ACTIVITY_OUTCOME'],axis=1) 75" " y = d ['PUBCHEM_ACTIVITY_OUTCOME'] X = X.d opna(axis=1) In [6]: om sklea n.model_selec ion impo G idSea chCV In [7]: X_ ain, X_ es , y_ ain, y_ es = ain_ es _spli (X, y, es _size=0.30, andom_s a e=42) sm = SMOTE( andom_s a e=12, a io = 0.7) X_ ain_ es, y_ ain_ es = sm. i _sample(X_ ain, y_ ain) In [11]: sco ing = [' ecall','accu acy', 'p ecision'] In [9]: cl 1 = Logis icReg ession( andom_s a e=1, C=10) cl 2 = RandomFo es Classi ie ( andom_s a e=1, n_es ima o s=25) cl 3 = GaussianNB() cl 4 = KNeighbo sClassi ie (n_neighbo s=5) cl 5 = SVC() cl 6 = MLPClassi ie (hidden_laye _sizes=(400)) cls = Vo ingClassi ie (es ima o s=[('l ', cl 1), (' ', cl 2), ('gnb', cl 3), ('knn ', cl 4), ('s m', cl 5), ('mlp', cl 6)], o ing='ha d') In [12]: sco es = c oss_ alida e(cls, X_ ain_ es, y_ ain_ es, sco ing=sco ing, c =10) In [13]: p in (' nA e age Accu acy %.2 +/- %.2 ' % (np.mean(sco es[' es _accu acy']), np.s d(sco es[' es _accu acy']))) p in (' nA e age Recall %.2 +/- %.2 ' % (np.mean(sco es[' es _ ecall']), np.s d(sc o es[' es _ ecall']))) p in (' nA e age P ecision %.2 +/- %.2 ' % (np.mean(sco es[' es _p ecision']), np. s d(sco es[' es _p ecision']))) A e age Accu acy 0.89 +/- 0.02 A e age Recall 0.87 +/- 0.04 A e age P ecision 0.87 +/- 0.01 In [15]: model = cls. i (X_ ain_ es, y_ ain_ es) In [16]: ain_acc = accu acy_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_ ec = ecall_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) ain_p ec = p ecision_sco e(y_ ue=y_ ain_ es, y_p ed=model.p edic (X_ ain_ es)) 76" " es _acc = accu acy_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _ ec = ecall_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) es _p ec = p ecision_sco e(y_ ue=y_ es , y_p ed=model.p edic (X_ es )) p edic ions = model.p edic (X_ es ) p in ('T aining accu acy: %.2 ' % ain_acc) p in ('T aining ecall: %.2 ' % ain_ ec) p in ('T aining p ecision: %.2 ' % ain_p ec) p in ('Tes accu acy: %.2 ' % es _acc) p in ('Tes ecall: %.2 ' % es _ ec) p in ('Tes p ecision: %.2 ' % es _p ec) p in (classi ica ion_ epo (y_ es ,p edic ions)) p in (con usion_ma ix(y_ es ,p edic ions)) T aining accu acy: 0.93 T aining ecall: 0.91 T aining p ecision: 0.91 Tes accu acy: 0.88 Tes ecall: 0.83 Tes p ecision: 0.84 p ecision ecall 1-sco e suppo 0 0.90 0.91 0.90 2497 1 0.84 0.83 0.83 1438 a g / o al 0.88 0.88 0.88 3935 [[2263 234] [ 244 1194]] 8.2. !PUBCHEM!FINGERPRINT! PubChem Subs uc u e Finge p in V1.3 h p://pubchem.ncbi.nlm.nih.go The PubChem Sys em gene a es a bina y subs uc u e inge p in o chemical s uc u es. These inge p in s a e used by PubChem o simila i y neighbo ing and simila i y sea ching. A subs uc u e is a agmen o a chemical s uc u e. A inge p in is an o de ed lis o bina y (1/0) bi s. Each bi ep esen s a Boolean de e mina ion o , o es o , he p esence o , o example, an elemen coun , a ype o ing sys em, a om pai ing, a om en i onmen (nea es neighbo s), e c., in a chemical s uc u e. The na i e o ma o he PubChem Subs uc u e Finge p in p ope y is bina y da a wi h a ou by e in ege p e ix, whe e his in ege p e ix indica es he leng h o he bi lis . Fo he ASN.1 and XML o ma ed da a, his p ope y is s o ed in a PC-In oDa a con aine , as desc ibed by he PCSubs ance ASN.1 de ini ion o XML schema: p:// p.ncbi.nlm.nih.go /pubchem/speci ica ions/ PC-In oDa a is able o handle a ious ypes o da a. Each PC- In oDa a has a PC-U n objec (u n = uni e sal esou ce name). Each p ope y has a unique io o "label", "name", and "da a ype" de ini ion (e.g., o PubChem Subs uc u e Finge p in , his is "Finge p in ", "SubS uc u e Keys", and " inge p in ", espec i ely). The inge p in bina y da a is hex-encoded, when 77" " p o ided in he XML o ex ual ASN.1 o ma s. When expo ing inge p in in o ma ion in he SD ile o ma , he SD ag o he PubChem Subs uc u e Finge p in p ope y is "PUBCHEM_CACTVS_SUBGRAPHKEYS". The PubChem Subs uc u e Finge p in is Base64 encoded o p o ide a ex ual ep esen a ion o he bina y da a. Fo a desc ip ion o he Base64 encoding and decoding algo i hm speci ica ion, go o: h p://www. aqs.o g/ cs/ c3548.h ml Below is he desc ip ion o each bi ep esen ed in he PubChem Subs uc u e Finge p in . Some inge p in bi desc ip ions a e w i en in SMILES o SMARTS no a ion. Fo addi ional in o ma ion on SMARTS and SMILES, please go o: h p://en.wikipedia.o g/wiki/Simpli ied_molecula _inpu _line_en y_ speci ica ion PubChem Subs uc u e Finge p in Desc ip ion Sec ion 1: Hie a chic Elemen Coun s - These bi s es o he p esence o coun o indi idual chemical a oms ep esen ed by hei a omic symbol. Bi Posi ion Bi Subs uc u e 0 >= 4 H 1 >= 8 H 2 >= 16 H 3 >= 32 H 4 >= 1 Li 5 >= 2 Li 6 >= 1 B 7 >= 2 B 8 >= 4 B 9 >= 2 C 10 >= 4 C 11 >= 8 C 12 >= 16 C 13 >= 32 C 14 >= 1 N 15 >= 2 N 16 >= 4 N 17 >= 8 N 18 >= 1 O 19 >= 2 O 20 >= 4 O 21 >= 8 O 22 >= 16 O 23 >= 1 F 24 >= 2 F 25 >= 4 F 26 >= 1 Na 27 >= 2 Na 28 >= 1 Si 29 >= 2 Si 30 >= 1 P 31 >= 2 P 32 >= 4 P 33 >= 1 S 34 >= 2 S 35 >= 4 S 36 >= 8 S 37 >= 1 Cl 38 >= 2 Cl 39 >= 4 Cl 40 >= 8 Cl 41 >= 1 K 42 >= 2 K 78" " 43 >= 1 B 44 >= 2 B 45 >= 4 B 46 >= 1 I 47 >= 2 I 48 >= 4 I 49 >= 1 Be 50 >= 1 Mg 51 >= 1 Al 52 >= 1 Ca 53 >= 1 Sc 54 >= 1 Ti 55 >= 1 V 56 >= 1 C 57 >= 1 Mn 58 >= 1 Fe 59 >= 1 Co 60 >= 1 Ni 61 >= 1 Cu 62 >= 1 Zn 63 >= 1 Ga 64 >= 1 Ge 65 >= 1 As 66 >= 1 Se 67 >= 1 K 68 >= 1 Rb 69 >= 1 S 70 >= 1 Y 71 >= 1 Z 72 >= 1 Nb 73 >= 1 Mo 74 >= 1 Ru 75 >= 1 Rh 76 >= 1 Pd 77 >= 1 Ag 78 >= 1 Cd 79 >= 1 In 80 >= 1 Sn 81 >= 1 Sb 82 >= 1 Te 83 >= 1 Xe 84 >= 1 Cs 85 >= 1 Ba 86 >= 1 Lu 87 >= 1 H 88 >= 1 Ta 89 >= 1 W 90 >= 1 Re 91 >= 1 Os 92 >= 1 I 93 >= 1 P 94 >= 1 Au 95 >= 1 Hg 96 >= 1 Tl 97 >= 1 Pb 98 >= 1 Bi 99 >= 1 La 100 >= 1 Ce 101 >= 1 P 102 >= 1 Nd 103 >= 1 Pm 104 >= 1 Sm 105 >= 1 Eu 106 >= 1 Gd 107 >= 1 Tb 108 >= 1 Dy 109 >= 1 Ho 110 >= 1 E 85" " 461 O-C-C=N 462 O-C-C=O 463 N:C-S-[#1] 464 N-C-C=C 465 O=S-C-C 466 N#C-C=C 467 C=N-N-C 468 O=S-C-N 469 S-S-C:C 470 C:C-C=C 471 S:C:C:C 472 C:N:C-C 473 S-C:N:C 474 S:C:C:N 475 S-C=N-C 476 C-O-C=C 477 N-N-C:C 478 S-C=N-[#1] 479 S-C-S-C 480 C:S:C-C 481 O-S-C:C 482 C:N-C:C 483 N-S-C:C 484 N-C:N:C 485 N:C:C:N 486 N-C:N:N 487 N-C=N-C 488 N-C=N-[#1] 489 N-C-S-C 490 C-C-C=C 491 C-N:C-[#1] 492 N-C:O:C 493 O=C-C:C 494 O=C-C:N 495 C-N-C:C 496 N:N-C-[#1] 497 O-C:C:N 498 O-C=C-C 499 N-C:C:N 500 C-S-C:C 501 Cl-C:C-C 502 N-C=C-[#1] 503 Cl-C:C-[#1] 504 N:C:N-C 505 Cl-C:C-O 506 C-C:N:C 507 C-C-S-C 508 S=C-N-C 509 B -C:C-C 510 [#1]-N-N-[#1] 511 S=C-N-[#1] 512 C-[As]-O-[#1] 513 S:C:C-[#1] 514 O-N-C-C 515 N-N-C-C 516 [#1]-C=C-[#1] 517 N-N-C-N 518 O=C-N-N 519 N=C-N-C 520 C=C-C:C 521 C:N-C-[#1] 522 C-N-N-[#1] 523 N:C:C-C 524 C-C=C-C 525 [As]-C:C-[#1] 526 Cl-C:C-Cl 527 C:C:N-[#1] 528 [#1]-N-C-[#1] 86" " 529 Cl-C-C-Cl 530 N:C-C:C 531 S-C:C-C 532 S-C:C-[#1] 533 S-C:C-N 534 S-C:C-O 535 O=C-C-C 536 O=C-C-N 537 O=C-C-O 538 N=C-C-C 539 N=C-C-[#1] 540 C-N-C-[#1] 541 O-C:C-C 542 O-C:C-[#1] 543 O-C:C-N 544 O-C:C-O 545 N-C:C-C 546 N-C:C-[#1] 547 N-C:C-N 548 O-C-C:C 549 N-C-C:C 550 Cl-C-C-C 551 Cl-C-C-O 552 C:C-C:C 553 O=C-C=C 554 B -C-C-C 555 N=C-C=C 556 C=C-C-C 557 N:C-O-[#1] 558 O=N-C:C 559 O-C-N-[#1] 560 N-C-N-C 561 Cl-C-C=O 562 B -C-C=O 563 O-C-O-C 564 C=C-C=C 565 C:C-O-C 566 O-C-C-N 567 O-C-C-O 568 N#C-C-C 569 N-C-C-N 570 C:C-C-C 571 [#1]-C-O-[#1] 572 N:C:N:C 573 O-C-C=C 574 O-C-C:C-C 575 O-C-C:C-O 576 N=C-C:C-[#1] 577 C:C-N-C:C 578 C-C:C-C:C 579 O=C-C-C-C 580 O=C-C-C-N 581 O=C-C-C-O 582 C-C-C-C-C 583 Cl-C:C-O-C 584 C:C-C=C-C 585 C-C:C-N-C 586 C-S-C-C-C 587 N-C:C-O-[#1] 588 O=C-C-C=O 589 C-C:C-O-C 590 C-C:C-O-[#1] 591 Cl-C-C-C-C 592 N-C-C-C-C 593 N-C-C-C-N 594 C-O-C-C=C 595 C:C-C-C-C 596 N=C-N-C-C 87" " 597 O=C-C-C:C 598 Cl-C:C:C-C 599 [#1]-C-C=C-[#1] 600 N-C:C:C-C 601 N-C:C:C-N 602 O=C-C-N-C 603 C-C:C:C-C 604 C-O-C-C:C 605 O=C-C-O-C 606 O-C:C-C-C 607 N-C-C-C:C 608 C-C-C-C:C 609 Cl-C-C-N-C 610 C-O-C-O-C 611 N-C-C-N-C 612 N-C-O-C-C 613 C-N-C-C-C 614 C-C-O-C-C 615 N-C-C-O-C 616 C:C:N:N:C 617 C-C-C-O-[#1] 618 C:C-C-C:C 619 O-C-C=C-C 620 C:C-O-C-C 621 N-C:C:C:N 622 O=C-O-C:C 623 O=C-C:C-C 624 O=C-C:C-N 625 O=C-C:C-O 626 C-O-C:C-C 627 O=[As]-C:C:C 628 C-N-C-C:C 629 S-C:C:C-N 630 O-C:C-O-C 631 O-C:C-O-[#1] 632 C-C-O-C:C 633 N-C-C:C-C 634 C-C-C:C-C 635 N-N-C-N-[#1] 636 C-N-C-N-C 637 O-C-C-C-C 638 O-C-C-C-N 639 O-C-C-C-O 640 C=C-C-C-C 641 O-C-C-C=C 642 O-C-C-C=O 643 [#1]-C-C-N-[#1] 644 C-C=N-N-C 645 O=C-N-C-C 646 O=C-N-C-[#1] 647 O=C-N-C-N 648 O=N-C:C-N 649 O=N-C:C-O 650 O=C-N-C=O 651 O-C:C:C-C 652 O-C:C:C-N 653 O-C:C:C-O 654 N-C-N-C-C 655 O-C-C-C:C 656 C-C-N-C-C 657 C-N-C:C-C 658 C-C-S-C-C 659 O-C-C-N-C 660 C-C=C-C-C 661 O-C-O-C-C 662 O-C-C-O-C 663 O-C-C-O-[#1] 664 C-C=C-C=C 88" " 665 N-C:C-C-C 666 C=C-C-O-C 667 C=C-C-O-[#1] 668 C-C:C-C-C 669 Cl-C:C-C=O 670 B -C:C:C-C 671 O=C-C=C-C 672 O=C-C=C-[#1] 673 O=C-C=C-N 674 N-C-N-C:C 675 B -C-C-C:C 676 N#C-C-C-C 677 C-C=C-C:C 678 C-C-C=C-C 679 C-C-C-C-C-C 680 O-C-C-C-C-C 681 O-C-C-C-C-O 682 O-C-C-C-C-N 683 N-C-C-C-C-C 684 O=C-C-C-C-C 685 O=C-C-C-C-N 686 O=C-C-C-C-O 687 O=C-C-C-C=O 688 C-C-C-C-C-C-C 689 O-C-C-C-C-C-C 690 O-C-C-C-C-C-O 691 O-C-C-C-C-C-N 692 O=C-C-C-C-C-C 693 O=C-C-C-C-C-O 694 O=C-C-C-C-C=O 695 O=C-C-C-C-C-N 696 C-C-C-C-C-C-C-C 697 C-C-C-C-C-C(C)-C 698 O-C-C-C-C-C-C-C 699 O-C-C-C-C-C(C)-C 700 O-C-C-C-C-C-O-C 701 O-C-C-C-C-C(O)-C 702 O-C-C-C-C-C-N-C 703 O-C-C-C-C-C(N)-C 704 O=C-C-C-C-C-C-C 705 O=C-C-C-C-C(O)-C 706 O=C-C-C-C-C(=O)-C 707 O=C-C-C-C-C(N)-C 708 C-C(C)-C-C 709 C-C(C)-C-C-C 710 C-C-C(C)-C-C 711 C-C(C)(C)-C-C 712 C-C(C)-C(C)-C Sec ion 7: Complex SMARTS pa e ns - These bi s es o he p esence o complex SMARTS pa e ns, ega dless o coun , bu whe e bond o de s and bond a oma ici y a e speci ic. Bi Posi ion Bi Subs uc u e 713 Cc1ccc(C)cc1 714 Cc1ccc(O)cc1 715 Cc1ccc(S)cc1 716 Cc1ccc(N)cc1 717 Cc1ccc(Cl)cc1 718 Cc1ccc(B )cc1 719 Oc1ccc(O)cc1 720 Oc1ccc(S)cc1 721 Oc1ccc(N)cc1 722 Oc1ccc(Cl)cc1 723 Oc1ccc(B )cc1 724 Sc1ccc(S)cc1 89" " 725 Sc1ccc(N)cc1 726 Sc1ccc(Cl)cc1 727 Sc1ccc(B )cc1 728 Nc1ccc(N)cc1 729 Nc1ccc(Cl)cc1 730 Nc1ccc(B )cc1 731 Clc1ccc(Cl)cc1 732 Clc1ccc(B )cc1 733 B c1ccc(B )cc1 734 Cc1cc(C)ccc1 735 Cc1cc(O)ccc1 736 Cc1cc(S)ccc1 737 Cc1cc(N)ccc1 738 Cc1cc(Cl)ccc1 739 Cc1cc(B )ccc1 740 Oc1cc(O)ccc1 741 Oc1cc(S)ccc1 742 Oc1cc(N)ccc1 743 Oc1cc(Cl)ccc1 744 Oc1cc(B )ccc1 745 Sc1cc(S)ccc1 746 Sc1cc(N)ccc1 747 Sc1cc(Cl)ccc1 748 Sc1cc(B )ccc1 749 Nc1cc(N)ccc1 750 Nc1cc(Cl)ccc1 751 Nc1cc(B )ccc1 752 Clc1cc(Cl)ccc1 753 Clc1cc(B )ccc1 754 B c1cc(B )ccc1 755 Cc1c(C)cccc1 756 Cc1c(O)cccc1 757 Cc1c(S)cccc1 758 Cc1c(N)cccc1 759 Cc1c(Cl)cccc1 760 Cc1c(B )cccc1 761 Oc1c(O)cccc1 762 Oc1c(S)cccc1 763 Oc1c(N)cccc1 764 Oc1c(Cl)cccc1 765 Oc1c(B )cccc1 766 Sc1c(S)cccc1 767 Sc1c(N)cccc1 768 Sc1c(Cl)cccc1 769 Sc1c(B )cccc1 770 Nc1c(N)cccc1 771 Nc1c(Cl)cccc1 772 Nc1c(B )cccc1 773 Clc1c(Cl)cccc1 774 Clc1c(B )cccc1 775 B c1c(B )cccc1 776 CC1CCC(C)CC1 777 CC1CCC(O)CC1 778 CC1CCC(S)CC1 779 CC1CCC(N)CC1 780 CC1CCC(Cl)CC1 781 CC1CCC(B )CC1 782 OC1CCC(O)CC1 783 OC1CCC(S)CC1 784 OC1CCC(N)CC1 785 OC1CCC(Cl)CC1 786 OC1CCC(B )CC1 787 SC1CCC(S)CC1 788 SC1CCC(N)CC1 789 SC1CCC(Cl)CC1 790 SC1CCC(B )CC1 791 NC1CCC(N)CC1 792 NC1CCC(Cl)CC1 90" " 793 NC1CCC(B )CC1 794 ClC1CCC(Cl)CC1 795 ClC1CCC(B )CC1 796 B C1CCC(B )CC1 797 CC1CC(C)CCC1 798 CC1CC(O)CCC1 799 CC1CC(S)CCC1 800 CC1CC(N)CCC1 801 CC1CC(Cl)CCC1 802 CC1CC(B )CCC1 803 OC1CC(O)CCC1 804 OC1CC(S)CCC1 805 OC1CC(N)CCC1 806 OC1CC(Cl)CCC1 807 OC1CC(B )CCC1 808 SC1CC(S)CCC1 809 SC1CC(N)CCC1 810 SC1CC(Cl)CCC1 811 SC1CC(B )CCC1 812 NC1CC(N)CCC1 813 NC1CC(Cl)CCC1 814 NC1CC(B )CCC1 815 ClC1CC(Cl)CCC1 816 ClC1CC(B )CCC1 817 B C1CC(B )CCC1 818 CC1C(C)CCCC1 819 CC1C(O)CCCC1 820 CC1C(S)CCCC1 821 CC1C(N)CCCC1 822 CC1C(Cl)CCCC1 823 CC1C(B )CCCC1 824 OC1C(O)CCCC1 825 OC1C(S)CCCC1 826 OC1C(N)CCCC1 827 OC1C(Cl)CCCC1 828 OC1C(B )CCCC1 829 SC1C(S)CCCC1 830 SC1C(N)CCCC1 831 SC1C(Cl)CCCC1 832 SC1C(B )CCCC1 833 NC1C(N)CCCC1 834 NC1C(Cl)CCCC1 835 NC1C(B )CCCC1 836 ClC1C(Cl)CCCC1 837 ClC1C(B )CCCC1 838 B C1C(B )CCCC1 839 CC1CC(C)CC1 840 CC1CC(O)CC1 841 CC1CC(S)CC1 842 CC1CC(N)CC1 843 CC1CC(Cl)CC1 844 CC1CC(B )CC1 845 OC1CC(O)CC1 846 OC1CC(S)CC1 847 OC1CC(N)CC1 848 OC1CC(Cl)CC1 849 OC1CC(B )CC1 850 SC1CC(S)CC1 851 SC1CC(N)CC1 852 SC1CC(Cl)CC1 853 SC1CC(B )CC1 854 NC1CC(N)CC1 855 NC1CC(Cl)CC1 856 NC1CC(B )CC1 857 ClC1CC(Cl)CC1 858 ClC1CC(B )CC1 859 B C1CC(B )CC1 860 CC1C(C)CCC1 91" " 861 CC1C(O)CCC1 862 CC1C(S)CCC1 863 CC1C(N)CCC1 864 CC1C(Cl)CCC1 865 CC1C(B )CCC1 866 OC1C(O)CCC1 867 OC1C(S)CCC1 868 OC1C(N)CCC1 869 OC1C(Cl)CCC1 870 OC1C(B )CCC1 871 SC1C(S)CCC1 872 SC1C(N)CCC1 873 SC1C(Cl)CCC1 874 SC1C(B )CCC1 875 NC1C(N)CCC1 876 NC1C(Cl)CC1 877 NC1C(B )CCC1 878 ClC1C(Cl)CCC1 879 ClC1C(B )CCC1 880 B C1C(B )CCC1 Decoding PubChem Finge p in s PubChem inge p in s a e cu en ly 881 bi s in leng h. Bina y da a is s o ed in one by e inc emen s. The inge p in is, he e o e, 111 by es in leng h (888 bi s), which includes padding o se en bi s a he end o comple e he las by e. A ou - by e p e ix, con aining he bi leng h o he inge p in (881 bi s), inc eases he s o ed PubChem inge p in size o 115 by es (920 bi s). When PubChem inge p in s a e encoded in base64 o ma , he base64-encoded inge p in s a e 156 by es in leng h. The las wo by es a e padding so ha he base64 leng h is di isible by ou (156 by es - 2 by es = 154 by es). Each base64 by e encodes six bina y bi s (154 by es * 6 bi s/by e = 924 bi s). The las ou bi s a e padding o comple e he las base64 by e (924 bi s - 4 bi s = 920 bi s). The esul ing 920 bina y bi s (115 by es) a e desc ibed in he p e ious pa ag aph. Documen Ve sion His o y V1.3 - 2009May01 - Upda ed in oduc ion o desc ibe how o iden i y he PubChem Subs uc u e Finge p in p ope y in a PubChem Compound eco d. V1.2 - 2007Aug30 - Added sec ion on decoding PubChem inge p in s. V1.1 - 2007Aug06 - Co ec ed and expanded documen a ion o bi s wi h SMARTS pa e ns used. V1.0 - 2005Dec02 - Ini ial elease. " " " " "