Mirnovo: genome-free prediction of micrornas from small RNA sequencing data and single-cells using decision forests

Transcription

1 Nlei ids Reserh, 17 1 doi: 1.193/nr/kx836 Mirnovo: enome-free predition of mirorns from smll RN seqenin dt nd sinle-ells sin deision forests Dimitrios M. Vitsios 1, Elissvet Kentepozido 1, Leonor Qintis 1, Eli Benito-tiérrez 2, Stijn vn Donen 1, Mtthew P. Dvis 1,* nd nton J. Enriht 1,* 1 Eropen Molelr Bioloy Lbortory Eropen Bioinformtis Institte, Wellome enome mps, Hinxton, mbride B1 1SD, K nd 2 Deprtment of Zooloy, niversity of mbride, Downin Street, mbride B2 3EJ, K Reeived Jly 27, 17; Revised September 6, 17; Editoril Deision September 1, 17; epted September, 17 BSTRT The disovery of mirorns (mirns) remins n importnt problem, prtilrly iven the rowth of hih-throhpt seqenin, ell sortin nd sinle ell bioloy. While lre nmber of mirns hve lredy been nnotted, there my well be lre nmbers of mirns tht re expressed in very prtilr ell types nd remin elsive. Seqenin llows s to qikly nd rtely identify the expression of known mirns from smll RN-Seq dt. The bioenesis of mirns leds to very speifi hrteristis observed in their seqenes. In brief, mirns slly hve well-defined 5 end nd more flexible 3 end with the possibility of 3 tilin events, sh s ridyltion. Previos pprohes to the predition of novel mirns slly involve the nlysis of strtrl fetres of mirn prersor hirpin seqenes obtined from enome seqene. We srmised tht it my be possible to identify mirns by sin these bioenesis fetres observed diretly from seqened reds, solely or in ddition to strtrl nlysis from enome dt. To this end, we hve developed mirnovo, mhine lernin bsed lorithm, whih is ble to identify known nd novel mirns in nimls nd plnts diretly from smll RN-Seq dt, with or withot referene enome. This method performs omprbly to existin tools, however is simpler to se with reded rn time. Its performne nd ry hs been tested on mltiple dtsets, inldin speies with poorly ssembled enomes, RNseIII (Drosh nd/or Dier) defiient smples nd sinle ells (t both embryoni nd dlt ste). INTRODTION The identifition nd nnottion of novel mirns from vrios speies, either nimls or plnts, hs been hllene in the field of smll non-odin RNs for mny yers. Trditionlly, novel mirn predition ws bsed on the identifition of short seqenes, mppin sh seqenes to the enome, nd serhin for those loi tht my prode the hrteristi hirpin strtre of premirn vi nlysis of derived strtrl fetres. However, we soht to explore the possibility of preditin novel mirns with hih ry withot reqirin referene enome in the proess. Or initil hypothesis is tht fetres of mirorn (mirn) seqenes, derived from their bioenesis my be sffiient to predit mirns de novo, i.e. withot sin referene enome. These bioenesis fetres (Fire 1) re lerly evident when one interrotes lre nmbers of mirn seqenin dtsets from mltiple speies. In order to perform enome-free fetre nlysis of mirn seqenes, one needs to tke n inpt set of smll RN seqenes nd loblly rop them into lsters of relted seqene. These lsters my then be mltiply lined nd filtered. This linment llows onsenss seqene to be onstrted nd bioenesis fetres to be ssessed. The dvntes of de novo disovery of mirns prely from seqenin dt re redily pprent: (i) it does not reqire referene enome, (ii) removin the enomi mppin nd RN seondry strtrl nlysis llows for fster ompttion nd (iii) it will prode smller set of novel ndidte seqenes, shold one wnt to do enomi fetre nlysis lter. To this end, we hve developed new method, mirnovo, whih llows for predition of novel mirns in nimls nd plnts, with or withot referene enome. * To whom orrespondene shold be ddressed. Tel: ; Fx: ; Emil: je@ebi..k orrespondene my lso be ddressed to Mtthew P. Dvis. Tel: ; Fx: ; Emil: mtdvis@ebi..k The thor(s) 17. Pblished by Oxford niversity Press on behlf of Nlei ids Reserh. This is n Open ess rtile distribted nder the terms of the retive ommons ttribtion Liense ( whih permits non-ommeril re-se, distribtion, nd reprodtion in ny medim, provided the oriinl work is properly ited. For ommeril re-se, plese ontt jornls.permissions@op.om Downloded from by Wellome Trst enome mps ser on 16 Otober 17

2 2 Nlei ids Reserh, 17 MTERILS ND METHODS Inpt dt Mirnovo epts s n inpt one zipped (.z) FSTQ or FST file for eh rn from either blk or sinle-ell smll RN-Seqenin dt. Inpt seqenes my hve lredy been pre-lened from their 3 dpters otherwise 3 dpter seqene needs to be provided by the ser. Mirnovo pipeline The 3 dpter from inpt dt is removed with reper (27) nd then lened seqenes re de-dplited with tlly (27). Initil lsterin of tllied seqenes is performed with vserh (28) sin n linment identity threshold of.9 by deflt. lsters refinement is hieved by merin similr onsenss seqenes with d-hit (2), bsed on 7-mer serhes nd sin.85 s the linment identity threshold. Mltiple-seqene linment for the refined lsters is performed with msle (3). Followin the mirn predition step, the onsenss seqenes of ll identified known nd/or novel mirns re mpped inst the referene enome (if pplible) sin bowtie2 (29) (seleted prmeters: -k 1, -D, -R 3, -N 1, -L, -i S,1,.5 rd 1,1 rf 1,1). The most stble hirpins, in terms of free enery, ssessed within 9nt window rond these seqenes, re seleted nd enomi fetres re llted for eh hirpin ndidte. Eventlly, p to 5 hirpins re reported s prlo prersors for eh mtre mirn in se the llted free eneries of these seondry strtres re below n empirilly defined threshold. Fetres definition The fll set of fetres sed for lssifition nd predition is desribed s follows: Twelve overe profile fetres: lster red depth, min body lenth, mismthes in min body, slin rte before 5, slin rte fter 3, ps before min body, mismthes in seed reion, vere ontent in min body, ps fter min body, vere ontent fter 3, linment identity inst the potentil reverse omplement nd vere seqene lenth. Twelve seqene omplexity fetres: +T skew (ts), + skew (s), p skew (p), omplexity by Wootton nd Federhen (3) (wf), entropy (e), omplexity s ompression rtio sin zip (z), omplexity s Mrkov model size of N {2,3} (m2, m3), Trifnov s omplexity (31) withordern {2,3} (t2, t3) nd linisti omplexity with order N {2,3} (l2, l3). Nine enomi fetres (hirpin foldin retrieved sin RNfold from the Viennpke (32)): hirpin size estimte, mtre mirn distne from stem loop, loop size estimte, nmber of loops in hirpin, minimm free enery of seondry strtre, mjority brkets in the entire foldin (prevlent of the two distinishin brket diretions, i.e. mx{nm[ (, ) ]}), mirn brket disrepny (K/N, where N is the totl nmber of brkets in the mirn nd K is the nmber of mjority brkets), mirn brket frtion (K/N,whereN is the mirn lenth nd K is the nmber of mjority brkets) nd nmber of nmthed nleotides from the mtre mirn seqene. Otpt The reslts from eh job ontin first of ll FST files for the predited known nd novel mirns (both for the mtre prodts nd their respetive hirpin prersors), nd for ny trn nd/or rrn identified hits. dditionlly, BED files with enomi oordintes of predited hirpins re provided lon with overe profiles for eh mtre mirn nd lso the seondry strtres of eh identified hirpin prlo. Frthermore, eh job is ssoited with tble of performne mesres with rerds to the mhine lernin preditions. The reported mesres re: preision = TP TP ; sensitivity = TP+ FP P ; speifiity = TN N, where TP: is the nmber of predited known mirns, FP: is the nmber of predited novel mirns, P: is the nmber of ll known mirns ontined in the inpt dt (bsed on the mirbse nnottion), TN: is the nmber of (orretly) predited non-mirn seqenes, N: is the nmber of ll non-mirn seqenes ontined in the inpt dt (bsed on the mirbse nnottion). Preditions re ompnied with ROR nd Preision-Rell (PR) rve, whih demonstrte the performne of the mhine lernin method with rerds to orretly identifyin known nd novel mirns, respetively. Finlly, the distribtions of ll fetre vles (overe, seqene omplexity nd enomi) for eh lss of predited mirn/non-mirn seqenes re vislised nd mde vilble s post-predition Q box-plots. Mhine lernin model trinin The ore mhine lernin lorithm sed for trinin ws bsed on Rndom Forests. The Rndom Forest implementtion ws provided by the rndomforest R pke (SVM nd rdient Boostin methods were lso tested sin the e171 nd bm R pkes, respetively). In order to finetne or model we tested its performne independently for vrios nmbers of rndomly seleted preditors (mtry) nd nmbers of trees (ntrees) on 65 mose smples downloded from EN (31). Optiml performne ws obtined for mtry = 6ndntrees = nd ths these prmeters were seleted for the trinin of eh lssifier (Spplementry Fire S2). Smples sed for trinin of the niml nd plnt speies models were lso downloded from EN (33). Spported speies Mirnovo n nlyse dtsets from ny speies, withot reqirin enome referene or mirbse nnotted mir- Ns. The option Not vilble shold be sed in this se in the ple of the Inpt speies. However, even hiher ry n be hieved by intertin the enomi fetres into predition. Ths, mirnovo hs interted enomi spport for 67 speies. This mens tht for those Downloded from by Wellome Trst enome mps ser on 16 Otober 17

3 Nlei ids Reserh, 17 3 B Fire 1. Mirnovo bioenesis fetres nd performne ross mltiple speies. () overe fetres definition for eh lster of similr seqenes. (B) Trinin model performnes with 1-fold ross-vlidtion ross 7 Plnt nd 8 niml Speies, with or withot sin referene enome. speies, the fll set of overe profile, seqene omplexity nd enomi fetres n be ompiled in order to identify known nd predit novel mirns. dditionlly, mirnovo spports mirn identifition nd predition for nother 1 speies with mirbse nnotted mirns, bt lkin enomi fetre spport. The ommnd-line version of or method thoh llows the ser to bild nd interte into the identifition proess ny stom referene enome. Trinin models per speies We hve trined individl models for 8 niml speies (pis mellifer, Bos trs, enorhbditis elens, nis fmiliris, Dnio rerio, Drosophil melnoster, Homo spiens, Ms msls) nd seven plnt speies (rbidopsis Thlin, hlmydomons reinhrdtii, lyine mx, Hordem vlre, Medio trntl, Oryz stiv jponi, Solnm lyopersim). These models offer optimised reslts when inpt files oriinte from one of those speies. dditionlly, we hve reted two niversl models, for nimls nd plnts respetively, tht n be sed enerilly for ny speies belonin to one of the two kindoms. These models hve been reted by smplin dt-points (refined seqene lsters) from the entire dtset of smll RN lsters from the forementioned niml nd plnt speies. The niversl models hve lso been trined sin trees (insted of ) sine the ddition of extr trees did not improve predition ry dt not shown bt only inresed the file size of the proded models. Frthermore, the ser is ble to selet whih rops of fetres re oin to be sed for mkin mirn preditions. The possible ombintions of sets of fetres sed for predition re: bioenesis-only (overe nd seqene omplexity), enomi-only, both bioenesis nd enomi. distint model hs been trined for eh of these ses, so three different trinin models were eventlly trined for eh speies to mth the sers preferenes in eh rn. Prmeter speifition MiroRN predition is performed by deflt sin ll 24 bioenesis fetres nd the nine enomi fetres (in se the referene enome is vilble). However, it is lso possible to ompletely disble enomi fetres, by seletin the Disble enomi fetres option, or se exlsively the enomi fetres for predition, by hekin the se only enomi fetres for predition option. Frthermore, mirnovo offers set of three prmeters in order to filitte seqene lsterin nd boost orret lssifition of predited mirns. Speifilly, when nlysin smples with hih red depth nd hih seqene omplexity (i.e. hih nmber of enerted lsters t the initil seqene lsterin of inpt dt with vserh), we notied tht in some ses preditions ontin n nexpetedly hih nmber of novel mirns, sometimes even hiher thn the nmber of predited known mirns (Spplementry Fire S16). In order to resolve this isse we introded, first of ll, the Rede inpt seqene omplexity option whih llows the ser to filter ot niqe seqenes from the inpt file with totl red depth below ertin threshold. For instne, by sin tlly-threshold of x3, ll niqe seqenes from the tllied file with depth of p to three reds re disrded from the rest of the nlysis. Followin the initil seqene lsterin, dditionl filterin is possible by retinin only those lsters tht hve totl depth eql to or reter thn the min-red-depth vle nd nmber of niqe isoforms within the lster t lest eql to the minvrints prmeter vle. We lso tested predition performne when sin pre-smpled inpt files, for vrios smplin thresholds (Spplementry Fire S17). We notied tht the hih seqene omplexity isse is resolved beyond ertin smplin threshold, similrly with the prmeter twekin whih is spported by mirnovo. Finlly, this sests tht mirn predition with mirnovo is in enerl fesible even with low seqenin depth. Downloded from by Wellome Trst enome mps ser on 16 Otober 17

4 4 Nlei ids Reserh, 17 Smplin test depths Sbsmples (Spplementry Fire S17) were mde for reds in the SRR (ession: PRJN147) nd H99 (18) (speifilly: H99 5 MI ) FSTQ files. For eh smple one of series of probbilities ws set for the smplin of inpt seqenes (.1,.5,.1,.2,.5,.1,.2,.3,.4,.5,.6,.7,.8,.9). Mirnovo ws sed to predit mirns in eh sbsmple nd the nsmpled dtset with the deflt prmeters, sin speies-speifi trinin set nd enome. Reslts were lssified ordin to mirbse nnottions. mirnovo web-server bkend Mirnovo ses one of the fstest demi networks in Erope for effiient file plod nd ll jobs re sbmitted to the EMBL-EBI hih performne omptin lster. Eh job proess is extensively prllelised with mltiple threds for lltions, proessin different sbsets of the dt nd dividin different sbtsks. The job s proress is vislised in rel-time throh onsole window t mirnovo s proress pe in the browser. mirnovo stnd-lone tool Mirnovo is vilble s stnd-lone pke lon-side the web-server version. The downlodble bndle ontins ll neessry sripts nd binries for exetion of mirnovo, providin seprte versions for either M OSX or Linx pltforms. The only reqired dependenies for the lol mhine re: Perl (v5.24.1), Python (v2.7.1), R (v3.2.2) nd bowtie2 (v2..6), with the reommended versions in prentheses. Refined mtre mirn qntifition with himir (21) Mirnovo is ble to predit both hirpins nd mtre mir- Ns, providin ont dt in the ltter se. However, inherent seqene lsterin steps (initil nd refined) of the mirnovo pipeline my be imperfet in some ses nd ths ffet, even t low level, the yielded expression dt. Ths, in order to extrt even more rte expression dt we hve expnded himir, method tht ws previosly pblished in or lb. In tht se, himir serves s mirnovo extension, llowin the ser to plod stom set of hirpin seqenes (e.. known nd/or novel hirpins predited by mirnovo) nd then lin their inpt files inst this referene set to et mtre mirn expression onts. ll ploded files re mered nd seqenes with n linment identity over.9 re ollpsed. s n dditionl fntionlity, himir is ble to enerte overe profiles of eh identified mtre mirn nd the seondry strtre of the orrespondin hirpin referene hit (sin the Vienn pke (32)). EVDIS dtset nlysis ll smples nder the ession nmber PRJEB3365 (PMID: 4868) were nlysed. The mjority of smples were rn sin the deflt mirnovo prmeters (lenth filter: 16 28nt, min-red-depth: 5, min-vrints: 1, vserh-id:.9). The Rede inpt seqene omplexity option with tlly-threshold of 3 ws sed only for 2% of ll dtsets in order to rede seqene omplexity within the smples nd ths optimise the initil seqene lsterin with vserh. The overe profiles nd hirpin prersors of ll predited novel mirns in the EVDIS smples re vilble t the followin link: mirnovo vs mirdeep2 benhmrkin mirdeep2 ws lwys provided with the hmn referene enome, ll known hmn hirpins, ll known hmn mtre mirn seqenes nd lso ll mtre mirns from two extr speies (D. melnoster nd. brise) for dditionl diversity. Mirnovo ws tested both with nd withot the referene enome. With rerds to the time benhmrkin, mirnovo is hihly-prllelised nd mltithreded method while mirdeep2 is serilly proessed method. Ths, we wnted the benhmrkin to reflet the rn time experiened by the end ser. Both methods rn on HP lsters onsisted of 32-proessor nodes eqipped with the Intel(R) Xeon(R) P E Hz P model. Mirnovo ws rn sin the deflt nmber of hosts tht is seleted for eh job (-n = 3) while mirdeep2 ws rn sin -n = 1 (ssinin -n = 3 hosts to mirdeep2 proved to be slihtly slower -dt not shown most likely de to synhronistion lteny mon the hosts, nd ths one host ws eventlly ssined for benhmrkin of the mirdeep2 rns). Both methods were provided with 8B of memory (-M 8192). nlysis of moth nd btterfly smples For eh smple the 3 dpter seqene ws identified sin minion (27) nd where possible onfirmed in the relevnt mnsript or dtbse methods. Eh smple ws nlysed sin mirnovo with either deflt or stom set of prmeters (Spplementry Tble S4). The smple ids tht were nlysed re: SRR35544 & SRR35546 (SE17965, PMID: ), SRR62599 (SE23292, PMID: 23292), SRR62 (SE23292, PMID: ), SRR & SRR (SE63644, PMID: ), nd SRR35545 (SE17965, PMID: 17965). relevnt enome ws sed for eh smple (Bombyx mori: , Helionis melpomene: Hmel2 v2 Relese 15113, merri ohridell: k51, Prre eeri: k51) nd Drosophil Melnoster (dme) trinin model for mirn preditions. To find the ortholoes, novel mtre mirn seqenes were ompred to ll mirbse seqenes (v21) sin swn (v17 96) reqirin t lest 9% identity mth ( key-vle prmeter). Drosh/Dier/XPO5-dependent nlysis Smples were normlised sin the sme strtey desribed in the oriinl mnsript (). Speifilly, we normlised the wild-type, Drosh nd XPO5 knokot smples bsed on the red onts of mir-3-3p ross ll replites, Downloded from by Wellome Trst enome mps ser on 16 Otober 17

5 Nlei ids Reserh, 17 5 sine its expression is independent of Drosh. The Dier knokot smples were respetively normlised bsed on the ombined trn nd rrn levels of the smples, whih shold remin nffeted in the knokot smples s well. In order to derive expression dt from ll smples with referene to the hirpins tht were identified nd/or predited by mirnovo, we expnded the lredy pblished method himir. The dditionl fetre in himir llows linment inst stom referene speies tht n be ploded s set of FST files by the ser. Novel mirn predition from sinle-ell RN-Seq dt Proessin of sinle-ell RN-Seq dt follows the sme ore pipeline s relr smll RN-Seq dt proessin. The only exeption is tht de to hih innte noise of sinleell dt, overe nd seqene omplexity fetres re not tken into onsidertion t the finl lssifition step, nd ths preditions re inferred by models tht hve been pretrined solely bsed on the enomi fetres. Ths, in order to mke preditions from sinle-ell dt the option se only enomi fetres for predition needs to be enbled. RESLTS The min methodoloy behind mirnovo lies in rph-bsed lsterin of red-to-red similrities obtined from rw seqenin dt (Spplementry Fire S1). Inpt reds et dpter lened, de-dplited nd lstered toether into rops of hihly similr seqenes (see Mterils nd Methods). Sbseqently, lsters re filtered bsed on the minimm nmber of isoform vrints they ontin nd their overll seqenin depth. onsenss seqene is then llted for eh lster nd ll lsters re lined inst Rfm (1) to identify likely rrns (or other ontminnts). lsterin is extremely rpid (see Methods), however inonsistenies my rise. Hene, n extr refinement step hs been introded in order to mere lsters with hihly similr onsenss seqenes sin d-hit (2). We then perform fst mltiple-seqene linment within eh lster sin msle (3). This is sed to extrt refined onsenss seqenes from mered lsters. In order to fl existin mir- Ns, we ompre these onsenss seqenes inst mir- Bse (4). This finl rop of filtered, mltiply lined reds is sed to ompte set of fetres for eh lster. Mhine lernin fetres We se set of 24 bioenesis fetres, roped into two teories: 12 overe profile fetres (Fire 1) nd set of 12 seqene omplexity fetres (see Methods). Optionlly, the ser n lso provide enomi seqene if desired, whih dds 9 enomi fetres bsed on predited RN seondry strtre from mpped onsenss seqenes (see Methods). Bsed on this fetre set, mirnovo ses mhine lernin lssifier to identify both known nd novel mtre mirns. Or initil hypothesis of bein ble to predit de novo mirns sin these 24 enome-free fetres bsed prely on bioenesis needed to be tested omprehensively. dditionlly, we wnted to diretly ompre mirn predition bsed on these 24 fetres or on the 33 fetres tht inlde enomi nd strtrl informtion. There re severl existin tools (5 12) tht ddress the novel mirn predition problem, sh s mirdeep2, mirtools nd mirnlyzer (Spplementry Tble S1). We seleted mirdeep2 s the most prominent tool (13 16), whih tilises enomi nd strtrl fetres, s omprison for or pproh. Trinin models performne Mhine lernin performne ws initilly ssessed with 1-fold ross-vlidtion sin lbeled set of fetre instnes derived from 65 mose smples. This llowed for evltion of fetre preditbility nd bis-free ssessment of the preditive power of the lssifition model in ontrolled dtset. We tested rne of mhine lernin pprohes tht inlded: Spport Vetor Mhines (SVMs), rdient Boostin nd Rndom Deision Forests (Spplementry Fire S2). The most effiient pproh in terms of disrimintive power (bsed on the re nder rve - - sores), with or withot sin the enomi fetres, trned ot to be Rndom Deision Forests (or Rndom Forests). Hene, we seleted this method to be interted into mirnovo s the primry predition lorithm. Overll, we hve trined models for 8 niml nd 7 plnt speies sin 2 66 smples with lbeled dt in eh se, mkin p 433 smples in totl (Spplementry Tbles S2 nd S3). The 1-fold ross-vlidtion demonstrted ry mesres of % withot referene enome sin model bilt from niml speies (Fire 1B). Interestinly, mirn preditions on plnt seqenes still mned ry between 7.7% nd 82.9%, despite their differenes in bioenesis ompred to nimls (17). Inspetion of the fetre importne sores for the ry of preditions (Spplementry Fire S3) yields some of the overe fetres (red depth, vere seqene lenth of mtre seqene, vere ontent nd vere T ontent fter 3 end) s the most ritil ones for orret lssifition, in both nimls nd plnts. Moreover, we n observe tht enomi fetres ply more predominnt role in nimls thn in plnts, most likely bese of the hih vribility of seondry strtres of mirn prersors in plnts. Besides, this vribility in plnt mirns n be seen in the hih vrine of fetre importne sores, in ontrst with the lower vrine respetive niml fetres. These initil reslts onfirmed tht withot intertin ny enome informtion it is still possible to relibly identify both known nd novel mirns diretly from seqenin dt. ddition of the extr 9 trditionl enomi fetres does improve ry, bt not by s mh s expeted. There ws 1.65% (±1.53%) nd.7% (±2.79%) improvement of predition ry for nimls nd plnts respetively ( %, % respetively). We now bild niversl-niml nd niversl-plnt model by smplin dt points (refined seqene lsters) from eh respetive pool of speies sh tht they n be sed niformly by ny speies oriintin from these kindoms (Spplementry Fire S4). Obtined ry for these niversl models ws 89.7 or 92% for nimls, nd 71.4 or 71.8% for plnts, withot nd with referene enome respetively. When sin only the enomi fetres for Downloded from by Wellome Trst enome mps ser on 16 Otober 17

6 6 Nlei ids Reserh, 17 mirn predition, ry sores derese to 58.4% nd 79.3% for plnts nd nimls, respetively (Spplementry Fire S4). This proves the essentil role of the bioenesis nd seqene omplexity fetres sed by or mhine lernin model for rte mirn predition. Lre-sle mirn preditions from EVDIS dtset (18) In order to more widely ssess mirnovo, we pplied lresle benhmrkin. We sed ll hmn smples from the EVDIS dtset (18) (derived from lymphoblstoid ell lines). The initil nlysis ws performed withot sin ny hmn referene enome dt (Spplementry Fire S5). The ry obtined vered 92.14% while sensitivity nd novel predition rtes were 69.7% nd 34.62%, respetively. fter introdin the referene enome nd the extr 9 orrespondin enomi fetres (Fire 2, Spplementry Fires S6 nd S7), performne is improved (ry nd sensitivity: 95.51% nd 78.8%), while novel predition rtes fll to 18.63%. This implies tht the se of enomi fetres boosts the predition lrity of rel mirns while t the sme time keepin the nmber of flse positive ssinments of novel mirns reltively low. We then ompred the performne of mirnovo to mirdeep2, the most widely sed tool for mirn disovery. Sine mirdeep2 reqires enomi dt, we lwys provided it with the hmn referene enome, known hmn hirpin nd mtre mirn seqenes. Mirnovo ws tested both with nd withot the referene enome in seprte rns (Fire 3, Spplementry Fires S8 nd S9). We observed tht mirnovo otperforms mirdeep2 in 92% of the ses for known mtre mirns identifition nd predits more novel mirns in 99.9% of the ses, when provided with enomi seqene. Benhmrkin inst mirdeep2 Even withot referene enome, mirnovo performs omprbly with mirdeep2 in terms of preditin known mir- Ns, even thoh mirdeep2 tilises enomi informtion (Fire 3 nd Spplementry Fire S9). With rerds to novel mirn predition, we observed hiher predition rte for mirnovo, whih indites likely more flse positive hits. However, mirnovo rns sinifintly fster thn mirdeep2 (see below). When the plyin fields re lebeled nd both methods tilize the referene enome (Fire 3 nd Spplementry Fire S8), we observe slihtly improved sensitivity bt notble impt on preision. This effetively redes the nmber of flsely predited novel mir- Ns (Fire 3, Spplementry Fires S6 nd S7). In smmry, mirnovo predited 2414 novel mtre mirns (see Mterils nd Methods) in the EVDIS dtset, oriintin from 3173 hirpin prersors (Spplementry Dt S1) inldin ny deteted prlos. Expression of novel mirns ws reltively blned between smples, nd similr to the expression of known mirns. The lenths of the mjority of predited novel mirns were lso within the expeted rne of 23nt (Spplementry Fire S1). dditionlly, we lso tested mirnovo nd mirdeep2 on eiht niml speies (Fire 3B), sin the referene enomes in both methods. in we observed tht mirnovo performed better thn mirdeep2 in the mjority of ses for both known nd novel mirn preditions. We then ssessed the ompttionl exetion time, for the E- VDIS benhmrkin rns. Mirnovo ws on vere 2.5 fster thn mirdeep2, with men exetion time bein 43 min (Spplementry Fire S11), s ompred to 1 h 49 min for mirdeep2. The mirnovo pipeline, bein fster nd more lihtweiht is lso esier to onfire nd set-p s ommnd-line tool nd lso lends itself extremely well s simple to se web bsed server. We ertinly see dvntes to rnnin both mirnovo nd mirdeep2 on seqenin dt nd believe tht mirnovo represents sinifint nd sefl ddition to this field. Performne in non-model ornisms with poor enomessembly enome-free performne of mirnovo ppers to prove or initil hypothesis tht bioenesis fetres exhibited on mirn seqenes llow rte de novo mirn predition. lthoh the performne is slihtly worse withot enomi informtion, this enbles de novo mirn disovery in the mltitdes of non-model ornisms withot enomes, or with low-overe dt. It lso presents trtble pproh for smll RN nlysis in metenomis dt. In order to test the potentil of the enomefree pproh in sh speies, we ssessed mirnovo s performne on seven smples from five different moth speies withot flly ssembled enomes (19), two of whih do not hve ny mirbse nnottion (Spplementry Tble S4). Mirnovo ws ble to retrieve known mirns from ll speies with mirbse nnottion (B. mori, H. melpomene melpomene, H. melpomene rosin) lon with hndreds of novel mirns (Spplementry Dt S2 S6). dditionlly, mirnovo predited 119 nd 192 mirns from. ohridell nd P. eeri respetively. These speies did not hve ny mirbse nnotted mirns (Spplementry Dt S7 nd S8). mon ll predited novel mirns,. ohridell nd P. eeri were the speies with the hihest nmber of mir- Ns linin with prlos from other speies. This is to be expeted sine the other three moth speies hve been stdied more extensively nd lredy hve mirn entries in mirbse. smll proportion of novel mirns were predited withot ny enomi evidene, bsed solely on fetres derived from their overe profiles. We believe tht this effetively demonstrtes nother strenth of mirnovo for inferene of mirns in non-model ornisms nd enbles reserh on non-odin RNs for mny new speies. s n dditionl step, we tested mirnovo s performne inst mireder (11) nd MirPlex (12) thtrelsoble to predit mirns withot referene enome t ll. Both of these methods impose notble restritions in mirn predition sine they reqire tht both strnds of the premirn dplex re detetble in the seqenin dt. However, this is not ommon for mirns sine, in most of the ses, only one strnd of the dplex beomes mtre mirn while the other one ets derded (34). Benhmrkin reslts show tht mirnovo performs either omprbly or better thn both of these methods (Spplementry Downloded from by Wellome Trst enome mps ser on 16 Otober 17

7 Nlei ids Reserh, 17 7 Mhine lernin performne sores from mirnovo preditions on EVDIS dtset, sin referene enome!"#$%"&$'#"()$*+ Spei ity: ry: Preision: % sore Sensitivity: 78.8 preision sensitivity speifiity ry FDR novel predition rte FDR: (novel predition rte) B 3 5 dtsets mirnovo overll performne on EVDIS dtset, with nd withot referene enome / : with referene enome / : withot referene enome Preision/Rell rve Rndom Forest.8 Preision Tre positive rte RO rve Rndom Forest Flse positive rte Rell Fire 2. Mirnovo mhine lernin performne on lre-sle nlysis involvin 491 smples from the EVDIS (16) dtset. () The distribtion of ry, speifiity, sensitivity, preision nd FDR (novel predition rte) sores is shown ross ll smples of the dtset. (B) ROR nd Preision Rell (PR) rves from mirnovo predition performne ross ll smples from the EVDIS dtset, with or withot sin referene enome. Fire S18), whih is expeted de to mirnovo s bility to detet mirns sin more flexible nd diverse riteri. MiroRN predition in RNse III-defiient ells Novel mirns re predited bsed on fetres onsistent with their proessin by smll RN bioenesis mhinery. Hene, if they re rel mirns, one wold expet to observe their dysreltion when key mirn bioenesis enzymes re missin or mtted. We tested this hypothesis sin pblished experimentl dt from Drosh, XPO5 nd Dier knokot smples (). These enzymes re responsible for leve of mirn primry trnsripts, their nler export nd proessin into fntionl mtre mirns respetively. We predited known nd novel h- Downloded from by Wellome Trst enome mps ser on 16 Otober 17

8 8 Nlei ids Reserh, 17 Homo spiens Nmber of predited hirpins mirnovo mirdeep2 known novel mirdeep2 (mtre mirns) mirnovo mirdeep2 smoothin spline mirnovo (mtre mirns) mirnovo (mtre mirns) mirdeep2 (mtre mirns) mirnovo: WITH enome mirnovo: WITHOT enome known novel mirnovo (mtre mirns) mirnovo (mtre mirns) mirdeep2 (mtre mirns) mirdeep2 (mtre mirns) mirnovo vs mirdeep2 benhmrkin (on EVDIS dtset) mirnovo vs mirdeep2 performne on 8 model ornisms B Bos Trs Ms Msls Rtts Norveis Dnio Rerio lls lls enorhbditis elens Drosophil melnoster Fire 3. Benhmrkin of mirnovo inst mirdeep2. () mirnovo-vs-mirdeep2 predition performne ross the EVDIS dtset: mirdeep2 ws lwys rn sin referene enome, ll known hmn hirpins, ll known hmn mtre mirns nd mtre mirns from nother two speies (D. melnoster nd. brise). Mirnovo ws rn either with or withot sin the referene enome. (B) mirnovo-vs-mirdeep2 performne on identifition nd predition of mirns from smples ross eiht model ornisms. mn hirpins from the wildtype () smples. We then lined ll nd Knokot (KO) smples sin himir (21) inst the predited known hirpins obtined from mirnovo (Fire 4 nd B, Spplementry Fires S12 nd S13). For this qntifition step we expnded himir s fntionlity, s mirnovo extension, in order to llow for linment inst stom set of referene hirpins, ploded by the ser. Or dt verified the observed minor effet of XPO5 knokot in mirn expression, sine mir- Ns re still bein expressed, jst in lower levels in some ses. The Dier knokot, s expeted, leds to notble derese in mirn expression. The bsene of Drosh is verified to be the most ritil one sine it reslts in extensive depletion of the mjority of mirns, in verifyin previosly reslts reported in the oriinl pper. Some previosly identified Dier dependent bt Drosh independent mirns (mirtrons) re lso observed from or dt. For or novel mirn preditions, we lined ll smples inst the list of predited novel hirpins (Fire 4, Spplementry Fires S12 nd S13). We then ssessed whih mirns were differentilly expressed (fold-hne > 2 nd P <.5) between the nd KO onditions nd fond three sets of novel mirns, dependent on different types of enzymes (Fire 4D nd Spplementry Dt S9). Overll, we find novel mirns sinifintly differentilly expressed both in Drosh nd Dier knokot smples (Spplementry Dt S1). This implies tht this set of novel mirns is dependent on the two most importnt enzymes for mirn bioenesis (Drosh nd Dier) nd ths re proessed by the nonil bioenesis pthwy. dditionlly, we notied tht 25 novel mirns were dependent only on Dier (likely mirtrons) nd 33 were Drosh-only dependent (Spplementry Dt S11 nd S12, respetively). This findin is in ordne with previos stdies (,22 24) tht mirns my be dependent on only one of the two key enzymes (either Drosh or Dier) possibly oriintin from other strtred nonodin RNs. These reslts in, provide vlidtion tht mirnovo is preditin moleles likely to be proessed by the nonil bioenesis mhinery yet n lso identify those mirns whih re independent of one or more of the key enzymes. MiroRN predition from sinle-ell RN-Seqenin dt Reently, sinle-ell RN seqenin hs beome both trtble nd n extremely tive topi of reserh. iven tht some mirns hve been shown to be extremely elltype speifi, sh dtsets represent n importnt re for novel mirn disovery. Hene, we wished to ssess the performne of mirnovo in nlysis of sinle-ell smll RN- Seq. We initilly ttempted predition sin ll sets of fetres (overe, seqene omplexity nd enomi) bt the extrted overe profiles nd extrted seqene omplexity sores were distortin preditions de to hih noise of inpt dt. We then tried mkin or preditions sin only enomi nd we observed ler improvement in ry sores, ths we followed this pproh for the nlysis Downloded from by Wellome Trst enome mps ser on 16 Otober 17

9 Nlei ids Reserh, 17 9 ll known (hmn) mirns B predited known mirns predited novel mirns DIER_KO DIER_KO DIER_KO DIER_KO DROSH_KO DIER_KO DROSH_KO DIER_KO DROSH_KO DROSH_KO DROSH_KO DROSH_KO D Predited novel mirns, dependent on their bioenesis DROSH & DIER - dependent DROSH - dependent DIER - dependent 5 5 onts 3 3 seqenin dt onts onts 5 3 seqenin dt onts seqenin dt Fire 4. Dependene of novel mirns on RNseIII enzymes (Drosh nd Dier) nd Exportin-5. MiroRN expression dt ross Wild-Type nd Drosh/Dier/XPO5 knokot onditions fter linment inst: () ll known hmn hirpins from mirbse, (B) ll known hmn hirpins predited by mirnovo in these smples, () ll novel hirpins predited from mirnovo in these smples. The ternry plots show the reltive expression of eh mirn ross the exmined onditions. Dt points (mirns) in the middle of the trinle re eqlly expressed mon ll smples while points proximl to ertin vertex re more hihly expressed t the orrespondin smple. (D) Predited novel mirns in hmn ell line (HT116) by mirnovo, dependent on both Drosh nd Dier, Drosh only or Dier only, respetively. of sinle-ell dt. This proves to be nother sefl fetre of mirnovo, sine the ser is lwys ble to swith off ertin sets of fetres in order to mke their preditions bsed on the speifi reqirements, qlity or noise of inpt dt. We re-nlysed 4 sh smples from HEK ells, nive hmn embryoni stem ells (hess) nd primed hess (25). Nive embryoni stem ells re obtined from preimplnttion embryos while primed ones re obtined from post-implnttion embryos (26). Mirnovo predited 4747 novel hirpin ndidtes overll from these smples, 356 of whih hve lso been predited from the EVDIS dtset on hmn lymphoblstoid ell lines. These initil findins reqire frther filterin bsed on mtre mirns expression, in order to ont for noise de to sinle-ell dt. Speifilly, we fond tht 3135 nd 361 mirns were expressed bove medin nd vere expression of ll novel mirns, respetively. We then lined ll smples inst the predited set of hirpins sin himir, nd obtined mtre mirn expression dt for eh ell smple. Novel mirn expression is qite blned between both types of ESs, with smll rop of mirns bein down-relted drin the trnsition from nive to primed ESs (Fire 5 nd Spplementry Fire S14). On the other hnd, both types of embryoni stem ells show notble differentil expression ompred to differentited HEK ells. Interestinly, hihly similr expression ptterns n be observed with rerds to known mirn expression ross these ell types (Fire 5B nd Spplementry Fire S14). We observed tht novel mirn expression vries ross different sttes of plripoteny nd/or development in Homo spiens, with more sinifint differene observed between embryoni stem ells verss flly differentited ell types. We performed hierrhil lsterin for ll ells bsed on their novel mirns expression. We identified five mjor rops of ells with similr novel mirn sintre (Fire 5). Two of those rops were exlsively omprised of HEK ells, two rops were primrily poplted by ES nive nd ES primed ells respetively. Finlly, the lst rop onsisted of ells from ll three ell types. Hierrhil lsterin of these ells bsed on known mirn Downloded from by Wellome Trst enome mps ser on 16 Otober 17

10 ES (nive) 26 ES (nive) 1 Nlei ids Reserh, 17 Di erentil expression of known nd novel mirns in sinle-ells Novel mirns B Known mirns ES (nive) ES (nive) ES (primed) ES (primed) ES (primed) ES (primed) HEK HEK 3 HEK 8 HEK 1 HEK 4 HEK 1 HEK 2 HEK 5 HEK 7 HEK 6 HEK 13 HEK 17 HEK 12 HEK 15 HEK 14 HEK 43 ES (nive) 48 ES (nive) 37 ES (primed) 6 ES (nive) 12 ES (nive) 2 ES (primed) 26 ES (primed) 43 ES (primed) 18 HEK 11 ES (primed) 42 ES (primed) 16 ES (primed) 8 ES (primed) 25 ES (primed) 9 HEK 27 ES (primed) 53 ES (primed) 28 HEK 57 ES (primed) 67 ES (primed) 44 ES (primed) 17 ES (primed) 5 ES (primed) 3 ES (primed) 28 ES (primed) 1 ES (primed) 73 ES (nive) 81 ES (nive) 78 ES (primed) 77 ES (primed) 71 ES (primed) 75 ES (primed) 21 ES (primed) 46 ES (primed) 33 ES (primed) 3 ES (primed) 11 HEK ES (primed) 38 ES (primed) 24 ES (nive) 56 ES (nive) 66 ES (nive) 84 ES (nive) 49 ES (primed) 7 ES (primed) 47 ES (primed) 13 ES (primed) 14 ES (primed) 19 ES (nive) 61 ES (nive) 4 ES (primed) 18 ES (primed) 19 ES (primed) Sinle-ell hierrhil lsterin bsed on novel mirns expression 58 ES (primed) ES (primed) 31 ES (primed) 51 ES (nive) 49 ES (nive) 55 ES (nive) 59 ES (nive) 39 ES (primed) ES (primed) 32 ES (primed) 15 ES (primed) ES (nive) 45 ES (primed) 59 ES (primed) 9 ES (primed) 56 ES (primed) 12 ES (primed) 84 ES (primed) 36 ES (primed) 72 ES (primed) 29 ES (primed) 7 ES (nive) 41 ES (primed) 1 ES (primed) 74 ES (nive) 23 ES (primed) 69 ES (nive) 29 ES (nive) 63 ES (nive) 32 ES (nive) 78 ES (nive) 7 ES (primed) ES (nive) 66 ES (primed) 35 ES (nive) 51 ES (primed) 69 ES (primed) 39 ES (nive) 54 ES (primed) 36 ES (nive) 68 ES (primed) 23 ES (nive) 82 ES (nive) 28 ES (nive) 7 ES (nive) 72 ES (nive) 83 ES (nive) 22 ES (primed) 79 ES (nive) 34 ES (primed) 54 ES (nive) 35 ES (primed) 76 ES (nive) 48 ES (primed) 53 ES (nive) 16 ES (nive) 71 ES (nive) 34 ES (nive) 5 ES (primed) 3 ES (nive) 3 ES (nive) 42 ES (nive) 74 ES (primed) 79 ES (primed) 26 HEK 4 ES (nive) 61 ES (primed) 16 HEK 2 ES (nive) 44 ES (nive) 83 ES (primed) 35 HEK 27 ES (nive) 32 HEK 33 HEK 36 HEK 27 HEK 25 HEK 1 ES (nive) 14 ES (nive) 13 ES (nive) 31 ES (nive) 38 ES (nive) ell type lsses HEK ES (nive) ES (primed) Mltiple 73 ES (primed) 55 ES (primed) 5 ES (nive) 15 ES (nive) 46 ES (nive) 22 ES (nive) 68 ES (nive) 64 ES (nive) 63 ES (primed) 62 ES (nive) 34 HEK 1 ES (nive) 31 HEK 9 ES (nive) 18 ES (nive) 45 ES (nive) 33 ES (nive) 58 ES (nive) 77 ES (nive) 52 ES (nive) 5 ES (nive) 3 HEK 29 HEK 21 HEK 22 HEK HEK 23 HEK 24 HEK 19 HEK 8 ES (nive) 17 ES (nive) 11 ES (nive) 21 ES (nive) 64 ES (primed) 52 ES (primed) 75 ES (nive) 65 ES (primed) 25 ES (nive) ES (nive) 41 ES (nive) 37 ES (nive) 47 ES (nive) ES (primed) 62 ES (primed) 65 ES (nive) 76 ES (primed) 67 ES (nive) 57 ES (nive) 6 ES (primed) 24 ES (primed) 81 ES (primed) 82 ES (primed) Fire 5. Novel mirn preditions from sinle-ell RN-Seqenin dt involvin three ell-types: nive ESs, primed ESs nd HEK ells. () Normlized expression of novel mirns, predited by mirnovo nd qntified by himir, ross the three smple onditions in pirwise plots. (B) Normlized expression of known mirns, predited by mirnovo nd qntified by himir, ross the three smple onditions in pirwise plots. () Hierrhil lsterin of nïve, primed hmn ESs nd HEK ells bsed on novel mirn expression. expression lso yield similr ropin of the smples bsed on their ell type (Spplementry Fire S15). This findin illstrtes tht individl ells my ontin niqe novel mirn sintre tht is hrteristi for the ell type of oriin while other ells my show lower deree of differentition nd ths retin more eneri mirn expression profile, rerdless of ell type. ONLSION We hve demonstrted tht mhine lernin bsed, enome-free disovery of mirns is possible from smll RN seqenin in niml nd plnt speies. Or pproh hs similr levels of ry to the most widely sed previosly pblished tool, whih tilises enomi informtion (mirdeep2). dditionlly, or pproh exeeds mirdeep2 s performne when enome informtion is vilble nd does so t sinifintly lower ompttionl ost. This pproh hs been extensively vlidted sin mltiple speies, trinin sets nd 1-fold ross vlidtion. Besides, or method hs been vlidted sin lre-sle dtsets nd mirn bioenesis mtnt dtsets tht elidte potentil novel mirn bioenesis pthwys, bsed on their dependeny on different types of RNseIII enzymes. We hve lso demonstrted the possibility of disoverin novel mirn ndidtes from sinle-ell dt, despite their inherent noise, nd ths frther enble the disovery of novel mirn moleles ssoited with very prtilr ell types nd/or onditions. Moreover, we observed hiher deree of onsisteny in preditin novel mirns in nimls thn in plnts, in terms of the fetres with the most disrimintive power, whih omplies with the more omplex mirn bioenesis mehnisms present in plnts. However, mirn preditions on plnts still mned hih levels of ry nd ths mirnovo n serve dditionlly s formidble nd esy-to-se resore for reserhers of the plnts ommnity. Or method, mirnovo, is simple to instll s ommndline tool nd my lso be sed s ser-friendly webbsed method. iven the qlity of reslts obtined withot enome dt, we believe this method old hve n importnt role for mirn disovery in non-model ornisms. We believe tht mirnovo represents sinifint new ontribtion to the mirn field nd in prtilr the predition of novel mirns. DT VILBILITY Or method is vilble s both web-pplition ( wwwdev.ebi..k/enriht-dev/mirnovo) nd stnd-lone tool ( The overe profiles nd hirpin prersors of ll predited novel mir- Downloded from by Wellome Trst enome mps ser on 16 Otober 17

11 Nlei ids Reserh, Ns in the EVDIS smples re vilble t the followin link: SPPLEMENTRY DT Spplementry Dt re vilble t NR Online. KNOWLEDEMENTS We thnk members of the Enriht lb (EMBL) for interestin nd sefl disssions nd spport. FNDIN EMBL ore fndin; MR methodoloy reserh fellowship [MR/L12367/1 to M.P.D.]. Fndin for open ess hre: EMBL ore fndin. onflit of interest sttement. None delred. REFERENES 1. Nwroki,E.P., Bre,S.W., Btemn,., Db,J., Eberhrdt,R.Y., Eddy,S.R., Floden,E.W., rdner,p.p., Jones,T.., Tte,J. et l. (15) Rfm 12.: pdtes to the RN fmilies dtbse. Nlei ids Res., 43, D13 D Limin,F., Beifn,N., Zhenwei,Z., Sito,W. nd Weizhon,L. (12). D-HIT: elerted for lsterin the next enertion seqenin dt. Bioinformtis, 28, Edr,R.. (4) MSLE: mltiple seqene linment with hih ry nd hih throhpt, Nlei ids Res., 32, riffiths-jones,s., Sini,H.K., vn Donen,S. nd Enriht,.J. (8) mirbse: tools for mirorn enomis. Nlei ids Res., 36, D154 D Friedländer,M.R., Mkowik,S.D., Li,N., hen,w. nd Rjewsky,N. (12) mirdeep2 rtely indentifies known nd hndreds of novel mirorn enes in seven niml ldes. Nlei ids Res.,, Hkenber,M., Rodríez-Ezpelet,N. nd rnsy,.m. (11) mirnlyzer: n pdte on the detetion nd nlysis of mirorns in hih-throhpt seqenin experiments. Nlei ids Res., 39, W132 W W,J., Li,Q., Wn,X., Zhen,J., Wn,T., Yo,M., Shen Sn,Z. nd Shi,Q. (13) mirtools 2. for non-odin RN disovery, profilin, nd fntionl nnottion bsed on hih-throhpt seqenin. RN Biol., 1, Hendrix,D., Levine,M. nd Shi,W. (1) mirtrp, ompttionl method for the systemti identifition of mirns from hih throhpt seqenin dt. enome Biol., 11, R Yn,X. nd Li,L. (11) mirdeep-p: ompttionl tool for nlyzin the mirorn trnsriptome in plnts. Bioinformtis, 27, Mthelier,. nd rbone,. (1) MIReN findin mirorns with hih ry nd no lernin t enome sle nd from deep seqenin dt. Bioinformtis, 26, Jh,. nd Shnkr,R. (13) mireder: disoverin novel mirns in speies withot seqened enome. PLoS One, 8, e Mpleson,D., Moxon,S., Dlmy,T. nd Molton,V. (13) MirPlex: tool for identifyin mirns in hih-throhpt srn dtsets withot enome. J. Exp. Zool. B Mol. Dev. Evol., 3, Lomte,P.R., Mhjn,N.S., Kle,S.M., pt,v.s. nd iri,.p. (14) Identifition nd expression profilin of Helioverp rmier mirorns nd their possible role in the reltion of diestive protese enes. Inset Biohem. Mol. Biol., 54, Friedländer,M.R., Lizno,E., Hoben,.J., Bezdn,D., Báñez-oronel,M., Kdl,., Mte-Herts,E., Kerber,B., onzález,j., hen,k.. et l. (14) Evidene for the bioenesis of more thn 1, novel hmn mirorns. enome Biol., 15,R Dhhbi,J.M., tmn,h., Boffelli,D., Mis,W., Spindler,S.R. nd Mrtin,D.I. (11) Deep seqenin revels novel mirorns nd reltion of mirorn expression drin ell senesene. PLoS One, 6, e Mrkmi,Y., Tnhshi,T., Okd,R., Toyod,H., Kmd,T., Enomoto,M., Tmori,., Kwd,N., Thi,Y.H. nd zm,t. (14) omprison of heptoelllr rinom mirn expression profilin s evlted by next enertion seqenin nd mirorry. PLoS One, 9, e hen,x. (5) MiroRN bioenesis nd fntion in plnts. FEBS Lett., 579, Lpplinen,T., Smmeth,M., Friedländer,M.R., t Hoen,P.., Monlon,J., Rivs,M.., onzàlez-port,m., Krbtov,N., riebel,t., Ferreir,P.. et l. (13) Trnsriptome nd enome seqenin novers fntionl vrition in hmns. Ntre, 51, Qh,S., Hi,J.H. nd Hollnd,P.W. (15) brst of mirn innovtion in the erly evoltion of btterflies nd moths. Mol. Biol. Evol., 32, Kim,Y.K., Kim,B. nd Kim,V.N. (16) Re-evltion of the roles of DROSH, Export in 5, nd DIER in mirorn bioenesis. Pro. Ntl. d. Si..S.., 113, E1881 E Vitsios,D.M. nd Enriht,.J. (15) himir: nlysis of smll RN seqenin dt nd mirorn modifitions. Bioinformtis, 31, Rby,J.., Jn,.H. nd Brtel,D.P. (7) Introni mirorn prersors tht bypss drosh proessin. Ntre, 448, helofi,s., Dos Sntos,.O., hon,m.m. nd Hnnon,.J. (1) dier-independent mirn bioenesis pthwy tht reqires o tlysis. Ntre, 465, ifentes,d., Xe,H., Tylor,D.W., Ptnode,H., Mishim,Y., helofi,s., M,E., Mne,S., Hnnon,.J., Lwson,N.D. et l. (1) Novel mirn proessin pthwy independent of Dier reqires ronte2 tlyti tivity. Siene, 328, Fridni,O.R., bdllyev,i., Hemnn-Jensen,M., Shell,J.P., Lnner,F. nd Sndber,R. (16) Sinle-ell seqenin of the smll-rn trnsriptome. Nt. Biotehnol., 34, Nihols,J. nd Smith,. (9) Nive nd primed plripotent sttes. ell Stem ell, 4, Dvis,M.P., vn Donen,S., bre-ooder,., Brtoniek,N. nd Enriht,.J. (13) Krken: set of tools for qlity ontrol nd nlysis of hih-throhpt seqene dt. Methods, 63, Rones,T., Flori,T., Nihols,B., Qine,. nd Mhé,F. (16) VSERH: verstile open sore tool for metenomis. PeerJ, 4, e Lnmed,B. nd Slzber,S.L. (12) Fst pped-red linment with Bowtie 2. Nt. Methods, 9, Wootton,J.. nd Federhen,S. (1993) Sttistis of lol omplexity in mino-id-seqenes nd seqene dtbses. ompt. hem., 17, Trifonov,E.N. (199) In: Srm,RH nd Srm,MH (eds). Mkin Sense of the Hmn enome. Strtre & Methods denine Press, lbny, Vol. 1, pp Lorenz,R., Bernhrt,S.H., Höner Z Siederdissen,., Tfer,H., Flmm,., Stdler,P.F. nd Hofker,I.L. (11) ViennRN Pke 2.. lorithms Mol. Biol., 6, Leinonen,R., khtr,r., Birney,E., Bower,L., erdeno-tárr,., hen,y., lelnd,i., Frqe,N., oodme,n., ibson,r. et l. (11) The Eropen nleotide rhive. Nlei ids Res., 39, D28 D H,M. nd Kim,N. (14) Reltion of mirorn bioenesis. Nt. Rev. Mol. ell Biol., 15, Downloded from by Wellome Trst enome mps ser on 16 Otober 17