Incorporating prior biological knowledge for network-based differential gene expression analysis using differentially weighted graphical LASSO

Transcription

1 Zuo et al. BMC Bonformatcs (2017) 18:99 DOI /s METHODOLOGY ARTICLE Open Access Incorporatng pror bologcal knowledge for network-based dfferental gene expresson analyss usng dfferentally weghted graphcal LASSO Ymng Zuo 1,2,3,YCu 2, Guoqang Yu 1, Rujang L 2 and Habtom W. Ressom 3* Abstract Background: Conventonal dfferental gene expresson analyss by methods such as student s t-test, SAM, and Emprcal Bayes often searches for statstcally sgnfcant genes wthout consderng the nteractons among them. Network-based approaches provde a natural way to study these nteractons and to nvestgate the rewrng nteractons n dsease versus control groups. In ths paper, we apply weghted graphcal LASSO (wglasso) algorthm to ntegrate a data-drven network model wth pror bologcal knowledge (.e., proten-proten nteractons) for bologcal network nference. We propose a novel dfferentally weghted graphcal LASSO (dwglasso) algorthm that bulds group-specfc networks and perform network-based dfferental gene expresson analyss to select bomarker canddates by consderng ther topologcal dfferences between the groups. Results: Through smulaton, we showed that wglasso can acheve better performance n buldng bologcally relevant networks than purely data-drven models (e.g., neghbor selecton, graphcal LASSO), even when only a moderate level of nformaton s avalable as pror bologcal knowledge. We evaluated the performance of dwglasso for survval tme predcton usng two mcroarray breast cancer datasets prevously reported by Bld et al. and van de Vjver et al. Compared wth the top 10 sgnfcant genes selected by conventonal dfferental gene expresson analyss method, the top 10 sgnfcant genes selected by dwglasso n the dataset from Bld et al. led to a sgnfcantly mproved survval tme predcton n the ndependent dataset from van de Vjver et al. Among the 10 genes selected by dwglasso, UBE2S, SALL2, XBP1 and KIAA0922 have been confrmed by lterature survey to be hghly relevant n breast cancer bomarker dscovery study. Addtonally, we tested dwglasso on TCGA RNA-seq data acqured from patents wth hepatocellular carcnoma (HCC) on tumors samples and ther correspondng non-tumorous lver tssues. Improved senstvty, specfcty and area under curve (AUC) were observed when comparng dwglasso wth conventonal dfferental gene expresson analyss method. Conclusons: The proposed network-based dfferental gene expresson analyss algorthm dwglasso can acheve better performance than conventonal dfferental gene expresson analyss methods by ntegratng nformaton at both gene expresson and network topology levels. The ncorporaton of pror bologcal knowledge can lead to the dentfcaton of bologcally meanngful genes n cancer bomarker studes. Keywords: Pror bologcal knowledge, Gaussan graphcal model, Weghted graphcal LASSO, Network-based dfferental gene expresson analyss *Correspondence: hwr@georgetown.edu 3 Lombard Comprehensve Cancer Center, Georgetown Unversty, Washngton, DC, USA Full lst of author nformaton s avalable at the end of the artcle The Author(s). 2017Open Access Ths artcle s dstrbuted under the terms of the Creatve Commons Attrbuton 4.0 Internatonal Lcense ( whch permts unrestrcted use, dstrbuton, and reproducton n any medum, provded you gve approprate credt to the orgnal author(s) and the source, provde a lnk to the Creatve Commons lcense, and ndcate f changes were made. The Creatve Commons Publc Doman Dedcaton waver ( apples to the data made avalable n ths artcle, unless otherwse stated.

2 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 2 of 14 Background Typcally, a dfferental gene expresson analyss (e.g., student s t-test, SAM, Emprcal Bayes, etc.) s performed to dentfy genes wth sgnfcant changes between bologcally dsparate groups [1 3]. However, ndependent studes for the same clncal types of patents often lead to dfferent sets of sgnfcant genes and had only few n common [4]. Ths may be attrbuted to the fact that genes are members of strongly ntertwned bologcal pathways and are hghly nteractve wth each other. Wthout consderng these nteractons, dfferental gene expresson analyss wll easly yeld based result and lead to a fragmented pcture. Network-based methods provde a natural framework to study the nteractons among genes [5]. Data-drven network model reconstructs bologcal networks solely based on statstcal evdence. Relevance network s one common data-drven network model [6, 7]. It uses correlaton or mutual nformaton to measure the relevance between genes and sets a hard threshold to connect hgh relevant pars. Relevance network has extensve applcaton due to ts smplcty and easy mplementaton. However, ts drawback becomes sgnfcant when the varable number ncreases: t confounds drect and ndrect assocatons [8]. For example, a strong correlaton for gene par X-Y and X-Z wll ntroduce a less strong but probably stll statstcally sgnfcant correlaton for gene par Y-Z. As a result, when the number of genes s large, relevance network tends to generate over-complcated networks that contan overwhelmng false postves. Bayesan network s another classc data-drven network model [9]. Unlke undrected graphs such as relevance networks, Bayesan networks generate drected acyclc graphs, n whch each edge ndcates a condtonal dependence relatonshp between two genes gven ther parents. The benefts of usng Bayesan networks are: 1) By modelng condtonal dependence relatonshp, Bayesan networks only dentfy drect assocatons; 2) Wth drectons n the graph, Bayesan networks allow to nfer causal relatonshp. However, t s challengng to apply Bayesan networks on hgh-throughput omc data snce learnng the structure of Bayesan networks for hgh dmensonal data s tme-consumng and can be statstcally unrelable. Addtonally, Bayesan network cannot model cyclc structures, such as feedback loops, whch are common n bologcal networks. Recently, Gaussan graphcal models (GGMs) have been ncreasngly appled on bologcal network nference [10 12]. Smlar to Bayesan network, GGMs can remove the effect of ndrect assocatons through estmaton of the condtonal dependence relatonshp. At the same tme, they generate undrected graphs and have no lmtaton on modelng only acyclc structures. In GGMs, a connecton between two nodes corresponds to a non-zero entry n the nverse covarance matrx (.e., precson matrx), whch ndcates a condtonal dependency between these two nodes gven the others. GGMs dates back to early 1970s when Dempster ntroduced covarance selecton problem [13]. The conventonal approach to solve ths problem reles on statstcal test (e.g., devaton tests) and forward/backward selecton procedure [14]. Ths s not feasble for hgh-throughput omc data when the number of genes s rangng from several hundred to thousands whle the number of samples are only tens to hundreds. In addton, the small n, largep scenaro for omc data (.e., sample sze s far less than the varable number), makes maxmum lkelhood estmaton (MLE) of precson matrx not to exst because the sample covarance matrx s rank defcent. To deal wth these ssues, Schäfer et al. proposed to combne Moore- Penrose pseudonverse and bootstrappng technque to approxmate the precson matrx [15]. Others appled l 1 regularzaton to get a sparse network [16 18]. Takng nto account of the sparsty property of bologcal networks and the computatonal burden of bootstrappng, l 1 regularzaton methods are preferred. Among varous l 1 regularzaton methods, Menshausen et al. performed l 1 regularzed lnear regresson (.e., LASSO) for each node to select ts neghbors [16]. Gven all ts neghbors, one node s condtonally ndependent wth the remanng ones. Snce LASSO s performed for each node, ths neghbor selecton approach may face a consstency problem. For example, whle gene X s selected as Y s neghbor, gene Y may not be selected as X s neghbor when performng LASSO for gene X and gene Y separately. Compared wth neghbor selecton method, a more reasonable approach s graphcal LASSO, whch drectly estmates precson matrx by applyng l 1 regulaton on the elements of the precson matrx to obtan a sparse estmated precson matrx [17, 18]. We wll pursut the extenson of graphcal LASSO n ths paper. In addtonal to data-drven network models, there are many publcly avalable databases such as STRING ( KEGG ( kegg), BoGRID( and ConsensusPathDB ( where one can extract varous types of nteractons ncludng proten-proten, sgnalng, and gene regulatory nteractons [19 22]. Bologcal networks reconstructed from these databases have been reported useful. For example, Chuang et al. reconstructed proten-proten nteracton (PPI) network from multple databases to help dentfy markers of metastass for breast cancer studes usng gene expresson data [23]. They overlad the gene expresson value on ts correspondng proten n the network and searched for sub-networks whose actvtes across all patents were hghly dscrmnatve of metastass. By dong ths, they found several hub genes related to

3 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 3 of 14 known breast cancer mutatons, whle these genes were not found sgnfcant by conventonal dfferental gene expresson analyss. They also reported that the dentfed sub-networks are more reproducble between dfferent breast cancer cohorts than ndvdual gene markers. However, databases are far from beng complete. Networks constructed purely based on the databases have a large number of false negatves. In addton, databases are seldom specfc to a certan dsease, so the nteractons that exst n the databases may not be reflectve of the patent populaton under study. In contrast, data-drven modelsarelkelytohavealargenumberoffalsepostves due to background nose. Consderng ths, an approprate approach to ntegrate the pror bologcal knowledge from databases and data-drven network model s desrable for more robust and bologcally relevant network reconstructon [24]. Prevously, pror bologcal knowledge has been ncorporated nto the neghbor selecton method [25]. It reles on the Bayesan nterpretaton of LASSO and assgns two dfferent pror dstrbutons for connectons that are present n the database and those are not. Recently, weghted graphcal LASSO (wglasso) has been proposed to ncorporate pror bologcal knowledge nto graphcal LASSO by assgnng dfferent weghts to the entres of precson matrx [26]. In ths work, we extend the orgnal wglasso algorthm, explan ths dea from a Bayesan perspectve, and perform comprehensve comparsons between wglasso and competng data-drven network models (e.g., neghbor selecton, graphcal LASSO). Addtonally, explorng the topologcal changes between bologcal dsparate groups may lead to new dscoveres that cannot be dentfed by conventonal dfferental gene expresson analyss [27 29]. For example, hgh-degree nodes (.e., hubs) that only exst n one of the bologcally dsparate groups may ndcate the regulatory rule of the hub genes only n that group. Knowledge-fused dfferental dependency network (KDDN) s a recently proposed method to construct knowledge ncorporated network that can show the rewrng connectons between two groups [29]. An opensource Cytoscape app s avalable for easy mplementaton [30]. In ths paper, we propose a novel algorthm called dfferentally weghted graphcal LASSO (dwglasso) for network-based dfferental gene expresson analyss. Ths s acheved by buldng separate networks for bologcally dsparate groups usng wglasso, explorng the topologcal changes between dfferent groups, and prortzng sgnfcant gene lst from conventonal dfferental gene expresson analyss as shown n Fg. 1. Other prevously reported methods nclude those that focus on ntegratng pror bologcal knowledge nto data-drven network model to dentfy sub-networks that are related to the dsease under study [31, 32]. Our work dffers wth these methods snce we compute a dfferental network score for each gene and prortze them for subsequent analyss rather than outputtng a sub-network lst for bologcal nterpretaton. Also, methods that drectly ncorporate gene networks or pror bologcal knowledge nto statstcal models for classfcaton and regresson tasks have Fg. 1 An overvew of dwglasso. The nput s gene expresson data (e.g., Mcroarray, RNA-seq data, etc.) and the output s a prortzed lst based on the dfferental network (DN) score defned wthn dwglasso

4 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 4 of 14 been reported [33, 34]. The ratonale s that functonally lnked genes tend to be co-regulated and co-expressed, and therefore should be treated smlarly n the statstcal model. Our work leaves the statstcal model untouched. Instead, t focuses on usng the best set of gene bomarkers as an nput to the statstcal model. Ths s consdered to have advantages over provdng multple lnked genes from the network whose expresson values have smlar patterns. We show the applcaton of dwglasso on two ndependent mcroarray datasets from breast cancer patents for survval tme predcton, and on TCGA RNA-seq data acqured from patents wth hepatocellular carcnoma (HCC) for classfcaton task between tumor samples and ther correspondng non-tumorous lver tssues. Therestofthepapersorganzedasfollows. Methods secton ntroduces the extended wglasso algorthm and the proposed dwglasso for network-based dfferental gene expresson analyss. Results and dscusson secton presents the results of wglasso and dwglasso based on smulaton, mcroarray and RNA-seq data. Fnally, Concluson secton summarzes our work and dscusses possble future extensons. Methods Network nference usng wglasso Consder ( a centered and scaled data matrx X n p.e., n =1 x j = 0, n =1 x 2 j ), = 1 t measures the ntenstes of p genes on n samples, from a p-dmensonal Gaussan dstrbuton wth zero means on each dmenson and postve defnte covarance matrx p p (.e., X N (0, )). Suppose the sample sze n s far less than the varable number p (.e., n p), then the MLE of the precson matrx (.e., = 1 ) does not exst snce the sample covarance matrx S s rank defcent. If we further assume s sparse, then a l 1 regularzaton term can be added to the negatve loglkelhood functon f (X ) = log det + tr(s ) for a sparse precson matrx estmaton as shown n Eq. (1). Graphcal LASSO s an algorthm to effcently solve Eq. (1) by usng block coordnate descent [8, 9]. Once the sparse precson matrx ˆ s obtaned, a non-zero element n ˆ (.e., ˆθ j = 0) ndcates a condtonal dependence between x and x j gven the others. For network G ={(, j);1 < j p},wehaveĝ ={(, j) : ˆθ j = 0}. arg mn log det + tr(s ) + λ 1 (1) 0 where s the precson matrx, 0 s the constrant that has to be postve defnte, S s the sample covarance matrx, tr denotes the trace, the sum of the dagonal elements n a matrx, 1 represents the l 1 norm of, the sum of the absolute values of all the elements n,and λ s the tunng parameter controllng the sparsty of. LASSO based estmates have a Bayesan nterpretaton [35]. ˆ s the maxmum a posteror (MAP) estmate for the posteror dstrbuton p( X) wth a Laplacan pror dstrbuton p( ) as shown n Eq. (2). The LASSO term λ 1 n Eq. (1) s now part of p( ) = exp( λ 1 ) wth zero means and a scalng parameter λ. Fromthe Bayesan perspectve, p( ) encodes the pror knowledge of the network topology. For a database that contans only bnary nformaton (connectng or not) for a gven gene par, a natural way s to assgn two dfferent scalng parameters λ 1 and λ 2 for connectng pars and those are not connected, as shown n Eq. (3). For connectng pars, ther Laplacan pror dstrbuton s dffused, whle for non-connectng pars ther Laplacan pror dstrbuton s concentrated (.e., λ 1 λ 2 ). In another word, a larger penalty wll be assgned to non-connectng pars to ncrease the chance of ther correspondng entres n to shrnk to zero. In realty, tunng λ 1 and λ 2 at the same tme nvolves two dmensonal grd search, whch s qute tme-consumng for hgh-dmensonal data. An extreme soluton to set λ 2 = 0 lnks all the connectng gene pars from the database n the graph, neglectng the fact that the database mght contan some spurous connectons for the dsease under study. p( X) = p(x )p( ) p(x) exp(log det tr(s )) exp( λ 1 ) (2) p( ) = exp( λ 1 non con 1 ) λ 2 con 1 ) (3) Instead of usng the bnary nformaton, a contnuous confdence score s more sutable to ncorporate pror bologcal knowledge nto graphcal LASSO. The confdence score can be obtaned from multple resources. For example, an estmated functonal assocaton score for PPIs s provded by STRING database. We scale ths confdence score nto the range [0,1] and create a weght matrx W p p.inw, 1 ndcates a complete trust for a gene par to be connected, 0 represents that no evdence supports a gene par to be connected. In ths way, we can assgn dfferent penaltes to dfferent gene pars as shown n Eq. (4). Compared to Eq. (3), (4) also gves larger penalty for less lkely connectng gene pars, but now there s only one tunng parameter λ. For a fxed λ, R package glasso can solve Eq. (4) effcently gven W [17]. arg mn log det + tr(s ) + λ (1 W) 1 (4) 0 where 1 s all 1 matrx, W stheweghtmatrxcontanng the confdence score for each gene par and represents the element-wse multplcaton between two matrces. For LASSO based optmzaton problem as shown n Eq. (4), tunng the parameter λ s crucal snce t con-

5 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 5 of 14 trols the sparsty of the output ˆ. Typcally, λ s tuned by cross-valdaton, Akake nformaton crteron (AIC), Bayesan nformaton crteron (BIC), or stablty selecton [36]. Consderng that AIC and BIC often lead to data under-fttng (.e., over-sparse network) and stablty selecton requres extensve computatonal tme, we prefer to use cross valdaton wth one standard error rule to select the optmal tunng parameter λ opt.byusngone standard error rule, we can acheve the smplest (most regularzed) model whose error s wthn one standard devaton of the mnmal error. Our wglasso algorthm s shown below. Algorthm 1 wglasso Input: A centered and scaled data matrx X n p ; AweghtmatrxW p p ; A regularzaton parameter set ; A cross valdaton fold number k. Output: Estmated precson matrx ˆ. 1: Randomly and equally dvde X nto k folds, gven by X 1, X 2,..., X k. 2: for each λ do 3: for each m {1, 2,..., k} do 4: Run graphcal LASSO algorthm wth nput X n =[..., X m 1, X m+1,... ], and regularzaton parameter λ (1 W) to obtan the estmated precson matrx ˆ λ m. 5: Calculate the negatve log-lkelhood ( functon as the model fttng error f X m ˆ λ m) = log det ˆ λ ) m ( S + tr m λ m. 6: end for ( 7: Calculate the standard error for f X 1 ˆ λ 1), ( f X 2 ˆ λ ( 2),..., f X k ˆ λ k) as SE( ˆ λ ) = ( ( var f X 1 ˆ λ ) ( 1,...,f X k ˆ λ )) k k. ( kl=1 f X l ˆ λ ) l 8: Compute the average model fttng error f (X ˆ λ ) = k. 9: end for 10: Obtan λ mn that acheves the mnmal model fttng error λ mn ={λ :mn f (X ˆ λ )}. λ 11: Move λ n the drecton of ncreasng regularzaton untl reachng to one standard error lmt λ opt ={λ : f (X ˆ λ ) = f (X ˆ λmn ) + SE( ˆ λmn )}. 12: Run graphcal LASSO algorthm wth nput X and regularzaton parameter λ opt (1 W) to obtan the fnal estmated precson matrx ˆ. Network-based dfferental gene expresson analyss usng dwglasso Fgure 2 shows the framework of the proposed dwglasso algorthm for network-based dfferental gene expresson analyss. dwglasso prortzes the sgnfcant lst obtaned from the conventonal dfferental gene expresson analyss based on the topologcal changes between the group-specfc networks bult by wglasso. Specfcally, dwglasso frst performs dfferental gene expresson analyss to obtan a lst of sgnfcant genes whose expresson values dffer between the two bologcally dsparate groups. Then based on these sgnfcant genes, dwglasso bulds group specfc networks usng wglasso. After the networks are constructed, dwglasso calculates a dfferental network score for each gene n the sgnfcant lst based on the topologcal changes between the two group-specfc networks. In calculatng the dfferental network score, dwglasso frst computes the node degree for each gene n both networks, meanng the number of neghbors each gene s connected wth. Then consderng the sze of the two networks are dfferent, the node degrees are scaled nto the range [0,1]. At last, the dfferental network score for one gene s computed as the absolute value of the dfference between the two assocated scaled node degrees from dfferent groups. Fnally, wth the dfferental network scores, dwglasso prortzes the sgnfcant lst from the conventonal dfferental gene expresson analyss n a decreasng order. The prortzed gene lst s used for subsequent analyss such as buldng classfcaton or regresson models. We beleve dwglasso can help classfcaton or regresson models to acheve better predcton performance snce the prortzed lst ntegrates nformaton at the gene expresson and network structure levels. More than that, the ncorporaton of pror bologcal knowledge s more lkely to dentfy bologcally meanngful genes. Detaled algorthm for dwglasso s shown below. Results and dscusson Smulaton data Bologcal networks are reported to be scale-free, whch means the degree dstrbuton of the network follows a power law [37]. We consdered ths scale-free property of bologcal network n generatng smulaton data usng R package huge [38]. Usng huge, a scale-free network was bult by nputtng the node number p. The sparsty of the network s s fxed, dependng on p. For example, when the node number s 100, the sparsty of the network s 0.02, ndcatng only 2% of all possble connectons (.e., p (p 1) 2 ) exst n the scale-free network. Once the scale-free network s bult, huge creates the true precson matrx true based on the network topology and the postve defnte constrant true 0 so that

6 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 6 of 14 Algorthm 2 dwglasso Input: The raw data matrx X raw n p ; AweghtmatrxW p p. Output: Prortzed sgnfcant lst L dwglasso. 1: Perform conventonal dfferental gene expresson analyss on X raw to obtan a sgnfcant lst L. 2: Get two centered and scaled group specfc data matrx X (1) n 1 p sg and X (2) n 2 p sg from X raw and L,pckng out only the sgnfcant genes. 3: Buld group specfc networks G (1) and G (2) by runnng wglasso algorthm wth {X (1), W} and {X (2), W} as nputs. 4: for each L do 5: Compute the node degree d (1) and d (2) from G (1) and G (2),respectvely. 6: end for 7: for each L do 8: Compute the scaled node degree sd (1) and sd (2) as sd (1) = sd (2) = max j L max j L ( ) d (1) mn d (1) j j L ( ) ( d (1) j mn d (1) j j L ( ) d (2) mn d (2) j j L ( ) ( d (2) j mn d (2) j j L 9: Compute the dfferental network score dns = sd (1) sd (2). 10: end for 11: Prortze L based on the dfferental network score n a decreasng order to obtan L dwglasso. true = ( true ) 1 exsts. At last, smulaton data X n p N (0, true ) was generated. We created smulaton datasets wth varous p and n, as seen n Table 1. The weght matrx W, whch contans pror bologcal knowledge, was constructed based on true. In realty, databases may also contan spurous connectons for the dsease under study. To evaluate how the ncorrect connectons n W wll mpact wglasso, we ntroduced an addtonal metrc, acc. Whenacc = 60%, we randomly reassgned 40% ncorrect connectons n W. Specfcally, W was created as follows. Intally, for zero entres n true, the correspondng entres n W were also zero; for non-zero entres n true, the correspondng entres n W were randomly generated from the unform dstrbuton U(0,1). Then, we randomly assgned ncorrect connectons nto W based on the acc value whle keepng the total connectons n W the same as those n true. Under the assumpton that ncorrect entres n W should have lower confdence scores compared to those ), ). of correct entres, we generated ncorrect entres from the unform dstrbuton U(0, 0.5). We estmated the true network topology by usng neghbor selecton, graphcal LASSO, and the proposed wglasso methods. For neghbor selecton method, two strateges were appled to deal wth the nconsstency problem. Neghbor selecton wth or operator accepted nconsstent connectons whle neghbor selecton wth and operator rejected them. To make a far comparson, we tuned the regularzaton parameter n each method to ensure the output network has the same sparsty as the true network (.e., s = 0.02 for p = 100, s = for p = 500). For each n and p scenaro, we regenerated X n p 100 tmes, calculated the false postves and false negatves of connectons for each method, and lsted ther means and standard devatons n Table 1. To evaluate how the ncorrect connectons n W would mpact the performance of wglasso, we randomly reassgned 40% (acc = 60%) and 60% (acc = 40%) ncorrect pror bologcal knowledge n W. From Table 1, we can conclude that the estmated network from wglasso has much less false postves and false negatves, compared wth those from neghbor selecton and graphcal LASSO methods. A decrease of acc n W would lead to more false postves and false negatves from wglasso, but t stll outperforms neghbor selecton and graphcal LASSO methods when the acc n W s only as moderate as 40%. To make more comprehensve comparson, we plotted precson recall curve to evaluate the performance of neghbor selecton, graphcal LASSO and wglasso methods. We ran the above methods wth p = 100, n = 50 and acc = 40% n W, computed the precson and recall, and generated the plot as shown n Fg. 3. From Fg. 3, wglasso dsplays a clear mprovement over neghbor selecton and graphcal LASSO methods. Ths agrees wth our expectaton snce wglasso consders whether the connecton has supportng evdence from database and how well t fts the data n the model. Mcroarray data We appled the proposed dwglasso algorthm on two breast cancer mcroarray datasets: Bld et al. and van de Vjver et al. datasets [39, 40]. The former ncludes 158 patents wth all ther survval records, and was used for tranng. We excluded patents wth less than 5-year follow-up tme. Among the remanng patents, 42 wth less than 5-year survval durng the follow-up tme were consdered to form hgh rsk group whle the other 60 formed the low rsk group. van de Vjver et al. dataset contans 295 breast cancer patents, together wth ther survval records, and was used for ndependent testng. Both datasets are avalable at PRECOG webste ( precog.stanford.edu), an onlne repostory for queryng cancer gene expresson and clncal data, and have been

7 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 7 of 14 Fg. 2 Framework for dwglasso preprocessed for subsequent statstcal analyss [41]. The raw Bld et al. and van de Vjver et al. datasets are also avalable at Gene Expresson Omnbus (GSE3143) and R package seventygenedata, respectvely [42]. Our nterest s to obtan a prortzed sgnfcant gene lst based on dwglasso for more accurate survval tme predcton. The workflow s shown n Fg. 4. We frst performed unvarate analyss on Bld et al. dataset to select a lst of statstcally sgnfcant genes based on concordance ndex between the expresson value and survval tme [43]. Ths lead to a total of 58 genes whose adjusted p-values were less than The nflaton of Type I error caused by multple testng was controlled by the false dscovery rate (FDR) usng the Benjamn-Hochberg procedure. The total 58 sgnfcant genes are ncluded n Addtonal fle 1: Table S1 along wth ther assocated adjusted p-values. We then appled wglasso algorthm to buld two separate networks usng the total 58 sgnfcant genes for the hgh rsk and low rsk groups, respectvely. The weght matrx W was constructed based on the confdence scores from STRING database after nputtng the 58 sgnfcant genes to nvestgate the PPIs among them. For gene pars wth no confdence scores from STRING, we assgned the correspondng entres n W to zeros. In wglasso, we performed 10-fold cross valdaton and chose the optmal tunng parameter λ opt by one standard error rule. Fg. 5 shows our chose of λ opt : λ opt = for hgh rsk group and λ opt = for low rsk group. From the Table 1 The mean and standard devaton (n parenthess) of false postves (FP) and false negatves (FN) for connectons from neghbor selecton (NS), graphcal LASSO (glasso) and weghted graphcal LASSO (wglasso) methods under dfferent node number (p) and sample sze (n) scenaros p n NS (or) NS (and) glasso wglasso (acc = 60%) wglasso (acc = 40%) FP FN FP FN FP FN FP FN FP FN (17) 151 (10) 166 (15) 157 (10) 154 (23) 148 (11) 112 (17) 104 (11) 129 (18) 122 (11) (16) 111 (15) 132 (17) 122 (16) 114 (20) 112 (15) 82 (15) 74 (13) 93 (16) 87 (12) (13) 59 (18) 78 (15) 72 (21) 79 (17) 63 (19) 51 (11) 39 (14) 58 (13) 50 (15) (42) 679 (77) 758 (43) 738 (82) 710 (48) 681 (77) 480 (36) 451 (66) 549 (39) 526 (60) (30) 453 (129) 473 (42) 493 (134) 431 (40) 468 (129) 277 (26) 290 (87) 330 (31) 313 (106) (22) 164 (117) 189 (27) 177 (118) 199 (28) 186 (126) 109 (18) 110 (76) 130 (21) 135 (88) The best performance s marked n bold

8 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 8 of 14 Fg. 3 Precson recall curves for neghbor selecton, graphcal LASSO and weghted graphcal LASSO methods under p = 100, n = 50 and acc = 40% networks, we( calculated ) the node degree for each gene n two groups d h, dl, scaled them based on the network ( ) sze sd h, sdl, and computed the dfferental network ) score (dns = sd h sd l. At last, we prortzed the 58 sgnfcant genes based on the network dfferental scores n a decreasng order. To evaluate whether dwglasso could lead to more accurate survval tme predcton, we tested the prortzed gene lst usng dfferent methods on the ndependent van de Vjver et al. dataset. The 295 patents were dvded nto hgh rsk and low rsk groups accordng to the rsk scores calculated usng multvarate Cox regresson from the top 10 sgnfcant genes based on dwglasso, a competng pror knowledge ncorporated network analyss method (.e., KDDN), and conventonal dfferental gene expresson analyss (.e., concordance ndex). Unlke dwglasso that bulds group-specfc networks, KDDN generates only one network wth all rewrng connectons. From the network constructed by KDDN, we computed the node degree for each gene to help prortze the sgnfcant gene lst. Kaplan-Meer survval analyss was then performed to evaluate the performance of the above three scenaros. The resultng survval curves are shown n Fgs. 6a, b, and d. To evaluate how much the ncorporaton of pror bologcal knowledge contrbutes to the mproved performance of dwglasso, we tested the top 10 sgnfcant genes selected based on dwglasso wth no pror bologcal knowledge ncorporated (.e., W = 0). The resultng survval curve s shown n Fg. 6c. As expected, dwglasso wth no pror bologcal knowledge ncorporated s equvalent to usng graphcal LASSO n buldng group specfc networks (Fg. 4). As llustrated n Fg. 6, the top 10 sgnfcant genes from dwglasso wth pror bologcal knowledge ncorporated yelded the best performance (p value = ,hazardrato= 3.325), compared to the top 10 sgnfcant genes from KDDN (p value = ,hazardrato= 3.304), the top 10 sgnfcant Fg. 4 Workflow of dwglasso for more accurate survval tme predcton on mcroarray data

9 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 9 of 14 Fg. 5 Error curves to choose optmal tunng parameter λ opt usng 10-fold cross valdaton by one standard error rule. The blue lne ndcates the one standard error for λ mn n the drecton of ncreasng regularzaton genes based on dwglasso wth no pror bologcal knowledge ncorporated (p value = , hazard rato = 2.316), and the top 10 sgnfcant genes based on concordance ndex (p value = 0.002, hazard rato = 2.037). We beleve the mproved performance acheved by dwglasso and KDDN are due to the extra nformaton provded from the topologcal changes between hgh rsk and low rsk groups. Also, dwglasso and KDDN beneft from ncorporatng pror bologcal knowledge to obtan more relable and bologcally relevant genes shared across ndependent datasets, leadng to better predcton performance than those that do not use pror bologcal knowledge (Fg. 6). Table 2 presents the top 10 sgnfcant genes selected based on concordance ndex and dwglasso wth pror bologcal knowledge ncorporated, together wth ther adjusted p-values. The top 10 genes from the other methods are presented n Addtonal fles 2: Table S2. Among the top 10 sgnfcant genes based on dwglasso n Table 2, UBE2S has been reported to be over-expressed n breast cancer [44]. The authors showed UBE2S knockdown suppressed the malgnant characterstcs of breast cancer cells, such as mgraton, nvason, and anchorage-ndependent growth. SALL2 has also been reported as a predctor of lymph node metastass n breast cancer [45]. Unlke UBE2S, SALL2 was dentfed as a tumor suppressor gene that can suppress cell growth when over-expressed [46]. Addtonally, XBP1 has been Fg. 6 Survval curves. a top 10 sgnfcant genes based on dwglasso wth pror bologcal knowledge ncorporated, b top 10 sgnfcant genes based on KDDN,c top 10 sgnfcant genes based on dwglasso wth no pror knowledge ncorporated, d top 10 sgnfcant genes based on concordance ndex

10 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 10 of 14 Table 2 The top 10 sgnfcant genes based on conventonal dfferental gene expresson analyss (.e., concordance ndex) and dwglasso wth pror bologcal knowledge ncorporated, along wth ther adjusted p-value Top 10 sgnfcant genes based on concordance ndex Top 10 sgnfcant genes based on dwglasso Gene symbol Adjusted p-value Gene symbol Adjusted p-value BTD SALL FKTN UBE2S LRRC RAB11FIP RAB11FIP KIAA EMX XBP HNRNPAB KIAA TKT EMX LANCL OAZ TFF NDC USF CCT Common genes are marked n bold reported to be actvated n trple-negatve breast cancer and has a pvotal role n the tumorgencty and progresson of ths breast cancer subtype [47]. KIAA0922 has also been reported as a novel nhbtor of Wnt sgnalng pathway, whch s closely related to breast cancer [48]. None of UBE2S, SALL2, XBP1 and KIAA0922 s among the top 10 sgnfcant genes based on concordance ndex accordng to Table 2. In Fg. 7, we showed the neghbors of UBE2S and SALL2 n the hgh rsk and low rsk groups based on the networks created by wglasso from Bld et al. dataset. UBE2S s over-expressed n the hgh rsk group whle SALL2 s under-expressed. Ths agrees wth that UBE2S s a promotng breast cancer gene whle SALL2 s a suppressor breast cancer gene [44, 46]. Addtonally, UBE2S has hgher scaled node degree n the hgh rsk group whle SALL2 ( has hgher scaled node degree n the low rsk group sdube2s h = 0.286, sdl UBE2S = 0.778, sdh SALL2 = 1.0, sd l SALL2 = ). Ths shows, as a promotng breast cancer gene, UBE2S s more actvely connected wth ts neghbors n the hgh rsk group whle, the suppressor breast cancer gene, SALL2 s more actvely connected wth ts neghbors n the low rsk group. In Fg. 7, yellow edges represent connectons that have been supported from STRING database. We can see that these connectons based on pror bologcal knowledge are not always showng up from the output of wglasso. Ths s a nce property snce pror bologcal knowledge only provdes evdence. We stll need the support from the data to make a connecton. Therefore, by ntegratng pror bologcal knowledge nto data-drven models, we expect to buld more robust and bologcally relevant networks. Table 3 shows the survval tme predcton performance when the top 5, top 10 and top 15 sgnfcant genes are selected by each of the four methods as the nputs to the multvarate Cox regresson model (Fg. 6). In all three cases, the proposed dwglasso algorthm wth pror bologcal knowledge ncorporated acheved the best performance, followed by KDDN and dwglasso wthout pror bologcal knowledge ncorporated. The method that reles purely on concordance ndex had the least performance. RNA-seq data Usng UCSC Cancer Genomcs Browser, we obtaned TCGA RNA-seq data (level 3) acqured from patents wth HCC [49]. The RNA-seq data was acqured by analyss of 423 lver tssues, ncludng 371 prmary tumor, 50 sold normal and 2 recurrent tumor samples based on Illumna HSeq 2000 RNA Sequencng platform and mapped onto the human genome coordnates usng UCSC cgdata HUGO probemap. Among the 371 prmary tumor samples, 50 of them can fnd ts correspondng sold normal samples. To evaluate dwglasso on RNA-seq data, we apply a workflow shown n Fg. 8. We frst pcked out the 100 samples whose tumor tssues and ther correspondng non-tumorous tssues can both be found. Randomly, we selected 60 of them (30 tumor samples and ther correspondng normal samples) as the tranng dataset. The remanng 40 samples (20 tumor samples and ther correspondng normal samples) were used as testng dataset 1. Consderng testng dataset 1 only contans 40 samples, we created testng dataset 2 by combnng the above 40 samples and the remanng 321 tumor samples whose correspondng normal samples cannot be found. Wth testng datasets 1 and 2, we evaluated the performance of dwglasso on both balanced and large sample sze datasets. Specfcally, we preprocessed RNA-seq data usng R package DESeq2 on the tranng dataset [50].

11 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 11 of 14 Fg. 7 Neghbors of UBE2S and SALL2 n two groups. a neghbors of UBE2S n the hgh rsk group, b neghbors of UBE2S n the low rsk group, c neghbors of SALL2 n the hgh rsk group, d neghbors of SALL2 n the low rsk group. Label colors represent over- (red) or under- (green) expresson n the hgh rsk group. Node shapesndcate unque (crcle) or shared(rectangle) genes between the two groups. Node colors show the sgnfcance ofthe gene expresson valuebetween the twogroups. Yellow edges represent nteractons recorded n the STRING database. Thckness of the edge ndcates the strength of the nteracton From DESeq2, we selected statstcally sgnfcant genes whose adjusted p-values were less than 0.01 for subsequent analyss. At ths step, the number of sgnfcant genes s typcally between 1000 and We prortzed the sgnfcant gene lst based on dwglasso. From the prortzed gene lst, the top 5 genes were selected to tran a logstc regresson classfer to dstngush tumor and normal samples. The traned logstc regresson classfer was fnally evaluated on testng datasets 1 and 2. To compare dwglasso wth other methods, we also prortzed the sgnfcant gene lst based on adjusted p-value from DESeq2, dwglasso wthout pror bologcal knowledge ncorporated and KDDN, bult logstc regresson classfer usng the top 5 genes on the prortzed lst and evaluated the traned classfer on the testng datasets 1 and 2. Table 3 The survval tme predcton performance (p-value and hazard rato) for the top 5, top 10 and top 15 sgnfcant genes based on concordance ndex: DEA, dwglasso wth no pror bologcal knowledge ncorporated: dwglasso (no pror), KDDN, and dwglasso wth pror bologcal knowledge ncorporated: dwglasso (pror) Top 5 sgnfcant genes Top 10 sgnfcant genes Top 15 sgnfcant genes Methods p-value Hazard rato p-value Hazard rato p-value Hazard rato DEA E E dwglasso (no pror) E E KDDN E E dwglasso (pror) E E The best performance s marked n bold when the gene number s fxed

12 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 12 of 14 Fg. 8 Workflow of dwglasso for more accurate classfcaton predcton on RNA-seq data The above procedure was repeated 100 tmes and the means and standard devatons for senstvty, specfcty and area under curve (AUC) were calculated usng testng datasets 1 and 2 as shown n Table 4. In agreement wth mcroarray data, network-based methods wth pror bologcal knowledge ncorporated yelded the best performance, followed by network-based method wthout pror bologcal knowledge ncorporated, and the conventonal dfferental gene expresson analyss method was the worst. Ths s expected snce both dwglasso and KDDN methods take nto account of the changes of genes at gene expresson and network topology levels, and ncorporate pror bologcal knowledge nto ther network models. Concluson In ths paper, we apply a novel network nference method, wglasso to ntegrate pror bologcal knowledge nto a data-drvenmodel.wealsoproposeanewnetwork-based dfferental gene expresson analyss method dwglasso for better dentfcaton of genes assocated wth bologcally dsparate groups. Smulaton results show that wglasso can acheve better performance n buldng bologcally relevant networks than purely data-drven models (e.g., neghbor selecton and graphcal LASSO) even when only a moderate level of nformaton s avalable as pror bologcal knowledge. We demonstrate the performance of dwglasso n survval tme predcton usng two ndependent mcroarray breast cancer datasets prevously publshed by Bld et al. and van de Vjver et al. The top 10 genes selected by dwglasso based on the dataset from Bld et al. dataset lead to a sgnfcantly mproved survval tme predcton on the dataset from van de Vjver et al., compared wth the top 10 sgnfcant genes obtaned by conventonal dfferental gene expresson analyss. Among the top 10 genes selected by Table 4 The mean and standard devaton (n parenthess) of senstvty, specfcty and area under curve (AUC) calculated for conventonal dfferental gene expresson analyss: DEA, dwglasso wth no pror bologcal knowledge ncorporated: dwglasso (no pror), KDDN, and dwglasso wth pror bologcal knowledge ncorporated: dwglasso (pror) Testng dataset 1 Testng dataset 2 Methods Specfcty Senstvty AUC Specfcty Senstvty AUC DEA (0.07) (0.06) (0.04) (0.07) (0.04) (0.01) dwglasso (no pror) (0.03) (0.11) (0.02) (0.03) (0.05) (0.01) KDDN (0.08) (0.04) (0.02) (0.08) (0.03) (0.01) dwglasso (pror) (0.03) (0.07) (0.03) (0.03) (0.03) (0.01) The best performance s marked n bold

13 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 13 of 14 dwglasso, UBE2S, SALL2, XBP1 and KIAA0922 have been prevously reported to be relevant n breast cancer bomarker dscovery study. We also tested dwglasso usng TCGA RNA-seq data acqured from patents wth HCC on tumors samples and ther correspondng nontumorous lver tssues. Improved senstvty, specfcty and AUC were observed when comparng dwglasso wth conventonal dfferental gene expresson analyss method. Future research work wll focus on applyng dwglasso on other omc studes such as proteomcs and metabolomcs. Addtonal fles Addtonal fle 1: Table S1: The total 58 sgnfcant genes along wth ther assocated adjusted p-values. (CSV1.09kb) Addtonal fle 2: Table S2: The top 10 sgnfcant genes based on KDDN and dwglasso wthout pror bologcal knowledge ncorporated. (CSV 4.00 kb) Abbrevatons AIC: Akake nformaton crteron; AUC: area under curve; BIC: Bayesan nformaton crteron; DEA: dfferental gene expresson analyss; DN: dfferental network; dwglasso: dfferentally weghted graphcal LASSO; FDR: false dscovery rate; FP: false postves; FN: false negatves; GGMs: Gaussan graphcal models; HCC: hepatocellular carcnoma; KDDN: Knowledge-fused dfferental dependency network; LASSO: least absolute shrnkage and selecton operator; MAP: maxmum a posteror; MLE: maxmum lkelhood estmaton; PPI: proten-proten nteracton; wglasso: weghted graphcal LASSO Acknowledgements None. Fundng Ths work s n part supported by the Natonal Insttutes of Health Grants U01CA185188, R01CA and R01GM awarded to HWR. Avalablty of supportng data The datasets supportng the results of ths artcle are ncluded wthn the artcle and ts addtonal fles, or from referenced sources. Authors contrbutons YZ desgned and mplemented the algorthms, conducted the synthetc smulaton and real data applcaton, and drafted the paper. YC collected the two mcroarray datasets and partcpated n generatng the results for the synthetc smulaton and real data applcaton. GY and RL provded expertse n dfferental expresson analyss. HWR drected the project and completed the paper. All authors revewed and approved the fnal manuscrpt. Competng nterests The authors declare that they have no competng nterests. Consent for publcaton Not applcable. Ethcs approval and consent to partcpate Not applcable. Author detals 1 Department of Electrcal and Computer Engneerng, Vrgna Polytechnc Insttute and State Unversty, Arlngton, VA, USA. 2 Department of Radaton Oncology, Stanford Unversty, Palo Alto, CA, USA. 3 Lombard Comprehensve Cancer Center, Georgetown Unversty, Washngton, DC, USA. Receved: 18 February 2016 Accepted: 31 January 2017 References 1. Tusher VG, Tbshran R, Chu G. Sgnfcance analyss of mcroarrays appled to the onzng radaton response. Proc Natl Acad Sc. 2001;98(9): Newton MA, Kendzorsk CM, Rchmond CS, Blattner FR, Tsu KW. On dfferental varablty of expresson ratos: mprovng statstcal nference about gene expresson changes from mcroarray data. J Comput Bol. 2001;8(1): Efron B, Tbshran R, Storey JD, Tusher V. Emprcal bayes analyss of a mcroarray experment. J Am Stat Assoc. 2001;96(456): En-Dor L, Zuk O, Domany E. Thousands of samples are needed to generate a robust gene lst for predctng outcome n cancer. Proc Natl Acad Sc. 2006;103(15): Zuo Y, Yu G, Zhang C, Ressom HW. A new approach for mult-omc data ntegraton. In: Bonformatcs and Bomedcne (BIBM), 2014 IEEE Internatonal Conference On; p IEEE. 6. Butte AJ, Kohane IS. Unsupervsed knowledge dscovery n medcal databases usng relevance networks. In: Proceedngs of the AMIA Symposum; p Amercan Medcal Informatcs Assocaton. 7. Butte AJ, Kohane IS. Mutual nformaton relevance networks: functonal genomc clusterng usng parwse entropy measurements. Pac Symp Bocomput. 2000;5: Cteseer. 8. Zuo Y, Yu G, Tadesse MG, Ressom HW. Bologcal network nference usng low order partal correlaton. Methods. 2014;69(3): Fredman N, Lnal M, Nachman I, Pe er D. Usng Bayesan networks to analyze expresson data. J Comput Bol. 2000;7(3-4): Toh H, Hormoto K. Inference of a genetc network by a combned approach of cluster analyss and graphcal gaussan modelng. Bonformatcs. 2002;18(2): Dobra A, Hans C, Jones B, Nevns JR, Yao G, West M. Sparse graphcal models for explorng gene expresson data. J Multvar Anal. 2004;90(1): Kshno H, Waddell PJ. Correspondence analyss of genes and tssue types and fndng genetc lnks from mcroarray data. Genome Inform. 2000;11: Dempster AP. Covarance selecton. Bometrcs. 1972; Edwards D. Introducton to Graphcal Modellng: Sprnger Scence & Busness Meda; Schäfer J, Strmmer K. An emprcal bayes approach to nferrng large-scale gene assocaton networks. Bonformatcs. 2005;21(6): Menshausen N, Bühlmann P. Hgh-dmensonal graphs and varable selecton wth the lasso. Ann Stat. 2006; Fredman J, Haste T, Tbshran R. Sparse nverse covarance estmaton wth the graphcal lasso. Bostatstcs. 2008;9(3): Mazumder R, Haste T. The graphcal lasso: New nsghts and alternatves. Electron J Stat. 2012;6: Snel B, Lehmann G, Bork P, Huynen MA. Strng: a web-server to retreve and dsplay the repeatedly occurrng neghbourhood of a gene. Nuclec Acds Res. 2000;28(18): Kanehsa M, Goto S. Kegg: kyoto encyclopeda of genes and genomes. Nuclec Acds Res. 2000;28(1): Stark C, Bretkreutz BJ, Reguly T, Boucher L, Bretkreutz A, Tyers M. Bogrd: a general repostory for nteracton datasets. Nuclec Acds Res. 2006;34(suppl 1): Kamburov A, Werlng C, Lehrach H, Herwg R. Consensuspathdb a database for ntegratng human functonal nteracton networks. Nuclec Acds Res. 2009;37(suppl 1): Chuang HY, Lee E, Lu Y-T, Lee D, Ideker T. Network-based classfcaton of breast cancer metastass. Mol Syst Bol. 2007;3(1):. 24. Zuo Y, Yu G, Ressom HW. Integratng pror bologcal knowledge and graphcal lasso for network nference. In: Bonformatcs and Bomedcne (BIBM), 2015 IEEE Internatonal Conference On; p IEEE. 25. Wang Z, Xu W, San Lucas FA, Lu Y. Incorporatng pror knowledge nto gene network study. Bonformatcs. 2013;29(20): L Y, Jackson SA. Gene network reconstructon by ntegraton of pror bologcal knowledge. G3: Genes Genomes Genetcs. 2015;5(6): Ha MJ, Baladandayuthapan V, Do K-A. Dngo: dfferental network analyss n genomcs. Bonformatcs. 2015;31(21):

14 Zuo et al. BMC Bonformatcs (2017) 18:99 Page 14 of Zhang B, L H, Rggns RB, Zhan M, Xuan J, Zhang Z, Hoffman EP, Clarke R, Wang Y. Dfferental dependency network analyss to dentfy condton-specfc topologcal changes n bologcal networks. Bonformatcs. 2009;25(4): Tan Y, Zhang B, Hoffman EP, Clarke R, Zhang Z, Shh IM, Xuan J, Herrngton DM, Wang Y. Knowledge-fused dfferental dependency network models for detectng sgnfcant rewrng n bologcal networks. BMC Syst Bol. 2014;8(1): Tan Y, Zhang B, Hoffman EP, Clarke R, Zhang Z, Shh IM, Xuan J, Herrngton DM, Wang Y. Kddn: an open-source cytoscape app for constructng dfferental dependency networks wth sgnfcant rewrng. Bonformatcs. 2015;31(2): We Z, L H. A markov random feld model for network-based analyss of genomc data. Bonformatcs. 2007;23(12): Chouvardas P, Kollas G, Nkolaou C. Inferrng actve regulatory networks from gene expresson data usng a combnaton of pror knowledge and enrchment analyss. BMC Bonforma. 2016;17(5): We P, Pan W. Incorporatng gene networks nto statstcal tests for genomc data va a spatally correlated mxture model. Bonformatcs. 2008;24(3): Bnder H, Schumacher M. Incorporatng pathway nformaton nto boostng estmaton of hgh-dmensonal rsk predcton models. BMC Bonforma. 2009;10(1): Tbshran R. Regresson shrnkage and selecton va the lasso. J R Stat Soc Seres B (Methodol). 1996; Menshausen N, Bühlmann P. Stablty selecton. J R Stat Soc Seres B (Stat Methodol). 2010;72(4): Barabas AL, Oltva ZN. Network bology: understandng the cell s functonal organzaton. Nat Rev Genet. 2004;5(2): Zhao T, Lu H, Roeder K, Lafferty J, Wasserman L. The huge package for hgh-dmensonal undrected graph estmaton n r. J Mach Learn Res. 2012;13(1): Bld AH, Yao G, Chang JT, Wang Q, Pott A, Chasse D, Josh MB, Harpole D, Lancaster JM, Berchuck A, et al. Oncogenc pathway sgnatures n human cancers as a gude to targeted therapes. Nature. 2006;439(7074): Van De Vjver MJ, He YD, van t Veer LJ, Da H, Hart AA, Voskul DW, Schreber GJ, Peterse JL, Roberts C, Marton MJ, et al. A gene-expresson sgnature as a predctor of survval n breast cancer. N Engl J Med. 2002;347(25): Gentles AJ, Newman AM, Lu CL, Bratman SV, Feng W, Km D, Nar VS, Xu Y, Khuong A, Hoang CD, et al. The prognostc landscape of genes and nfltratng mmune cells across human cancers. Nat Med. 2015;21(8): Marchonn L, Afsar B, Geman D, Leek JT. A smple and reproducble breast cancer prognostc test. BMC Genomcs. 2013;14(1): Pencna MJ, D Agostno RB. Overall c as a measure of dscrmnaton n survval analyss: model specfc populaton value and confdence nterval estmaton. Stat Med. 2004;23(13): Ayesha AK, Hyodo T, Asano E, Sato N, Mansour MA, Ito S, Hamaguch M, Senga T. UBE2S s assocated wth malgnant characterstcs of breast cancer cells. Tumor Bol. 2016;37(1): Huang E, Cheng SH, Dressman H, Pttman J, Tsou MH, Horng CF, Bld A, Iversen ES, Lao M, Chen CM, et al. Gene expresson predctors of breast cancer outcomes. Lancet. 2003;361(9369): Lu H, Adler AS, Segal E, Chang HY. A transcrptonal program medatng entry nto cellular quescence. PLoS Genet. 2007;3(6): Chen X, Ilopoulos D, Zhang Q, Tang Q, Greenblatt MB, Hatzapostolou M, Lm E, Tam WL, N M, Chen Y, et al. Xbp1 promotes trple-negatve breast cancer by controllng the hf1 [agr] pathway. Nature. 2014;508(7494): Maharz N, Parett V, Nelson E, Dent S, Robledo-Sarmento M, Setterblad N, Parceler A, Pla M, Sgaux F, Gluckman JC, et al. Identfcaton of tmem131l as a novel regulator of thymocyte prolferaton n humans. J Immunol. 2013;190(12): Zhu J, Sanborn JZ, Benz S, Szeto C, Hsu F, Kuhn RM, Karolchk D, Arche J, Lenburg ME, Esserman LJ, et al. The ucsc cancer genomcs browser. Nat Methods. 2009;6(4): Love MI, Huber W, Anders S. Moderated estmaton of fold change and dsperson for rna-seq data wth deseq2. Genome Bol. 2014;15(12):1. Submt your next manuscrpt to BoMed Central and we wll help you at every step: We accept pre-submsson nqures Our selector tool helps you to fnd the most relevant journal We provde round the clock customer support Convenent onlne submsson Thorough peer revew Incluson n PubMed and all major ndexng servces Maxmum vsblty for your research Submt your manuscrpt at