 Research
 Open access
 Published:
MultiDCoX: Multifactor analysis of differential coexpression
BMC Bioinformatics volumeÂ 18, ArticleÂ number:Â 576 (2017)
Abstract
Background
Differential coexpression (DCX) signifies change in degree of coexpression of a set of genes among different biological conditions. It has been used to identify differential coexpression networks or interactomes. Many algorithms have been developed for singlefactor differential coexpression analysis and applied in a variety of studies. However, in many studies, the samples are characterized by multiple factors such as genetic markers, clinical variables and treatments. No algorithm or methodology is available for multifactor analysis of differential coexpression.
Results
We developed a novel formulation and a computationally efficient greedy search algorithm called MultiDCoX to perform multifactor differential coexpression analysis. Simulated data analysis demonstrates that the algorithm can effectively elicit differentially coexpressed (DCX) gene sets and quantify the influence of each factor on coexpression. MultiDCoX analysis of a breast cancer dataset identified interesting biologically meaningful differentially coexpressed (DCX) gene sets along with genetic and clinical factors that influenced the respective differential coexpression.
Conclusions
MultiDCoX is a space and time efficient procedure to identify differentially coexpressed gene sets and successfully identify influence of individual factors on differential coexpression.
Background
Differential coexpression of a set of genes is the change in their degree of coexpression among two or more relevant biological conditions [1], illustrated in Fig. 1 for two conditions. Differential coexpression signifies loss of control of factor(s) over the respective downstream genes in a set of samples compared to the samples in which the gene set is coexpressed or variable influence of a factor in one set of samples over the other. This could also be due to a latent factor which had a significant influence on gene expression in a particular condition [2].
Since the proposal by Kostka & Spang [1], many algorithms have been developed to identify differentially coexpressed (referred as DCX throughout the paper) gene sets and quantify differential coexpression. The algorithms can be classified based on two criteria: (1) method of identification of DCX gene sets (targeted, semitargeted and untargeted); and (2) scoring method of differential coexpression (gene set scoring and genepair scoring).
Based on the method of identification, similar to the one described by Tesson et al. [3], the algorithms can be classified into targeted, semitargeted and untargeted algorithms. The Targeted algorithms [4] perform differential coexpression analysis on predefined sets of genes. The candidate gene sets may be obtained from public databases such as GO categories and KEGG pathways. They do not find novel DCX gene sets. Another disadvantage of targeted methods is their reduced sensitivity if only a subset of the given gene set is differentially coexpressed which results in the DCX signal diluted. In addition, the DCX gene sets that are composed of genes of multiple biological processes or functions may not be identified at all [2]. The semitargeted algorithms [5, 6] work on the observation that the DCX genes are coexpressed in one group of samples. Hence, they perform clustering of genes in one set of samples, identify gene sets tightly coexpressed and test for their differential coexpression using the remaining group of samples. Although semitargeted algorithms can identify novel gene sets, their applicability is limited to the coexpressed sets identified by the clustering algorithm. In addition, this approach also may suffer from lower sensitivity due to diluted DCX signal, similar to in targeted approach. On the other hand, the untargeted algorithms [1, 3, 7, 8] assume no prior candidate sets of genes and instead find the gene sets de novo and therefore have a high potential to identify novel gene sets without diluting DCX signal. The major drawback of untargeted approach is higher false discovery rate and computational requirements.
The second aspect of DCX gene set identification algorithms is the methodology employed in scoring differential coexpression of a given gene set: (1) gene set scoring or setwise method, and (2) gene pair scoring. In gene set scoring, all genes are considered in the scoring at once such as in the linear modelling used in Kostka & Spang [1] and Prieto et al. [7]. On the other hand, genepair scoring, as used in DiffFNs [8] and DiffCoEx [3], computes differential correlation of each pair of genes in the gene set and summarizes them to obtain DCX score for the gene set. Gene pair scoring is intuitive and amenable to network like visualization and interpretation in single factor analysis settings. The first few methods (e.g. Kostka & Spang [1] and Prieto et al. [7]) are untargeted setwise methods, while DiffFNs [8] is an untargeted genepair scoring method. However, many later methods, including an early method (DCA [5]) are predominantly targeted or semitargeted algorithms using gene pair scoring. Differential coexpression has been used in various disease studies and identified many interesting changed interactomes of genes among different disease conditions. DiffFNs [8], Differential coexpression analysis [9], TSPG [10], and Topologybased cancer classification [11] were applied for the classification of tumor samples using interactome features identified using differential coexpression and shown good results over using individual gene features. The application of Ray and Zhangâ€™s coexpression network using PCC and topological overlap on Alzheimerâ€™s data helped identify gene sets whose coexpression changes in Alzheimerâ€™s patients [12]. The multigroup timecourse study on ageing [13] has identified gene sets whose coexpression is modulated by ageing. Application on data of Shewanella oneidens identified a network of transcriptional regulatory relationships between chemotaxis and electron transfer pathways [14]. Many other studies have also shown the significant utility of application of differential coexpression analysis [15,16,17,18]. However, none of the existing algorithms allow direct multifactor analysis of differential coexpression, i.e. deconvolving and quantifying the influence of different biological, environmental and clinical factors of relevance on the change in coexpression of gene sets. Multifactor differential coexpression analysis is important in many practical settings since each sample is characterized by many factors (a.k.a. cofactors) such as environmental variables, genetic markers, genotypes, phenotypes and treatments. For example, a lung cancer sample may be characterized by EGFR expression [19], smoking status of the patient, KRAS mutation and age. Similarly, ageing of skin may depend on age, exposure to sun, race and sex [20]. Deconvolving and quantifying the effects of these factors on gene setâ€™s coexpression and eliciting relevant regulatory pathways is an important task towards understanding the change in the cellular state and the underlying biology of interest. In such a case, singlefactor differential coexpression analysis suffers from multitude of tests and the interpretation of the gene sets may be cumbersome and misleading. Hence, we propose a very first methodology for such purpose called MultiFactor Analysis of Differential CoeXpression or MultiDCoX, a gene set scoring based untargeted algorithm. MultiDCoX performs greedy search for gene sets that maximize absolute coefficients of cofactors (as suggested in our earlier work [21]) in a linear model, while minimizing residuals for each geneset. The analysis of several simulated datasets demonstrate that the algorithm can be used to reliably identify DCX gene sets, and deconvolve and quantify the influence of multiple cofactors on the coexpression of a DCX geneset in the background of large set of nonDCX gene sets. The algorithm performed well even for genesets with weak signaltonoise ratio. The analysis of a breast cancer gene expression dataset revealed interesting biologically meaningful DCX gene sets and their relationship with the relevant cofactors. Furthermore, we have shown that the coexpression of CXCL13 is not only due to the Grade of the tumor as identified in [22], but also could be influenced by ER status. Similarly, MMP1 appears to play role in two different contexts defined by more than one cofactor. These together demonstrate the importance of multifactor analysis.
Methods
MultiDCoX formulation and algorithm
MultiDCoX procedure consists of two major steps: (1) identifying DCX gene sets and obtaining respective DCX profiles; and (2) identifying covariates that influence differential coexpression of each DCX gene set. The formulation essential to carry out these two steps is as follows.
Let E _{ im } denote expression of gene g _{ i } in sample S _{ m }. The cofactor vector characterizing S _{ m } is denoted by B _{ m } =â€‰(B _{ m1 } , B _{ m2 } , B _{ m3 } ,â€¦,B _{ mz } ) where B _{ mk } is the value of k ^{th} factor for S _{ m } which is either a binary or an ordinal variable. A categorical cofactor can be converted into as many binary variables as one less the number of categories of the factor. A real valued cofactor can be discretized into reasonably number of levels and be treated as ordinal variable.
We define a new variable A _{ mn } (I) to summarize coexpression of gene set I between sample pair S _{ m } and S _{ n } for which B _{ m } =â€‰B _{ n } as
A _{ mn } (I) measures square of mean change of expression of all genes in I from S _{ m } to S _{ n, } i.e. measuring correlation between two samples over geneset I. Most of A _{ mn } (I)â€˜s are expected to be nonzero among a group of samples in which I is coexpressed. On the other hand, if genes in I are not coexpressed in a group of samples then A _{ mn } (I)â€˜s tend to be closer to zero as illustrated in Fig. 2.
We quantify the influence of the cofactors by fitting a linear model between A _{ mn } (I)s and B _{ mn }s. In other words, A _{ mn } (I)s are the instances of the response variable A(I), B _{ mn }s form design matrix (B) and factors in the B _{ mn }s are explanatory variables or cofactors (F).
The coefficient vector obtained from the above modelling (Eq2) is called differential coexpression profile of the gene set I, denoted by F(I). A(I), B and F are of ax1, axz and zx1 dimensions respectively. Where â€˜aâ€™ is number of sample pairs which satisfy the condition in Eq1 or subset of these sample pairs sampled for modelling, whichever is lower; z is number of factors in the model.
The MultiDCoX algorithm identifies DCX gene sets by iteratively optimizing coefficient of a cofactor as outlined in Fig. 3: (1) setting significance threshold for cofactor coefficients; (2) choosing seed pairs of genes that demonstrate significant coefficient for the cofactor under consideration, i.e. the gene pairs may be differentially coexpressed for the cofactor; (3) expanding each chosen seed gene pair into a conservative multigene set by optimizing the respective coefficient; (4) augmenting the geneset to increase sensitivity or reduce false negatives while keeping the respective cofactor coefficient significant; and, (5) filtering out weak contributing genes from each geneset to increase specificity or reduce false positives. Each of these steps is explained in detail below.
1. Setting threshold of significance for cofactor coefficients: We generate the distribution of coefficients of the cofactors in F by random sampling of gene pairs: randomly sample large number of gene pairs, fit the linear model in Eq2 for each pair and obtain the coefficients in the linear models. Pool absolute values of coefficients of all factors of all gene pairs, and set half of the m ^{th} (mâ€‰=â€‰10 in our experiments) highest value as absolute threshold of significance for all cofactors. In other words,
where F _{ k } (I _{ l } ) is coefficient on gene set (a pair of genes in this case) I _{ l } for k ^{th} factor.
T _{ oi } is the threshold for cofactor â€˜iâ€™ for geneset I and â€˜oâ€™ stands for â€˜originalâ€™, derived from C _{ T } as follows
The division by 2 is necessary to avoid damagingly strict threshold and lay wider net at the beginning of the algorithm. mâ€‰>â€‰1 is required as some of the sampled gene pairs could belong to DCX genesets which may overestimate the threshold and reduce sensitivity of the algorithm.
2. Identifying DCX seed gene pairs: For each gene, search is performed throughout the dataset to find its partner gene whose pair can result in a linear model (Eq2) with at least one significant cofactor. A cofactor is considered to be significant if its linear model Ftest pvalue is <0.01 and absolute value of its coefficientâ€‰>â€‰C _{ T }. If no partner gene could be found, then the gene will be filtered out from the dataset to improve the computational speed at later stages of the algorithm. We have implemented this step using the procedure: (a) batch application of qr.coef() in Rpackage which computes only linear model coefficients using one QR decomposition, (b) filter out gene pairs whose linear model coefficients are in the range [âˆ’C _{ T } , C _{ T }], (c) apply lm() on the gene pairs remaining after step â€˜bâ€™ to compute Ftest pvalues, and (d) further filter out gene pairs which do not meet requirements for the coefficient pvalue. The batch application of qr.coef() is multifold faster than lm(). We use similar strategy in the steps 3.A3.C below to reduce computational requirements compared to the direct application of lm().
3. Identifying DCX gene sets: We optimize coefficient of each significant cofactor for each gene pair in the direction, in positive or negative direction, depending on the sign of the coefficient i.e. if the coefficient is negative (positive) its minimized (maximized). To do so, for each factor, the steps 3.A3.C are iterated until all seed pairs for which the factor is significant are exhausted from the seed pairs obtained in the step 2.
3.A. Expanding top gene pair to a multigene set: We choose the gene pair whose constituent genes are not part of any of the multigene sets identified and whose linear model fit resulted in the highest coefficient for the cofactor of interest. It will be expanded to multigene set by adding genes that improve the coefficient of that cofactor in the direction of its coefficient for the gene pair. A sequential search is performed from first gene to the last gene in the data (the order of the genes will be randomized prior to this search). A gene is added to the set if it improved the coefficient of the cofactor under consideration i.e. the threshold to add a gene thereby the stringency increases as the search proceeds. The final set obtained at the end of this step is denoted by J. This step results in a most conservative DCX gene set. Factor profile FP(J) of J is defined as set of (f _{ i } ,h _{ i }) pairs as follows:
Where f _{ i } is factor â€˜iâ€™ and h _{ i } denotes whether it is positively (h _{ i } =â€‰1) or negatively (h _{ i } =â€‰âˆ’1) significant or insignificant (h _{ i } =â€‰0):
F _{ i } (J) is coefficient of factor f _{ i } for gene set J.
Pval _{ i } (J) is pvalue of F _{ i } (J).
3.B. Augmenting gene set J: As we tried to improve the coefficient of the cofactor for each addition of a gene in the expansion step (3.A), we may have missed many true positives which are not as strong constituents of J, but could be significant contributors. Therefore, we perform augmentation step to elicit some of the potential notso strong constituents of J while preserving the factor profile of J. As the gene set identified in step (3.A) is conservative, we set a new threshold T _{ ni } (J) or simply T _{ ni } for the coefficient F _{ i } (J) of each f _{ i }as
T _{ ni } (J) will be as stringent as T _{ oi } and at most equal to F _{ i } (J) which is the coefficient obtained at the end of step (3.A). Moreover, we define centroid E _{ C } (J)â€‰=â€‰{E _{ Cm } (J)} of J as
E _{ C } (J)â€‰=â€‰[E _{ c1 } (J), E _{ c2 } (J),â€¦,E _{ cs } (J)] is treated as a representative gene expression profile of J and find a gene sub set K such that each gene in K, g _{ k } , the pair K _{ k } =â€‰(g _{ k } , E _{ C } (J)) satisfies the condition
Then the augmented set Lâ€‰=â€‰J â‹ƒ K as new DCX gene set.
3. C. Filtering gene set L: The set L obtained after the step (3.B) may contain false positives which can be filtered out as follows: As in the augmentation step, we compute E _{ C } (L _{ k } ), L _{ k } =â€‰L{k}, and evaluate each gene pair Q_{k} âˆˆ {(g_{k}, E _{ C } (L _{ k } ))  g _{ k } âˆˆ L} for F(Q _{ k } ). g _{ k } is removed from the set if F _{ i } (Q _{ k } )â€‰>â€‰F _{ i } (L) for all hiâ€‰=â€‰1. Then the final gene set Râ€‰=â€‰{g _{ k }  g _{ k } âˆˆ L and F(Q _{ k } )â€‰<â€‰F(L)}. R is the final set output for the run.
4. Identifying cofactors significantly influencing DCX of each gene set: It is important to identify the factors influencing the DCX of a geneset (i.e. FP(R)) to elicit underlying biology. The Ftest pvalue obtained for each cofactor by the linear model fit (in Eq2) in the above procedure need to be further examined owing to the dependencies among the gene sets explored. Therefore, we mark a cofactor to be influential (hi =1) on coexpression of R if it satisfies the following two criteria:

(a)
Effect size criterion: We pool coefficients of all factors on all gene sets identified (denoted as C _{ R }) and examine their distribution. The valleys close to zero on either side of the central peak are chosen as the significance threshold T _{ f+ } and T _{ f }, see Fig. 4 for illustration. The central peak is the result of the tests that signify chance association between the respective cofactor and coexpression of genesets. Whereas, the peaks on either side of the central peak signify coefficients of significant effects in testing/modelfitting. The valleys are identified by T _{ f+ } and T _{ f }, which are good thresholds to call coefficients significant i.e. F _{ i } (R) is considered to be significant if it is > T _{ f+ } or < T _{ f }. The underlying assumption is that not all factors influence all gene sets and the coefficients of the cofactors with no or little influence on certain gene sets will be suggestive of the distribution of the coefficients under null hypothesis.

(b)
Permutation pvalue criterion: We permute the factor values of a DCX gene set (i.e. permute columns of B_{mk} matrix) and fit the linear model in Eq2 for each gene set R. We repeat this procedure for a predefined number of iterations. A factor is said to be noninfluential on the coexpression of the gene set under consideration if a minimum predefined fraction of permutations (0.01 in this paper) resulted in a fit in which the coefficient is better than F _{ i } (R) and its Ftest pvalue is better than the Ftest pvalue of the coefficient without permutation or 0.01 whichever is lower.
Finally, the gene sets with at least one significant cofactor and of predefined size (i.e. at least 6 genes in the set) will be output as DCX gene sets along with their factor profiles.
Reducing computational and space requirements
Computational and space requirements can be further reduced using the following strategies: (1) Filter out genes with no detectable signals among almost all samples and genes that demonstrate very little variance across the samples. This can filter out up to 50% of the genes from the analysis. As a result, we can accomplish modest reduction in space requirement and substantial reduction in computational requirement as the search procedure is at least of quadratic complexity in time; (2) Further reduction in computational time can be achieved in the step 2 i.e. identifying seed gene pairs. Randomly split the genes into two halves and search for possible pairs where one belongs to one half and the other belongs to the other half, instead of all possible gene pairs. As many DCX genesets are expected to be sufficiently large, >10 genes, each split set is expected to contain >2 genes from each DCX geneset. This reduces computational time to find seed gene pairs by 2 fold. (3) Another possibility is to consider only a subset of sample pairs by randomly sampling a small fraction of (m,n)s for the linear model, it could be as small as 10% of all (m,n)s. These three strategies put together with the optimization described in the step 2 of MultiDCoX can massively reduce the space and computational requirement by several folds and make the algorithm practical.
Results
Simulation results
To evaluate efficacy of MultiDCoX, we analyzed simulated datasets of varying degrees of signaltonoise ratio and sample size. Each simulated dataset consists of 50,000 probes as in a typical microarray and three factors of 12 stratums. Sample sizes were chosen to be either 60 or 120 or 240 i.e. 5, 10 and 20 samples per stratum respectively. Two factors B1 and B2 were binary (âˆˆ {âˆ’1, 1}) and the other (B3) is an ordinal variable of three levels (âˆˆ {âˆ’1, 0, 1}). Sample labels were randomly chosen for each factor and gene expression (E _{ im }) was simulated as described below:
B1 _{ im } =â€‰B1 _{ m } ~ N(0,1) if S _{ m } is in coexpressed group of B1 and g _{ i } is in DCX gene set for the factor B1, 0 otherwise. Similar interpretation holds for the remaining factors, B2 and B3, too. O _{ im } =â€‰O _{ m } ~ N(0,1) indicates coexpression over all samples if g _{ i } belongs to set of genes coexpressed across all samples irrespective of the factor values. E _{ im } ~ N(0,Ïƒ ^{2} ) is noise term and Ïƒ ^{2} is the extant of noise in the data.
We simulated 20 genes which show coexpression for B1 _{ m } =â€‰1 and B2 _{ m } =â€‰1, 20 genes coexpressed for B1 _{ m } =â€‰âˆ’1 only, and another 20 genes with O _{ i } =â€‰1 only. With this we have two sets of negative controls: large number of genes with no coexpression and a set of 20 genes coexpressed across all samples. Ideally, a DCX geneset identification algorithm should be able to discriminate the first two sets of genes from the two control (negative) sets. Furthermore, we have tested our MultiDCoX for three different values of Ïƒ âˆˆ {0.2, 0.5, 0.8} i.e. from low noise to the noise comparable to the signal. We carried out 10 simulations for each choice of Ïƒ.
The simulation results are summarized in the panel of plots in Fig. 5: plots of average numbers of false positives (FPs) and false negatives (FNs) over 10 independent simulation runs for each choice of Ïƒ and sample size. MultiDCoX performed well in terms of both false positives and false negatives for low to medium values of Ïƒ. Moreover, the algorithm exhibited reasonable performance even at the noise (Ïƒ) comparable to the signal (i.e. Ïƒâ€‰=â€‰0.8). The simulation results also demonstrate that MultiDCoX is sensitive even at small sample size for low to medium noise level. The failure rate of identifying genesets and their profiles are dependent not only on the sample size and noise level, but also on the type of set identified, especially for low sample size and high noise: the single factor influenced geneset has better chance of being identified with right factor profile, whereas the set influenced by 2 factors has higher chance of being identified. The effect of noise on FNR also depended on the number of factors influencing the DCX gene set. However, FDR is less dependent on both noise level and the number of factors influencing coexpression. Number of simulations that identified false gene sets increased with increased noise and reduced sample size. It is the lowest for 5 samples/stratum and high noise (Ïƒâ€‰=â€‰0.8). The computational time for MultiDCoX analysis, to optimize each cofactor in both directions (maximization and minimization), was ~12â€“15 h for one simulated data of 240 samples using 1 node of a typical HPC cluster.
MultiDCoX analysis of breast tumor data
We analyzed a breast tumor gene expression data published by Miller et al. [23]. It contains expression profiles of tumors from 258 breast cancer patients on U133A and U133B Affymetrix arrays i.e. ~44,000 probes. Tumors were annotated for their oestrogen receptor (ER) status (1 for recognizable level of ER or ER+, âˆ’1 otherwise or ER), p53 mutational status (1 for mutation or p53+, and âˆ’1 for wild type or p53) and grade of tumor (âˆ’1 for grade 1, 0 for grade 2, and 1 for grade 3). ER and p53 status are important markers used to guide treatment and prognosis of breast cancer patients. Hence it is important to identify the genesets regulated and thereby coexpressed by these factors while accounting for the effect of the tumor status as indicated by its grade and strong association between these three cofactors. For example, p53mutant tumors are typically of higher grade (grades 2 or 3) tumors with correlation of ~ 63% [24] and ERpositive tumors are typically of low grade (grade 1) [25]. In the presence of these correlations among the cofactors, it is important to identify and quantify their effects on coexpression of gene sets. We applied MultiDCoX on this dataset using ER status, p53 mutational status and tumor grade as cofactors. We discuss a few DCX genesets here and the remaining DCX gene sets are given in the Additional file 1.
Coexpression of ER pathway and the genes associated with relevant processes is modulated in p53 mutated tumors: A DCX gene set is shown in Table 1. The set is coexpressed only in p53 mutant tumors. The coexpression plot of p53 mutant tumors is shown in Fig. 6.
The set includes ESR1 (which encodes ERÎ±), its cofactor GATA3 and pioneering factor FOXA1 [26] along with ER downstream targets CA12, SPDEF and AGR2. We retrieved a total of 1349 p53 binding sitesâ€™ associated genes data from Botcheva K et al. [27] and Wei CL et al. [28]. p53 binding sites are reported to be close to the promoters of ESR1 [29] as well as GATA3. Furthermore, GATA3 binds to FOXA1 [30]. Our finding reinforces the observations made by Rasti et al. [29] that different p53 mutations may have varying effect on the expression of ESR1 gene, itâ€™s cofactor GATA3, pioneering factor FOXA1 and SAMdependent Mythyltransferase & p53 interacting GAMT which could have resulted in the differential coexpression of the ER pathway. In addition, comodulation of chromatin structure alternating & ER promoter stimulating TOX3 and Protein transfer associated REEP6 appears to be required to modulate ER pathway by p53.
Genes coexpressed with BRCA2 in ERnegative tumors are associated with Her2neu status:
Another gene set of interest is coexpressed in ERnegative tumors only and its details are given in Table 2. The coexpression plot of the gene set in ERnegative tumors is shown in Fig. 7. The gene set includes tumor suppressor gene BRCA2. We have investigated ER binding sites published by Carroll et al. [31] and Lin et al. [32] for ER binding sites close (within Â±35Kb from TSS) to these genes. The ~4800 binding sites mapped to ~1500 genes. Significantly, 10 of the 21 genes in this DCX gene set have ER binding sites mapped to them which is statistically significant at Ftest pvalue <0.01. Interestingly, most of these genes have not been identified to be ER regulated in the earlier studies using differential expression methodologies, possibly owing to the complexity of regulatory mechanisms. However, many of these genes are down regulated in ERnegative tumors. Testing for association of expression of this set with Her2neu status revealed that higher expression in ERnegative tumors is associated with Her2neu positivity which must have led to coexpression in ER negative tumors. Odds ratio of such an association is 18 which is much higher than that of ER positive tumors (ORâ€‰=â€‰4).
DCX of CXCL13 is modulated by grade as well as ER status
Analysis of Grade1 and Grade3 tumors using GGMs [22] helped identify CXCL13 in breast cancer as hub gene. It emerged as one of the hub genes in our analysis too, contributing to multiple DCX gene sets (see Additional file 1 , sheet:maxGrade). Although they are significant for Grade, they are significant for ER status too. It shows that CXCL13â€™s differential coexpression appears to be influenced by ER status, in addition to Grade. This couldnâ€™t be identified in the previous study as it was restricted to singlefactor (Grade) analysis.
DCX of MMP1 is modulated by factor subspace associated with poor survival
MMP1 is another gene we have examined whose family of genes are associated with poor survival [33]. MMP1 is coexpressed among tumors which are P53+ (mutant) and ERnegative or higrade tumors which are ERpostive (see Additional file 1 , sheets: maxP53, maxGrade and minER). Both these categories are known to be associated with poor survival of patients. This couldnâ€™t have been revealed in single factor analyses.
DCX Modulated by Multiple Factors
Coexpression of many genesets is modulated by more than one factor. The genesets discussed for MMP1 and CXCL13 are examples of such multifactor DCX i.e. coexpression of these genesets is modulated by ER status and Grade of the tumors. One such set is shown in the 1st row of Table 3. In addition, we presented one geneset whose coexpression is modulated by all factors (covariates): ER status, p53 mutational status and Grade of tumors (ER+ & P53 & Grade+); and, another gene set whose coexpression is modulated by ER status and p53 status (ER & p53+), Table 3.
Functional analysis of DCX profiles
To elucidate the biological function of different DCX profiles (ER+, ER & p53+, etc.), we pooled all genes from gene sets of same DCX profile and used ClusterProfiler [34], Huang et al. [35] to identify GO terms and pathways enriched. Results for coexpression influenced by individual factors as well as selected two factor combinations are tabulated in Additional file 2. It shows a clear distinction of GO functional categories and pathways enriched between different DCX profiles. For example, many pathways and GO terms are uniquely enriched for single factors. Strikingly, numerous pathways and biological processes/functions are modulated by more than one factor. It couldnâ€™t have been easily deciphered by univariate analyses. Both these observations assert the need for multivariate analysis of coexpression and such need met by MultiDCoX.
Conclusions
MultiDCoX is a space and time efficient algorithm which successfully elicits quantitative influence of cofactors on coexpression of gene sets. It required only 12hr of computation on a typical HPC node to identify DCX gene sets for each factor for a dataset of 240 samples and ~44,000 probes. The simulation results demonstrated that MultiDCoX has tolerable false discovery rates even at 5 samples/stratum and noise (Ïƒ) of 0.8. However, false negative rate (FNR) was affected by both sample size and noise level: FNR is very low for large sample size (20 samples per stratum) and low noise level (Ïƒâ€‰=â€‰0.2). Interestingly, both FDR and FNR did not greatly depended on the type of the gene set to be discovered, or whether it is influenced by single factor or multifactors. The discovery of a gene set whose DCX is driven by two cofactors is less affected by noise and sample size than the gene sets influenced by a single cofactor. On the other hand, at low sample size and high noise, the set influenced by 2cofactors has higher likelihood of arriving at the incomplete profile compared to that of a 1cofactor driven DCX. Occurrence of false DCX sets increased substantially at high noise level and small sample size. This is a major issue to be addressed in the future improvements of MultiDCoX. Moreover, the performance of the algorithm needs to be studied for varying parameter settings and further reductions in computational time. It is possible to reduce the computational time by 2fold by filtering out 50% of probes of low variance in expression. Though we have not used this strategy as we needed to study its impact on the discovery and profiling of DCX gene sets, the current implementation could complete the analysis within half a day of computing for each factor. The massive parallel processing allows us to complete all analyses within a day.
Though the current implementation of MultiDCoX is limited to linear model, we can easily augment the implementation to use any link function to transform A(I) and then using linear function. However, we need to test the performance of the algorithm for various link functions.
By MultiDCoX formulation, we identify DCX gene sets exhibiting Btype coexpression only [22]. The other two types of differential coexpression may be identified using multivariate differential expression analysis followed by clustering.
MultiDCoX algorithm can be applied to different clinical data to quantify the influence of cofactors on the coexpression and its associated phenotypes.
Multiple aspects of the formulation and the algorithm need to be studied in our future improvements: Robustness of A _{ mn } (I) to outliers is an important aspect of the performance of the algorithm and impact of the thresholds used in the algorithm also to be studied. However, without tuning, the choice of parameters appears to be effective enough for both simulated and real data sets.
The application of MultiDCoX on a breast cancer data has revealed interesting sets of DCX genes: the set of ESR1, its cofactors along with downstream genes of ESR1 and genes associated with relevant ESR1 dependent transcriptional regulation; the set of genes containing ER binding site in their cis region. Furthermore, we have shown that the coexpression of gene sets that contain CXCL13 and the gene sets that contain MMP1 is affected by ER status too, in addition to tumor grade which couldnâ€™t have been elicited in a typical univariate DCX analysis. The utility of MultiDCoX is further demonstrated by revelation of coexpression modulated by multiple factors for numerous genesets and pathways.
Abbreviations
 DCX:

Differential Coexpression/Differentially Coexpressed
 DE:

Differential Expression
 ER:

Oestrogen Receptor
 FDR:

False Discovery Rate
 FNR:

False Negative Rate
 FNs:

False Negatives
 FPR:

False Positive Rate
 FPs:

False Positives
 GO:

Gene Ontology
 HPC:

Hi Performance Computing
 KEGG:

Kyoto Encyclopaedia of Genes and Genomes
 OR:

Odds Ratio
References
Kostka D, Spang R. Finding disease specific alterations in the coexpression of genes. Bioinformatics. 2004;20:il94â€“9.
Hsu CL, Juan HF, Huang HC. Functional analysis and characterization of differential Coexpression networks. Sci Rep. 2015;5:13295. 2015
Tesson BM, Breitling R, Jansen RC. DiffCoEx: a simple and sensitive method to find differentially coexpressed gene modules. BMC Bioinformatics. 2010;11:497.
Choi Y, Kendziorski C. Statistical methods for gene set coexpression analysis. Bioinformatics. 2009;25(21):2780â€“6.
Ihmels J, Bergmann S, Berman J, Barkai N. Comparative gene expression analysis by differential clustering approach: application to the Candida Albicans transcription program. PLoS Genet. 2005;1(3):e39.
Michael Watson, CoXpress: differential coexpression in gene expression data. BMC Bioinformatics 2006, 7:509.
C. Prieto, M.J. Rivas, J.M. SaÂ´ nchez, J. Lo pezFidalgo: Algorithm to find gene expression profiles of deregulation and identify families of diseasealtered genes. Bioinformatics 2006, 22(9):11031110.
R. Krishna Murthy Karuturi, Silvia Wong, WingKin Sung and Lance D. Miller: Differential Friendly Neighbors Algorithm for Differential Relationships Based Gene Selection and Classification using Microarray Data. Proc. The 2006 International Conference on Data Mining (DMINâ€™06) 2006, Las Vegas, NV.
Choi JK, Yu U, Yoo OJ, Kim S. Differential coexpression analysis using microarray data and its application to human cancer. Bioinformatics. 2005;21(24):4348â€“55.
Xu L, Geman D, Winslow RL. Largescale integration of cancer microarray data identifies a robust common cancer signature. BMC Bioinformatics. 2007;8:275.
Liu CC. Topologybased cancer classification and related pathway mining using microarray data. Nucleic Acids Res. 2006;34(14):4069â€“80.
Ray M, Zhang W. Analysis of Alzheimerâ€™s disease severity across brain regions by topological analysis of gene coexpression network. BMC Syst Biol. 2010;4:136.
Gillis J, Pavlidis P. A methodology for the analysis of differential coexpression across the human lifespan. BMC Bioinformatics. 2009;10:306.
Tai SK, Wu GI, Yuan S, Li KC. Genomewide expression links the electron transfer pathway of Shewanella oneidensis to chemotaxis. BMC Genomics. 2010;11:319.
De la Fuente A. From â€˜differential expressionâ€™ to â€˜differential networkingâ€™ identification of dysfunctional regulatory networks in diseases. Trends Genet. 2010;26(7):326â€“33.
Fuller TF, Ghazalpour A, Aten JE. Weighted gene coexpression network analysis strategies applied to mouse weight. Mamm Genome. 2007;18(6â€“7):463â€“72.
Southworth LK, Owen AB, Kim SK. Aging mice show a decreasing correlation of gene expression within genetic modules. PLoS Genet. 2009;5(12):e1000776.
Van Nas A, Guhathakurta D, Wang SS, Yehya N, Horvath S, Zhang B. Elucidating the role of gonadal hormones in sexually dimorphic gene coexpression networks. Endocrinology. 2009;150(3):1235â€“49.
Veale D, Ashcroft T, Marsh C, Gibson GJ, Harris AL. Epidermal growth factor receptors in nonsmall cell lung cancer. Br J Cancer. 1987;55:513â€“6.
Fang M, Wee SA, Ronski K, Fan H, Tao S, Lin Q. Evidence of EGR1 as a differentially expressed gene among proliferative skin diseases. Genomic Med. 2007;1(1â€“2):75â€“85.
Burton KH Chia, R Krishna Murthy Karuturi: Differential coexpression framework to quantify goodness of biclusters and compare biclustering algorithms. Algorithms Mol Biol 2010, 5(1), art. no.23.
Chu Jh, Lazarus R, Carey VJ, Raby BA. Quantifying differential gene connectivity between disease states for objective identification of diseaserelevant genes. BMC Syst Biol. 2011;5:89.
Miller LD, Smeds J, George J. An expression signature for p53 status in human breast cancer predicts mutation status, transcriptional effects, and patient survival. PNAS. 2005;102(38)
J.J. Sirvent, M.T. SalvadÃ³, M. SantafÃ©, p53 in breast cancer. Its relation to histological grade, lymphnode status, hormone receptors, cellproliferation fraction (ki67) and cerbB2. lmmunohistochemical study of 153 cases. Histol Histopathol 1995, 10: 531 539.
Fisher ER, Osborne CK, McGuire WL, Redmond C. Correlation of primary breast cancer histopathology and estrogen receptor content. Breast Cancer Res Treat. 1981;1(1):37â€“41.
Lupien M, Eeckhoute J, Meyer CA. FoxA1 translates epigenetic signatures into enhancerdriven lineagespecific transcription. Cell. 2008;132(6):958â€“70.
Botcheva K, McCorkle SR, McCombie WR. Distinct p53 genomic binding patterns in normal and cancerderived human cells. Cell Cycle. 2011;15(10(24)):4237â€“49.
Wei CL, Wu Q, Vega V. A global map of p53 transcriptionfactor binding sites in the human genome. Cell. 2006;124:207â€“19.
Rasti M, Arabsolghar R, Khatooni Z, MostafaviPour Z. p53 binds to estrogen receptor 1 promoter in human breast cancer cells. Pathol Oncol Res. 2012;18(2):169â€“75.
Theodorou V, Stark R, Menon S, Carroll JS. GATA3 acts upstream of FOXA1 in mediating ESR1 binding by shaping enhancer accessibility. Genome Res. 2012; doi:10.1101/gr.139469.112.
Caroll JS, Meyer CA, Song J. Genomewide analysis of estrogen receptor binding sites. Nat Genet. 2006;38:1289â€“97.
Lin CY, Vega VB, Thomsen JS. Wholegenome cartography of estrogen receptor Î± binding sites. PLoS Genet. 2007;3(6):e87.
BostrÃ¶m P, SÃ¶derstrÃ¶m M, Vahlberg T, SÃ¶derstrÃ¶m KO, Roberts PJ, CarpÃ©n O, HirsimÃ¤k P. MMP1 expression has an independent prognostic value in breast cancer. BMC Cancer. 2011;11:348.
Yu G, Wang L, Han Y and He Q (2012): clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS J Integrative Biol, (5), pp. 284â€“287.
Huang DW, Sherman BT, Lempicki RA. Bioinformatics enrichment tools: paths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Res. 2009;37(1):1â€“13.
Acknowledgements
The authors would like to thank Drs Joshy George and Foo Jia Nee for their helpful comments on the manuscript. We thank Swarna for helping in the early discussion of the project, Drs Juntao, Sigrid and Huaien for comments and discussion.
Funding
The publication cost of this article was paid by The Jackson Laboratory. The research presented in this study was supported by resources and technical expertise from the Genome Institute of Singapore, Singapore and The Jackson Laboratory, USA.
Availability of data and materials
The breast cancer dataset is available from the publication Miller et al. [22]. Analysis results are available as Additional file 1. Software is available at https://github.com/lianyh/MultiDCoX.
Author information
Authors and Affiliations
Contributions
RKMK conceived the project. RKMK and JCR guided HL to code the algorithm and carry out analysis. HL coded the algorithm. Both RKMK and HL drafted the manuscript and carried out all analyses. JCR revised the manuscript. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare that they have no competing interests.
Consent for publication
Not applicable. The data used for the study was obtained from public repositories.
Ethics approval and consent to participate
Not applicable
About this supplement
This article has been published as part of BMC Bioinformatics Volume 18 Supplement 16, 2017: 16th International Conference on Bioinformatics (InCoB 2017): Bioinformatics. The full contents of the supplement are available online at https://bmcbioinformatics.biomedcentral.com/articles/supplements/volume18supplement16
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Additional files
Additional file 1:
Results of Analysis of Breast Cancer Data. Contains all differentially coexpressed genesets with respective differential coexpression model fit (Ftest pvalue, coefficient value), gene counts, and permutation results over three factors (ER, p53 and Grade) in breast cancer data. Remarks: Gradeâ€‰+â€‰indicates higher grade tumor i.e. 2 and 3, while Gradeâ€“ indicates lower grade tumour i.e. 1. (XLS 804Â kb)
Additional file 2:
Functional analysis of joint and individual influence of cofactors on coexpression of genesets. Summary of GO terms and pathways enriched for joint and individual influence of different cofactors on coexpression of genests. Joint influence of cofactors is evident from the number of pathways and GO terms enriched for genesets whose coexpression is affected by more than one cofactor. (DOC 66Â kb)
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Cite this article
Liany, H., Rajapakse, J.C. & Karuturi, R.K.M. MultiDCoX: Multifactor analysis of differential coexpression. BMC Bioinformatics 18 (Suppl 16), 576 (2017). https://doi.org/10.1186/s1285901719637
Published:
DOI: https://doi.org/10.1186/s1285901719637