 Research
 Open Access
 Published:
A fast and high performance multiple data integration algorithm for identifying human disease genes
BMC Medical Genomicsvolume 8, Article number: S2 (2015)
Abstract
Background
Integrating multiple data sources is indispensable in improving disease gene identification. It is not only due to the fact that disease genes associated with similar genetic diseases tend to lie close with each other in various biological networks, but also due to the fact that genedisease associations are complex. Although various algorithms have been proposed to identify disease genes, their prediction performances and the computational time still should be further improved.
Results
In this study, we propose a fast and high performance multiple data integration algorithm for identifying human disease genes. A posterior probability of each candidate gene associated with individual diseases is calculated by using a Bayesian analysis method and a binary logistic regression model. Two prior probability estimation strategies and two feature vector construction methods are developed to test the performance of the proposed algorithm.
Conclusions
The proposed algorithm is not only generated predictions with high AUC scores, but also runs very fast. When only a single PPI network is employed, the AUC score is 0.769 by using F_{2} as feature vectors. The average running time for each leaveoneout experiment is only around 1.5 seconds. When three biological networks are integrated, the AUC score using F_{3} as feature vectors increases to 0.830, and the average running time for each leaveoneout experiment takes only about 12.54 seconds. It is better than many existing algorithms.
Background
The term disease broadly refers to any condition that impairs normal conditions of part or all of an organism. Among various diseases, genetic disorders are those related to disfunction of one or multiple genes in the human genome. A genetic disorder may arise from or lead to mutations of one or more genes, or associate with over/under expression of one or more genes [1]. This phenomenon is also a reflection of the module characteristic of real biological systems [2], where genes, proteins or other molecules often interact with each other to perform majority cellular processes [3–5]. Even disfunction of a single kind of gene may lead to disassembling some protein complexes or disturb a whole normal cellular pathway, thereby resulting in genetic disorders.
The issue of disease gene identification is to find those genetic disorder related genes, or called disease genes for short, for each specific genetic disease. Various kinds of evidence have shown that disease genes are not randomly distributed, but rather tend to lie close to each other in many biological networks if they are associated the same or similar diseases [1, 2, 6, 7].
Various kinds of biological data sources have shown their power for identifying disease genes. Oti et al. [2] use several sets of proteinprotein interaction (PPI) data to predict disease genes. They argue that the use of PPI data can greatly increase the prediction performance for disease gene identifications. Fraser et al. [8] investigate both yeast and human functional genomic data and argue that protein complexes contain valuable information which is helpful for detecting disease genes. Li et al. [9] investigate genetic diseases from a pathway based point of view. They find that individual pathways often enrich genes related to the same or similar diseases. Ma et al. [10] propose a combining gene expression and protein interaction (CGI) method to prioritize genes associated with a specific phenotype or trait. Ganegoda et al. [11] and Li et al. [12] use tissuespecific data together with PPI information to predict disease genes within individual tissues. Li et al. [13] also use Gene Ontology (GO) annotations to identify disease genes by combining topological features of PPI networks.
Besides different data sources, many different computational methods have also been employed for identifying disease genes. Lage et al. [14] propose a Bayesian method to analyze a phenomeinteractome network. Wu et al. [7] use a linear regression method to calculate the concordance score between a PPI network and a phenotype network. A tool called CIPHER is developed to predict disease genes based on those concordance scores. Vanunu et al. [15] formulate a smoothnessrelated prioritization function in a PPI network, which predicts not only disease genes but also disease associated protein complexes. Zhang et al. [16] develop a Bayesian regression approach to explain similarities of disease phenotypes by using diffusion kernels of one or several PPI networks. Köhler et al. [17] propose a random walk with restart (RWR) algorithm to detect disease genes by using a global network distance measure and random walk analysis.
Among those algorithms, the RWR algorithm [17] often yields better performance than other algorithms in terms of the prediction accuracy and the running time. However, the RWR algorithm can only take a single network as the input. When multiple kinds of biological networks need to be integrated, the RWR algorithm can only simply merge them into a mixed network as the input. Although this strategy can integrate useful information from different data sources, it integrates noises from them as well. Predictions of the RWR algorithm from a mixed network do not always perform better than those from individual networks. To improve the data integration method, Chen et al. [18] define a data integration rank (DIR) score to select the most informative evidence among a set of data sources. Chen et al. [19, 20] recently propose two improved Markov random field (MRF) algorithms, which can automatically assign weights to different data sources by using Gibbs sampling processes. They often yield better performance than those using only single data source, and the MRF algorithms are even more better than the the DIR method in terms of the prediction accuracy. However, the DIR algorithm is too timeconsuming due to the calculating of a normalized similarity measure for all gene pairs, while the MRF algorithms spend more time to maintain a long Markov chain for every gene during the Gibbs sampling processes.
In paper [21], we have proposed a logistic regression based algorithm to reduce the computational time of the MRF algorithm. It directly formulates the issue of disease gene identification as a binary logistic regression problem by using similar feature vectors as the MRF algorithm. No Markov chains need to be maintained for all genes, which makes the algorithm runs very fast. However, the logistic regression based algorithm in [21] is only a single network based algorithm, and the feature vector construction method is limited to using information of only direct neighbors. In this paper, we propose a fast and high performance multiple data integration algorithm to generalize the logistic regression based algorithm in [21]. Two aspects of generalization are proposed: (1) the generalization of the feature vector construction method; and (2) the extension of the application scope for using multiple data integrations. To be more specific, we first theoretically introduce how binary logistic regression model is used to formulate the disease gene identification issue. Then, the feature vector construction method is generalized by using not only direct neighbors but also higherorder neighborhood information in a network. After that, the logistic regression based algorithm is extended to the multiple data integration case, where the parameters (weights) of different data sources can be tuned automatically. A prior probability estimation method is also proposed by using protein complex information, together with a validation method and evaluation criteria. The numerical experiments show that the proposed algorithm not only achieves high AUC score, but also runs very fast even in the multiple data integration case. It outperforms many existing algorithms for identifying human disease genes.
Methods and materials
Problem formulation
Let H be a bipartite graph consisting of two disjoint sets of vertices, where one set represents all known human genes {g_{1}, g_{2}, . . . , g_{ N }}, while the other set represents all known genetic diseases {d_{1}, d_{2}, . . . , d_{ r }}. The associations between those genes and genetic diseases can be obtained from either the Online Mendelian Inheritance in Man (OMIM) database [22] or similar databases.
Although a disease d_{ k } may associate with one or several genes, the number of all known disease genes m is much smaller than N . Hence, associations of most other genes are still not known and need to be analyzed. Without loss of generality, we can reorder the set of all human genes as a vector (g_{1}, g_{2}, . . . , g_{N} ), according to a set of given genedisease associations, where g_{n+1}, g_{n+2}, . . . , g_{n+m}are genes associated with at lease one known disease (disease genes), and g_{1}, g_{2}, . . . , g_{ n }are others. Here N = n + m, and n is the number of all genes that are not known to associate with any diseases and they are called unknown genes in this paper.
For a specific disease d_{ k }, the issue of disease gene identification is to find a set of candidate genes which may have associations with d_{ k }. To achieve this, let ${x}^{k}=\left({x}_{1}^{k},\phantom{\rule{0.3em}{0ex}}{x}_{2}^{k},...,{x}_{n+m}^{k}\right)$ be a vector of binary class labels (i.e. taking the value zero or one) defined on all genes, where ${x}_{i}^{k}=1$ represents gene g_{ i } being a disease gene of d_{ k } , and ${x}_{i}^{k}=0$ otherwise. Since we have to address each genetic disease one by one, we take d_{ k }for example, and ignore the superscript k in the vector x^{k}for simplicity as x = (x_{1}, x_{2}, . . . , x_{n+m}), hereafter. Therefore, the identification of disease genes is equivalent to find labels of x_{ i } for all unknown genes. Identification of disease genes for other diseases can be similarly conducted by changing d_{ k } to another disease.
In this paper, the issue of disease gene identification is formulated as a twoclass classification problem by using Bayesian analysis and logistic regression. The conditional probability p(x_{ i } = 1 Φ) for each unknown gene is first calculated in an inference stage, and a decision score is then obtained according to this probability in a decision stage [23]. Here Φ represents the information used to make the inference, such as a vector of prior labels of x, the connectivity of the bipartite graph H, the neighborhood relationships of g_{1}, g_{2}, . . . , g_{ N } , and the similarity relationship between d_{1}, d_{2}, . . . , d_{ r } . The flow diagram of the proposed algorithm is depicted in Figure 1.
Prior label estimation
The logistic regression based algorithm needs a vector of prior labels for x. For those known disease genes, one can directly assign 1 or 0 according to the known genedisease associations. For those unknown disease genes, a prior probability of each gene get the label 1 should be first estimated.
The simplest way is to assign the prior probability as 0 for all unknown genes. The prediction results in this case is denoted as P_{0} hereafter.
However, one can make it better by using additional prior information, such as protein complex data, to estimate prior probabilities for unknown genes. This is not only due to the fact that they are naturally available from various databases, but also due to their capability to describe the module characteristic of disease genes. If a disease is resulted from the disfunction of a protein complex, then any component of the complex should associate with the disease with a high probability.
Similar to the method used in [19, 20], if a gene g_{ i } encodes a protein in a complex, then let
be its prior probability, where A is the number of disease genes of the specific disease in the complex, and B is the number of all disease genes in the complex. If g_{ i } appears in multiple protein complexes, we use the maximum value as its prior probability. If g_{ i } does not belong to any protein complex, let
be its prior probability, where C is the number of all currently known disease genes of the specific disease, and D is the total number of genes in human genome.
Once a prior probability ${\hat{p}}_{i}$ is estimated for g_{ i }, its prior label of ${\hat{x}}_{i}$ can be obtained as follows. First, generate a random number following the standard uniform distribution. If the value of the random number is large than ${\hat{p}}_{i}$, then assign 0 as the prior label for g_{ i }. Otherwise, assign 1 as the prior label for g_{ i }. Repeat this step for all unknown genes, one can obtain prior labels for all of them. The prediction results generated by using those prior labels is denoted as P_{ c } hereafter.
Logistic regression
For a twoclass classification problem, each gene is labelled with either 1 or 0. A vector of all binary values of x is called a configuration. In the previous MRF algorithm [19, 20], the configuration x is formulated as an MRF which follows a Gibbs distribution. However, the Markovianity characteristic of the MRF model makes it only considering direct neighbors to construct feature vectors, which limits the capability of the method to use other topological attributes in a biological network. It is also very time consuming to maintain Markov chains for all unknown genes. In [21], we have introduced a logistic regression based algorithm to directly estimate the configuration by using the same feature vectors. However, the application of the logistic regression based algorithm is still limited to single biological network. A multiple data integration method should be further investigated. To generalize the formulation of feature vectors by using other topological attributes and extend its applicability to multiple data integration, we propose an improved logistic regression based algorithm in this study as follows.
Let C_{1} be a set of genes with label 1 and C_{0} be a set of genes with label 0. Suppose the following four kinds of probabilities are given: the classconditional densities p(xC_{1}) and p(xC_{0}), which indicate the probability of the configuration x conditional on C_{1} and C_{0}, respectively, and the class prior densities p(C_{1}) and p(C_{0}), which indicate the prior probability of genes in C_{1} and C_{0} being labelled with 1 and 0, respectively.
According to the Bayes' rule, the posterior probabilities of those genes in C_{1} that are labelled with 1 can be described as a logistic sigmoid function [23, 24]
and the posterior probabilities of those genes in C_{0} that are labelled with 0 can be similarly written as
where the variable t is defined as
which is related to the four kinds of probabilities.
Although t is often unavailable for a real problem, under general assumptions [23], t can be formulated as a function of different features t = f (·) associated with the integrated networks. To be more specific, let $x$ be a prior configuration of all human genes and f be a function. For any given gene g_{ i }, let ϕ_{ i } be the feature vector of g_{ i } that is related to the prior configuration $\hat{x}$. The posterior probability that the specific gene g_{ i } has label 1 and 0 are
and
respectively. Note that the sum of these two probabilities (6) and (7) must equal to 1 in this twoclass classification problem. A linear function f (ϕ_{ i }) = w^{T} ϕ_{ i } with variables (feature vectors) ϕ_{ i } and coefficients (parameters) w is the most commonly used function to ensure the calculation of the posterior probability not too complex.
The key step of the proposed algorithm is the construction of feature vectors. In the previous methods [19, 21], the numbers of direct neighbors that connects to disease genes and nondisease genes are employed as the feature vector for each gene. Take g_{ i } for example, its feature vector can be written as
where ϕ_{i1 }and ϕ_{i0 }are the number of direct neighbors of g_{ i } that connected to vertices with labels 1 and 0, respectively. It is a three dimensional vector, where the first element represents the constant term. All feature vectors of individual genes together form a feature matrix as
where N is the number of all human genes. The corresponding parameters are w = (w_{ 0 }, w_{ 1 }, w_{ 2 })^{T}. Predictions generated by using (9) are denoted as F_{1} hereafter.
In this study, two extended feature vector construction methods are proposed as follows. Firstly, in a single biological network, not only the number of direct neighbors of g_{ i }, but also the number of its second order neighbors are employed to construct the feature vector as
where ϕ_{i1 }and ϕ_{i0 }are the numbers of direct neighbors of g_{ i } connected to vertices with labels 1 and 0, respectively, and ${\varphi}_{i1}^{\prime}$and ${\varphi}_{i0}^{\prime}$ are the numbers of the second order neighbors of g_{ i } connected to vertices with labels 1 and 0, respectively. The contribution of those indirect neighbors has been investigated for predicting disease genes in [20, 25, 26]. The feature matrix in this situation can be written as
The corresponding parameter vector w = (w_{0}, w_{1}, w_{2},w_{3}, w_{4})^{T} is a five dimensional vector. Predictions generated by using (11) are denoted as F_{2} hereafter.
Secondly, in the multiple data integration situation, suppose there are l biological networks. Let ${\varphi}_{i1}^{j},{\varphi}_{i0}^{j}$ be the number of direct neighbors of g_{ i } connected to vertices with labels 1 and 0 in the j^{th} network, respectively. The feature vector obtained from those l networks
is a 2l + 1 dimensional vector. All those feature vectors together form a feature matrix as
The corresponding parameter vector w = (w_{0}, w_{1}, w_{2},. . . , w_{2l−1}, w_{2l})^{T} is a 2l + 1 dimensional vector, and N is the number of all human genes. Predictions generated from (13) by integrating multiple networks is denoted as F_{3} hereafter.
Parameter estimation
Parameter estimation can be conducted on a training set consists of known disease genes, where known genes associated with d_{ k } are labelled with 1 and known genes associated with other diseases are labelled with 0. However, as we discussed in [19, 21], the exclusion of most unknown genes reduces the number of vertices with label 0 significantly, thereby making the estimation of parameters inaccurate. Predictions from those inaccurate parameters are unreliable in disease gene identification.
It is noteworthy that the majority of human genes should not be disease genes associated with d_{ k }. Hence, the inclusion of all unknown genes with prior labels as the training set will make the training set more reasonable, where the number of vertices with label 0 is significantly increased, while the number of vertices with label 1 does not change too much. Such a training set, which consists of both known genes and unknown genes, has proved its powerful and efficient to estimate meaningful parameters in [19–21].
Given a prior configuration $\widehat{x}$ for all vertices, a maximumlikelihood estimation (MLE) method can be employed to estimate the parameter vector w. The likelihood function can be written as
where x_{ i } is the label of g_{ i }, ϕ_{ i } is its feature vector that is calculated according to $\widehat{x}$, f is a linear function of ϕ_{ i } with the form as f (ϕ_{ i }) = w^{T} ϕ_{ i }, and N is the number of all human genes. The log likelihood of (14) is
ln L (w; x_{1}, x_{2}, . . . , x_{ N } )
The log likelihood (15) is a convex function [27]. Hence, we can find an unique global optimal solution by solving a convex optimization problem. In this study, the standard MATLAB function fminunc() is employed to find a numerical solution of (15) (by calculating the minimum of − ln L (w; x_{1}, x_{2}, . . . , x_{ N })). The the initial value of w is simply set as zero for the fminunc() function.
Decision score and evaluation methods
The logistic regression based algorithm returns a set of posterior probabilities during the inference stage. One can directly use those probabilities to make decisions in the following decision stage. However, the posterior probabilities do not always work well due to the hardness to set a threshold for a genetic disease. Inspired by the DIR method [18], we propose to use a percentage value of a posterior probability as the decision score for each gene. The decision score is calculated as follows
where {p_{1}, p_{2}, . . . , p_{ n }} is the posterior probabilities of individual unknown genes, and q_{ i } is the top percentage value of p_{ i } among all those posterior probabilities. A candidate gene is more likely to be associated with d_{ k }, if its decision score is larger than majority of others.
To evaluate the performance of the proposed algorithm, the leaveoneout cross validation paradigm is employed by using above decision scores. The receiver operating characteristic (ROC) curve is employed as one of the evaluation criteria, which shows the relationship between the true positive rate (TPR) and the false positive rate (FPR) by varying a threshold for determining positives. The area under the ROC curve (AUC) is employed to show the overall performance of algorithms.
The positive control genes are those known disease genes associated with d_{ k }. For those negative control genes, although they are indispensable to calculate false positives and true negatives, it is generally hard to obtain a true negative dataset [28]. In this study, the negative control genes are randomly selected from known disease genes that do not associate with d_{ k }. Since those genes have been widely studied as disease genes for other genetic diseases, it is less likely for them to be disease genes for a different specific disease. If there are s known disease genes associated with d_{ k }, we randomly select $\u230a\frac{s}{2}\u230b$ such genes as a negative control set. Each gene belonging to the negative control set is also validated by using the leaveoneout cross validation paradigm.
The proposed algorithm is compared with four previous algorithms: (1) the initial logistic regression based algorithm proposed in [21]; (2) the RWR algorithm proposed in [17]; (3) the MRF algorithm proposed in [19]; and (4) the DIR algorithm proposed in [18]. The first algorithm is applicable to a single network. The second and the third algorithms are applicable to both single network and multiple data integration. The fourth algorithm works only for multiple data integration. All those algorithms identify disease genes with high prediction performance and they work better than many previous methods [17–19, 21].
Algorithm
The stepbystep description of the proposed logistic regression based algorithm is given as follows.
Input: The vector of all human genes (g_{1}, . . . , g_{n+m}), where (g_{1}, . . . , g_{ n }) are unknown genes, and (g_{n+1}, . . . , g_{n+m}) are known genes; l integrated biological networks G_{1}, G_{2}, . . . , G_{ l }; a set of protein complexes; and a set of genedisease associations.
Output: The vector of decision score for each unknown gene for each disease.

1:
For a specific disease d_{ k } , calculate prior probabilities for all human genes, where the prior probability of unknown genes ${\widehat{p}}_{1},...,{\widehat{p}}_{n}$ are calculated according to (1) and (2).

2:
For each known gene g_{n+i}, i = 1, . . . , m, if g_{n+i}is known to be associated with d_{ k } , let ${\widehat{p}}_{n+i}=1.$ Otherwise, let ${\widehat{p}}_{n+i}=0.$

3:
Assign prior labels $\widehat{x}=\left({\widehat{x}}_{1},{\widehat{x}}_{2},...,{\widehat{x}}_{n},{\widehat{x}}_{n+1},...,{\widehat{x}}_{n+m}\right)$ for all genes according to the prior probabilities $\left({\widehat{p}}_{1},...,{\widehat{p}}_{n+m}\right),$ respectively.

4:
Calculate the feature vector ϕ_{ i } for each g_{ i } according to the integrated biological networks and $\widehat{x}$.

5:
Estimate parameters $\u0175$ by maximizing the log likelihood ln $\mathcal{L}\left(w;{x}_{1},{x}_{2},...,{x}_{N}\right)$ in (15) based on $\widehat{x}$ and ϕ_{ i }, i = 1, . . . , n + m. A binary logistic regression is performed here by taking the vector $\widehat{x}$ as the categorical dependent variables and those labelrelated feature vectors ϕ_{ i } as predictor variables. Here i = 1, . . . , N.

6:
Calculate the posterior probability p_{1}, . . . , p_{ n } for each unknown gene according to (6) by using $\u0175$ and ϕ_{ i }.

7:
Calculate the decision scores q_{1}, . . . , q_{ n } according to (16).

8:
Repeat all the steps for another disease until every disease is checked.
Results and discussion
Data sources
We use the same datasets as [19] in order to directly compare with previous methods. To be more specific, genedisease associations are collected from the Morbid Map list of the Online Mendelian Inheritance in Man (OMIM) [22]. Since a disease generally only associates with a few disease genes, it is hard to perform a logistic regression based on such small amount of positive samples. Hence, merging similar diseases into a disease class, and identifying disease genes associated with the disease class can circumvent this problem to some extent. Goh et al. [1] manually classify all diseases in OMIM into 22 primary disease classes. The dataset contains 1284 genetic diseases and 1777 disease genes. In this study, we use twelve disease classes that consist of 815 genes to test the performance of the proposed algorithm.
The PPI dataset is derived from the database of HPRD (Release 9) [29]. Duplicated edges between the same pair of vertices and selfloop edges are deleted. The final PPI network consists of 9465 vertices and 37039 edges. Two another PPI datasets are derived from the database of BioGrid (Release 3.2.108) [30] and the database of IntAct (downloaded on Jan 26, 2014) [31], respectively, which are used to select edges of biological networks.
The pathway datasets are obtained from the database of KEGG [32], Reactome [33], PharmGKB [34], and PIN [35]. There are 280, 1469, 99 and 2679 pathways in those datasets, respectively. The total number of proteins/genes consisting of those pathways is 8614. A pathway coexisting network is constructed by taking individual proteins/genes as vertices. Edges are constructed between two vertices, if they coexist in any pathway.
The human gene expression profiles are obtained from BioGPS (GSE1133) [36, 37], which contains 79 human tissues in duplicates, measured using the Affymetrix U133A array. Pairwise Pearson correlation coefficients (PCC) are calculated. A pair of genes are linked by an edge if the PCC value is large than 0.5, similar to the method used in [1, 18] to construct the gene coexpression network.
The human protein complexes are collected from the database of CORUM [38] and PCDq [39]. There are 1677 and 1103 protein complexes in datasets with at least two proteins, respectively. There are in total 3881 proteins in those protein complexes.
In summary, three kinds of biological networks are constructed and all protein (or gene) IDs are mapped onto the form of gene symbol. In order to test the performance of multiple data integration of our method, we selected those vertices that appear at least four times in all five biological networks (three PPI networks, a pathway coexisting network and a gene coexpression network). The final datasets consist of 7311 human genes, 815 out of which are known associated with 12 disease classes. The details of those datasets used in this study can be found in the "Availability of supporting data" section.
Comparisons between different priors
If there is no prior information available for the application of the proposed algorithm, zero prior P_{0} still works in most situations. However, if there is general prior information available in practice (such as the protein complex information), the proposed algorithm should work better than that using P_{0}.
Figure 2 compares the logistic regression based algorithm by using either the zero prior P_{0} or the protein complex prior P_{ c }. We can see from Figure 2 that P_{ c } always works better than P_{0} in all three kinds of feature vectors in terms of the AUC score. The highest improvement is achieved when F_{1} is employed, where the AUC score increases from 0.737 to 0.765. There is only slight improvement when F_{3} is employed in multiple data integration, where the AUC score increases from 0.821 to 0.830. This may due to the fact that F_{1} using P_{0} achieves the lowest prediction AUC score for identifying disease genes. It has the highest potential to be improved. While F_{3} using P_{0} in the multiple data integration already achieves a very high AUC score. There is only a little room for it to be further improved by using additional prior information.
Although the improvement of the protein complex information is not so significant for F_{2} and F_{3}, the increased AUC score still indicates that additional knowledge is helpful for improving the prediction performance. This characteristic makes the proposed algorithm very promising, since it is flexible in terms of the usage of different prior information. Any prior knowledge related to genedisease associations can be employed to estimate the prior labels.
Comparisons between different feature vectors
Figure 3 compares the logistic regression based algorithm by using different feature vectors. F_{1} and F_{2} are tested on the single HPRD PPI network, and F_{3} is tested by integrating the following three biological networks: (1) the HPRD PPI network, (2) the pathway coexisting network and (3) the gene coexpression network. They are the same experimental results as Figure 2 shows, but from a different point of view.
We can see from Figure 3 that F_{3} achieves the highest AUC score in both P_{0} and P_{ c }, while F_{1} always obtains the lowest AUC score. In the zero prior situation P_{0}, F_{1} reaches the AUC score at only 0.737, F_{2} on the same single PPI network reaches that at 0.766, while F_{3} by integrating three networks achieves the AUC score at 0.821. In the protein complex prior situation P_{ c }, the AUC score of F_{1} is 0.765. It increases to 0.769 by using F_{2} on the same single PPI network, and it continually rises to 0.830 by using F_{3} in the multiple data integration. Both F_{2} and F_{3} proposed in this study work better than the initial feature vector F_{1}.
Comparing with previous algorithms
To test the efficiency of the proposed algorithm, four previous algorithms are employed as comparison in either single network or multiple data integration. The initial logistic regression works only in single network, while the DIR algorithm works only in multiple data integration.
The comparison is first conducted in terms of the computational time. All those tests are conducted on a Windows 7 professional computer (Inter(R) Core(TM) i7 CPU, 3.07 GHz, 8.0 GB RAM, 64bit OS). The MATLAB version is 7.10.0.499 (R2010a), 64bit (win 64). Each algorithm is evaluated by using the leaveoneout cross validation paradigm, where each known gene is left out once. The probabilities of all unknown genes (include the left out one) are calculated by using each algorithm. All algorithms are conducted on the same datasets and the same computational conditions. Figure 4 illustrates the average computational time for each leaveoneout experiment among different algorithms.
We can see from Figure 4 that the MRF algorithm is the slowest algorithm. A leaveoneout experiment spends around 95.72 seconds for the single network, and it increases to about 273.77 seconds when three biological networks are integrated. The initial logistic regression based algorithm F_{1} runs very fast. It only spends approximately 0.54 seconds in the single network. The improved logistic regression based algorithms F_{2} and F_{3} also runs very fast. It only takes around 1.5 seconds when F_{2} is used, and it increases to about 12.54 seconds when three biological networks are integrated by using F_{3}, which is almost the same as the DIR algorithm (11.52 seconds). The RWR algorithm also runs very fast, and it does not vary too much in both situations. It is due to the fact that the RWR algorithm uses the mixed network as input. No matter how many networks are integrated, it combines them together as a single mixed network. Hence, the number of integrated networks does not affect the computational time significantly.
The comparison is then conducted in terms of the AUC scores. When only the single HPRD PPI network is employed, as illustrated in Figure 5(a), the proposed logistic regression based algorithm using F_{2} works better than all of the previous single network based algorithms. The AUC score is 0.766, which achieves 2.9%, 4.5% and 2.9% improvements compared with the initial logistic regression based algorithm using F_{1}, the MRF algorithm and the RWR algorithm, respectively. When three biological networks are employed, as illustrated in Figure 5(b), the proposed logistic regression based algorithm using F_{3} achieves the highest AUC score among all these multiple data integration algorithms. The AUC score is 0.830 when protein complex prior P_{ c } is used, which is 9.9%, 11.9% and 11.4% improvements compared with the MRF algorithm, the RWR algorithm and the DIR algorithm under the same situation, respectively.
The comparison is finally conducted in terms of the AUC scores for each of the 12 disease classes. It can be seen from the Figure 6(a) that the proposed logistic regression based algorithm (F_{2}P_{ c }) is the most stable algorithm in the single network. Its AUC score is larger than the other three algorithms in many cases. When three biological networks are integrated, the proposed logistic regression based algorithm (F_{3}P_{ c }) achieves the highest AUC score in all cases, which makes the algorithm very promising in terms of multiple data integration.
Conclusions
In this paper, we have proposed an improved logistic regression based algorithm to identify disease genes by using either a single network or multiple networks. A Bayesian analysis method is first used to formulated the disease gene identification issue as a twoclass classification problem. A binary logistic regression model is then employed to calculate the posterior probability of each unknown gene obtained the label 1. Parameters of the model are estimated based on the whole gene set, and the final decision scores are obtained by using the percentage values of individual posterior probabilities.
Compared with previous algorithms, the proposed logistic regression based algorithm not only runs fast, but also generates predictions with high AUC scores. It only takes around 1.50 seconds in the single PPI network, and the AUC score is larger than all of the three single network based competing algorithms. Although the running time for the multiple networks is a little longer than the RWR algorithm and the DIR algorithm, it is still comparable, and the AUC score of the proposed algorithm is much better than those two algorithms. Compared with the MRF algorithm, the computational time has been significantly reduced, while the predictive performance becomes much better in terms of the AUC score. The best AUC score of the proposed algorithm is 0.766 in the single network, and it increases to 0.830 if three networks are integrated. The high prediction performance and the short computation time make the proposed algorithm very promising for identifying human disease genes.
Availability of supporting data
The Matlab code of the proposed algorithm with data can be found in https://www.dropbox.com/s/bs0ekmu718u4sea/Package15.zip
Abbreviations
 AUC:

area under the ROC curve
 CGI:

combining gene expression and protein interaction
 DIR:

data integration rank
 FPR:

false positive rate
 MRF:

Markov random field
 OMIM:

online Mendelian inheritance in man
 PCC:

Pearson correlation coefficient
 PPI:

proteinprotein interaction
 ROC:

receiver operating characteristic
 RWR:

random walk with restart
 TPR:

true positive rate.
References
 1.
Goh KI, Cusick ME, Valle D, Childs B, Vidal M, Barabási AL: The human disease network. Proc Natl Acad Sci USA. 2007, 104 (21): 86858690. 10.1073/pnas.0701361104.
 2.
Oti M, Brunner HG: The modular nature of genetic diseases. Clin Genet. 2007, 71 (1): 111.
 3.
Snel B, Bork P, Huynen MA: The identification of functional modules from the genomic association of genes. Proc Natl Acad Sci USA. 2002, 99 (9): 58905895. 10.1073/pnas.092632599.
 4.
Barabási AL, Oltvai ZN: Network biology: understanding the cell's functional organization. Nat Rev Genet. 2004, 5 (2): 101113. 10.1038/nrg1272.
 5.
Chen B, Fan W, Liu J, Wu FX: Identifying protein complexes and functional modules from static PPI networks to dynamic PPI networks. Brief Bioinform. 2014, 15 (2): 177194. 10.1093/bib/bbt039.
 6.
Oti M, Snel B, Huynen MA, Brunner HG: Predicting disease genes using proteinprotein interactions. J Med Genet. 2006, 43 (8): 691698. 10.1136/jmg.2006.041376.
 7.
Wu X, Jiang R, Zhang MQ, Li S: Networkbased global inference of human disease genes. Mol Syst Biol. 2008, 4: 189
 8.
Fraser HB, Plotkin JB: Using protein complexes to predict phenotypic effects of gene mutation. Genome Biol. 2007, 8 (11): R25210.1186/gb2007811r252.
 9.
Li Y, Agarwal P: A PathwayBased View of Human Diseases and Disease Relationships. PLoS One. 2009, 4 (2): e434610.1371/journal.pone.0004346.
 10.
Ma X, Lee H, Wang L, Sun F: CGI: a new approach for prioritizing genes by combining gene expression and proteinprotein interaction data. Bioinformatics. 2007, 23 (2): 215221. 10.1093/bioinformatics/btl569.
 11.
Ganegoda G, Wang J, Wu FX, Li M: Prediction of disease genes using tissuespecified genegene network. BMC Syst Biol. 2014, 8 (Suppl 3): S310.1186/175205098S3S3.
 12.
Li M, Zhang J, Liu Q, Wang J, Wu FX: Prediction of diseaserelated genes based on weighted tissuespecific networks by using DNA methylation. BMC Med Genomics. 2014, 7 (Suppl 2): S410.1186/175587947S2S4.
 13.
Li M, Li Q, Ganegoda G, Wang J, Wu FX, Pan Y: Prioritization of orphan diseasecausing genes using topological feature and GO similarity between proteins in interaction networks. Sci China Life Sci. 2014, 57 (11): 10641071. 10.1007/s1142701447476.
 14.
Lage K, Karlberg EO, Størling ZM, Olason PI, Pedersen AG, Rigina O, et al: A human phenomeinteractome network of protein complexes implicated in genetic disorders. Nat Biotechnol. 2007, 25 (3): 309316. 10.1038/nbt1295.
 15.
Vanunu O, Magger O, Ruppin E, Shlomi T, Sharan R: Associating genes and protein complexes with disease via network propagation. PLoS Comput Biol. 2010, 6 (1): e100064110.1371/journal.pcbi.1000641.
 16.
Zhang W, Sun F, Jiang R: Integrating multiple proteinprotein interaction networks to prioritize disease genes: a Bayesian regression approach. BMC Bioinformatics. 2011, 12 (Suppl 1): S1110.1186/1471210512S1S11.
 17.
Köhler S, Bauer S, Horn D, Robinson PN: Walking the interactome for prioritization of candidate disease genes. Am J Hum Genet. 2008, 82 (4): 949958. 10.1016/j.ajhg.2008.02.013.
 18.
Chen Y, Wang W, Zhou Y, Shields R, Chanda SK, Elston RC, et al: In silico gene prioritization by integrating multiple data sources. PLoS One. 2011, 6 (6): e2113710.1371/journal.pone.0021137.
 19.
Chen B, Wang J, Li M, Wu FX: Identifying disease genes by integrating multiple data sources. BMC Medical Genomics. 2014, 7 (Suppl 2): S210.1186/175587947S2S2.
 20.
Chen B, Li M, Wang J, Wu FX: Disease gene identification by using graph kernels and Markov random fields. Sci China Life Sci. 2014, 57 (11): 10541063. 10.1007/s1142701447458.
 21.
Chen B, Li M, Wang J, Wu FX: A logistic regression based algorithm for identifying human disease genes. 2014 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). 2014, Belfast. IEEE, 2014: 197200. 25 Nov.
 22.
McKsick VA: Mendelian Inheritance in Man and its online version, OMIM. Am J Hum Genet. 2007, 80 (4): 588604. 10.1086/514346.
 23.
Bishop CM: Pattern Recognition and Machine Learning. 2006, Singapore: Springer
 24.
Shi J, Chen B, Wu FX: Unifying protein inference and peptide identification with feedback to update consistency between peptides. Proteomics. 2013, 13 (2): 239247. 10.1002/pmic.201200338.
 25.
Lee H, Tu Z, Deng M, Sun F, Chen T: Diffusion kernel based logistic regression models for protein function prediction. OMICS. 2006, 10 (1): 4055. 10.1089/omi.2006.10.40.
 26.
Li SZ: Markov Random Field Modeling in Image Analysis. 2009, London: Springer
 27.
Boyd SP, Vandenberghe L: Convex optimization. 2004, New York: Cambridge University Press
 28.
Mackay JP, Sunde M, Lowry JA, Crossley M, Matthews JM: Response to Chatraryamontri et al.: Protein interactions: to believe or not to believe?. Trends Biochem Sci. 2008, 33 (6): 242243. 10.1016/j.tibs.2008.04.003.
 29.
Keshava Prasad TS, Goel R, Kandasamy K, Keerthikumar S, Kumar S, Mathivanan S, et al: Human Protein Reference Database  2009 update. Nucleic Acids Res. 2009, 37 (Database): D767D772. 10.1093/nar/gkn892.
 30.
Stark C, Breitkreutz BJ, Reguly T, Boucher L, Breitkreutz A, Tyers M: BioGRID: a general repository for interaction datasets. Nucleic Acids Res. 2006, 34 (Database): D535539.
 31.
Kerrien S, AlamFaruque Y, Aranda B, Bancarz I, Bridge A, Derow C, et al: IntAct  open source resource for molecular interaction data. Nucleic Acids Res. 2007, 35 (Database): D561D565. 10.1093/nar/gkl958.
 32.
Kanehisa M, Goto S: KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res. 2000, 28 (1): 2730. 10.1093/nar/28.1.27.
 33.
Vastrik I, D'Eustachio P, Schmidt E, Gopinath G, Croft D, de Bono B, et al: Reactome: a knowledge base of biologic pathways and processes. Genome Biol. 2007, 8 (3): R3910.1186/gb200783r39.
 34.
WhirlCarrillo M, McDonagh EM, Hebert JM, Gong L, Sangkuhl K, Thorn CF, et al: Pharmacogenomics knowledge for personalized medicine. Clin Pharmacol Ther. 2012, 92 (4): 414417. 10.1038/clpt.2012.96.
 35.
Schaefer CF, Anthony K, Krupa S, Buchoff J, Day M, Hannay T, et al: PID: the Pathway Interaction Database. Nucleic Acids Res. 2009, 37 (Database): D674D679. 10.1093/nar/gkn653.
 36.
Wu C, Orozco C, Boyer J, Leglise M, Goodale J, Batalov S, et al: BioGPS: an extensible and customizable portal for querying and organizing gene annotation resources. Genome Biol. 2009, 10 (11): R13010.1186/gb20091011r130.
 37.
Su AI, Wiltshire T, Batalov S, Lapp H, Ching KA, Block D, et al: A gene atlas of the mouse and human proteinencoding transcriptomes. Proc Natl Acad Sci USA. 2004, 101 (16): 60626067. 10.1073/pnas.0400782101.
 38.
Ruepp A, Waegele B, Lechner M, Brauner B, DungerKaltenbach I, Fobo G, et al: CORUM: the comprehensive resource of mammalian protein complexes  2009. Nucleic Acids Res. 2010, 38 (Database): D497D501. 10.1093/nar/gkp914.
 39.
Kikugawa S, Nishikata K, Murakami K, Sato Y, Suzuki M, AltafUlAmin M, et al: PCDq: human protein complex database with quality index which summarizes different levels of evidences of protein complexes predicted from hinvitational proteinprotein interactions integrative dataset. BMC Syst Biol. 2012, 6 (Suppl 2): S710.1186/175205096S2S7.
Declarations
The publication costs for this article were supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), the National Natural Science Foundation of China (Grant No. 61332014 and 61370024), and the fundamental research program of the Northwestern Polytechnical University (Grant No. G2015KY0104 and G2015KY0302).
This article has been published as part of BMC Medical Genomics Volume 8 Supplement 3, 2015: Selected articles from the IEE International Conference on Bioinformatics and Biomedicine (BIBM 2014): Medical Genomics. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcmedgenomics/supplements/8/S3.
Author information
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
FXW initiated this study, FXW and BC designed algorithms and experiments. BC performed the experiments, analyzed the results, and drafted the manuscript. FXW, ML, JW and XS revised the manuscript. All authors have read and approved the final manuscript.
Rights and permissions
About this article
Published
DOI
Keywords
 disease gene
 Bayesian analysis
 logistic regression
 multiple data integration
 feature vector