A biclustering algorithm based on a Bicluster Enumeration Tree: application to DNA microarray data

1756-0381-2-9 1756-0381 Research A biclustering algorithm based on a Bicluster Enumeration Tree: application to DNA microarray data Ayadi Wassim wassim.ayadi@gmail.com Elloumi Mourad mourad12345678@yahoo.com Hao Jin-Kao hao@info.univ-angers.fr

UTIC, Higher School of Sciences and Technologies of Tunis, 1008 Tunis, Tunisia

LERIA, Université d'Angers, 2 Boulevard Lavoisier, 49045 Angers, France

BioData Mining 1756-0381 2009 2 1 9 http://www.biodatamining.org/content/2/1/9 10.1186/1756-0381-2-9 20015398 20 7 2009 16 12 2009 16 12 2009 2009 Ayadi et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

Background

In a number of domains, like in DNA microarray data analysis, we need to cluster simultaneously rows (genes) and columns (conditions) of a data matrix to identify groups of rows coherent with groups of columns. This kind of clustering is called biclustering. Biclustering algorithms are extensively used in DNA microarray data analysis. More effective biclustering algorithms are highly desirable and needed.

Methods

We introduce BiMine, a new enumeration algorithm for biclustering of DNA microarray data. The proposed algorithm is based on three original features. First, BiMine relies on a new evaluation function called Average Spearman's rho (ASR). Second, BiMine uses a new tree structure, called Bicluster Enumeration Tree (BET), to represent the different biclusters discovered during the enumeration process. Third, to avoid the combinatorial explosion of the search tree, BiMine introduces a parametric rule that allows the enumeration process to cut tree branches that cannot lead to good biclusters.

Results

The performance of the proposed algorithm is assessed using both synthetic and real DNA microarray data. The experimental results show that BiMine competes well with several other biclustering methods. Moreover, we test the biological significance using a gene annotation web-tool to show that our proposed method is able to produce biologically relevant biclusters. The software is available upon request from the authors to academic users.

Background

DNA microarray technology is a revolutionary method enabling the measurement of expression levels of at least thousands of genes in a single experiment under diverse experimental conditions. This technology has found numerous applications in research and applied areas like biology, drug discovery, toxicological study and diseases diagnosis.

DNA microarray data is typically represented by a matrix where each cell represents the gene expression level of a gene under a particular experimental condition. One important analysis task of microarray data concerns the simultaneous identification of groups of genes that show similar expression patterns across specific groups of experimental conditions (samples) 1. Such an application can be addressed by a biclustering process whose aim is to discover coherent biclusters. That is, a bicluster is a subset of genes and conditions of the original expression matrix where the selected genes present a coherent behavior under all the experimental conditions contained in the bicluster.

More generally, biclustering has also applications in other domains such as text mining 23, target marketing 45, markets search 6, search in databases 78 and analyzing foreign exchange data 9.

Formally, let I = {1, 2, ..., n} denote the index set of n genes and J = {1, 2, ..., m} the index set of m conditions, a data matrix M(I, J) associated with I and J is a n*m matrix where the i^throw, i ∈ I, represents the i^thgene or attribute and the j^th, j ∈ J, column represents the j^thcondition or individual and m_ijof the i^throw and the j^thcolumn represents the value of the j^thcondition for the i^thgene. A bicluster in a data matrix M(I, J) is a couple (I', J') such that I'⊆ I and J'⊆ J. The biclustering problem can be formulated as follows: Given a data matrix M, construct a bicluster B_optassociated with M such that:

where f is an objective function measuring the quality, i.e., degree of coherence, of a group of biclusters and BC(M) is the set of all the possible groups of biclusters associated with M.

Clearly, biclustering is a highly combinatorial problem with a search space of order of O(2^|I|+|J|). In the general case, biclustering is known to be NP-hard 1. Consequently, most of the algorithms used to discover biclusters are based on heuristics to explore partially the combinatorial search space. The existing algorithms for biclustering can roughly be classified into two large families: systematic search methods and stochastic search methods (also called metaheuristic methods). Representative examples of systematic search methods include, among others, greedy algorithms 11011121314, divide and conquer algorithms 715 and enumeration algorithms 161718. On the other hand, among the metaheuristic methods, we can mention neighbourhood-based algorithms like simulated annealing 19, GRASP 20, evolutionary and hybrid algorithms 21222324. A recent review of various biclustering algorithms for biological data analysis is provided in 25.

Since the biclustering problem is a NP-hard problem and no single existing algorithm is completely satisfactory for solving the problem. It is useful to seek more effective algorithms for better solutions. In this paper, we introduce a new enumeration algorithm for biclustering of DNA microarray data, called BiMine. Our algorithm is based on three original features. First, BiMine relies on a new evaluation function called Average Spearman's rho (ASR) which is used to guide effectively the exploration of the search space. Second, BiMine uses a new tree structure, called Bicluster Enumeration Tree (BET), to represent conveniently the different biclusters discovered during the enumeration process. Third, to avoid the combinatorial explosion of the search tree, BiMine introduces a parametric rule that allows the enumeration process to cut tree branches that cannot lead to good biclusters.

To assess the performance of the proposed BiMine algorithm, we show computational results obtained on both synthetic and real datasets and compare our results with those from four state-of-the-art biclustering algorithms. Moreover, to evaluate the biological relevance of our resulting biclusters, we carry out a practical validation with respect to a specific Gene Ontology (GO) annotation with the help of a popular web tool.

Methods

A New Evaluation Function of Biclustering

Like any search algorithm, BiMine needs an evaluation function to assess the quality of a candidate bicluster. One possibility is to use the so-called Mean Squared Residue (MSR) function 1. Indeed, since its introduction, MSR has largely been used by biclustering algorithms, see for instance 11132021222627. However, MSR is known to be deficient to assess correctly the quality of certain types of biclusters 142829. In a recent work, Teng and Chan 14 proposed another function for bicluster evaluation called Average Correlation Value (ACV). However, the performance of ACV is known to be sensitive to errors 13.

In this paper, we propose a new evaluation function called Average Spearman's rho (ASR) based on Spearman's rank correlation. Let and be two vectors, the Spearman's rank correlation 30 expresses the dependency between the vectors X_iand X_j(denoted by ρ_ij) and is defined as follows:

where (resp. ) is the rank of (resp. ).

Let (I', J') be a bicluster in data matrix M(I, J), the ASR evaluation function is then defined by:

where:

ρ_{i, j}(i ≠ j) is the Spearman's rank correlation associated with the row indices i and j in the bicluster (I', J'). ρ_{k, l}(k ≠ l) is the Spearman's rank correlation associated with the column indices k and l in the bicluster (I', J').

Proposition 1: Let (I', J') be a bicluster in a data matrix M(I, J). We have:

Proof: Let us first show that:

Indeed, we have Spearman's rank correlations to calculate. According to 30, a Spearman's rank correlation belongs to [-1..1], we have then:

i.e.

It is easy to show in the same way that:

Hence:

i.e.:

With Spearman's rank correlation, a high (resp. low) value, close to 1 (resp. close to -1), indicates that the data is strongly (resp. weakly) correlated between two vectors 30. As shown above, ASR also takes values from [-1..1]. A high (resp. low) ASR value, close to 1 (resp. close to -1), indicates that the genes/conditions of the bicluster are strongly (resp. weakly) correlated.

Furthermore, in the next subsection, we want to assess the quality of the proposed ASR evaluation function in comparison with two popular functions MSR and ACV.

Studies of the ASR Evaluation Function

We compare the ASR evaluation function with Mean Squared Residue (MSR) 1. As mentioned previously, MSR is probably the most popular evaluation function and largely used in the literature. As a second reference function, we use Average Correlation Value (ACV) which was proposed very recently in 14.

For the comparison, we apply the evaluation functions (without using any algorithms), i.e., ASR, MSR and ACV, on seven matrices (biclusters) denoted by M₁to M₇(Figure 1). These matrices are employed in 1425 and represent all typical biclusters. They are defined as follows. M₁is a constant bicluster, M₂has constant rows, M₃has constant columns, M₄is composed of coherent values (additive model), M₅represents coherent values (multiplicative model), M₆contains coherent values (multiplicative model, where the first row of M₅is multiplied by 10) and M₇represents a coherent evolution.

Figure 1

Different typical Biclusters

Different typical Biclusters. Data matrix M₁represents a constant bicluster, M₂represents a constant rows bicluster, M₃represents a constant column bicluster, M₄represents coherent values (additive model), M₅represents coherent values (multiplicative model), M₆represents coherent values (multiplicative model, where the first row of M₅is multiplied by 10) and M₇represents a coherent evolution.

The values of ASR versus MSR and ACV are illustrated by Table 1 where the values of MSR and ACV were taken from 14.

Table 1

ASR versus MSR and ACV.

Biclusters

M ₁

M ₂

M ₃

M ₄

M ₅

M ₆

M ₇

Evaluation Functions

MSR

0.62

2.425

131.87

ACV

0.84

ASR

0.99

Concerning MSR, a low (resp. high) value, close to 0 (resp. higher than a fixed threshold), indicates that the genes/conditions of the bicluster are strongly (resp. weakly) correlated.

Concerning ACV, a high (resp. low) value, close to 1 (resp. close to 0), indicates that the genes/conditions of the bicluster are strongly (resp. weakly) correlated.

According to Table 1, the ASR, ACV and MSR functions are perfect to assess the quality of biclusters M₁, M₂, M₃and M₄. However, MSR is deficient on M₆and M₇, confirming the claim that MSR may have trouble on certain types of biclusters 142829. On the other hand, ASR and ACV are perfect to assess the quality of biclusters M₅and M₆but ASR is slightly better than ACV when applied on M₇.

BiMine Algorithm

We present now our biclustering algorithm called BiMine which uses ASR as its evaluation function and a new structure, called Bicluster Enumeration Tree (BET) to represent the different biclusters associated with a data matrix. We describe first the main procedure for building biclusters and then show an illustrative example to ease the understanding of the algorithm.

Let M be a data matrix, by using our algorithm, we operate in three steps: During the first step, we preprocess the data matrix M. During the second step, we construct a BET associated with M. Finally, during the last step, we identify the best biclusters.

Preprocessing

In the clustering area, preprocessing is often used to eliminate insignificant attributes (genes). For the biclustering, the preprocessing step aims to remove irrelevant expression values of the data matrix M that do not contribute in obtaining pertinent results. A value m_ijof M is considered to be insignificant if we have:

where avg_iis the average over the non-missing values in the i^throw, m_ijrepresents the intersection of row i with column j and δ is a fixed threshold. Equation 4 is applied for each value of M. See Tables 2 and 3 for an example.

Table 2

Data matrix M'.

C ₁

C ₂

C ₃

C ₄

C ₅

C ₆

I₁

I₂

I₃

I₄

I₅

Table 3

Data matrix M after preprocess.

C ₁

C ₂

C ₃

C ₄

C ₅

C ₆

I₁

I₂

I₃

I₄

I₅

By considering only non-missing values, we minimize the loss of information in the data matrix. This way of preprocessing missing values should be contrasted with other techniques. For instance, in 31, where the whole row is removed if the row contains at least one missing value or in 32, where the whole column is removed if it contains at least 5% of missing values. Furthermore, BiMine operates directly on the raw data matrix without resorting to a discretization of data, reducing thus the risk of loss of information.

Building Bicluster Enumeration Tree

After the preprocessing step, we construct a Bicluster Enumeration Tree (BET) that represents every possible bicluster that can be made from M. Compared to other data structure, BET permits to represent the maximum number of significant biclusters and the links that exist between these biclusters. Since the number of possible biclusters (nodes of BET) increases exponentially, BiMine employs parametric rules to help the enumeration process to close (or cut) a tree node. Intuitively, a node is cut down if the quality of the bicluster represented by this node is below a fixed threshold.

To describe formally our BiMine algorithm, let us define in the following the needed notations:

n_i: ith node order containing biclusters.

n_i.g_i: genes of n_i.

n_i.Cg_i: conditions of n_i.

bic: bicluster.

δ: threshold used in Equation 4.

Threshold: quality threshold according to ASR.

The BiMine algorithm (Figure 2 (Algorithm 1)) uses a first function to built an initial tree (Init_BET) which is recursively extended by a second function (BET-tree). Init_BET (Figure 2 (Function 1)) generates thus the different biclusters from data matrix M with one gene and significant conditions after using Equation 4. The root of BET is the empty bicluster (Line 1). The nodes at level one are the possible biclusters with one gene (Line 2-4). Notice that each node n_iis composed of two part n_i.g_i(genes) and n_i.Cg_i(significant conditions after the filter preprocessing with Equation 4). From these initial biclusters, new and larger biclusters are recursively built while pruning as soon as possible any bicluster if its ASR value doesn't reach a fixed Threshold. This is the role of the next function BET-tree.

Figure 2

BiMine algorithm

BiMine algorithm.

BET-tree (Figure 2 (Function 2)) creates recursively the BET (Line 13) and generates the set of the best biclusters. The i^thchild of a node is made up, on the one hand, of the union of the genes of the father node and the genes of the i^thuncle node, starting from the right side of the father. On the other hand, it is made up of the intersection of the conditions of the father and those of the i^thuncle starting from the right side of the father (Line 4-12). If the ASR value associated with the i^thchild is smaller than or equal to the given Threshold, then this child will be ignored (Line 6-11).

Notice that this parametric pruning rule based on a quality threshold is fully justified in this context. Indeed, if the current bicluster is not good enough, then it is useless to keep it because expanding such a bicluster leads certainly to biclusters of worse quality. From this point of view, the pruning rule shares similar principles largely applied in optimization methods like Dynamic Programming. In addition, this pruning rule is essential in reducing the tree size and remains indispensable for handling large datasets.

Finally, the union of the leaves of the constructed BET that are not included in other leaves and have at least two genes represents a good group of biclusters (Line 8-9).

Proposition 2: Time complexity of BiMine is O(2ⁿmlog(m)), where n is the number of rows and m is the number of columns of the data matrix.

Proof: Time complexity of the first step of BiMine is O(nm). Indeed, this step is achieved via a scanning of the whole data matrix M that is of size nm.

Time complexity of the second step of BiMine is O(2ⁿmlog(m)). Actually, in the worst case, we have 2ⁿnodes in the BET, representing the possible clusters of genes, each of which is associated with m conditions. On the other hand, since the conditions of the node are sorted, the construction of the intersection of two subsets of conditions of size m boils down to the search of m elements in a sorted array of size m. This can be done via a dichotomic search with a time complexity O(mlog(m)). Hence, the time complexity of the second step of BiMine is O(2ⁿmlog(m)). Thus, The time complexity of BiMine is O(2ⁿmlog(m)).

Illustrative Example

Let M' a data matrix (Table 2). During the first step, we make a preprocessing of M' to obtain the data matrix M (Table 3). The character "-" represents a removed insignificant value. During the second step, we construct a BET that represents every possible bicluster that can be made from M. Let us set δ = 0.1 and threshold of ASR = 1. The first level of the BET is made up of the nodes that represent the possible biclusters with one gene. Each node represents a row of data matrix M (Figure 3).

Figure 3

First level of BET

First level of BET.

The second level of the BET is made up of nodes that are the union of genes and the intersection of conditions in the first level.

In the Figure 4, we explain the construction of the children of node I₁. Each dashed edges without cross represents a valid combination between two nodes (with ASR = 1). First, we perform the union of genes of node labeled I₁with those of I₂(first uncle), and the intersection of {c₁, c₂, c₃, c₄, c₅} of I₁with those of {c₁, c₂, c₃, c₄, c₆} of I₂. The ASR of the obtained bicluster (I₁, I₂; c₁, c₂, c₃, c₄) is 1; hence we insert it as a first child of I₁. After that, we process I₁with node labeled I₃(second uncle). We obtain the bicluster (I₁, I₃; c₂, c₃, c₄, c₅) with ASR lower than 1, hence, this child bicluster of I₁is discarded. We carry out the same process with node I₄. We obtain the bicluster (I₁, I₄; c₁, c₂, c₃, c₄) with ASR equal to 1. We insert it as child of I₁. Finally, with I₅we obtain the bicluster (I₁, I₅; c₁, c₃, c₄, c₅) with ASR lower than 1; hence we don't insert it.

Figure 4

Children construction of the first node of the second level of BET

Children construction of the first node of the second level of BET.

We repeat the same process for the node I₂, I₃, I₄and I₅. This completes the second level of the BET (Figure 5).

Figure 5

Second level of BET

Second level of BET.

The third level of the BET is made up of nodes that are the union of genes and the intersection of conditions in the second level (Figure 6).

Figure 6

Last level of BET

Last level of BET.

At each level of the BET, we keep only nodes whose ASR is equal to 1. The union of the leaves of the constructed BET that are not included in other leaves is { (I₁, I₂, I₄; c₁, c₂, c₃, c₄), (I₃, I_5;c₃, c₄, c₅, c₆) }. This constitutes the group of biclusters (Figure 7).

Figure 7

Extracted biclusters are presented with bold line

Extracted biclusters are presented with bold line.

Results

In this section, we assess the BiMine algorithm on both synthetic and real DNA microarray data. We have implemented our algorithm in Java programming language. We compare BiMine results with the results of four prominent biclustering algorithms used by the community, named as: CC 1, OPSM 10, ISA 33 and Bimax 15. For these reference algorithms, we have used Biclustering Analysis Toolbox (BicAT) which is a recent software platform for clustering-based data analysis that integrates all these biclustering algorithms 34.

Synthetic Data

Data Sets

According to 141935, we generated randomly two types of synthetic datasets of size (I, J) = (200, 20). Different types of biclusters are embedded like constant columns, additive, multiplicative and coherent evolution biclusters. The first (resp. second) dataset contains biclusters without (resp. with) overlapping. To obtain statistically stable results, for each type of datasets, we generated 10 problem instances by randomly inserting the biclusters at different places in the data matrix.

Comparison Criteria

Following 35, we have used the following two ratios to evaluate our biclustering algorithm:

with

S_cb= Portion size of biclusters correctly extracted

Tot_size= Total size of correct biclusters

with

S_ncb= Portion size of biclusters not correctly extracted

Tot_size= Total size of corrected biclusters

The ratio θ_Shared(resp. θ_NotShared) expresses the percent of shared (resp. not shared) biclusters volume which corresponds (resp. not corresponds) with the real biclusters. In fact, when θ_Shared(resp. θ_NotShared) is equal to 100% the algorithm extracts the corrected (resp. not corrected) biclusters. A perfect solution have θ_Shared= 100% and θ_NotShared= 0%.

Protocol for Experiments

For our biclustering algorithm, we have fixed δ = 0.2 and threshold of ASR = 0.85. The parameter settings used for the four reference algorithms are the default values as used in 12. We run all the algorithms and we select the 4 biclusters obtained by each algorithm which best fit the 4 real biclusters. We compute the θ_Sharedand the θ_NotSharedfor each algorithm to show the averaged percentage of volume of the resulting biclusters which is shared and not shared with the real biclusters. The objective of this experiment is to determine which algorithm is able to extract all implanted biclusters.

Table 4 shows the best biclusters provided by each algorithm for the first dataset.

Table 4

BiMine results and comparison with other algorithms in synthetic data without overlapped biclusters.

Algorithms

θ _Shared

θ _NotShared

18.21%

36.57%

OPSM

46.39%

74.42%

ISA

39.38%

5.31%

Bimax

58.18%

21.39%

BiMine

100%

33.03%

As we can see in Table 4, BiMine can extract 100% of implanted biclusters with an extra volume that represent 33,03% of implanted biclusters. In fact, to obtain a new bicluster, combining two biclusters provide an extra volume only on conditions but give exactly the correct number of genes. However, the best of the studied algorithms, i.e., Bimax, can extract only 58.18% of implanted biclusters with 21.39% of extra volume. CC uses the MSR function of the selected elements as the biclustering criterion. When the signal of the implanted biclusters is weak, the greedy nature of CC may delete some rows and columns of the implanted biclusters in the beginning of the algorithm and miss the deleted rows and columns in the output biclusters. ISA uses only up-regulated and down-regulated constant expression values in its biclustering algorithm. When coherent biclusters exist, ISA may miss some rows and columns of the implanted biclusters. OPSM seeks only up and down regulation expression values with coherent evolution. Its performance decreases when there exist scenarios constant biclusters. The discretization preprocessing used by Bimax cannot identify the elements in the coherent biclusters. Hence, the algorithm cannot find exactly the implanted biclusters.

Table 5 illustrates the best biclusters provided by each algorithm for the second dataset.

Table 5

BiMine results and comparison with other algorithms in synthetic data with overlapped biclusters.

Algorithms

θ _Shared

θ _NotShared

9.21%

47.94%

OPSM

42.87%

49.31%

ISA

23.28%

23.97%

Bimax

34.07%

3.43%

BiMine

85.35%

41.78%

As we can see in Table 5, the results with BiMine present the highest coverage of the correct biclusters. In fact, BiMine can extract 85.35% of implanted biclusters with an extra volume that represent 41.78% of implanted biclusters. However, the best of the studied algorithms, i.e., OPSM, can extract only 42.87% of implanted biclusters with 49.31% of extra volume. To find overlapped biclusters in a given matrix, some algorithms, e.g., CC, need to mask the discovered biclusters with random values which is not necessary for BiMine. ISA and OPSM are sensitive to overlapping biclusters. They use the normalization step in the first preprocessing step of their algorithms. With overlapping biclusters, the expression value range after normalization becomes narrower. Table 5 shows that BiMine is marginally affected by the implanted overlap biclusters. We can conclude that BiMine can extract all implanted biclusters unlike other algorithms that can extract only certain types of biclusters.

Real data

Data Sets

We applied our approach to the well-known yeast cell-cycle dataset. This dataset is publicly available from 36 and described in 37 and processed in 1. It contains the expression profiles of more than 6000 yeast genes measured at 17 conditions over two complete cell cycles. In our experiments we use 2884 genes selected by 1.

Comparison Criteria

Two criteria are used. First, in order to evaluate the biological relevance of our proposed biclustering algorithm, we compute the p-values to indicate the quality of the extracted biclusters. Second, we identify the biological annotations for the extracted biclusters.

Protocol for Experiments

For our biclustering algorithm, we have fixed δ = 0.1 and threshold of ASR = 0.85. The parameter settings used for the different reference biclustering algorithms are the default settings as used in 12. For the first experiment, we run all the algorithms and we compute the p-value for extracted biclusters. With BiMine (resp. Bimax), we have obtained more than 1800 (resp. 3700) biclusters. Since a biological analysis on 1800 (resp. 3700) biclusters was not feasible, only the 100 biggest biclusters with high ASR were selected for analysis like Christinat et al. 38. Post-filtering was applied for all algorithms in order to eliminate insignificant biclusters like Cheng et al. 13. With the others algorithms, we obtained 10 biclusters for CC, 45 biclusters for ISA and 14 biclusters for OPSM. For the second experiment, we use a well-known web-tool to search for the significant shared Gene Ontology terms of the groups of genes.

Biological relevance

In order to evaluate the biological relevance of our proposed biclustering algorithm, we compare it with the results of CC, ISA, Bimax, OPSM on yeast cell-cycle dataset. The idea is to determine whether the set of genes discovered by biclustering algorithms shows significant enrichment with respect to a specific Gene Ontology (GO) annotation. We use the web-tool FuncAssociate 39 to evaluate the discovered biclusters. FuncAssociate computes the adjusted significance scores for each bicluster. Indeed, the adjusted significance scores assess genes in each bicluster by computing adjusted p-values, which indicates how well they match with the different GO categories. Note that a smaller p-value, close to 0, is indicative of a better match 37. Table 6 represents the different values of significant scores p-value for each algorithm over the percentage of total extracted biclusters. In fact with BiMine, 100% of tested biclusters have p-value = 5%. The same result is obtained with p-value = 1%. With p-value equals to 0.5% (resp. 0.1%), BiMine has 93% (resp. 82%) of biclusters. On the other hand, the best results (with the p-value is equals to 0.5% and 0.1% respectively) among the compared algorithms are obtained by Bimax with 89% (resp. 79%) of extracted biclusters. Finally, 51% of extracted biclusters with BiMine have p-value = 0.001% while those of Bimax have 64%. We note that BiMine performs well for all p-values compared to CC, ISA and OPSM. Also, BiMine performs well for four cases of p-value (p-value = 5%, p-value = 1%, p-value = 0.5% and p-value = 0.1%) over five compared to Bimax. Best results are obtained by BiMine and Bimax.

Table 6

Proportions of Biclusters significantly enriched by GO annotations.

p-value

0.5%

0.1%

0.001%

Algorithms

BiMine

100

OPSM

100

Bimax

100

ISA

Furthermore, in order to identify the biological annotations for the extracted biclusters we use GOTermFinder http://db.yeastgenome.org/cgi-bin/GO/goTermFinder which is a tool available in the Saccharomyces Genome Database (SGD). GOTermFinder is designed to search for the significant shared GO terms of the groups of genes and provides users with the means to identify the characteristics that the genes may have in common.

We present the significant shared GO terms (or parent of GO terms) used to describe the two selected set of genes (extracted by BiMine) with 11 genes × 11 conditions and 12 genes × 13 conditions in each bicluster with ASR equal to 0.8690 and 0.8873 respectively, for biological process, molecular function and cellular component. As 40, we report the most significant GO terms shared by these biclusters. For example, with the first bicluster (Table 7), the genes (YDL003W, YDL164C, YDR097C, YDR440W, YKL113C, YLL002W, YLR183C, YNL102W) are particularly involved in the process of cellular response to DNA damage stimulus, response to DNA damage stimulus, cellular response to stress, cellular response to stimulus, response to stress and response to stimulus.

Table 7

Most significant shared GO terms (process, function, component) for two biclusters on Yeast data.

Bicluster volume (genes × conditions)

Process Ontology

Function Ontology

Component Ontology

(12 × 13)

cellular response to DNA damage stimulus (66.7%, 1.87e-08)

response to DNA damage stimulus (66.7%, 6.30e-08)

cellular response to stress(66.7%, 2.12e-07)

cellular response to stimulus(66,7%, 3.25e-07)

DNA repair(50%, 2.58e-05)

response to stress(66.7%, 2.98e-05)

chromatin binding (25%,0.00037)

microtubule organizing center part(16.7%, 0.00742)

(11 × 11)

cell cycle process (63.6%, 2.93e-05)

cell cycle (63.6%, 6.85e-05)

GTPase activator activity (18.2%,0.00994)

microtubule cytoskeleton (45.5%, 6.33e-06)

microtubule organizing center (36.4%,4.97e-05)

spindle pole body (36.4%, 4.97e-05)

spindle pole (36.4%, 6.77e-05)

The values within parentheses after each GO term in Table 7, such as (66.7%, 1.87e-08) in the first bicluster, indicate the cluster frequency and the statistical significance. The cluster frequency (66.7%) shows that out of 12 genes in the first bicluster 8 belong to this process, and the statistical significance is provided by a p-value of 1.87e-08 (highly significant).

According to 414243, in microarray data analysis, genes are considered to be in the same cluster if their trajectory patterns of expression levels are similar across a set of conditions. Figure 8 shows the biclusters of Table 7 found by BiMine algorithm on the yeast dataset. From a visual inspection of the biclusters presented, we can notice that the genes present a similar behaviour under a subset of conditions. In Additional File 1, we show the best bicluster found by each compared algorithm using GoTermFinder. Also, we show their gene expression profiles drawn by BicAT. We notice that BiMine and Bimax have a high p-value. CC (resp. OPSM) cannot identify any component ontology (resp. function ontology) and ISA have p-value lower than BiMine.

Figure 8

Two Biclusters found by BiMine on Yeast dataset

Two Biclusters found by BiMine on Yeast dataset. (a): Bicluster of size (12 × 13) with ASR = 0.8873. (b): Bicluster of size (11 × 11) with ASR = 0.8690.

Additional file 1

The best bicluster obtained by each compared algorithm. This file illustrates the best bicluster found by each compared algorithm using GoTermFinder. The gene expression profile of each best bicluster is drawn using BicAT.

Click here for file

All these experiments show that for this dataset, the proposed approach is able to detect biologically significant and functionally enriched biclusters with low p-value. Furthermore, BiMine gives a good degree of homogeneity.

Discussion

BiMine algorithm has several interesting features. First, with BiMine, we avoid using a discretization of the data matrix. Indeed, classifying the gene expression values using intervals often leads to bad results 44. Also, the discretization may limit the performance of an algorithm to discover a biological model because of noises which are inherent in most experiences of microarrays 31. Thus, to discretize biological data we must have a good knowledge of these data to assign good values. However, this is not always possible.

Second, the BiMine algorithm can enumerate all possible cases of attributes while reducing the tree size. In fact, the parametric rule based on ASR threshold allows the enumeration process to prune tree branches that cannot lead to good biclusters.

Third, the BiMine algorithm provides naturally multiple biclusters of variable sizes. The number of the desired biclusters can be determined by tuning the ASR threshold. These multiple solutions of different sizes and different characteristics may be of interest for biological investigations.

Forth, the new ASR evaluation function can be applied by other biclustering algorithm in replacement of MSR or ACV. It can also be used as a complementary function to these previously ones.

Finally, in 45, it has been shown that Spearman's rank correlation is less sensitive to the presence of noise in the data. Since our evaluation function ASR is based on Spearman rank correlation, ASR would also be less sensitive to the presence of noise in the data.

Conclusions

In this paper, we described BiMine, a new algorithm for biclustering of DNA microarray data. Compared with existing biclustering algorithms, BiMine distinguishes itself by a number of original features. First, BiMine operates directly on the raw data matrix without resorting to a discretization of data, reducing thus the risk of loss of information. Second, with BiMine, it is not necessary to fix a minimum or maximum number of genes or conditions, enabling the generation of diversified biclusters. Third, using a convenient tree structure for representing biclusters with a parametric and effective branch pruning rule, BiMine is able to explore effectively the search space. Notice that ASR can also be used by other biclustering algorithm as an alternative evaluation function.

The performance of the BiMine algorithm is tested and assessed on a set of synthetic data as well as a real microarray data (yeast cell-cycle). Computational experiments showed highly competitive results of BiMine in comparison with four other popular biclustering algorithms for both types of datasets. In addition, a biological validation of the selected genes within the biclusters for yeast cell-cycle has been provided based on a publicly available Gene Ontology (GO) annotation tool. Notice that although we presented BiMine with the context of DNA microarray data analysis, it should be clear that the algorithm can be applied or adapted to other biclustering problems.

Finally, let us mention that the proposed algorithm is computational time expensive; one of our ongoing works aims to find new heuristics to speed up the enumeration process. In particular, it would be possible to define other heuristic rules to improve the branch pruning in order to further reduce the size of the explored search tree.

Competing interests

The authors declare that they have no competing interests.

Authors' contributions

WA implemented the system, conducted the experimentations and wrote the draft manuscript. ME and JKH supervised the project and co-wrote the manuscript. All authors read and approved the final manuscript.

Acknowledgements

The authors are grateful to Dr. Jason Moore and Dr. Federico Divina for their insightful comments and questions that helped us to improve the work.

Biclustering of expression data Cheng Y Church GM Proceedings of the Eighth International Conference on Intelligent Systems for Molecular Biology AAAI Press 2000 93 103 Information-theoretical coclustering Dhillon IS Mallela S Modha DS Proc. 9th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'03) 2003 89 98 full_text RCV1: A new benchmark collection for text categorization research Lewis DD Yang Y Rose T Li F Journal of Machine Learning Research 2004 5 361 97 Latent Class Models for Collaborative Filtering Hofmann T Puzicha J Proc. International Joint Conference on Artificial Intelligence 1999 668 693 Clustering by pattern similarity in large data sets Wang H Wang W Yang J Yu P SIGMOD '02: Proceedings of the international conference on Management of data ACM SIGMOD, New York, NY, USA 2002 394 405 full_text A new algorithm for two-mode clustering Gaul W Schader M Data Analysis and Information Systems Springer 1996 15 23 Direct clustering of a data matrix Hartigan JA Journal of the American Statistical Association 1978 67 337 123 129 10.2307/2284710 Automatic subspace clustering of high dimensional data for data mining applications Agrawal R Gehrke J Gunopulus D Raghavan P Proc. 1st ACM/SIGMOD International Conference on Management of Data 1998 94 105 Plaid models for gene expression data Lazzeroni L Owen A Statistica Sinica 2002 12 61 86 Discovering local structure in gene expression data: the order-preserving submatrix problem Ben-Dor A Chor B Karp R Yakhini Z J Comput Biol 2003 10 373 384 10.1089/10665270360688075 12935334 Enhanced biclustering on expression data Yang J Wang H Wang W Yu P Proceedings of the Third IEEE Symposium on Bioinformatics and Bioengineering (BIBE'03) 2003 1 7 Computing the maxim <it>um similarity bi-clusters of gene expression data</it> Liu X Wang L Bioinformatics 2007 23 1 50 56 10.1093/bioinformatics/btl560 17090578 Identification of coherent patterns in gene expression data using an efficient biclustering algorithm and parallel coordinate visualization Cheng K Law N Siu W Liew A BMC Bioinformatics 2008 9 210 10.1186/1471-2105-9-210 2396181 18433478 Discovering biclusters by iteratively sorting with weighted correlation coefficient in gene expression data Teng L Chan L J Signal Process Syst 2008 50 3 267 280 10.1007/s11265-007-0121-2 A systematic comparison and evaluation of biclustering methods for gene expression data Prelic A Bleuler S Zimmermann P Buhlmann P Gruissem W Hennig L Thiele L Zitzler E Bioinformatics 2006 22 9 1122 1129 10.1093/bioinformatics/btl060 16500941 Discovering statistically significant biclusters in gene expression data Tanay A Sharan R Shamir R Bioinformatics 2002 18 S136 S144 12169541 Op-cluster: Clustering by tendency in high dimensional space Liu J Wang W Proc.3rd IEEE International Conference on Data Mining 2003 187 194 Exhaustive Search Method of Gene Expression Modules and Its Application to Human Tissue Data Okada Y Okubo K Horton P Fujibuchi W IAENG International Journal of Computer Science 2007 34 1 16 Application of simulated annealing to the biclustering of gene expression data Bryan K Cunningham P Bolshakova N IEEE Transactions on Information Technology on Biomedicine 2006 10 3 519 525 10.1109/TITB.2006.872073 Biclustering of gene expression data using reactive greedy randomized adaptive search procedure Dharan A Nair AS BMC Bioinformatics 2009 10 Suppl 1 S27 10.1186/1471-2105-10-S1-S27 2648745 19208127 An EA framework for biclustering of gene expression data Bleuler S Prelic A Zitzler E Proceedings of Congress on Evolutionary Computation 2004 1 166 173 Multi-objective evolutionary biclustering of gene expression data Mitra S Banka H Pattern Recognition 2006 2464 2477 10.1016/j.patcog.2006.03.003 A Multi-Objective Approach to Discover Biclusters in Microarray Data Divina F Aguilar-Ruiz A Proceedings of the 9th annual conference on Genetic and evolutionary computation 2007 385 392 full_text Microarray Biclustering: A Novel Memetic Approach Based on the PISA Platform Gallo C Carballido J Ponzoni I EvoBIO: Proceedings of the 7th European Conference on Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics 2009 44 55 full_text Biclustering algorithms for biological data analysis: A survey Madeira SC Oliveira AL IEEE/ACM Transactions on Computational Biology and Bioinformatics (TCBB) 2004 1 1 24 45 10.1109/TCBB.2004.2 Mining deterministic biclusters in gene expression data Zhang Z Teo A Ooi BC Tan KL Proceedings of the Fourth IEEE Symposium on Bioinformatics and Bioengineering (BIBE'04) 2004 283 292 full_text Random walk biclustering for microarray data Angiulli F Cesario E Pizzuti C Journal of Information Sciences 2008 1479 1497 10.1016/j.ins.2007.11.007 Shifting and scaling patterns from gene expression data Aguilar-Ruiz JS Bioinformatics 2005 21 3840 3845 10.1093/bioinformatics/bti641 16144809 Virtual error: A new measure for evolutionary biclustering Pontes B Divina F Giraldez R Aguilar-Ruiz J Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics 2007 217 226 full_text Nonparametrics: Statistical Methods Based on Ranks Lehmann EL D'Abrera HJM rev. ed Englewood Cliffs, NJ: Prentice-Hall 1998 292 323 An efficient biclustering algorithm for finding genes with similar patterns in time-series expression data Madeira SC Oliveira AL Proc. of the 5th Asia Pacific Bioinformatics Conference, Series in Advances in Bioinformatics and Computational Biology Imperial College Press 2007 5 67 80 full_text Strategies for identifying statistically significant dense regions in microarray data Yip A Ng M Wu E Chan T IEEE/ACM Trans Comput Biol Bioinformatics 2007 4 3 415 429 10.1109/TCBB.2007.1022 Defining transcription modules using large-scale gene expression data Bergmann S Ihmels J Barkai N Bioinformatics 2004 13 1993 2003 Bicat: a biclustering analysis toolbox Barkow S Bleuler S Prelic A Zimmermann P Zitzler E Bioinformatics 2006 22 10 1282 1283 10.1093/bioinformatics/btl099 16551664 Possibilistic approach for biclustering microarray data Cano C Adarve L López J Blanco A Computers in Biology and Medicine 2007 37 1426 1436 10.1016/j.compbiomed.2007.01.005 17346690 Biclustering of expression data. (supplementary information) Cheng Y Church GM Technical report 2006 http://arep.med.harvard.edu/biclustering Systematic determination of genetic network architecture Tavazoie S Hughes JD Campbell MJ Cho RJ Church GM Nature Genetics 1999 22 281 285 10.1038/10343 10391217 Gene Expression Data Analysis Using a Novel Approach to Biclustering Combining Discrete and Continuous Data Christinat Y Wachmann B Zhang L IEEE/ACM Transactions on Computational Biology and Bioinformatics 2008 5 4 583 593 10.1109/TCBB.2007.70251 Charactering gene sets with FuncAssociate Berriz GF King OD Bryant B Sander C Frederick P Bioinformatics 2003 19 2502 2504 10.1093/bioinformatics/btg363 14668247 Combining Pareto-optimal clusters using supervised learning for identifying co-expressed genes Maulik U Mukhopadhyay A Bandyopadhyay S BMC Bioinformatics 2009 10 27 10.1186/1471-2105-10-27 2657792 19154590 Gene selection and clustering for time-course and dose-response microarray experiments using order-restricted inference Peddada SD Lobenhofer EK Li L Afshari CA Weinberg CR Umbach DM Bioinformatics 2003 19 834 841 10.1093/bioinformatics/btg093 12724293 Using hidden Markov models to analyze gene expression time course data Schliep A Schonhuth A Steinhoff C Bioinformatics 2003 19 i255 i263 10.1093/bioinformatics/btg1036 12855468 Clustering of time-course gene expression data using a mixed-effects model with B-splines Luan Y Li H Bioinformatics 2003 19 474 482 10.1093/bioinformatics/btg014 12611802 Improved biclustering of microarray data demonstrated through systematic performance tests Turner H Bailey T Krzanowski W Journal of Computational Statistics and Data analysis 2005 48 235 254 10.1016/j.csda.2004.02.003 Clustering of gene expression data using a local shape-based similarity measure Balasubramaniyan R llermeier H Weskamp E Kamper J Bioinformatics 2005 21 1069 1077 10.1093/bioinformatics/bti095 15513997