 Methodology
 Open Access
 Published:
The nonnegative matrix factorization toolbox for biological data mining
Source Code for Biology and Medicine volume 8, Article number: 10 (2013)
Abstract
Background
Nonnegative matrix factorization (NMF) has been introduced as an important method for mining biological data. Though there currently exists packages implemented in R and other programming languages, they either provide only a few optimization algorithms or focus on a specific application field. There does not exist a complete NMF package for the bioinformatics community, and in order to perform various data mining tasks on biological data.
Results
We provide a convenient MATLAB toolbox containing both the implementations of various NMF techniques and a variety of NMFbased data mining approaches for analyzing biological data. Data mining approaches implemented within the toolbox include data clustering and biclustering, feature extraction and selection, sample classification, missing values imputation, data visualization, and statistical comparison.
Conclusions
A series of analysis such as molecular pattern discovery, biological process identification, dimension reduction, disease prediction, visualization, and statistical comparison can be performed using this toolbox.
Background
Nonnegative matrix factorization (NMF) is a matrix decomposition approach which decomposes a nonnegative matrix into two lowrank nonnegative matrices [1]. It has been successfully applied in the mining of biological data.
For example, Ref. [2, 3] used NMF as a clustering method in order to discover the metagenes (i.e., groups of similarly behaving genes) and interesting molecular patterns. Ref. [4] applied nonsmooth NMF (NSNMF) for the biclustering of gene expression data. Leastsquares NMF (LSNMF) was proposed to take into account the uncertainty of the information present in gene expression data [5]. Ref. [6] proposed kernel NMF for reducing dimensions of gene expression data.
Many authors indeed provide their respective NMF implementations along with their publications so that the interested community can use them to perform the same data mining tasks respectively discussed in those publications. However, there exists at least three issues that prevent NMF methods from being used by the much larger community of researchers and practitioners in the data mining, biological, health, medical, and bioinformatics areas. First, these NMF softwares are implemented in diverse programming languages, such as R, MATLAB, C++, and Java, and usually only one optimization algorithm is provided in their implementations. It is inconvenient for many researchers who want to choose a suitable NMF method or mining task for their data, among the many different implementations, which are realized in different languages with different mining tasks, control parameters, or criteria. Second, some papers only provide NMF optimization algorithms at a basic level rather than a data mining implementation at a higher level. For instance, it becomes hard for a biologist to fully investigate and understand his/her data when performing clustering or biclustering of his data and then visualize the results; because it should not be necessary for him/her to implement these three data mining methods based on a basic NMF. Third, the existing NMF implementations are applicationspecific, and thus, there exists no systematic NMF package for performing data mining tasks on biological data.
There currently exists NMF toolboxes (which we discuss in this paragraph), however, none of them addresses the above three issues altogether.
NMFLAB[7] is MATLAB toolbox for signal and image processing which provides a userfriendly interface to load and process input data, and then save the results. It includes a variety of optimization algorithms such as multiplicative rules, exponentiated gradient, projected gradient, conjugate gradient, and quasiNewton methods. It also provides methods for visualizing the data signals and their components, but does not provide any data mining functionality. Other NMF approaches such as semiNMF and kernel NMF are not implemented within this package.
NMF:DTU Toolbox[8] is a MATLAB toolbox with no data mining functionalities. It includes only five NMF optimization algorithms, such as multiplicative rules, projected gradient, probabilistic NMF, alternating least squares, and alternating least squares with optimal brain surgery (OBS) method.
NMFN: Nonnegative Matrix Factorization[9] is an R package similar to NMF:DTU but with few more algorithms.
NMF: Algorithms and framework for Nonnegative Matrix Factorization[10] is another R package which implements several algorithms and allows parallel computations but no data mining functionalities.
Text to Matrix Generator (TMG) is a MATLAB toolbox for text mining only.
Ref. [11] provides a NMF plugin for BRBArrayTools. This plugin only implements the standard NMF and semiNMF and for clustering gene expression profiles only.
Coordinated Gene Activity in Pattern Sets (CoGAPS) [12] is a new package implemented in C++ with R interface. In this package, the Bayesian decomposition (BD) algorithm is implemented and used in place of the NMF method for factorizing a matrix. Statistical methods are also provided for the inference of biological processes. CoGAPS can give more precise results than NMF methods [13]. However, CoGAPS uses a Markov chain Monte Carlo (MCMC) scheme for estimating the BD model parameters, which is slower than the NMFs optimization algorithms implemented with the blockcoordinate gradient descent scheme.
In order to address the lack of data mining functionalities and generality of current NMF toolboxes, we propose a general NMF toolbox in MATLAB which is implemented in two levels. The basic level is composed of the different variants of NMF, and the top level consists of the diverse data mining methods for biological data. The contributions of our toolbox are enumerated in the following:

1.
The NMF algorithms are relatively complete and implemented in MATLAB. Since it is impossible and unnecessary to implement all NMF algorithms, we focus only on wellknown NMF representatives. This repository of NMFs allows users to select the most suitable one in specific scenarios.

2.
Our NMF toolbox includes many functionalities for mining biological data, such as clustering, biclustering, feature extraction, feature selection, and classification.

3.
The toolbox also provides additional functions for biological data visualization, such as heatmaps and other visualization tools. They are pretty helpful for interpreting some results. Statistical methods are also included for comparing the performances of multiple methods.
The rest of this paper is organized as below. The implementations of the basis level are first discussed in the next section. After that, examples of implemented data mining tasks at a high level are described. Finally, we conclude this paper and give possible avenues for future research directions.
Implementation
As mentioned above, this toolbox is implemented at two levels. The fundamental level is composed of several NMF variants and the advanced level includes many data mining approaches based on the fundamental level. The critical issues in implementing these NMF variants are addressed in this section. Table 1 summarizes all the NMF algorithms implemented in our toolbox. Users (researchers, students, and practitioners) should use the command help nmfrule, for example, in the command line, for help on how to select a given funtion and set its parameters.
StandardNMF
The standardNMF decomposes a nonnegative matrix \mathit{X}\in {\mathbb{R}}^{m\times n} into two nonnegative factors \mathit{A}\in {\mathbb{R}}^{m\times k} and \mathit{Y}\in {\mathbb{R}}^{k\times n} (where k< min{m,n}), that is
where, E is the error (or residual) and M_{+} indicates the matrix M is nonnegative. Its optimization in the Euclidean space is formulated as
Statistically speaking, this formulation is obtained from the loglikelihood function under the assumption of a Gaussian error. If multivariate data points are arranged in the columns of X, then A is called the basis matrix and Y is called the coefficient matrix; each column of A is thus a basis vector. The interpretation is that each data point is a (sparse) nonnegative linear combination of the basis vectors. It is wellknown that the optimization objective is a nonconvex optimization problem, and thus, blockcoordinate descent is the main prescribed optimization technique for such problem. Multiplicative update rules were introduced in [14] for solving Equation (2). Though simple to implement, this algorithm is not guaranteed to converge to a stationary point [15]. Essentially the optimizations above, with respect to A and Y, are nonnegative least squares (NNLS). Therefore we implemented the alternating NNLS algorithm proposed in [15]. It can be proven that this algorithm converges to a stationary point. In our toolbox, functions nmfrule and nmfnnls are the implementations of the two algorithms above.
SemiNMF
The standard NMF only works for nonnegative data, which limits its applications. Ref. [16] extended it to semiNMF which removes the nonnegative constraints on the data X and basis matrix A. It can be expressed in the following equation:
SemiNMF can be applied to the matrix of mixed signs, therefore it expands NMF to many fields. However, the gradientdescentbased update rule proposed in [16] is slow to converge (implemented in function seminmfrule in our toolbox). Keeping Y fixed, updating A is a least squares problem which has an analytical solution
where Y^{‡}=Y^{T}(Y Y^{T})^{−1} is MoorePenrose pseudoinverse. Updating Y while fixing A is a NNLS problem essentially as above. Therefore we implemented the fast NNLS based algorithm to optimize semiNMF in function seminmfnnls.
SparseNMF
The standard NMF and semiNMF have the issues of scalevariance and nonunique solutions, which imply that the nonnegativity constrained on the least squares is insufficient in some cases. Sparsity is a popular regularization principle in statistical modeling [17], and has already been used in order to reduce the nonuniqueness of solutions and also and enhance interpretability of the NMF results. The sparseNMF proposed in [3] is expressed in the following equation
where, y_{ i } is the ith column of Y. From the Bayesian perspective, this formulation is obtained from the logposterior probability under the assumptions of Gaussian error, Gaussiandistributed basis vectors, and Laplacedistributed coefficient vectors. Keeping one matrix fixed and updating the other matrix can be formulated as a NNLS problem. In order to improve the interpretability of the basis vectors and speed up the algorithm, we implemented the following model instead:
We optimize this using three alternating steps in each iteration. First, we optimize the following task:
then, A is updated as follows:
and then, the columns of A are normalized to have unit l_{2} norm. The first and second steps can be solved using nonnegative quadratic programming (NNQP), whose general formulation is
where, z_{ i } is the ith column of the variable matrix Z. It is easy to prove that NNLS is a special case of NNQP. For example, Equation (7) can be rewritten as
The implementations of the method in [3] and our method are given in functions sparsenmfnnls and sparseNMFNNQP, respectively. We also implemented the sparse semiNMF in functionl sparseseminmfnnls.
Versatile sparse matrix factorization
When the training data X is of mixed signs, the basis matrix A is not necessarily constrained to be nonnegative; this depends on the application or the intentions of the users. However, without nonnegativity, A is not sparse any more. In order to obtain sparse basis matrix A for some analysis, we may use l_{1}norm on A to induce sparsity. The drawback of l_{1}norm is that correlated variables may not be simultaneously nonzero in the l_{1}induced sparse result. This is because l_{1}norm is able to produce sparse but nonsmooth results. It is known that l_{2}norm is able to obtain smooth but nonsparse results. When both norms are used together, then correlated variables can be selected or removed simultaneously [18]. When smoothness is required on Y, we may also use l_{2}norm on it in some scenarios. We thus generalize the aforementioned NMF models into a versatile form as expressed below
where, parameters: α_{1}≥0 controls the sparsity of the basis vectors; α_{2}≥0 controls the smoothness and the scale of the basis vectors; λ_{1}≥0 controls the sparsity of the coefficient vectors; λ_{2}≥0 controls the smoothness of the coefficient vectors; and, parameters t_{1} and t_{2} are boolean variables (0: false, 1: true) which indicate if nonnegativity needs to be enforced on A or Y, respectively. We can call this model versatile sparse matrix factorization (VSMF). It can be easily seen that the standard NMF, semiNMF, and the sparseNMFs are special cases of VSMF.
We devise the following multiplicative update rules for the VSMF model in the case of t_{1}=t_{2}=1 (implemented in function sparsenmf2rule):
where, A∗B and \frac{\mathit{A}}{\mathit{B}} are the elementwise multiplication and division operators of matrices A and B, respectively. Alternatively, we also devise an activeset algorithm for VSMF (implemented in function vsmf). When t_{1}(ort_{2})=1, A (or Y) can be updated by NNQP (this case is also implemented in sparsenmf2nnqp). When t_{1}(ort_{2})=0, A (or Y) can be updated using 1_{1}QP.
KernelNMF
Two features of a kernel approach are that i) it can represent complex patterns, and ii) the optimization of the model is dimensionfree. We now show that NMF can also be kernelized.
The basis matrix is dependent on the dimension of the data, and it is difficult to represent it in a very high (even infinite) dimensional space. We notice that in the NNLS optimization, updating Y in Equation (10) needs only the inner products A^{T}A, A^{T}X, and X^{T}X. From Equation (4), we obtain A^{T}A=(Y^{‡})^{T}X^{T}X Y^{‡}, A^{T}X=(Y^{‡})^{T}X^{T}X. Therefore, we can see that only the inner product X^{T}X is needed in the optimization of NMF. Hence, we can obtain the kernel version, kernelNMF, by replacing the inner product X^{T}X with a kernel matrix K(X,X). Interested readers can refer to our recent paper [6] for further details. Based on the above derivations, we implemented the kernel semiNMF using multiplicative update rule (in kernelseminmfrule) and NNLS (in kernelseminmfnnls). The sparse kernel semiNMFs are implemented in functions kernelsparseseminmfnnls and kernelSparseNMFNNQP which are equivalent to each other. The kernel method of decomposing a kernel matrix proposed in [19] is implemented in kernelnmfdecom.
Other variants
Ref. [16] proposed the ConvexNMF, in which the columns of A are constrained to be the convex combinations of data points in X. It is formulated as X_{±}=X_{±}W_{+}Y_{+}+E, where M_{±} indicates that matrix M is of mixed signs. X W=A and each column of W contains the convex coefficients of all the data points to get the corresponding column of A. It has been demonstrated that the columns of A obtained with the convexNMF are close to the real centroids of clusters. ConvexNMF can be kernelized as well [16]. We implemented the convexNMF and its kernel version in convexnmfrule and kernelconvexnmf, respectively.
The basis vectors obtained with the above NMFs are nonorthogonal. Alternatively, orthogonal NMF (orthoNMF) imposes the orthogonality constraint in order to enhance sparsity [20]. Its formulation is
where, the input X is nonnegative, S absorbs the magnitude due to the normalization of A and Y. Function orthnmfrule is its implementation in our toolbox. OrthoNMF is very similar with the nonnegative sparse PCA (NSPCA) proposed in [21]. The disjoint property on orthoNMF may be too restrictive for many applications, therefore this property is relaxed in NSPCA. OrthoNMF does not guarantee the maximumvariance property which is also relaxed in NSPCA. However NSPCA only enforces nonnegativity on the basis vectors, even when the training data have negative values. We plan to devise a model in which the disjoint property, the maximumvariance property, the nonnegativity and sparsity constraints can be controlled on both basis vectors and coefficient vectors.
There are two efficient ways of applying NMF on data containing missing values. First, the missing values can be estimated prior to running NMF. Alternatively, weightedNMF[22] can be directly applied to decompose the data. WeightedNMF puts a zero weight on the missing elements and hence only the nonmissing data contribute to the final result. An expectationmaximization (EM) based missing value estimation during the execution of NMF may not be efficient. The weightedNMF is given in our toolbox in function wnmfrule.
Results and discussion
Based on the various implemented NMFs, a number of data mining tasks can be performed via our toolbox. Table 2 lists the data mining functionalities we provide in this level. These mining tasks are also described along with appropriate examples.
Clustering and biclustering
NMF has been applied for clustering. Given data X with multivariate data points in the columns, the idea is that, after applying NMF on X, a multivariate data point, say x_{ i } is a nonnegative linear combination of the columns of A; that is x_{ i }≈A y_{ i }=y_{1i}a_{1}+⋯+y_{ k i }a_{ k }. The largest coefficient in the ith column of Y indicates the cluster this data point belongs to. The reason is that if the data points are mainly composed with the same basis vectors, they should therefore be in the same group. A basis vector is usually viewed as a cluster centroid or prototype. This approach has been used in [2] for clustering microarray data and in order to discover tumor subtypes. We implemented function NMFCluster through which various NMF algorithms can be selected. An example is provided in exampleCluster file in the folder of our toolbox.
The task of interpreting both the basis matrix and the coefficient is equivalent to simultaneously clustering the rows and columns of matrix X. This is biclustering and the interested readers can refer to [23] for an excellent survey on biclustering algorithms and to [4] for a biclustering method based on NMF. We implemented a biclustering approach based on NMF in biCluster function. The biclusters can be visualized via the function NMFBicHeatMap. We applied NMF to simultaneously grouping the genes and samples of a leukemia data set [2] which includes tumor samples of three subtypes. The goal is to find strongly correlated genes over a subset of samples. A subset of such genes and a subset of such samples form a bicluster. The heatmap is shown in Figure 1. Readers can find the script in exampleBiCluster file of our toolbox.
Basis vector analysis for biological process discovery
We can obtain interesting and detailed interpretations via an appropriate analysis of the basis vectors. When applying NMF on a microarray data, the basis vectors are interpreted as potential biological processes [3, 13, 24]. In the following, we give one example for finding biological factors on genesample data, and two examples on timeseries data. Please note they only serve as simple examples. Fine tuning of the parameters of NMF is necessary for accurate results.
First example
We ran our VSMF on the ALLAML genesample data of [2] with the settings k=3, α_{1}=0.01, α_{2}=0.01, λ_{1}=0, λ_{2}=0.01, t_{1}=1, and t_{2}=1. Next, we obtain 81, 37, and 448 genes for the three factors, respectively. As in [3], we then performed gene set enrichment analysis (GSEA) by applying OntoExpress [25] on each of these sets of genes. Part of the result is shown in Table 3. We can see that the factorspecific genes selected by NMF correspond to some biological processes significantly. Please see file exampleBioProcessGS in the toolbox for details. GSEA can also be done using other tools, such as MIPS [26], GOTermFinder [27], and DAVID [28, 29].
Second example
We used NMF to cluster a timeseries data of yeast metabolic cycle in [30]. Figure 2 shows the heatmap of NMF clustering, and Figure 3 shows the three basis vectors. We used nmfnnls function to decompose the data and NMFHeatMap to plot the heatmap. The detailed script is given in the exampleBioProcessTSYeast file in the toolbox. We can clearly see that the three periodical biological processes corresponds exactly to the Ox (oxidative), R/B (reductive, building), and R/C (reductive, charging) processes discovered in [30].
Third example
We used NMF to factorize a breast cancer timeseries data set, which includes wild type MYCN cell lines and mutant MYCN cell lines [31]. The purpose of this example is to show that NMF is a potential tool to finding cancer drivers. One basic methodology is in the following. First, basis vectors are produced applying NMF on a timeseries data. Then factorspecific genes are identified by computational or statistical methods. Finally, the regulators of these factorspecific genes are identified from any prior biological knowledge. This data set has 8 time points (0, 2, 4, 8, 12, 24, 36, 48 hr.). The zero time point is untreated and samples were collected at the subsequent time points after treatment with 4hydroxytamoxifen (4OHT). In our computational experiment, we use our VSMF implementation (function vsmf). We set k=2. Because this data set has negative values we set t_{1}=0 and t_{2}=1. We set α_{1}=0.01, α_{2}=0, λ_{1}=0, and λ_{2}=0.01. The basis vectors of both wildtype and mutant data are compared in Figure 4. From the wildtype timeseries data, we can successfully identify two patterns. The rising pattern corresponds to the induced signature and the falling pattern corresponding to the repressed signature in [31]. It is reported in [31] that the MYC target genes contributes to both patterns. From the mutant timeseries, we can obtain two flat processes, which are reasonable. The source code of this example can be found in exampleBioProcessMYC. We also recommend the readers to see the methods based on matrix decompositions which are proposed in [13, 32] and devised for identifying signaling pathways.
Basis vector analysis for gene selection
The columns of A for a gene expression data set are called metasamples in [2]. They can be interpreted as biological processes, because their values imply the activation or inhibition of some the genes. Gene selection aims to find marker genes for disease prediction and to understand the pathways they contribute to. Rather than selecting genes on the original data, the novel idea is to conduct gene selection on the metasamples. The reason is that the discovered biological processes via NMF are biologically meaningful for class discrimination in disease prediction, and the genes expressed differentially across these processes contribute to better classification performance in terms of accuracy. In Figure 1 for example, three biological processes are discovered and only the selected genes are shown. We have implemented the informationentropybased gene selection approach proposed in [3] in function featureFilterNMF. We give an example on how to call this function in file exampleFeatureSelection. It has been reported that it can select meaningful genes, which has been verified with gene ontology analysis. Feature selection based on supervised NMF will also be implemented.
Feature extraction
Microarray data and mass spectrometry data have tens of thousands of features but only tens or hundreds of samples. This leads to the issues of curse of dimensionality. For example, it is impossible to estimate the parameters of some statistical models since the number of their parameters grow exponentially as the dimension increases. Another issue is that biological data are usually noisy; which crucially affects the performances of classifiers applied on the data. In cancer study, a common hypothesis is that only a few biological factors (such as the oncogenes) play a crucial role in the development of a given cancer. When we generate data from control and sick patients, the highdimensional data will contain a large number of irrelevant or redundant information. Orthogonal factors obtained with principal component analysis (PCA) or independent component analysis (ICA) are not appropriate in most cases. Since NMF generates nonorthogonal (and nonnegative) factors, therefore it is much reasonable to extract important and interesting features from such data using NMF. As mentioned above, training data X_{m×n}, with m features and n samples, can be decomposed into k metasamples A_{m×k} and Y_{k×n}, that is
where, Y_{tr} means that Y is obtained from the training data. The k columns of A span the kdimensional feature space and each column of Y_{tr} is the representation of the corresponding original training sample in the feature space. In order to project the p unknown samples S_{m×p} into this feature space, we have to solve the following nonnegative least squares problem:
where, Y_{uk} means the Y is obtained from the unknown samples. After obtaining Y_{tr} and Y_{uk}, the learning and prediction steps can be done quickly in the kdimensional feature space instead of the mdimensional original space. A classifier can learn over Y_{tr}, and then predicts the class labels of the representations of unknown samples, that is Y_{uk}.
From the aspect of interpretation, the advantage of NMF over PCA and ICA is that the metasamples are very useful in the understanding of the underlying biological processes, as mentioned above.
We have implemented a pair of functions featureExtractionTrain and featureExtractionTest including many linear and kernel NMF algorithms. The basis matrix (or, the inner product of basis matrices in the kernel case) is learned from the training data via the function featureExtractionTrain, and the unknown samples can be projected onto the feature space via the function featureExtractionTest. We give examples of how to use these functions in files exampleFeatureExtraction and exampleFeatureExtractionKernel.
Figure 5 shows the classification performance of SVM without dimension reduction and SVM with dimension reduction using linear NMF, kernel NMF with radial basis function (RBF) kernel, and PCA on two data sets, SRBCT [33] and Breast [34]. Since ICA is computationally costly, we did not include it in the comparisons. The bars represent the averaged 4fold crossvalidation accuracies using support vector machine (SVM) as classifier over 20 runs. We can see that NMF is comparable to PCA on SRBCT, and is slightly better than PCA on Breast data. Also, with only few factors, the performance after dimension reduction using NMF is at least comparable to that without using any dimension reduction. As future work, supervised NMF will be investigated and implemented in order to extract discriminative features.
Classification
If we make the assumption that every unknown sample is a sparse nonnegative linear combination of the training samples, then we can directly derive a classifier from NMF. Indeed, this is a specific case of NMF in which the training samples are the basis vectors. Since the optimization process within NMF is a NNLS problem, we call this classification approach the NNLS classifier[35]. A NNLS problem is essentially a quadratic programming problem as formulated in Equation (9), therefore, only inner products are needed for the optimization. We thus can naturally extend the NNLS classifier to kernel version. Two features of this approach are that: i) the sparsity regularization help avoid overfitting the model; and ii) the kernelization allows a dimensionfree optimization and also linearizes the nonlinear complex patterns within the data. The implementation of the NNLS classifier is in file nnlsClassifier. Our toolbox also provides many other classification approaches including SVM classifier. Please see file exampleClassification for demonstration. In our experiment of 4fold crossvalidation, accuracies of 0.7865 and 0.7804 are respectively obtained with linear and kernel (RBF) NNLS classifier on Breast data set. They achieved accuracies of 0.9762 and 0.9785, respectively, over SRBCT data.
Biological data are usually noisy and sometimes contain missing values. A strength of the NNLS classifier are that it is robust to noise and to missing values, making NNLS classifiers quite suitable for classifying biological data [35].
In order to show its robustness to noise, we added a Gaussian noise of mean 0 and variance from 0 to 4 with increment 0.5 on SRBCT. Figure 6 illustrates the results of NNLS, SVM, and 1nearest neighbor (1NN) classifiers using this noisy data. It can be seen that as the noise increases, NNLS outperforms SVM and 1NN significantly.
To deal with the missing value problem, three strategies are usually used: incomplete sample or feature removal, missing value imputation (i.e., estimation), and ignoring missing values. Removal methods may delete important or useful information for classification and particularly when there is a large percentage of missing values in the data. Imputation methods may create false data depending on the magnitude of the true estimation errors. The third method ignores using the missing values during classification. Our approach in dealing with the missing value problem is also to ignore them. The NNLS optimization needs only the inner products of pairs of samples. Thus, when computing the inner product of two samples, say x_{ i } and x_{ j }, we normalize them to have unit l_{2}norm using only the features present in both samples, and then we take their inner product. As an example, we randomly removed between 10% to 70% of data values in STBCT data. Using such incomplete data, we compared our method with the zeroimputation method (that is, estimating all missing values as 0). In Figure 7, we can see that the NNLS classifier using our missing value approach outperforms the zeroimputation method in the case of large missing rate. Also, the more sophisticated knearest neighbor imputation (KNNimpute) method [36] will fail on data with in high percentage of missing values.
Statistical comparison
The toolbox provides two methods for statistical comparisons and evaluations of different methods. The first is a twostage method proposed in [37]. The importance of this method is that it can estimate the datasize requirement for attaining a significant accuracy and extrapolate the performance based on the current available data. Generating biological data is usually very expensive and thus this method can help researchers to evaluate the necessity of producing more data. At the first stage, the minimum data size required for obtaining a significant accuracy is estimated. This is implemented in function significantAcc. The second stage is to fit the learning curve using the error rates of large data sizes. It is implemented in function learnCurve. In our experiments, we found that the NNLS classifier usually requires fewer number of samples for obtaining a significant accuracy. For example on SRBCT data, NNLS requires only 4 training samples while SVM needs 19 training samples. The fitted learning curves of NNLS and SVM classifiers are shown in Figure 8. We provide an example of how to plot this figure in file exampleFitLearnCurve.
The second method is the nonparametric Friedman test coupled with posthoc Nemenyi test to compare multiple classifiers over multiple data sets [38]. It is difficult to draw an overall conclusion if we compare multiple approaches in a pairwise fashion. Friedman test has been recommended in [38] because it is simple, safe and robust, compared with parametric tests. It is implemented in function FriedmanTest. The result can be presented graphically using the crucial difference (CD) diagram as implemented in function plotNemenyiTest. CD is determined by significant level α. Figure 9 is an example of the result of the Nemenyi test for comparing 8 classifiers over 13 high dimensional biological data sets. This example can be found in file exampleFriedmanTest. If the distance of two methods is greater than the CD then we conclude that they differ significantly.
Conclusions
In order to address the issues of the existing NMF implementations, we propose a NMF Toolbox written in MATLAB, which includes a basic NMF optimization level and an advanced data mining level. It enable users to analyze biological data via NMFbased data mining approaches, such as clustering, biclustering, feature extraction, feature selection, and classification.
The following are the future works planned in order to improve and augment the toolbox. First, we will include more NMF algorithms such as nsNMF, LSNMF, and supervised NMF. Second, we are very interested in implementing and speeding up the Bayesian decomposition method which is actually a probabilistic NMF introduced independently in the same period as the standard NMF. Third, we will include more statistical comparison and evaluation methods. Furthermore, we will investigate the performance of NMF for denoising and for data compression.
Availability and requirements
Project name: The NMF Toolbox in MATLABProject home page:https://sites.google.com/site/nmftool and http://cs.uwindsor.ca/~li11112c/nmfOperating system(s): Platform independentProgramming language: MATLABOther requirements: MATLAB 7.11 or higherLicense: GNU GPL Version 3Any restrictions to use by nonacademics: Licence needed
References
Lee DD, Seung S: Learning the parts of objects by nonnegative matrix factorization. Nature. 1999, 401: 788791. 10.1038/44565.
Brunet J, Tamayo P, Golub T, Mesirov J: Metagenes and molecular pattern discovery using matrix factorization. PNAS. 2004, 101 (12): 41644169. 10.1073/pnas.0308531101.
Kim H, Park H: Sparse nonnegatice matrix factorization via alternating nonnegativityconstrained least aquares for microarray data analysis. SIAM J Matrix Anal Appl. 2007, 23 (12): 14951502.
CarmonaSaez P, PascualMarqui RD, Tirado F, Carazo JM, PascualMontano A: Biclustering of gene expression data by nonsmooth nonnegative matrix factorization. BMC Bioinformatics. 2006, 7: 7810.1186/14712105778.
Wang G, Kossenkov A, Ochs M: LSNMF: A modified nonnegative matrix factorization algorithm utilizing uncertainty estimates. BMC Bioinformatics. 2006, 7: 17510.1186/147121057175.
Li Y, Ngom A: A new kernel nonnegative matrix factorization and its application in microarray data analysis. CIBCB. 2012, IEEE CIS Society Piscataway: IEEE Press, 371378.
Cichocki A, Zdunek R: NMFLAB  MATLAB toolbox for nonnegative matrix factorization. 2006, Tech. rep., [http://www.bsp.brain.riken.jp/ICALAB/nmflab.html]
The NMF: DTU toolbox. Tech. rep., Technical University of Denmark [http://cogsys.imm.dtu.dk/toolbox/nmf]
Liu S: NMFN: nonnegative matrix factorization. 2011, Tech. rep., Duke University,, [http://cran.rproject.org/web/packages/NMFN]
Gaujoux R, Seoighe C: A flexible R package for nonnegative matrix factorization. BMC Bioinformatics. 2010, 11: 36710.1186/1471210511367. [http://cran.rproject.org/web/packages/NMF]
Qi Q, Zhao Y, Li M, Simon R: nonnegative matrix factorization of gene expression profiles: A plugin for BRBArrayTools. Bioinformatics. 2009, 25 (4): 545547. 10.1093/bioinformatics/btp009.
Fertig E, Ding J, Favorov A, Parmigiani G, Ochs M: CoGAPS: an R/C++ package to identify patterns and biological process activity in transcriptomic data. Bioinformatics. 2010, 26 (21): 27922793. 10.1093/bioinformatics/btq503.
Ochs M, Fertig E: Matrix factorization for transcriptional regulatory network inference. CIBCB, IEEE CIS Society. Piscataway: IEEE Press;. 2012, 387396.
Lee D, Seung S: Algorithms for nonnegative matrix factorization. NIPS. 2001, Cambridge: MIT Press, 556562.
Kim H, Park H: Nonnegative matrix factorization based on alternating nonnegativity constrained least squares and active set method. SIAM J Matrix Anal Appl. 2008, 30 (2): 713730. 10.1137/07069239X.
Ding C, Li T, Jordan MI: Convex and seminonnegative matrix factorizations. TPAMI. 2010, 32: 4555.
Tibshirani R: Regression shrinkage and selection via the lasso. J R Stat Soc Ser B (Methodol). 1996, 58: 267288.
Zou H, Hastie T: Regularization and variable selection via the elastic Net. J R Stat Soc Ser B: Stat Methodol. 2005, 67 (2): 301320. 10.1111/j.14679868.2005.00503.x.
Zhang D, Zhou Z, Chen S: Nonnegative matrix factorization on kernels. LNCS. 2006, 4099: 404412.
Ding C, Li T, Peng W, Park H: Orthogonal nonnegative matrix trifactorizations for clustering. KDD. 2006, New York: ACM, 126135.
Zass R, Shashua A: Nonnegative sparse PCA. NIPS. 2006, Cambridge: MIT Press
Ho N: Nonnegative matrix factorization algorithms and applications. PhD thesis,. LouvainlaNeuve: Belgium; 2008
Madeira S, Oliveira A: Biclustering algorithms for biological data analysis: A survey. IEEE/ACM Trans Comput Biol Bioinformatics. 2004, 1: 2445. 10.1109/TCBB.2004.2.
Kim P, Tidor B: Subsystem identification through dimensionality reduction of largescale gene expression data. Genome Res. 2003, 13: 17061718. 10.1101/gr.903503.
Draghici S, Khatri P, Bhavsar P, Shah A, Krawetz S, Tainsky M: Ontotools, the toolkit of the modern biologist: ontoexpress, ontocompare, ontodesign and ontotranslate. Nucleic Acids Res. 2003, 31 (13): 37753781. 10.1093/nar/gkg624.
Mewes H, Frishman D, Gruber C, Geier B, Haase D, Kaps A, Lemcke K, Pfeiffer F, Schuller C, Stocker S, Mannhaupt G: MIPS: A database for genomes and protein sequences. Nucleic Acids Res. 2000, 28: 3740. 10.1093/nar/28.1.37.
Boyle E, Weng S, Gollub J, Jin H, Botstein D, Cherry J, Sherlock G: GO::TermFinder – open source software for accessing gene ontology information and finding significantly enriched gene ontology terms associated with a list of genes. Bioinformatics. 2004, 20: 37103715. 10.1093/bioinformatics/bth456.
Sherman B, Huang D: Systematic and integrative analysis of large gene lists using DAVID Bioinformatics resources. Nature Protoc. 2009, 4: 4457.
Huang D, Sherman B, Lempicki R: Bioinformatics enrichment tools: paths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Res. 2009, 37: 113. 10.1093/nar/gkn923.
Tu B, Kudlicki A, Rowicka M, McKnight S: Logic of the yeast metabolic cycle: temporal compartmentalization of cellular processes. Science. 2005, 310: 11521158. 10.1126/science.1120499.
Chandriani S, Frengen E, Cowling V, Pendergrass S, Perou C, Whitfield M, Cole M: A core MYC gene expression signature is prominient in basallike breast cancer but only partially overlaps the core serum response. PloS ONE. 2009, 4 (5): e6693.
Ochs M, Rink L, Tarn C, Mburu S, Taguchi T, Eisenberg B, Godwin A: Detection of treatmentinduced changes in signaling pathways in sastrointestinal stromal tumors using transcripttomic data. Cancer Res. 2009, 69 (23): 91259132. 10.1158/00085472.CAN091709.
Khan J: Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med. 2001, 7 (6): 673679. 10.1038/89044.
Hu Z: The molecular portraits of breast tumors are conserved across microarray platforms. BMC Genomics. 2006, 7: 9610.1186/14712164796.
Li Y, Ngom A: Classification approach based on nonnegative least squares. Neurocomputing. 2013,, in press
Troyanskaya O, Cantor M, Sherlock G, Brown P, Hastie T, Tibshirani R, Botstein D, Altman R: Missing value estimation methods for DNA microarrays. Bioinformatics. 2001, 17 (6): 520525. 10.1093/bioinformatics/17.6.520.
Mukherjee S, Tamayo P, Rogers S, Rifkin R, Engle A, Campbell C, Golub T, Mesirov J: Estimating dataset size requirements for classifying DNA microarray data. J Comput Biol. 2003, 10 (2): 119142. 10.1089/106652703321825928.
Demsar J: Statistical comparisons of classifiers over multiple data sets. J Mach Learn Res. 2006, 7: 130.
Acknowledgements
This research has been partially supported by IEEE CIS Walter Karplus Summer Research Grant 2010, Ontario Graduate Scholarship 2011–2013, and Canadian NSERC Grants #RGPIN228117–2011.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
YL did comprehensive survey on NMF, implemented the toolbox, and drafted this manuscript. AN supervised the whole project, provided constructive suggestions on this toolbox, and wrote the final manuscript. All authors read and approved the final manuscript.
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
This article is published under license to 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.
About this article
Cite this article
Li, Y., Ngom, A. The nonnegative matrix factorization toolbox for biological data mining. Source Code Biol Med 8, 10 (2013). https://doi.org/10.1186/17510473810
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/17510473810
Keywords
 Nonnegative matrix factorization
 Clustering
 Biclustering
 Feature extraction
 Feature selection
 Classification
 Missing values