An Improved Fuzzy Based Missing Value Estimation in DNA Microarray Validated by Gene Ranking

1Department of Computer Science and Engineering, Heritage Institute of Technology, Kolkata 700107, India 2Department of Computer Science and Engineering, Netaji Subhash Engineering College, Kolkata 700152, India 3Department of Computer Science and Engineering, University of Engineering & Management, Kolkata 700156, India 4Department of Computer Science and Engineering, University of Calcutta, Kolkata 700098, India


Introduction
Microarray expression analysis is a widely used technique for profiling mRNA expression.The mRNA carries genetic information from DNA to the ribosome, where they specify the amino acid sequence of the protein products of gene expression.Microarray datasets often contain missing values which may occur due to various reasons including imperfections in data preparation steps (e.g., poor hybridization and chip contamination by dust and scratches) that create erroneous and low-quality values, which are usually discarded and referred to as missing.It is common for gene expression data to contain at least 5% missing values [1].Most of the microarray data analysis algorithms, such as gene clustering, disease (experiment) classification, and gene network design, require the complete information, that is, the entire gene expression matrix without any missing values.Hence, different imputation techniques should be used which would accurately impute multiple missing data values.Numerous imputation algorithms have been proposed to estimate the missing values.At first, we have applied modified version of our existing imputation technique LRFDVImpute [2] that first finds a subset of similar genes using the fuzzy difference vector (FDV) algorithm used in [3] where gene expression profiles have been considered as continuous time series curves and then use linear regression on the subset to estimate the missing value.We have considered estimating only those genes with one, two, or three missing values since these genes constitute 5-10% of the entire dataset.Absolute error has been calculated from the difference between the 2

Microarray gene dataset with missing values
Find subset of similar genes using fuzzy difference vector (FDV) algorithm Estimate missing values using linear regression on the subset of similar genes original value and the estimated value.Root Mean Square Error (RMSE) of those absolute errors is then determined.The workflow for the first phase has been shown in Figure 1.
After that we rank those genes to find the top ranked genes [4].We have used a hypothesis test, Wilcoxon rank sum test [5], to sort the features (genes) and rank them in order of their  values and select top  genes from them, thereby reducing the dimensionality, where  is the population size that has been used later for GA.The reduced set of genes has then been ranked by our GA method.The two ranks, one by Wilcoxon method and the other by our GA method, are then compared.The top  genes (value of  defined by the user) selected by our method are then used for classification using support vector machine (SVM) classifiers.The performance of classification justifies the efficiency of the ranking method used.Figure 2 shows the workflow for this phase.
Once this is done, we then forcibly make some cells missing in the top ranked genes and again estimate them using the same missing value estimation technique.Finally, we rank them once more to find the top ranked genes.Results show that most of the top ranked genes remain the same, which validates the proposed missing value estimation technique biologically as far as the estimation is concerned.

Present State of the Art
As discussed earlier, various statistical and analytical methods used for gene expression analysis are not robust to missing values and require the complete gene expression matrix for providing accurate results.Hence, it is necessary to devise accurate methods which would impute data values when they are missing.Many imputation methods have been proposed.The earliest method, named as row averaging or filling with zeroes, used to fill in the gaps for the missing values in gene dataset with zeroes or with the row average.
KNNImpute method proposed in [1] selects genes with expression profiles similar to the gene of interest to impute missing values.After experimenting with a number of metrics to calculate the gene similarity, such as Pearson correlation, Euclidian distance, and variance minimization, it was found that Euclidian distance was a sufficiently accurate norm.
The SVDImpute method, proposed in [1], uses Singular Value Decomposition of matrices to estimate the missing values of a DNA microarray.This method works by decomposing the gene data matrix into a set of mutually orthogonal expression patterns that can be linearly combined to approximate the expression of all genes in the dataset.These patterns, which in this case are identical to the principle components of the gene expression matrix, are further referred to as eigengenes [6,7].
Another method named as LLSImpute [8] represents a target gene with missing values as a linear combination of similar genes.The similar genes are chosen by -nearest neighbours or  coherent genes that have large absolute values of correlation coefficients followed by least square regression and estimation.
BPCAImpute method, proposed in [9], uses a Bayesian estimation algorithm to predict missing values.BPCA suggests using the number of samples minus 1 as the number of principal axes.Since BPCA uses an EM-like repetitive algorithm to estimate missing values, it needs intensive computations to impute missing values.
Another algorithm for time series gene expression analysis is presented in [10] that permits the principled estimation of unobserved time points, clustering, and dataset alignment.Each expression profile is modelled as a cubic spline (piecewise polynomial) that is estimated from the observed data and every time point influences the overall smooth expression curve.The alignment algorithm uses the Advances in Fuzzy Systems 3 same spline representation of continuous time series gene expression profiles.
FDVImpute method, proposed in [11], incorporates some fuzziness to estimate the missing value of a DNA microarray.The first step selects nearest (most similar) genes of the target gene (whose some component is missing) using fuzzy difference vector algorithm.Then the missing cell is estimated by using least square fit on the selected genes in the second step.
FDVSplineImpute, presented in [3], takes into account the time series nature of gene expression data and permits the estimation of missing observations using B-splines of similar genes from fuzzy difference vectors.
Another method, LRFDVImpute, proposed in [2], estimates multiple missing observations by first finding the most similar genes of the target gene and then applying the linear regression on those similar genes.This approach works in two stages.At the first stage, it estimates the real missing cells of SPELLMAN COMBINED dataset and at the later stage, it makes some cells miss forcefully of the same dataset and then using the estimated results from the first step, this approach estimates those missed cells using the same approach used earlier.Absolute error has been calculated from the difference between the original value and the estimated value.Root Mean Square Error (RMSE) of those absolute errors is then determined.
Extracting relevant information from microarray data is also difficult because of the inherent characteristics of the datasets, where there are the thousands of variables (genes) and very few numbers of samples.Finding out the set of significant genes or, in other words, the most differentially expressed genes, by studying data from tissues affected or unaffected by cancer cells, is an important task.This problem can be termed as gene selection.Several techniques have been used to rank genes and find out the most significant ones.
In [12], the algorithm used discriminant partial least squares (DPLS) and fuzzy clustering methods to interpret the gene expression patterns of acute leukemia and identify leukemia subtypes.
In [13], the proposed method used Mann-Whitney test and -sample Kruskal-Wallis ANOVA test to rank genes.Dimension reduction was done using -means clustering and PCA and classification performed using ANN trained during 8-fold cross-validation with recursive feature elimination (RFE) and leave-one-out testing.
In [14], the algorithm proposed a gene selection method based on Wilcoxon rank sum test and SVM.Wilcoxon rank sum test was used to select a subset of genes and then each selected gene is trained and tested using SVM classifier with linear kernel separately, and genes with high testing accuracy rates were chosen to form the final reduced gene subset.Classification was performed on two datasets: Breast Cancer [15] and ALL/AML Leukemia [16] using leave-one-out crossvalidation (LOOCV).
A hybrid GA/SVM approach is proposed for gene selection in [17], where a fuzzy logic based preprocessing tool is used to reduce dimensionality, GA for finding out the most frequent genes, and a SVM classifier used for classification.Experiments were performed on two well-known cancer datasets, Leukemia [16] and Colon [18], and results were compared with six other methods.
A multiobjective genetic approach is proposed in [19] for simultaneous clustering and gene ranking where a method to simultaneously optimize the feature ranking and clustering has been used.NSGA-II (Nondominated Sorting Genetic Algorithm-II) [20] has been used as a multiobjective evolutionary algorithm to optimize the chromosomes.
In [21], the proposed algorithm uses feature selection method based on genetic algorithms (GAs) and classification methods focusing on constructive neural networks (CNNs), C-Mantec.Several comparison results on six public cancer databases are provided using other feature selection strategy (Stepwise Forward Selection method) and different classification techniques (LDA, SVM, and Naive Bayes).
A PSO based graph theoretic approach, proposed in [22], is used for identifying the nonredundant gene markers from microarray gene expression data.The microarray data is first converted into a weighted undirected complete feature graph where the nodes represent the genes having gene's relevance as node weights and the edges are weighted in order of correlation among the genes.The densest subgraph having minimum average edge weight (similarity) and maximum average node weight (relevance) is then identified from the original feature graph.Binary particle swarm optimization is then applied for minimizing the average edge weight (correlation) and maximizing the average node weight (gene relevance) through a single objective function.
A web based tool DWFS, proposed in [23], is used to select significant features for a variety of problems efficiently.The search strategy is implemented using Parallel Genetic Algorithm.DWFS also applies various filtering methods as a preprocessing step in the feature selection process.It also uses three classifiers, like KNN classifier, Naive Bayes Classifier, and the combination of these two.Experiments using datasets taken from different biomedical applications show the efficiency of DWFS and lead to a significant reduction of the number of features without sacrificing performance as compared to several widely used existing methods.

Missing Value Estimation Using Linear
Regression.This phase of the work modifies an existing method LRFDVImpute for estimating missing values present in the microarray dataset using linear regression.Earlier version of LRFDVImpute inserts the newly estimated gene into the training data after estimation of each target gene.In this way, the newly estimated gene is taken into consideration while estimating the next target gene.This process has the risk of increasing the error while estimating the subsequent genes since the error term is cumulatively multiplied.To overcome this problem, modified LRFDVImpute does not add the target gene to the training data after it has been estimated.This way, the training gene set size remains constant and with increasing membership values of , the size of training data reduces.The effects of modifications have been studied and results are shown in the experimental results section.In our problem, the genes with missing values in the ( × ) ( is the number of genes and  is the number of samples) dataset are to be estimated.The method of finding a similar gene as used in [3] using fuzzy difference vector (FDV) algorithm is described below.
Target Row/Testing Data.The row whose missing value is being estimated: a target row may have multiple missing values but in a single run, a single value is estimated.
Similar Rows/Training Data.The rows that are similar to the target row: in this case only those rows are selected that have no missing values.Before applying the similarity measures all the columns from the complete matrix are removed that correspond to missing values in target row.
Let  1 ,  2 , . . .,   be the set of genes in the dataset.Let th   be the target gene, that is, the gene with  missing values.We remove the columns having missing values from the entire dataset.Let the resultant matrix contain ( − ) columns.Each target gene  is compared with each of the similar rows in the dataset.For the th gene   , the difference vector   of   is calculated as follows: Once the difference vectors are calculated for each of the target rows and the similar rows, say DifferenceTable 1 (for target row) and DifferenceTable 2 (for similar row), we then calculate Membership() to obtain the number of matches between difference vectors DifferenceTable 1 and DifferenceTable 2 for each target gene   .A match in the th component of the vectors DifferenceTable 1 and DifferenceTable 2 is determined by whether the signs of DifferenceTable 1 (, ) and DifferenceTable 2 (, ) are the same or not.Membership() defines the degree of match between the distribution of the target gene and the similar gene.We then define a membership grade for   as follows: The genes in the training data that have a membership value greater than a chosen membership grade  are considered to be a part of the similar genes.The steps for estimation can be summarized below: (1) Load the dataset with missing values.
(2) Calculate the missing number of columns for each gene and start with the first row with the least number of missing values (for our dataset it is 1).
(3) Compute the corresponding membership grade for the target gene from the training data using the FDV algorithm as shown above.
(4) Estimate the missing value using linear regression.
(5) Obtain coefficients of the regression from the linear model object lmObj.
(6) Add a bias of 1 at the beginning of the target row to allow for the bias parameter.
(7) Perform a vector multiplication between the modified target row and the coefficients of regression and add the obtained vector's elements together to get the estimated value.
(8) Replace the missing value with the estimated value.
(9) Go to step (2) and repeat the above steps to fill the missing values unless the mentioned "least number of missing values" in step (2) is less than or equal to 3.
Although we mentioned here that we go on filling the missing value till a point, it is not true.In between we stop this filling in process to do assessment of our algorithm.
After we have filled in all the missing values corresponding to rows with single missing values we select a particular collection of row-column positions corresponding to rows that did not have missing values initially and deliberately treat the values at these positions as missing and use the exact same process to estimate the values.
The same collection of row-column positions are again used when the algorithm has filled up all the rows up to two missing vales and then when it has filled up missing values existing in rows with up to three missing values.

Gene Ranking Using Genetic Algorithm.
In phase 2 of the proposed work, the result of the missing value estimation procedure carried out in phase 1 is biologically validated by ranking the genes using GA.Since a characteristic of gene expression microarray data is that the number of variables (genes) far exceeds the number of samples , we must reduce its dimension.Executing GA on the original dataset is quite impractical and time consuming.As a preprocessing step, we have reduced the dimension using Wilcoxon rank sum test.

Dimension Reduction Using Wilcoxon Rank Sum Test (WRST).
The inputs to the Wilcoxon rank sum test function are the two gene sets, the diseased set and the normal set, both of which have individually undergone the missing value estimation procedure (if there was any missing value).The two gene sets may have different number of samples.Let us consider that the diseased set is a ( ×  1 ) sized gene expression data, where  is the number of genes and  1 is the number of samples, and the normal set has a size ( ×  2 ), where  2 is the number of samples.The Wilcoxon rank sum function processes the two datasets in order to find out for which genes the null hypothesis is accepted or rejected.It returns two values,  value and ℎ-value, as discussed earlier.
The null hypothesis for our problem is that the genes are not differentially expressed; that is, either all the samples have come from diseased patients or they have come from normal patients.The alternative hypothesis can be that genes are differentially expressed.We record the  values and ℎ-values for each gene.
In the next step, we consider only those genes for which the alternative hypothesis holds (ℎ = 1) at the significance level alpha and sort the genes according to the  values thereby ranking the genes.We then select the topmost  genes, where  is the population size that has been used for GA later.Thus, we have two reduced populations, one representing diseased and the other representing normal tissues.Let (  ) × 1 be the diseased set, where  is the reduced set of genes and  1 is the number of samples, respectively, and let (  ) × 2 be the normal set, where  2 is the number of samples.

Chromosome Representation and Initial Population for
GA.The reduced gene sets (  ) × 1 and (  ) × 2 serve as the initial population for the genetic algorithm step.They contain pop size number of genes which is preselected by the user.We use real value encoding to represent each chromosome; that is,   and   are the measurements recorded for the th gene and th sample for each population, respectively.

Fitness Calculation.
The fitness for each gene in the reduced gene sets is again calculated by a method similar to that used in [14] where gene expression profiles have been considered as continuous time series curves.
In our problem, we have two populations, one for the diseased tissues and the other for the normal tissues.The two populations contain the same number of genes  but may have different number of samples.In that case, we consider the minimum of the two and extract the same number of samples from each set.
Let  1 ,  2 , . . .,   be the reduced set of genes in each population.If  = min( 1 ,  2 ), then for each population, the difference vector   of   is calculated using (1).Once the difference vectors are calculated for each of the two populations, say DifferenceTable 1 (for diseased) and DifferenceTable 2 (for normal), the number of matches between the difference vectors Membership() and the membership grade for   is computed using (2).
The fitness of gene   is the reciprocal of memgrade() and is calculated as This signifies that the more similar the distributions of gene   in the two populations are, the less differentially expressed the gene is, and vice versa.Thus, a fitter gene will have different distributions in the two populations.We then rank the genes in order of their fitness.

Elitism.
We have used an elitist version of GA where the best chromosomes are carried forward to the next generation unchanged; that is, the crossover and mutation operators are not applied on the best chromosomes.This technique ensures faster convergence of the process by keeping track of the best solutions.

Selection.
For selection, we have used a roulette wheel technique where genes are selected based on their relative fitness values.The better the chromosomes are, the more chances to be selected they have.Let count be the number of elite children.We construct a roulette wheel as follows [22]: (i) Calculate the fitness value fit() for each chromosome   , count + 1 ≤  ≤ .
(ii) Find the total fitness of the population  = ∑ pop size =count+1 fit(  ).(iii) Calculate the probability of selection   for each chromosome   , count + 1 ≤  ≤ : (iv) Calculate a cumulative probability   for each chromosome   , count + 1 ≤  ≤ : We now spin the wheel (pop size − count) times and select a single chromosome as follows: (i) Generate a random number (float)  between 0 and 1.
(ii) If  <  1 , we select the first chromosome Some chromosomes get selected more than once.According to Schema Theorem [24], the best chromosomes get more copies, the average stay even, and the worst die off.

Crossover.
For crossover, we proceed as follows.
For each chromosome   in the population, (i) generate a random number (float)  between 0 and 1, (ii) if  <  cross (crossover probability), we select the given chromosome for crossover.
We have used single point crossover where the crossover site is also generated randomly in the range where  is the number of samples.Thus after crossover, a pair of parent chromosomes generates a pair of offspring chromosomes [25].The new population obtained after crossover contains the new generation produced by crossover as well as the elite children that did not undergo crossover.This new population is used in the mutation process.

Mutation.
A nonuniform mutation operator as proposed in literature [25] has been used here.The new operator is defined as follows: (i) A random experiment is carried out which produces an outcome which is either 0 or 1.
(ii) Another random number pos is generated in the range where  is the number of samples, to select the mutation site.
We have used  = 2 for our experiment.The entire genetic transformation has been performed on one population with respect to the other.We made the diseased gene set to undergo genetic transformation while fitness evaluation has been made with respect to the normal gene set.The opposite transformation will produce similar results.
Once the genetic transformations are done, we obtain a final population set (here, genetically transformed diseased gene set) which have been ranked in order of their fitness.We compare the two ranks, one by the Wilcoxon method and the other by our GA method.A threshold of ±2 has been considered while comparing the two ranks.Results show that there is a good percentage of matches in the two ranks.Moreover, we find out the top ranked genes produced by both methods and the significant genes produced by the two methods are also similar.This also validates the result of the missing value estimation method carried out in phase 1.

Gene Classification Using SVM.
In order to prove the significance of ranking by our GA method, we perform classification.The top ranked  genes, n' {5, 10, 15, 20, 25}, ranked by our GA method are used for the purpose.We use -fold LOO cross-validations, where  is varied from one dataset to another depending on the number of samples.For cross-validation, we have divided our dataset into two sets, a training set and a testing set, in 80 : 20 ratio.The reason behind taking this ratio is that 80 : 20 is a commonly occurring ratio, which is often referred to as Pareto Principle.So, if there are  samples in the training set and  −  samples in the test set, where  is the total number of samples, the training set is divided into  equal sized subsets.Of the  subsets, one subset is retained for validation and the remaining −1 subsets are used as training data.Thus,  SVM classifiers with linear kernel are trained using the  training subsets.The classification accuracy rates are recorded and the classifier with the best accuracy rate is used to test the  −  samples.

Datasets Used.
The missing value estimation part of the proposed modified LRFDVImpute technique has been evaluated on the publicly available yeast cell cycle time series dataset from Spellman et al. [26] described in Table 1.
After the experiments on Spellman dataset are done, the combined gene ranking and classification portion of the proposed method are evaluated on four publicly available datasets: Colorectal Cancer tumours dataset (GDS4382), Breast Cancer dataset (GSE349-350), Prostate Cancer dataset, and Leukaemia Cancer dataset (DLBCL-FL).

Platform Used.
All algorithms have been implemented using MATLAB R2013a in Windows 8.1.

Results of Missing Value Estimation Part.
We perform the initial estimation using modified version of LRFDVImpute with a membership grade  = 0.55.After the initial estimation is over, we forcibly treat cells at specified locations as missing and estimate them using different membership values of  and both earlier and modified versions.This has been carried out only once, after estimating rows with single missing values and the corresponding RMSE values have been recorded.We have performed our experiments only on alpha, cdc15, and elu data of Spellman dataset.The number of missing values is too large for cdc28; that is why we ignore that segment.The results for the alpha, cdc15, and elu datasets using both methods are shown in Tables 2-4.Figures 3-5 show the corresponding plots of RMSE versus membership grade  for each of the four datasets.Table 5 compares the performance of both versions of LRFDVImpute method to that of some other existing methods, like SVDImpute, LLSImpute, FDVLLSImpute, FDVSPLINEImpute, and so forth, and the results show that modified version of LRFDVImpute outperforms the other existing methods as far as RMSE value is concerned.

Combined Results.
We test the significance of our proposed missing value estimation technique using the gene ranking method.We have not found any state-of-the-art work on gene ranking so far where Spellman dataset is used.That is why we use four more publicly available reallife gene expression datasets, like Colorectal Cancer dataset (GDS4382), Breast Cancer dataset (GSE349-350), Prostate Cancer dataset, and Leukaemia Cancer dataset (DLBCL-FL) [4,[27][28][29][30][31][32], to perform steps such as missing values estimation and gene ranking and analyze the results.We start with the microarray dataset containing missing values and apply our proposed missing value estimation technique to estimate the genes with missing values (if any).We rank them using proposed gene ranking method and find the top ranked genes.We then forcibly insert missing values in the top ranked genes and again estimate them using the same missing value estimation technique.Finally, we rank them once more to find the top ranked genes.Results show that most of the top ranked genes remain the same, which implies that the proposed missing value estimation technique has been accurate in estimating the unknown values.We have normalized most of the datasets using -score normalization method in order to bring the data values to a common scale.
Tables 6, 8, 10, and 13 show the estimated values for the four datasets, Tables 7, 9, 11, and 14 show the common gene indices before and after the estimation, and Tables 12 and 15 compare the performance of the proposed approach with two state-of-the-art methods [22,23] for Prostate and Leukaemia dataset on the basis of accuracy, sensitivity, specificity, 1score, and -mean metrics.We have found that Prostate and Leukaemia are the common dataset on which both the existing methods have done their experiments.The results show that the proposed gene ranking approach performs far better compared to those existing approaches, where one is a PSO based graph theoretic approach [22] and the other is a web based tool DWFS, which uses KNN and NBC classifiers [23] as far as those metrics are concerned.

Conclusion and Future Scope
The proposed modified version of LRFDVImpute technique has been tested on the dataset from Spellman et al. [26] and has shown impressive results.It outperforms some stateof-the-art methods.The plots of RMSE versus membership grade  show that modified version is equivalent to or better than earlier version for the alpha and cdc15 datasets.However, for the cdc28 dataset, earlier version has shown better results.For the elu datasets, both have reached 0 error margin.For both versions, a membership grade between 0.55 and 0.65 produces minimum error and any value in this range can be considered as a threshold to be used for fresh experiments.
The validation of the missing value estimation shows that most of the top ranked genes remain the same, before and after imputation, which implies that the proposed modified LRFDVImpute technique has been accurate in estimating the unknown values.
As a future scope, we would like to analyze the effects of using quadratic regression for estimation of missing values and the use of data cleaning techniques before imputation which may remove outliers if any and may further reduce the error margin.For gene ranking, we wish to analyze the effects of different parameter settings for GA and observe the ranking and classification results using SVM with other kernels and also compare results with the ones mentioned in literature.We would also wish to modify our algorithms so as to make this ranking more efficient and find out the most significant genes that would correctly identify the subtypes of a particular type of cancer.For the Leukemia dataset [16], this could be identifying the B-cell and Tcell lineages for the acute lymphoblastic leukemia (ALL) samples.

Figure 1 :Figure 2 :
Figure 1: Workflow of the proposed missing value estimation technique.

Figure 3 :
Figure 3: Plot of RMSE versus membership grade  for alpha dataset.

Table 8 :
Estimated values for Breast Cancer dataset.

Table 5 :
Performance comparison with other existing methods.

Table 6 :
Estimated values for Colorectal Cancer GDS4382.

Table 7 :
Top 25gene indices before and after estimation for GDS4382.

Table 9 :
Top 25 gene indices before and after estimation for Breast Cancer dataset.Figure 4: Plot of RMSE versus membership grade  for cdc15 dataset.

Table 10 :
Estimated values for Prostate Cancer dataset.

Table 11 :
Top 25 gene indices before and after estimation for Prostate Cancer dataset.
Figure 5: Plot of RMSE versus membership grade  for elu dataset.

Table 12 :
Performance comparison for Prostate dataset.

Table 14 :
Top 25gene indices before and after estimation for DLBCL-FL.