MicroRNA identification using linear dimensionality reduction with explicit feature mapping
© Shakiba and Rueda; licensee BioMed Central Ltd. 2013
Published: 20 December 2013
microRNAs are a class of small RNAs, about 20 nt long, which regulate cellular processes in animals and plants. Identifying microRNAs is one of the most important tasks in gene regulation studies. The main features used for identifying these tiny molecules are those in hairpin secondary structures of pre-microRNA.
A new classifier is employed to identify precursor microRNAs from both pseudo hairpins and other non-coding RNAs. This classifier achieves a geometric mean G m = 92.20% with just three features and 92.91% with seven features.
This study shows that linear dimensionality reduction combined with explicit feature mapping, namely miLDR-EM, achieves high performance in classification of microRNAs from other sequences. Also, explicitly mapping data onto a high dimensional space could be a useful alternative to kernel-based methods for large datasets with a small number of features. Moreover, we demonstrate that microRNAs can be accurately identified by just using three properties that involve minimum free energy.
MicroRNAs are a class of small non-coding RNAs that play a crucial role in gene regulation by perfectly or imperfectly binding into three prime untranslated regions in messenger RNAs, and cause repression of translating mRNAs into proteins or their cleavage. Researchers have estimated that about one third of the human genes are regulated by microRNAs . MicroRNAs perform many cellular tasks in cells including controlling cell developmental timing, cell death and stem cell characterization . In addition, many studies show that malfunction of microRNAs may have devastating impacts on cell life and may cause different types of cancer, heart disease and nervous system disorder . Accordingly, identification of microRNA genes is an essential process in discovering microRNA functions and its role in cellular processes.
Earlier, microRNAs were only identified by using experimental methods. Traditional experimental approaches to microRNA discovery include cloning and sequencing , and can detect novel microRNAs. Since microRNAs usually express at low levels and depend on tissue and conditions of the cell, these methods may be unable to identify new microRNAs . Recently high-throughput sequencing approaches, in particular, 454 sequencing, have become popular for discovering new microRNAs .
Another category of approaches for identifying microRNAs are computational methods. The main idea behind these methods is to analyze hairpin secondary structures of precursor microRNAs (pre-microRNA). At first, non-coding genes are transcribed into primary microRNAs (pri-microRNA) by RNA Polymerase II . Then, the RNase III enzyme, named Drosha and its co-factor, Pasha, processes pri-microRNA to release pre-microRNA, which is about 80 nt long and folds into a hairpin secondary structure. This product is then ready for further processes to change into mature microRNA. Secondary structure of pre-microRNA allows researchers to propose computational methods that can distinguish these sequences from other sequences in the genome.
There are two subcategories for computational methods: comparative and non-comparative ap-proaches. Biologists believe that microRNAs are highly conserved in related genomes . Therefore, some methods use this property of microRNAs and introduce candidate microRNAs which fold into hairpin secondary structure and are conserved in related genomes. The other subcategory of computational methods includes non-comparative methods. These methods are mainly based on analyzing secondary structure of pre-microRNA and are based on extracted precursor micro RNA features, to classify them from other sequences. There are still some major obstacles to overcome. First, there are thousands of other genome sequences which fold into hairpin secondary structure, called "pseudo hairpins" . Second, many other non-coding RNAs such as YRNAs, snRNAs and tRNAs fold into hairpin secondary structure as well. Therefore, the main challenge is to extract features from the sequences in such a way that pre-microRNAs can be distinguished from other non-coding RNAs and pseudo hairpins to classify them effectively.
Previously, many computational methods were proposed for finding novel microRNAs. The first algorithm for finding microRNA genes, miRscan , is a comparative method that uses a sliding window to compare and analyze sequences with previously known microRNA sequences. Later, many other comparative and non-comparative methods were proposed - we only mention a few of them here. miRabela  uses 40 sequential and structural features to classify microRNAs. Triplet-SVM  is a support vector machine (SVM) classifier which uses "contiguous structure-sequence" features to identify microRNAs. Based on Triplet-SVM, miREncoding SVM  and MiPred  were proposed to improve the performance of Triplet-SVM. miREncoding SVM added 11 global features and performed feature selection to find the best subset of features. MiPred added two thermodynamical features and also replaced the SVM by the random forest learning algorithm. miPred SVM  is another microRNA identification system that uses 29 "global and intrinsic folding measures" as features. MicroPred  was proposed by Batuwita and Palade and uses 29 features introduced in miPred in addition to 19 newly introduced features. MicroPred uses human pre-microRNAs as the positive class and both genome pseudo hairpins and other non-coding RNAs as the negative class for feeding its SVM. The feature selection method used in  is a filter method that does not consider a classifier for selecting a subset of features. Instead, they rely on some discriminant measures before training the SVM-based classifier. Also, Wang et al.  proposed a feature selection method that is based on GA-SVM. Lately, Xuan et al.  proposed another pre-microRNA identification method that selects a subset of samples instead of the whole data set for solving the class imbalance problem.
On the other hand, linear dimensionality reduction (LDR) has been shown to be successfully used in pattern recognition and machine learning . However, LDR methods may not be very efficient and powerful, especially when the data is highly complex and non-linear. For some LDR methods, kernel tricks were proposed to improve classification performance [17–19]. The kernel trick aims to implicitly map the data that is not linearly separable onto higher dimensions hoping that the data become linearly separable or at least more "separable" than in the original space. Mapping implicitly is not feasible in all cases due to the complexity of kernelizing some LDR methods. Instead, the data could be explicitly mapped onto the target space and then LDR can be used on the mapped data.
In this paper, LDR combined with mapping the data onto higher dimensions is employed to classify precursor microRNAs from both pseudo hairpins and other non-coding RNAs. As discussed later, mapping the data onto higher dimensions can significantly improve the performance of the classifiers. In addition, using LDR can resolve the class imbalance problem since it takes the distribution of the data into consideration. As opposed to this, SVM only considers data near the margin. In addition, a feature selection method is proposed for selecting a subset of features instead of employing the whole feature vector, yielding very good results.
The proposed classifier is able to distinguish human pre-microRNAs from both pseudo hairpins and other non-coding RNAs. The training dataset includes pre-microRNA sequences as the positive class and pseudo hairpins and other non-coding RNA sequences as the negative class. This dataset is available as stated in  as supplementary material. Detailed information about the datasets is given below.
Known human pre-microRNAs: This dataset includes 691 non-redundant human pre-microRNA sequences, which are obtained from http://microrna.sanger.ac.uk/sequences/[20, 21]. At first, 695 sequences were downloaded from miRBase and then after removing redundant sequences, 691 sequences were obtained which fold into hairpin structures. Some of these sequences fold into multi-branched loops at default parameters, showing sequences with multi-branched secondary structures that can also be identified by our classifier. We do not make any assumption about pre-microRNA secondary structures.
Pseudo hairpins: The negative dataset is composed of 8,494 human pseudo hairpin sequences which were previously used in Triplet-SVM, MiPred, miPred and microPred. These sequences were obtained originally from RefSeq genes .
Other non-coding RNAs: This dataset contains all the non-coding RNA sequences, except microRNA sequences to make the classifier able to distinguish microRNA from other kinds of RNAs. This dataset contains 754 non-redundant sequences which are no longer than 150 nt. This dataset is known to be the best available ncRNA dataset for the human genome according to the authors of . It includes 334 snoRNAs, 327 tRNAs, 53 snRNAs, 32 YRNAs, 5 5S-rRNAs, and three more sequences from other types of RNAs.
Linear dimensionality reduction
Linear dimensionality reduction (LDR) is a well-known technique in pattern recognition. The basic idea of LDR is to represent an object of dimension n as a lower-dimensional vector of dimension d, achieving this by performing a linear transformation. The advantage of using a linear transformation is that, although the derivation of the underlying transformation may be slower, the classification is extremely fast as it performs linear-time operations to reduce to dimensions, typically, much lower than the original one.
We consider two classes, ω1 and ω2, which correspond to human pre-microRNAs as the positive class and pseudo hairpins and other ncRNAs as the negative class. The classes are represented by two normally distributed random vectors x1 ~ N (m1, S1) and x2 ~ N (m2, S2), respectively, with p1 and p2 the a priori probabilities. After the LDR is applied, two new random vectors y1 = Ax1 and y2 = Ax2, where y1 ~ N (Am1; AS1A t ) and y2 ~ N (Am2; AS2A t ) with m i and S i being the mean vectors and covariance matrices in the original space, respectively. The aim of LDR is to find a linear transformation matrix A in such a way that the new representations of the classes (y i = Ax i ) are as separable as possible. Various criteria have been proposed to measure this separability . We consider three LDR criteria: the well-known Fisher's discriminant analysis (FDA) [24, 25], the heteroscedastic discriminant analysis (HDA) approach , and the Chernoff discriminant analysis (CDA) approach . A brief discussion of these three follows.
and taking the d eigenvectors whose eigenvalues are the largest ones. The resulting d eigenvectors are the rows of A.
and choosing the d eigenvectors whose corresponding eigenvalues are the largest ones.
In , a gradient-based algorithm was proposed, which maximizes the function of (5) iteratively. LDR is applied to each input vector and classification is performed via a Bayesian classifier.
Non-linear mapping of linear dimensionality reduction
While LDR methods have been widely used in machine learning and pattern recognition due to their simplicity, there are some drawbacks in using linear transformations, especially when the data are non-linear and complex. Linear classifiers are usually inefficient compared to more sophisticated classifiers such as SVM.
Kernels have been used extensively in pattern recognition methods such as SVM, PCA and others. Also, it has been shown that FDA using kernels significantly improves the performance of LDR [17–19]. The main idea of kernel-based methods is to implicitly map input data to a higher dimension hoping the data become linearly separable. For some methods, the kernel trick allows one to solve the problem of mapping and classifying without explicitly mapping data to a higher dimensional space. However, this approach is not possible for LDR methods such as CDA and HDA for which an implicit solution is far from trivial.
On the other hand, explicit mapping is a good alternative in some scenarios. The problem that we work on is that the data are represented in a few dimensions (features), and a large number of samples. In fact, if implicit mapping solutions were available, the size of the kernel matrices would be in the order of the number of samples. In our dataset, it would be around 10,000 × 10,000 in which basic algebraic operations on the matrices are time and space consuming.
Another advantage for using LDR methods is also related to dealing with a large number of samples in a lower dimensional space. Even LDR methods are usually affected by the problem of singular matrices, it is not the case in our datasets. As a result of doing feature selection, we deal with a very small number of features, and even after mapping to a higher dimensional space, the number of new features does not increase to more than a hundred (depending on the choice of parameters), which is still very low compared to the number of samples (a few thousand).
Therefore, we propose to explicitly map the data onto a higher dimensional space and then apply LDR methods on the mapped data. Explicitly mapping the data can be a challenging task, since finding the actual mapping function ϕ(x) of kernels can be far from trivial, especially for the radial basis function (RBF) kernel. This is because RBF implicitly maps the data to an infinite dimensional Hilbert space. In this study, mapping functions that are extracted from polynomial and RBF kernels are used.
where ϕ(x) is of dimension C(n + d, d). For explicit mapping, we need to use the simplified function to map the data and apply the LDR method on the new space.
As mentioned earlier, finding the mapping function of RBF is not a straightforward task. Rahimi and Recht  proposed a mapping mechanism that maps input samples to a randomized feature space with low dimensions, in which the inner product of any two mapped data is equivalent to a Gaussian RBF kernel of the two input data. Scikit-learn provides a package written in Python for approximating radial basis function kernel . The package takes two parameters, m which is number of dimensions of the transformed features, and gamma, which is the parameter of the RBF. We use their package in our implementation.
One of the most important aspects of designing a classifier is to extract the most relevant features which empower the classifier to distinguish between classes of data. Therefore, extracting the set of appropriate features from the dataset is very momentous. In addition, since the datasets used in the training and testing processes contain multi-branched sequences, the extracted features should contain relevant information that can make the classifier able to succeed.
In our study, 48 features are used in which 29 of them were previously introduced in miPred and 19 features were proposed in microPred. While we provide here a brief description about the features, full details can be found in the supplementary material section of microPred . These features can be categorized into two groups: thermodynamical and base pair-related features. miPred introduced 17 composition features including 16 dinucleotide frequencies and |g + c|% ratio, 6 folding measure features, including modified base pairing propensity dP, modified minimum free energy (MFE) dG, modified base paire distance dD, modified Shannon entropy dQ, and MFE indices 1 and 2 (MFEI1 and MFEI2, respectively).
In addition, microPred introduced another 19 features including MFE index 3 and 4 (MFEI3 and MFEI4), normalized ensemble free energy (NEFE), MFE structure frequency (Freq), structural diversity (Diversity) and Diff = |MFE - EFE|/L, where EFE is the ensemble free energy and L is the length of the sequence. Also, some thermodynamical features were introduced, such as structure entropy dS and dS/L, structure enthalpy dH and dH/L, melting energy of the structure Tm and Tm/L. Additionally, a few base-pair-related features such as |A - U|/L, |G - C|/L,|G - U|/L, average base pairs per stem, where a stem is a structural motif in the secondary structure, and also %|A - U|/nstems, %|G - C|/nstems, %|G - U|/nstems, where nstem is number of stems in the secondary structure. All these features can be calculated and extracted from sequences of pre-microRNAs.
The class imbalance problem
The class imbalance problem occurs when there is a large difference in prior class probabilities or when the disproportionate numbers of samples in positive and negative classes lead to poor performance of the classifier with respect to the smallest class . As mentioned earlier, the number of positive samples in comparison to the negative samples is small, with a ratio of about 1:13. In this case, standard classifiers such as SVMs have tendency to classify well the largest class, while ignoring the smallest class.
Here, TP represents pre-microRNAs which are correctly classified as pre-microRNA (true positives), TN represents non pre-microRNAs (pseudo hairpins and other ncRNAs) which are correctly classified as non pre-microRNA (true negatives), FP represents non pre-microRNAs which are misclassified as pre-microRNA (false positives) and FN represents pre-microRNAs which are incorrectly classified as non pre-microRNA (false negatives). Using LDR as the classifier is a good strategy to overcome the class imbalance problem. This is because LDR methods take the distribution of the whole data into account, in order to optimize the prediction function. This is in contrast with the criterion followed by the SVM that uses only the "support vectors" to find the most efficient prediction function. Although some SVM schemes have improved this by incorporating the concept of soft margin, the SVM still relies on the vectors on (or next to) the margin, ignoring the contribution of the other samples to a more efficient classification rule. The latter feature is, indeed, intrinsic to LDR techniques.
Feature selection is important for a variety of reasons: increasing the generalization power, speeding up the training and testing processes, improving classification performance, and result comprehensibility . Feature selection algorithms can be widely categorized into two groups: filter and wrapper methods. Filter methods evaluate the "goodness" of the feature subset by using the intrinsic characteristics of the data. They are computationally cheap, since they do not involve the induction algorithm. However, they also take the risk of selecting subsets of features, which may not match the chosen induction algorithm. Wrapper methods, on the contrary, directly use the induction algorithm to evaluate the feature subsets. They generally outperform filter methods in terms of prediction accuracy, but are computationally more intensive . Brute-force search is a method that evaluates the performance of the classifier based on different subsets of features. In this method, the performance of all possible subsets of features are compared to each other. In other words, the performance of all possible two-feature pairs are compared to the performance of all possible subsets of three features and so on. Despite the fact that brute-force search guarantees the highest accuracy, it is extremely time-consuming and impractical - brute-force search should find the best subset of features among the 2 n possible subsets, where n is the number of features. Thus, the search space is extremely large that it is not possible to run this method for more than a few features. Another feature selection method is forward search which is a greedy algorithm to find a sub-optimal subset of features . This algorithm starts with the null set and selects features to be added to the set one at a time, based on the performance of the classifier with the currently selected features in addition to a potential selected feature. This algorithm is very fast and usually has an acceptable performance, but does not guarantee the best subset of features.
In this study, we introduce a systematic feature selection method that is based on floating forward search feature selection and aims to improve the performance of the basic algorithm. The improvement relies on searching a larger feature space compared to the basic forward search approach. In our approach, the best 10 pairs of features among all the pairs (2-tuples) of features are selected. Then, all combinations of pairs with a third feature are evaluated and the best 10 triplets (3-tuples) of features are selected. This procedure is continued with k-tuples, k = 4, 5, . . . until a criterion is satisfied. The criterion can be a certain number of features being selected or selecting a new feature that does not improve the performance significantly. In our approach, the feature selection process is continued until 12 features are evaluated.
As mentioned earlier, since our dataset is unbalanced, G m is used to measure the performance of the classifiers to ensure that class imbalance does not mislead feature selection algorithms to select the best subset of features, regardless of different proportions between the number of samples in each class. Combining feature selection with explicit mapping followed by LDR and Bayesian classification yields the prediction model proposed in this paper - we call it miLDR-EM.
Results and discussion
In this study different experiments have been conducted on the dataset described in the Methods section. In order to evaluate the performance of the classifiers and feature selection mechanisms, 10-fold cross validation is used. The partitioning is done in such a way that the ratio between the number of samples in the positive class and the number of samples in the negative class remain the same. At each stage, one of these groups is selected as the test set and the other nine subsets are used as the training set. In this method, no sample from the test set is used in the training process and only the test set can be used for evaluating the performance of the classifier.
Comparison of different mapping functions for LDR-based methods.
In order to evaluate the feature selection method, a comparison between the performance of the proposed system using the whole dataset versus a subset of features was performed. The aim is to select a subset of features and measure the goodness of that subset using a particular classifier. If mapping the data is required, all the samples with the selected features will be mapped to the target high dimensional space. Then, using 10-fold cross validation, the data is partitioned into ten subsets and the LDR-based classifier will evaluate the performance of the system. At this stage, all the intermediate results are stored in a database to be used for next round of feature selection.
Results for feature selection.
No. of Features
Results for optimized parameters.
No. of Features
Comparison with other methods
The results included in Table 4 correspond to those reported in , since that paper shows the geometric mean as a comparison, while the others focus on accuracy. For the reasons discussed earlier in this paper, classification accuracy is expected to be much higher than G m , at the expense of misclassifying most samples in the smallest class, if not all of them. Triplet-SVM used a complete dataset of 193 positive and 8,494 negative samples. Of these, several samples were taken randomly (without replacement) to create non-overlapping training and test sets, containing 163 positive and 168 negative, and 30 positive and 1,000 negative samples respectively. From these numbers, it is obvious that the classifier will not face the imbalance problem on the training data. However, randomly picking samples from the complete dataset could produce a biased classifier toward some particular distribution (as no cross-validation is used). The same datasets were also used for testing miPred and microPred, but with different numbers of features (as indicated in the table). In addition, as pointed out in , no clear details were given in  on how the training and test sets were obtained. The performance of that method is very low though, and it uses 40 features. microPred used the same dataset as that of our experiments, with a 5-fold cross-validation process and a separate dataset of randomly chosen samples for test purposes.
Although the performance of miLDR-EM is slightly lower than that of microPred, miLDR-EM uses only three features and performs very well compared to microPred, which uses 21 features and achieves a G m value of just 1.38% higher than miLDR-EM. Even better performance is achieved by miLDR-EM with seven features, 0.67% lower than that of microPred with 21 features. The other advantage that miLDR-EM shows with respect to other methods is that the difference between SE and SP is much smaller, showing the power of the proposed scheme for solving the class imbalance problem. Only miPred and Triplet-SVM achieve a better balance between SE and SP, but yield a smaller G m , and use a much larger subset of features, 34 and 32 respectively. Although microPred and miLDR-EM achieve a similar balance between SE and SP, miLDR-EM uses fewer features. In addition, the plot of Figure 1 reveals that selecting eight features does not lower the performance substantially and keeps an almost perfect balance between SE and SP.
Overall, due to the small difference between the two methods, we can assert that miLDR-EM is the best choice, since it can achieve comparable results with a much smaller subset of features, yields a smaller difference between SE and SP, and can also provide an additional biological insight (discussions at the end of this section). This may provide not only an improvement on the computational task for classification, but also a more meaningful insight on the RNA structural properties that are suitable for prediction of pre-microRNA and pseudo hairpins. Note that we have tested miLDR-EM with one and two features, and the performance lowers to a G m value under 90%. For this reason, selecting three features is the best choice.
To conclude this section, a discussion on the selected features is presented. The three features that miLDR-EM uses for classification are the following: dG, zG and NEFE. dG represents the normalized free energy of folding in the sequences, normalized by the length of the sequence. zG is the normalized variant (z-score) for feature dG. NEFE measures the normalized ensemble free energy of the sequence. As dG, NEFE is also normalized by the length of the sequence. In a nutshell, it can be inferred that the identification of microRNA is merely based on minimum free energy (and its normalized z-score), and the normalized ensemble free energy. The other 45 features are much less relevant or not relevant at all in the prediction problem, and could eventually be disregarded.
We have proposed a classification approach for identifying microRNA sequences and classifying pre-microRNAs from pseudo hairpins and other non-coding RNAs, which we call miLDR-EM. We have used linear dimensionality reduction (LDR) as the main classification engine. The proposed approach explicitly maps the data onto a higher dimensional space and classification takes place on the mapped data. While using different functions for mapping the data, the results show that RBF helps the classifier perform better. Also, a feature selection method is applied on the dataset in order to find a subset of relevant features. The results turn out to be very good compared to previously proposed methods. With only three features, miLDR-EM performs slightly lower than microPred which uses 21 features. Explicitly mapping the data onto a higher dimensional shows significant improvement in performance. This study shows that in cases in which the number of samples is very large, explicit mapping can be used instead of kernel-based methods. In addition, we propose a specialized feature selection method in which we keep a balance between exhaustive search and greedy search for finding a significantly smaller subset of features.
This work has been partially supported by NSERC, the Natural Science and Engineering Council of Canada. This work has also been made possible by the facilities of the Shared Hierarchical Academic Research Computing Network (SHARCNET:http://www.sharcnet.ca) and Compute/Calcul Canada.
The publication costs for this article were funded by the Natural Sciences and Engineering Council of Canada, NSERC.
This article has been published as part of BMC Proceedings Volume 7 Supplement 7, 2013: Proceedings of the Great Lakes Bioinformatics Conference 2013. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcproc/supplements/7/S7.
- Bartel DP: MicroRNAs: Genomics, Biogenesis, Mechanism, and Function. Cell. 2004, 116 (2): 281-297. 10.1016/S0092-8674(04)00045-5.View ArticlePubMedGoogle Scholar
- Cai Y, Yu X, Hu S, Yu J: A Brief Review on the Mechanisms of miRNA Regulation. Genomics, Proteomics & Bioinformatics. 2009, 7 (4): 147-154. 10.1016/S1672-0229(08)60044-3.View ArticleGoogle Scholar
- Chen K, Rajewsky N: The evolution of gene regulation by transcription factors and microRNAs. Nat Rev Genet. 2007, 8 (2): 93-103.View ArticlePubMedGoogle Scholar
- Esquela-Kerscher A, Slack FJ: The age of high-throughput microRNA profiling. Nat Methods. 2004, 1 (2): 106-107. 10.1038/nmeth1104-106.View ArticlePubMedGoogle Scholar
- Esquela KA, Slack FJ: Oncomirs - microRNAs with a role in cancer. Nat Rev Cancer. 2006, 6 (4): 259-269. 10.1038/nrc1840.View ArticleGoogle Scholar
- Bentwich I, Avniel A, Karov Y, Aharonov R, Gilad S, Barad O, Barzilai A, Einat P, Einav U, Meiri E, Sharon E, Spector Y, Bentwich Z: Identification of hundreds of conserved and nonconserved human microRNAs. Nat Genet. 2005, 37 (7): 766-770. 10.1038/ng1590.View ArticlePubMedGoogle Scholar
- Lim LP, Glasner ME, Yekta S, Burge CB, Bartel DP: Vertebrate MicroRNA Genes. Science. 2003, 299 (5612): 1540-1540. 10.1126/science.1080372.View ArticlePubMedGoogle Scholar
- Sewer A, Paul N, Landgraf P, Aravin A, Pfeffer S, Brownstein M, Tuschl T, van Nimwegen E, Zavolan M: Identification of clustered microRNAs using an ab initio prediction method. BMC Bioinformatics. 2005, 6: 267-10.1186/1471-2105-6-267.PubMed CentralView ArticlePubMedGoogle Scholar
- Xue C, Li F, He T, Liu GP, Li Y, Zhang X: Classification of real and pseudo microRNA precursors using local structure-sequence features and support vector machine. BMC Bioinformatics. 2005, 6: 310-10.1186/1471-2105-6-310.PubMed CentralView ArticlePubMedGoogle Scholar
- Dalkilic M, Kim S, Yang J, Zheng Y, Hsu W, Lee M, Wong L: Exploring Essential Attributes for Detecting MicroRNA Precursors from Background Sequences. 2006, Springer Berlin/Heidelberg, 4316: 131-145.Google Scholar
- Jiang P, Wu H, Wang W, Ma W, Sun X, Lu Z: MiPred: classification of real and pseudo microRNA precursors using random forest prediction model with combined features. Nucleic Acids Research. 2007, 35 (suppl 2): W339-W344.PubMed CentralView ArticlePubMedGoogle Scholar
- Ng KLS, Mishra SK: De novo SVM classification of precursor microRNAs from genomic pseudo hairpins using global and intrinsic folding measures. Bioinformatics. 2007, 23 (11): 1321-1330. 10.1093/bioinformatics/btm026.View ArticlePubMedGoogle Scholar
- Batuwita R, Palade V: microPred: effective classification of pre-miRNAs for human miRNA gene prediction. Bioinformatics. 2009, 25 (8): 989-995. 10.1093/bioinformatics/btp107.View ArticlePubMedGoogle Scholar
- Wang Y, Chen X, Jiang W, Li L, Li W, Yang L, Liao M, Lian B, Lv Y, Wang S, Wang S, Li X: Predicting human microRNA precursors based on an optimized feature subset generated by GA-SVM. Genomics. 2011, 98 (2): 73-78. 10.1016/j.ygeno.2011.04.011.View ArticlePubMedGoogle Scholar
- Xuan P, Guo MZ, Wang J, Wang CY, Liu XY, Liu Y: Genetic algorithm-based efficient feature selection for classification of pre-miRNAs. Genetics and molecular research. 2011, 10 (2): 588-603. 10.4238/vol10-2gmr969.View ArticlePubMedGoogle Scholar
- Rueda L, Herrera M: Linear Dimensionality Reduction by Maximizing the Chernoff Distance in the Transformed Space. Pattern Recognition. 2008, 41 (10): 3138-3152. 10.1016/j.patcog.2008.01.016.View ArticleGoogle Scholar
- Mika S, Rätsch G, Muller KR: A Mathematical Programming Approach to The Kernel Fisher Algorithm. Proc Neural Information Processing Systems. 2001, MIT Press, 591-597.Google Scholar
- Mika S, Ratsch G, Weston J, Scholkopf B, Mullers K: Fisher discriminant analysis with kernels. Neural Networks for Signal Processing IX, 1999. 1999, 41-48.Google Scholar
- Kim SJ, Magnani A, Boyd S: Optimal kernel selection in Kernel Fisher discriminant analysis. Proceedings of the 23rd international conference on Machine learning. 2006, ACM, 465-472.Google Scholar
- Griffiths-Jones S, Grocock RJ, van Dongen S, Bateman A, Enright AJ: miRBase: microRNA sequences, targets and gene nomenclature. Nucleic Acids Res. 2006, 34 (Database issue): D140-D144.PubMed CentralView ArticlePubMedGoogle Scholar
- Griffiths-Jones S, Saini HK, van Dongen S, Enright AJ: miRBase: tools for microRNA genomics. Nucleic Acids Research. 2008, 36 (suppl 1): D154-D158.PubMed CentralPubMedGoogle Scholar
- Pruitt KD, Maglott DR: RefSeq and LocusLink: NCBI gene-centered resources. Nucleic Acids Research. 2001, 29: 137-140. 10.1093/nar/29.1.137.PubMed CentralView ArticlePubMedGoogle Scholar
- Griffiths-Jones S: The microRNA Registry. Nucleic Acids Research. 2004, 32 (suppl 1): D109-D111.PubMed CentralView ArticlePubMedGoogle Scholar
- Duda R, Hart P, Stork D: Pattern Classification. 2000, New York, NY: John Wiley and Sons, Inc, 2ndGoogle Scholar
- Fisher R: The Use of Multiple Measurements in Taxonomic Problems. Annals of Eugenics. 1936, 7: 179-188. 10.1111/j.1469-1809.1936.tb02137.x.View ArticleGoogle Scholar
- Loog M, Duin P: Linear Dimensionality Reduction via a Heteroscedastic Extension of LDA: The Chernoff Criterion. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2004, 26 (6): 732-739. 10.1109/TPAMI.2004.13.View ArticlePubMedGoogle Scholar
- Chang YW, Hsieh CJ, Chang KW, Ringgaard M, Lin CJ: Training and Testing Low-degree Polynomial Data Mappings via Linear SVM. Journal of Machine Learning Research. 2010, 11: 1471-1490.Google Scholar
- Rahimi A, Recht B: Random Features for Large-Scale Kernel Machines. Advances in Neural Information Processing Systems 20. Edited by: Platt J, Koller D, Singer Y, Roweis S. 2008, MIT Press, 1177-1184.Google Scholar
- Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, Duchesnay E: Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. 2011, 12: 2825-2830.Google Scholar
- Akbani R, Kwek S, Japkowicz N: Applying Support Vector Machines to Imbalanced Datasets. Machine Learning: ECML 2004, Volume 3201 of Lecture Notes in Computer Science. Edited by: Boulicaut JF, Esposito F, Giannotti F, Pedreschi D. 2004, Springer, 39-50.Google Scholar
- Zhu Z, Ong YS, Dash M: Wrapper-Filter Feature Selection Algorithm Using a Memetic Framework. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on. 2007, 37: 70-76. FebView ArticleGoogle Scholar
- Guyon I, Elisseeff A: An introduction to variable and feature selection. J Mach Learn Res. 2003, 3: 1157-1182.Google Scholar
- Theodoridis S, Koutroumbas K: Pattern Recognition. 2006, Elsevier Academic Press, thirdGoogle Scholar
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. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.