Literature DB >> 33281862

Identification of miRNA Biomarkers for Diverse Cancer Types Using Statistical Learning Methods at the Whole-Genome Scale.

Jnanendra Prasad Sarkar1,2, Indrajit Saha3, Adrian Lancucki4, Nimisha Ghosh5, Michal Wlasnowolski6, Grzegorz Bokota7,8, Ashmita Dey2, Piotr Lipinski4, Dariusz Plewczynski6,8.   

Abstract

Genome-wide analysis of miRNA molecules can reveal important information for understanding the biology of cancer. Typically, miRNAs are used as features in statistical learning methods in order to train learning models to predict cancer. This motivates us to propose a method that integrates clustering and classification techniques for diverse cancer types with survival analysis via regression to identify miRNAs that can potentially play a crucial role in the prediction of different types of tumors. Our method has two parts. The first part is a feature selection procedure, called the stochastic covariance evolutionary strategy with forward selection (SCES-FS), which is developed by integrating stochastic neighbor embedding (SNE), the covariance matrix adaptation evolutionary strategy (CMA-ES), and classifiers, with the primary objective of selecting biomarkers. SNE is used to reorder the features by performing an implicit clustering with highly correlated neighboring features. A subset of features is selected heuristically to perform multi-class classification for diverse cancer types. In the second part of our method, the most important features identified in the first part are used to perform survival analysis via Cox regression, primarily to examine the effectiveness of the selected features. For this purpose, we have analyzed next generation sequencing data from The Cancer Genome Atlas in form of miRNA expression of 1,707 samples of 10 different cancer types and 333 normal samples. The SCES-FS method is compared with well-known feature selection methods and it is found to perform better in multi-class classification for the 17 selected miRNAs, achieving an accuracy of 96%. Moreover, the biological significance of the selected miRNAs is demonstrated with the help of network analysis, expression analysis using hierarchical clustering, KEGG pathway analysis, GO enrichment analysis, and protein-protein interaction analysis. Overall, the results indicate that the 17 selected miRNAs are associated with many key cancer regulators, such as MYC, VEGFA, AKT1, CDKN1A, RHOA, and PTEN, through their targets. Therefore the selected miRNAs can be regarded as putative biomarkers for 10 types of cancer.
Copyright © 2020 Sarkar, Saha, Lancucki, Ghosh, Wlasnowolski, Bokota, Dey, Lipinski and Plewczynski.

Entities:  

Keywords:  KEGG pathway; cancer; cox regression; feature selection; gene ontology; machine learning; next generation sequencing; stochastic neighbor embedding

Year:  2020        PMID： 33281862      PMCID： PMC7691578          DOI： 10.3389/fgene.2020.00982

Source DB:  PubMed          Journal:  Front Genet        ISSN： 1664-8021            Impact factor:   4.599


1. Introduction

MicroRNAs (miRNAs) belong to the non-coding RNA family. They consist of 19–25 nucleotides and play an important role in the regulation of gene silencing. These non-coding RNAs are present in every eukaryotic cell and can also be encoded by a viral genome (Ray and Maiti, 2015; Bruscella et al., 2017). The miRNAs are formed by RNA polymerase II in the cell nucleus and are then transferred to the cytoplasm (Bartel, 2009) for biological activities such as cell cycle control, apoptosis, and oncogenesis. They interact with the complementary strand of mRNAs and lead to the degradation of the corresponding mRNAs; they also interfere with protein production by suppressing protein synthesis (Valencia-Sanchez et al., 2006). A miRNA molecule can bind one or more targets, thus forming a complex underlying regulatory network. These networks have a profound impact on cancer signaling pathways (Wang et al., 2017). Previously, low-throughput and high-cost technologies were the main obstacle to answering systems-level biological questions. However, recent advancements in next generation sequencing (NGS) have enabled researchers to address such complex problems (van Dijk et al., 2014). Moreover, new sequencing technologies and genomic datasets have helped us to gain better understanding of the biological complexities related to genomic abnormalities in cancer. The considerable achievements in sequencing techniques have made high-throughput techniques a fundamental platform for miRNA, RNA, and DNA research. Generally, miRNAs are involved in a wide range of diseases, including neurological disease, heart disease, and cancer (Giza et al., 2014; Paul et al., 2018). In many cases of cancer in humans, dysregulation of miRNA expression has been observed, and it is well-known that miRNAs can serve as potential cancer biomarkers (Lu et al., 2005; Jacobsen et al., 2013; Wong et al., 2017). In this regard, scientific communities are also trying to understand the role of miRNAs in paring with mRNAs (Zhang et al., 2014; Shrestha et al., 2017), in different cancer types by ranking miRNAs (Li et al., 2014) to elucidate their effects and drug resistance (Ma et al., 2010; Li and Yang, 2014; Cheerla and Gevaert, 2017). To reduce the time taken for clinical trials, and to provide better and more accurate treatments while avoiding unnecessary interventions, the proper selection of miRNAs as biomarkers is crucial. For this purpose, miRNAs are often used as features in statistical learning methods viz. clustering, classification, and regression in order to identify potential biomarkers (Song et al., 2016; Yang et al., 2017; Yokoi et al., 2017). Song et al. (2016) performed a clustering analysis on breast cancer data in order to find miRNAs that could be prognostic biomarkers; these miRNAs are up-regulated in this type of cancer and are linked to local relapse, distant metastasis, and poor clinical outcomes. Similarly, Yang et al. (2017) used clustering to find miRNA biomarkers for breast cancer. The identified miRNAs have higher specificity and sensitivity than single-gene biomarkers. On the other hand, Yokoi et al. (2017) used a classification task to inform the development of a predictive model to distinguish patients with ovarian cancer tumors from healthy subjects. In this study, eight miRNAs were found as biomarkers for ovarian cancer. Jacob et al. (2017) conducted a study on colon cancer and identified 16 miRNA signatures, which act as prognostic biomarkers at cancer stages II and III. Apart from the aforementioned works, regression analysis has been used to predict the survival rates of patients with different types of cancer (Liang et al., 2018). Liang et al. (2018) used Cox regression analysis for pancreatic cancer and identified five miRNAs as independent prognostic factors. The progress in this regard can be found in the literature (e.g., Peng and Croce, 2016; Hosseinahli et al., 2018). Generally, in miRNA-based cancer studies, statistical learning algorithms viz. clustering, classification, and regression are used separately for different cancer types, as described above and in the literature (Akhtar et al., 2015; Ang et al., 2016; Li et al., 2016; Lin and Lane, 2017). However, to leverage the advantages of the different algorithms, it may be useful to integrate them into a single method for identifying potential biomarker miRNAs. Besides, little work exists on multi-class classification of diverse cancer types using NGS data. These two facts motivated us to develop the method described in this paper, which can not only classify 10 types of cancer (bladder, breast, colon, glioblastoma, head and neck squamous cell, kidney renal clear cell, lung adenocarcinoma, lung squamous cell, ovarian, and uterine corpus endometrial carcinoma) but also find putative miRNAs that are highly associated with these cancer types. The proposed two-part wrapper-based feature selection method, referred to as the stochastic covariance evolutionary strategy with forward selection (SCES-FS), uses stochastic neighbor embedding (SNE) (Hinton and Roweis, 2003) in conjunction with the covariance matrix adaptation evolutionary strategy (CMA-ES) (Hansen et al., 2003) and a simple classifier, either random forest (RF) (Breiman, 2001), support vector machine (SVM) (Cortes and Vapnik, 1995), naive Bayes (NB) classifier (George and Langley, 1995), K-nearest-neighbors (K-NN) classifier (Altman, 1992), or decision tree (DT) (Quinlan, 1986), in the first part. Here SNE is used to reorder the features by performing an implicit clustering, such that neighboring features are highly correlated. Then, from these clusters of features, a subset of features is selected randomly in order to perform the multi-class classification task for the diverse cancer types. However, as the features are randomly selected, this classification task is treated as an underlying optimization problem for CMA-ES to find the features automatically. Hence, the final set of features/miRNAs is obtained by using forward selection (Whitney, 1971). In the second part of the method, survival analysis is performed using Cox regression (Cox, 1972), where the levels of miRNA expression and corresponding clinical data are used. The experiment is conducted with data collected from The Cancer Genome Atlas (TCGA) for 10 different types of cancer. The performance of the proposed wrapper-based feature selection method is compared with the following methods in terms of classification accuracy, with the top 17 miRNAs selected as putative biomarkers: ensemble SVM-recursive feature elimination (ESVM-RFE) (Anaissi et al., 2016), the least absolute shrinkage and selection operator (LASSO) (Tibshirani, 1996), the non-dominated sorting genetic algorithm II-based stacked ensemble (NSGA-II-SE) (Saha et al., 2017), the SVM-wrapped multi-objective genetic algorithm (MOGA) (Mukhopadhyay and Maulik, 2013), SVM-based novel recursive feature elimination (SVM-nRFE) (Peng et al., 2009), SVM recursive feature elimination (SVM-RFE) (Guyon et al., 2002), conditional mutual information (CMIM) (Fleuret, 2004), interaction capping (ICAP) (Jakulin, 2005), smoothly clipped absolute deviation (SCAD) (Fan and Li, 2001), joint mutual information (JMI) (Bennasar et al., 2015), conditional infomax feature extraction (CIFE) (Brown et al., 2012), minimum redundancy maximum relevance (mRMR) (Peng et al., 2005), feature selection with Cox regression (FSCOX) (Kim et al., 2014), double-input symmetrical relevance (DISR) (Brown et al., 2012), signal-to-noise ratios (SNRs) (Mishra and Sahu, 2011), and the Wilcoxon rank-sum test (RankSum) (Troyanskaya et al., 2002). Thereafter, the significance of the 17 selected miRNAs to 10 different cancer types is determined using Cox regression analysis. Finally, survival analysis, network analysis, expression analysis using hierarchical clustering in the form of heatmaps, KEGG pathway analysis (Kanehisa and Goto, 2000), gene ontology (GO) enrichment analysis (Kuleshov et al., 2016), and protein-protein interaction (PPI) network analysis (Szklarczyk et al., 2019) are performed to assess the biological significance of the selected miRNAs. Additionally, a web-based cancer predictor application is developed to predict 10 different types of cancer given the expression of 17 miRNAs.

2. Materials and Methods

In this section, we briefly describe SNE (Hinton and Roweis, 2003), CMA-ES (Hansen, 2006), and Cox regression analysis (Cox, 1972). More details about classification techniques and feature selection methods are given in the Supplementary Material for this article. This section also describes the proposed method, which consists of two parts. The first part is the wrapper-based SCES-FS. In the second part, the selected features, such as expression of miRNAs and clinical data, are used in survival analysis to assess the importance of the selected miRNAs in different types of cancer and to evaluate the effectiveness of the selected miRNAs. Figure 1 shows the flowchart of the proposed method.
Figure 1

Flowchart for the proposed method.

Flowchart for the proposed method.

2.1. Stochastic Neighbor Embedding

Let denote a set of n observations, where . SNE (Hinton and Roweis, 2003) constructs a low-dimensional embedding that recreates in a space of lower dimension as , where . In SNE, both and are represented as discrete probability distributions P and Q, where that model pairwise distances between data points. The values of var ∈ ℝ are adjusted in such a way that the entropies of all distributions P are equal. The mismatch between P and Q is reduced through minimization of Kullback-Leibler (KL) divergence objective , by altering Q with gradient-based optimization methods. Optimization is difficult due to the existence of multiple local optima, and entirely different embeddings may be obtained with different initial Q distributions.

2.2. Covariance Matrix Adaptation Evolution Strategy

Evolutionary strategies are black-box optimization algorithms which belong to a broader group of evolutionary algorithms. In such methods, a set of candidate solutions is maintained. In successive iterations of the procedure, these candidate solutions are perturbed and evaluated, and in each iteration the best solution is left unchanged and carried over to the next set of candidate solutions. CMA-ES (Hansen and Ostermeier, 1996) is an evolutionary strategy where the set of candidate solutions is modeled and sampled from a multivariate Gaussian distribution . The covariance matrix C represents pairwise relationships between attributes. The objective function of CMA-ES maximizes two likelihoods: (a) the likelihood of having the best individuals in previous iterations, and (b) the likelihood of taking the best search steps in previous iterations. At the end of each iteration, (a) guides updates of the mean m as a weighted average of μ best solutions, where is the ith best solution in iteration g + 1 and w is its weight, and (b) updates the covariance matrix C as follows: Finally, is a vector that amplifies the updates in favorable directions: where c1, cμ, c ∈ ℝ are weights, σ( ∈ ℝ is an adaptive step size, which is dependent on the iteration, and z ∈ ℝ is a normalizing constant. Details of the parameters of CMA-ES can be found in Hansen and Ostermeier (1996) and Hansen et al. (2003).

2.3. Cox Regression Analysis

The Cox regression model (Cox, 1972) is a proportional hazards regression model in which the hazard ratio is constant but other contents have the same baseline hazard function. Based on this assumption, the survival function is calculated as where H0(τ) represents the cumulative baseline hazard function at time τ and S0(τ) = exp(−H0(τ)) is the baseline survival function; H0(τ) is taken to be Breslow's estimator (Breslow, 1974), which is the most widely used and given by As the Cox model is based on the proportional hazards assumption, it is represented as for an given instance i = 1, 2, 3, …, n, where the baseline hazard function h0(τ) can be an arbitrary negative function of time, and x = (x, x, …, x) is the corresponding covariate vector for instance i and is the coefficient vector. The Cox model is a semi-parametric algorithm where the baseline hazard function h0(τ) is unspecified. For any two instances x1 and x2, the hazard ratio is given by This means that the hazard ratio is independent of the baseline hazard function.

2.4. Wrapper-Based Feature Selection Integrating SNE and CMA-ES

The first part of our proposed method performs the task of miRNA selection for diverse cancer types, which is considered a multi-class classification problem here. SNE is used to reorder the features by performing an implicit clustering such that neighboring features are highly correlated. Then, the underlying multi-class classification task is performed using well-known classifiers and treated as an optimization problem for which CMA-ES is used to find the miRNAs automatically. The miRNAs thus found are further refined using forward selection (FS). Therefore, we call this wrapper-based feature selection method as stochastic covariance evolutionary strategy with forward selection (SCES-FS). Algorithm 1 presents the SCES-FS method in detail. It starts with the dataset , which denotes a set of n observations with , the class label , the population size λ, the maximum number of generations N, the classifier , and the number of runs as inputs. In the dataset, each feature is characterized by expression levels of the samples. The features of the original dataset are reordered using SNE in the ConstructEmbedding step, producing the embedding dataset whose size is the same as that of the original dataset. The parameters are initialized in the SetParamsCMAES step. The individuals/vectors in CMA-ES are encoded as a simple threshold weight vector, with a single weight for each miRNA. An individual x ∈ ℝ encodes a weight vector w ∈ ℝ and a threshold t ∈ ℝ, where H ≤ D is the number of weights. Only those features whose weights exceed the threshold are eventually selected into a feature set S: where t ∈ ℝ is a threshold that is chosen carefully. The population of vectors is drawn in the DrawPopulationCMAES step from , where m is the mean, C is the covariance matrix, and σ is the step-size control parameter of CMA-ES. In the ConstructFeatureSets step, each individual x is translated to a feature set S according to Equation (8). Pseudo-code of the SCES-FS In the ScoreFeatureSet step, each feature set is evaluated by training a classifier on a dataset restricted to only the selected features. The objective function of CMA-ES combines the accuracy obtained from the classifier for each individual x and the size of the feature set S as follows: where D is the dimension of the dataset, Υ is the target number of features, and α ∈ ℝ+ balances the penalty for the excessive number of features. The objective function is designed in this way so that higher classification accuracy can be achieved for a small number of features, and it incorporates L0 regularization term as penalty, which is bounded by Υ in order to have redundancy in selected subsets of features. The optimization with CMA-ES allows inter-feature relationships with covariance matrices. Finally, the parameters are updated according to CMA-ES update rules in the UpdateParamsCMAES step, and the best set of features/miRNAs in a particular run are kept in SRBest.

2.5. Preparation of the Final Set of Features

The SCES-FS is random in nature to reduce the probability of returning sub-optimal solutions. Therefore, a single run of SCES-FS does not guarantee a reliable solution. To overcome this limitation, SCES-FS is executed up to a maximum number of runs, . In each run, a set of best features/miRNAs are collected into a list, L. After completion of the maximum number of runs, RankFrequency sorts the cumulative set of features in descending order according to the frequency of occurrence in each run and produces a modified list, L′. Thereafter, ForwardSelection applies the forward feature selection method, using the classifier to evaluate the feature set iteratively to obtain the best feature set, Sbest.

2.6. Justification for miRNA Selection

Because of the random nature of the algorithm, there is a chance of false positives or false negatives occurring in the selection of miRNAs. To reduce the probability of having false positives or false negatives in the selected set of miRNAs, the SCES-FS algorithm is run 50 times, and then the miRNAs are ranked based on their frequencies of occurrence in 50 different sets of features. Thus, a stable set of miRNAs is selected on the basis of maximum classification accuracy, which is computed by considering the miRNAs cumulatively from the top of the list. By this process, 17 distinct miRNAs are selected. This procedure does not, however, ensure the absence of false positives or false negatives, so additional measures are taken. The presence of false positives is made unlikely by the sorting procedure. In fact, since a given false positive is supposed to occur less frequently than all the true positives, it will appear in the tail after sorting. Consequently, it is likely that this false positive will be filtered out when selecting the final list of miRNAs. On the other hand, the presence of false negatives is related to the choice of the number of runs and the corresponding expected number of miRNAs belonging to the sorted list. In this regard, a mathematical argument is given below to justify that the SCES-FS does not exhibit random behavior. Suppose that at each run the SCES-FS randomly selects 10 miRNAs from the whole collection of miRNAs. We first compute the expected number of distinct miRNAs reported after all the runs, and then we compare this number with the results of our experiments. We have the following parameters: D = 199 is the number of miRNAs, S = 10 is the number of miRNAs returned after each run, and is the number of runs. For i = 1, …, D, let V be a Bernoulli-distributed indicator variable, where V = 1 if the miRNA m never shows up. The probability that m is selected in one run is S/D, so the probability that it is never selected is 1−S/D. Since each of the runs is independent, the following equation can be written: Let be the random variable that counts the number of miRNAs that do not belong to the final set of miRNAs reported at least once. By linearity of the expectation, we obtain the equation Hence, the number of expected miRNAs reported at least once is Substituting the above-mentioned values for the parameters, we obtain that the expected number of miRNAs reported at least once after 50 iterations is 183; and the number of new miRNAs added in a further iteration would be 1. As the number of sorted miRNAs in our experiment is 39 (see the Supplementary Material), this difference suggests that 50 runs are enough to conclude that all the true positives are included in the sorted list and so false negatives are unlikely.

2.7. Cox Regression Analysis for Evaluating miRNAs

The second part of the proposed method is for the evaluation of selected miRNAs using Cox regression analysis. The primary objective of this stage is to assess the importance of the miRNAs selected in the first part of the method, and this is done using Cox regression analysis for survival in 10 different cancer types. Expression data of the selected miRNAs and the associated clinical data are used for the Cox regression analysis. In the clinical data, the vital status of each patient, indicating whether the patient is still alive or has died, and the number of days since the last followup are taken into account when performing the survival analysis. Based on the expression levels and the clinical data of the selected miRNAs in each cancer type, the Cox coefficient, hazard ratio, and p-value are computed. A higher value of the Cox coefficient signifies greater importance of that miRNA to the respective cancer type. Moreover, the up- and down-regulation of all the selected miRNAs are observed to understand their behavior with respect to that particular cancer type based on change in expression in tumor and normal samples.

2.8. Complexity Analysis

Let D be the number of features and n the number of samples in the input dataset. As the available approximations may considerably lower the overall complexity, we discuss the complexity of each building block separately. A single step of SNE requires computing relations between all data points. We embed a transposed dataset, making an optimization step in O(D2) time. Computing SNE usually involves performing principal component analysis for preliminary dimension reduction, though its cost is negligible. The internal complexity of CMA-ES is estimated as O(D2), due to sampling and updating of the covariance matrix. The matrix needs to be factorized, which can be done by eigen decomposition in O(D3) time. Factorization does not happen in every generation, which gives O(D2) amortized time. Empirical evidence suggests that the sufficient number of objective function evaluations usually scales sub-quadratically with D (Ros and Hansen, 2008). The computation time is similar, since the vast majority of it is time spent training similar classifiers. On the other hand, the time complexity of Cox regression analysis is O(nD2) for a single run (Kelley, 1999).

3. Results and Discussion

The performance of the proposed method (SCES-FS) was tested on real miRNA expression datasets for 10 different cancer types and compared with the results of 16 existing methods, namely ESVM-RFE (Anaissi et al., 2016), LASSO (Tibshirani, 1996), NSGA-II-SE (Saha et al., 2017), MOGA (Mukhopadhyay and Maulik, 2013), SVM-nRFE (Peng et al., 2009), SVM-RFE (Guyon et al., 2002), CMIM (Fleuret, 2004), ICAP (Jakulin, 2005), SCAD (Fan and Li, 2001), JMI (Bennasar et al., 2015), CIFE (Brown et al., 2012), mRMR (Peng et al., 2005), FSCOX (Kim et al., 2014), DISR (Brown et al., 2012), SNRs (Mishra and Sahu, 2011), and RankSum (Troyanskaya et al., 2002) (see section 1 for the full names of these methods), as well as the results with all features (i.e., without feature selection).

3.1. Dataset Preparation and Parameter Setting

The miRNA expression and clinical datasets of bladder, breast, colon, glioblastoma, head and neck squamous cell, kidney renal clear cell, lung adenocarcinoma, lung squamous cell, ovarian, and uterine corpus endometrial carcinoma were obtained from TCGA. These 10 cancer types have also been studied previously (see, e.g., Jacobsen et al., 2013). Moreover, Hoadley et al. (2018) found that the characteristics of certain cancers out of 33 types provided in TCGA are overlapping in nature. As a result, 10–15 distinct groups of cancer were reported in Hoadley et al. (2018), which are similar to the cancer types studied in the present article. Our choice of cancer types was based on: (1) careful review of the literature, (2) the availability of tissue-specific tumor and normal samples to avoid the class imbalance problem in classification, and (3) the availability of common miRNA expression data for different cancer types and their corresponding clinical data. Thus, 10 cancer types were selected for our study. The expression data were generated using an Illumina high-throughput sequencing machine in the form of read counts of 199 miRNAs, normalized to reads per million, while the clinical data contain gender, age, days since last followup, and vital status. After removing miRNAs that contain more than 60% zeros in each cancer type and taking the miRNAs common to all the cancer types, we found 199 miRNAs and considered their expression in different cancer types. The number of samples and other details for each cancer type are given in Table 1; we also included 333 normal samples in the analysis with the expression of same miRNAs. For convenience, the expression datasets containing reads per million are further normalized onto a log2 scale. These processed datasets can be downloaded from the Supplementary Material website. To construct the final ranking of miRNAs, the SCES-FS algorithm was run 50 times. Five-fold cross-validation was applied during each classification to avoid the issue of overfitting or underfitting. The parameters used in the experiments are shown in Table 2 and were obtained either experimentally or from the literature (Latinne et al., 2001; Oshiro et al., 2012; Hansen, 2016).
Table 1

Details of the data for 10 different cancer types.

Cancer typeCodeNo. of tumor samplesGenderAverage ageAverage no. days from last followup
MaleFemale
Bladder urothelial carcinomaBLCA94672767.07416.97
Breast invasive carcinomaBRCA255025558.281297.82
Colon adenocarcinomaCOAD119576270.91616.80
Glioblastoma multiformeGBM38201862.78376.81
Head and neck squamous cell carcinomaHNSC2982188060.981038.31
Kidney renal clear cell carcinomaKIRC146915559.991236.55
Lung adenocarcinomaLUAD61293265.62743.09
Lung squamous cell carcinomaLUSC89642564.781296.87
Ovarian serous cystadenocarcinomaOV509050959.851020.87
Uterine corpus endometrial carcinomaUCEC9809862.321066.01
Table 2

Values of parameters.

SymbolValueDescription
CICMA-ES initial covariance matrix of size H
σ0.3Initial value of step-size control parameter
λ200Population size
μ100Number of parents
R50Number of runs
N200Maximum number of generations
t0.5Threshold for calculating subsets Si
α0.5Excessive attributes penalty term
Υ10Target number of miRNAs
γ0.05SVM RBF kernel parameter
C1.0SVM C constant
K5Value of K in K-NN
M50Number of trees in RF
Details of the data for 10 different cancer types. Values of parameters.

3.2. Experimental Outcomes

The problem of finding miRNAs that can correctly distinguish different cancer types were posed as a multi-class classification task using expression data of miRNAs, and the importance of the selected miRNAs was evaluated using Cox regression analysis with the help of clinical data. The classification results of SCES-FS using the classifiers RF, SVM, NB, K-NN, and DT over 50 runs are reported in Table 3 and compared with the results of ESVM-RFE, LASSO, NSGA-II-SE, MOGA, SVM-nRFE, SVM-RFE, CMIM, ICAP, SCAD, JMI, CIFE, mRMR, FSCOX, DISR, SNR, and RankSum, as well as the results with all features (i.e., without feature selection). These results show that SCES-FS with RF achieved the highest classification accuracy, 96.881, with a standard deviation of 0.039, while the accuracy results of SCES-FS with SVM, NB, K-NN, and DT are 96.332 ± 0.194, 96.251 ± 0.168, 96.132 ± 0.369, and 94.232 ± 0.057, respectively, still better than those of the other existing feature selection methods.
Table 3

Number of features and classification accuracy of feature selection methods for five classifiers with five-fold cross-validation.

MethodNumber of featuresRFSVMNBK-NNDT
SCES-FS1796.881 ± 0.03996.332 ± 0.19496.251 ± 0.16896.132 ± 0.36994.232 ± 0.057
ESVM-RFE2295.684 ± 0.03195.902 ± 0.19392.672 ± 0.16191.382 ± 0.36990.429 ± 0.056
LASSO4895.601 ± 0.03895.547 ± 0.19192.582 ± 0.16491.251 ± 0.36790.241 ± 0.051
NSGA-II-SE2695.587 ± 0.03395.537 ± 0.19592.538 ± 0.16691.183 ± 0.36690.229 ± 0.052
MOGA2495.391 ± 0.03695.531 ± 0.19492.293 ± 0.16190.338 ± 0.36289.142 ± 0.055
SVM-nRFE2695.224 ± 0.03295.321 ± 0.19192.281 ± 0.16790.106 ± 0.36188.993 ± 0.051
SVM-RFE2895.048 ± 0.03895.159 ± 0.19992.116 ± 0.16589.889 ± 0.36688.691 ± 0.053
CMIM2894.299 ± 0.03792.029 ± 0.19390.683 ± 0.16189.374 ± 0.36989.161 ± 0.052
ICAP2793.721 ± 0.03192.951 ± 0.19290.874 ± 0.16390.643 ± 0.36287.057 ± 0.053
SCAD2591.972 ± 0.03490.839 ± 0.19490.003 ± 0.16589.918 ± 0.36687.495 ± 0.058
JMI2891.718 ± 0.03190.602 ± 0.19689.986 ± 0.16688.639 ± 0.36987.057 ± 0.051
CIFE3290.886 ± 0.03489.072 ± 0.19988.389 ± 0.16287.261 ± 0.36286.205 ± 0.056
mRMR2891.063 ± 0.03289.753 ± 0.19587.402 ± 0.16187.254 ± 0.36185.208 ± 0.059
FSCOX2389.298 ± 0.03888.529 ± 0.19887.505 ± 0.16987.287 ± 0.36885.498 ± 0.058
DISR2989.286 ± 0.03188.276 ± 0.19687.580 ± 0.16787.321 ± 0.36985.858 ± 0.053
SNR3087.866 ± 0.03586.712 ± 0.19385.749 ± 0.16385.364 ± 0.36584.042 ± 0.059
RankSum3286.633 ± 0.03385.556 ± 0.19985.466 ± 0.16684.669 ± 0.36684.322 ± 0.055
Without feature selection19986.428 ± 0.03685.294 ± 0.19185.183 ± 0.16284.552 ± 0.36784.118 ± 0.053
Number of features and classification accuracy of feature selection methods for five classifiers with five-fold cross-validation. It is to be noted that our results have been verified with FSCOX, which computes important miRNAs via Cox regression on all miRNAs and subsequently uses them for the classification. It has been observed that the overall classification accuracy of FSCOX is less than the accuracy attained by SCES-FS as reported in Table 3. Other omics-based analyses, such as protein array, copy number variation (CNV), and methylation studies, also have potential applications in pan-cancer classification, as explained in Zhang et al. (2015, 2016, 2019), where the overall accuracy achieved in the range of 93–97%. The cancer dataset used in Zhang et al. (2015, 2016, 2019) was carefully selected for protein array, CNV, and methylation studies and is not directly suitable for experimenting with miRNAs, as it creates a class imbalance problem when selecting miRNA expression data for both tumor and normal cases. However, the overall accuracy of our method is higher than 96%, which is on the higher side of the accuracy range reported by Zhang et al. (2015, 2016, 2019), suggesting that our selected miRNAs can also be considered potential markers for pan-cancer classification. The results of SCES-FS have been further validated by survival analysis, including Cox regression, network analysis, expression analysis using hierarchical clustering in the form of heatmaps, KEGG pathway analysis, GO enrichment analysis, and PPI network analysis, as described in the following.

3.2.1. Survival Analysis

Table 4 reports the results of the Cox regression analysis, i.e., the Cox coefficient and hazard ratio values, for 10 cancer types in order to evaluate the importance of the 17 selected miRNAs with respect to these cancer types. The Cox regression analysis was performed by integrating the miRNA expression and clinical data to assess the effect of miRNA expression on cancer survival. Higher Cox coefficient and hazard ratio values indicate greater influence of the miRNA on the cancer type. For example, the miRNA hsa-mir-375 has the highest Cox coefficient, 0.9882, and hazard ratio, 1.7879, for the GBM cancer type. To help visualize the importance of each miRNA to the different cancer types, a circos plot of Table 4 is shown in Figure 2. In the figure, a broader band signifies a stronger association between the miRNA and the particular cancer type. Table 5 summarizes the cancer types that are most highly associated with the selected miRNAs. In this table, for each miRNA, the cancer type for which that miRNA has the highest Cox coefficient is shown, along with the associated p-value, false discovery rate (FDR), and up- or down-regulation of the miRNA expression in that cancer type. Similarly, Supplementary Table 2 gives the up- and down-regulations along with the corresponding FDR values of the selected miRNAs for all cancer types. The up- and down-regulations of the miRNAs are computed after evaluating the change in expression of the selected miRNAs between tumor samples and normal samples or the population for a particular cancer type using the ANOVA test; this is discussed further in section 3.2.3. Here, the change in expression is considered significant if the p < 0.05, and the up- or down-regulation is computed if the change in population expression is positive or negative, respectively. In summary, we find that hsa-mir-205, hsa-mir-10a, hsa-mir-196b, hsa-mir-10b, hsa-mir-375, hsa-mir-143, hsa-let-7c, hsa-mir-107, hsa-mir-378, hsa-mir-133a, hsa-mir-1, hsa-mir-30c, hsa-mir-16, hsa-mir-30a, hsa-let-7i, hsa-mir-24, and hsa-mir-95 are highly associated with GBM, UCEC, OV, BLCA, COAD, and BRCA.
Table 4

Results of Cox regression analysis of each selected miRNA.

miRNABLCABRCACOADGBMHNSCKIRCLUADLUSCOVUCEC
Cox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratioCox coefficientHazard ratio
hsa-mir-205−0.03440.9661−0.07950.92340.11091.11720.92332.51770.05861.06000.01871.01890.12921.1379−0.04490.9560−0.02240.9778−0.00390.9960
hsa-mir-10a0.06181.0637−0.07500.9277−0.16440.84830.48811.6293−0.04280.9580−0.01220.98780.11931.12670.16191.17580.05691.0586−0.40750.6652
hsa-mir-196b−0.03330.96720.00891.0090−0.22360.79960.25731.2935−0.00950.99050.13671.14650.11241.1189−0.02820.97210.01461.01470.45811.5811
hsa-mir-10b0.08281.0864−0.22130.80140.40421.49810.07951.08270.05971.0615−0.27430.76000.16261.17660.01761.0178−0.02180.97830.50801.6621
hsa-mir-375−0.02260.97760.13341.1427−0.03130.96910.98821.78790.05251.0539−0.00510.9948−0.09670.9078−0.03400.9665−0.07680.92600.02431.0246
hsa-mir-1430.11581.1228−0.24720.7809−0.07590.9268−0.06540.9366−0.06300.9389−0.04330.9575−0.14860.86190.00531.00540.15041.16230.07701.0801
hsa-let-7c0.17311.1890−0.18250.83310.07021.0727−0.13490.87370.04301.0440−0.11580.8905−0.15440.85690.06651.06870.06181.06370.00901.0090
hsa-mir-107−0.03950.96110.11271.11930.04481.04590.53631.7096−0.03890.96170.04051.04130.08851.0925−0.09260.9114−0.09190.91210.06091.0628
hsa-mir-378−0.02860.9717−0.13550.87320.15201.16410.55111.73520.06881.0712−0.18810.8284−0.12280.88430.17481.19100.05661.0583−0.15960.8524
hsa-mir-133a0.06101.06290.16041.1740−0.10250.90250.26831.4481−0.03930.96140.03281.03340.08101.0844−0.05890.94270.44221.55610.21601.2411
hsa-mir-10.06301.0651−0.11500.8913−0.06370.93820.37021.3078−0.02950.9708−0.02080.97930.00891.00900.09541.10010.15801.17120.00841.0084
hsa-mir-30c−0.00230.9976−0.05500.94640.50531.6575−0.01680.98330.06211.0641−0.01210.9879−0.44430.64120.14381.1547−0.00560.9943−0.61180.5423
hsa-mir-16−0.03260.96780.09471.0993−0.06340.93850.14091.1513−0.03700.96360.02661.02700.04711.0482−0.07320.92930.00581.00580.04471.0457
hsa-mir-30a0.08451.0882−0.15890.85300.80832.24410.05651.05820.08081.0841−0.07790.9250−0.16800.84520.07641.07940.07371.0765−0.70610.4935
hsa-let-7i0.13111.14010.43831.55010.19151.21100.23271.2621−0.03540.96510.13941.14960.25741.29360.07021.0728−0.10470.90050.03531.0360
hsa-mir-240.01911.0193−0.08630.91720.06591.06810.48851.62980.03931.0401−0.03450.9660−0.07820.92460.06661.06880.05031.0516−0.11750.8890
hsa-mir-95−0.03450.96600.11851.1258−0.02820.97210.28721.33260.01841.01860.13271.14190.03201.03250.01021.01020.04671.0478−0.12470.8827
Figure 2

Circos plot of Cox regression analysis results: Cox coefficient values are used to graphically visualize the association of 17 miRNAs with ten cancer types; a broader band signifies a stronger association between the miRNA and the particular cancer type.

Table 5

Cancer type most strongly associated with each selected miRNA, based on Cox coefficient.

miRNACox coefficientHazard ratioCancer typep-valueFDRRegulation (up ↑/down ↓)PubMed ID
hsa-mir-2050.92332.5177GBM5.14E-035.46E-03↓23054677
hsa-mir-10a0.48811.6293GBM6.17E-241.87E-23↓20444541
hsa-mir-196b0.45811.5811UCEC2.84E-481.21E-47↑–
hsa-mir-10b0.50801.6621UCEC4.71E-034.71E-03↓–
hsa-mir-3750.98821.7879GBM1.80E-142.36E-14↓29110584
hsa-mir-1430.15041.1623OV3.38E-571.79E-56↓25304686
hsa-let-7c0.17311.1890BLCA3.85E-381.31E-37↓21464941
hsa-mir-1070.53631.7096GBM5.42E-241.87E-23↑24213470
hsa-mir-3780.55111.7352GBM5.04E-207.79E-20↓29088758
hsa-mir-133a0.44221.5561OV2.78E-204.73E-20↑24944666
hsa-mir-10.37021.3078GBM9.18E-081.04E-07↑–
hsa-mir-30c0.50531.6575COAD4.48E-176.34E-17↓–
hsa-mir-160.14091.1513GBM5.42E-241.87E-23↑25864039
hsa-mir-30a0.80832.2441COAD2.88E-569.80E-56↓22287560
hsa-let-7i0.43831.5501BRCA1.33E-432.83E-43↑26378051
hsa-mir-240.48851.6298GBM5.42E-241.87E-23↓25864039
hsa-mir-950.28721.3326GBM6.59E-241.87E-23↑28155650
Results of Cox regression analysis of each selected miRNA. Circos plot of Cox regression analysis results: Cox coefficient values are used to graphically visualize the association of 17 miRNAs with ten cancer types; a broader band signifies a stronger association between the miRNA and the particular cancer type. Cancer type most strongly associated with each selected miRNA, based on Cox coefficient.

3.2.2. Network Analysis

For the 17 selected miRNAs, miRTarBase (Huang et al., 2020) was used to find their targets in order to elucidate their role in the different cancer types. To identify the most correlated targets, we computed the Pearson correlation between the expression values of miRNAs and mRNAs obtained from TCGA for the 10 cancer types and took the negative correlation value used in Zhou et al. (2015) as indicating strong association. The top five negatively correlated mRNAs associated with each of the 17 miRNAs are reported in Table 6; the rest are reported in Supplementary Material. To construct the interaction network, the miRNAs and their targets were ranked based on the cumulative negative correlation score and their presence in different cancer types as indicated by the association number. These results are reported in Supplementary Table 3. For example, hsa-mir-16 and its target, PHYHIP, are related to six cancer types and the cumulative negative correlation score is −4.142. Similarly, hsa-mir-24 is correlated with C1QTNF6 in another six cancer types, with a cumulative negative correlation score of −3.906. The subset of such targeted mRNAs is used to construct the interaction network shown in Figure 3. The network reveals that the miRNAs hsa-mir-205, hsa-mir-10a, hsa-mir-107, hsa-mir-378, and hsa-mir-16 and their targets {CYR61, STARD8, TNFSF8}, {TTYH3, CARHSP1, LILRA2}, {CPEB3, TGFBR3, FGF2}, {NME4, NWD1, ORAI2}, and {PHYHIP, CPEB3, CTDSPL} play a crucial role in different types of cancer. The red, blue, pink, dark green, light green, and black edges in Figure 3 signify that the number of cancer types associated with the corresponding pair of miRNA and target mRNA is 6, 5, 4, 3, 2, and 1, respectively. The targets of the miRNAs are investigated further using KEGG pathway, GO enrichment, and PPI network analysis in the following subsections in order to see their impact on the different types of cancer.
Table 6

Association of the 17 selected miRNAs and their top five targets in 10 cancer types.

BLCABRCACOADGBMHNSCKIRCLUADLUSCOVUCEC
Corr.Corr.Corr.Corr.Corr.Corr.Corr.Corr.Corr.Corr.
miRNAmRNAscoremRNAscoremRNAscoremRNAscoremRNAscoremRNAscoremRNAscoremRNAscoremRNAscoremRNAscore
ZEB2−7.45E2F1−3.29SLC7A2−9.95PDLIM5−5.28RCAN2−4.89MAF−4.68ANGPTL7−4.15TRPV2−6.43PARD6B−2.54SLC7A2−5.22
ZEB1−7.05SATB2−2.94TIMP1−9.63ZNF707−4.61STARD8−4.61VEGFA−4.58ITM2A−3.88ALPK3−6.20PLCXD2−2.10SRC−3.77
hsa-mir-205SYT11−7.02RAB11FIP3−2.88SESN3−9.43ANKRD50−4.25ESRRG−4.60SLC37A4−3.83CPEB3−3.54STARD8−6.16C11orf74−2.08PARD6B−3.55
RCAN2−6.99ZFHX3−2.70SAMD8−8.99ERBB3−4.25ZEB1−4.55FLCN−3.79FAM19A1−3.47ENPP4−6.04FGF2−2.05PISD−3.23
LRRK2−6.99CDK1−2.63HOXA11−8.78YES1−4.14CTGF−4.40FGFR1OP−3.55CTGF−3.17LPCAT1−6.02SLC39A14−1.97BCL9L−3.13
NACC2−5.33E2F1−3.93H3F3C−9.62SLC2A3−5.45PANX1−3.16TTYH3−6.29TAF1D−3.10YOD1−4.23ARSK−2.11IRGQ−5.32
CLIC4−5.29BIRC5−3.76FHL2−9.11KIAA1143−5.41CARHSP1−3.02SCD−6.28DVL1−2.93ANP32E−3.46RIOK2−2.04FEM1A−4.80
hsa-mir-10aCOL6A2−5.04PPM1G−3.69MTR−8.99YOD1−5.36FHL2−2.79CD3D−6.19AHCYL2−2.85CHMP1B−3.37ZBTB10−2.03E2F1−4.78
RAP1A−4.79TIMM50−3.60SFT2D2−8.82BCL6−5.18TFAP2A−2.67KLHL6−6.12PABPC1−2.77HNRNPF−3.20CHL1−2.01NF2−4.77
TGFB3−4.77TPI1−3.54NF2−8.79DUSP3−5.12EBNA1BP2−2.44COL6A2−6.03YAP1−2.65NOP16−3.13TRA2B−1.99CHRNA5−4.71
GATA6−5.61REEP5−2.67KCTD21−9.89MAP2K2−5.92PRUNE2−3.57HSD17B10−2.28MEIS1−3.96TGFBR2−7.15MYC−2.45HOXB7−3.46
PRUNE2−5.26DCTN4−2.18ACER2−9.74MYC−4.55BEST3−3.05CALM1−2.16TBRG1−3.85LAMB2−5.98MARS2−1.97HOXB8−3.17
hsa-mir-196bTGFBR2−4.74PBX1−2.03BCAR3−9.68PRKACA−4.42IGDCC4−2.82PBX1−1.86REEP5−3.80TRPC3−5.67GATA6−1.91SLC23A2−2.95
NR4A3−4.65SUOX−1.81IGF2BP3−9.43IARS−4.11KLHDC8B−2.59C14orf37−1.43TRPC3−3.53NR4A3−5.52ALDOA−1.78TGFBR3−2.88
SNX9−4.59TLE3−1.78HIST1H2BD−9.10HMGA1−4.07NR4A3−2.33SUOX−1.41TGFBR2−3.32GATA6−5.11GGA3−1.63GATA6−2.61
TPM1−4.58PLK1−7.27INHBA−9.91CMPK1−6.23RNF2−2.50TUBA1B−5.41MARVELD3−3.34LILRA2−4.68SDC1−2.79ASCL2−3.78
SFRP1−4.13BUB1−6.83MBNL3−9.82PDK3−4.45TTYH3−2.43HTATIP2−5.11NPEPPS−2.92AHCYL2−4.52INHBA−2.72GLB1L3−3.53
hsa-mir-10bMBNL1−4.08CCNA2−6.76OPA3−9.61EXOSC2−4.29UBE2Z−2.20LILRB2−4.95TRIM2−2.87S1PR2−4.17CMPK1−2.69PLA2G2C−2.95
PPP3CB−3.86MELK−6.76SLC2A3−9.55HNRNPF−4.28SLC5A5−2.15LILRA2−4.86FAHD1−2.59PAG1−3.95TMED5−2.26GPCPD1−2.94
SGCD−3.78POC1A−6.70PPP1R13B−9.35EIF1−4.21FZD2−2.09PLK1−4.86ALKBH4−2.57FGD4−3.74SLC2A3−2.14MSTO1−2.73
CTGF−4.17FAM89A−6.15SON−9.94REEP3−5.94CALU−4.71HEY1−3.62JAG1−5.77PIK3CA−4.28DPYSL3−2.98KLHDC8B−4.97
FSTL3−4.09RHOQ−5.97LIMD2−9.81BAK1−5.02COL12A1−4.50ZNF785−3.38KLF4−5.58NUP54−3.67CLDN1−2.50ASAP2−4.73
hsa-mir-375TNS1−4.09CFL2−5.87ATG7−9.80SPRED1−4.97EXT1−4.28CDCA7L−3.28MBD2−5.43ARNTL2−3.66IL1RAP−2.06SFT2D2−4.48
SH3D19−4.01ACSL4−5.86JAK2−9.58CARD8−4.64CCDC88A−4.12CARD8−3.15AKAP7−5.21JAG1−3.58COL12A1−2.03CLDN1−4.22
SAMD4A−3.99CELF2−5.71ESPNL−9.54ZNF799−4.58NETO2−4.04NETO2−3.13CRIM1−5.14USP46−3.46SEC23A−1.98TSC22D2−4.20
OTUB1−5.34CENPM−5.21NKPD1−9.87HIST1H2BG−4.67SDC1−2.79FSD2−2.98TIMM8A−5.13COX6B1−4.67MAT2A−2.64FHIT−3.62
CAPZA1−4.88PIK3R2−4.73TIAL1−9.81ANG−4.30RAB22A−2.43DTNB−2.81RAB10−5.01TRUB2−4.34SLC25A33−2.35SNX22−2.89
hsa-mir-143SYNPO2L−4.88STXBP2−4.69TUBD1−9.51RER1−3.97TMEM40−2.39TMEM120B−2.48PRMT3−5.00RPS19−4.06RACGAP1−2.26C4orf19−2.85
STXBP2−4.79LMNB2−4.68PHAX−9.48GLB1L−3.90RAB10−2.38MRPS25−2.47PTCD3−4.93AKT2−4.01GPSM2−2.22RDH10−2.39
KCNA7−4.65AP1S1−4.66TTC38−9.44ADCY2−3.88NKPD1−2.30QPRT−2.40C15orf48−4.89OTUB1−3.94CAPZA1−2.17ZNF117−2.32
YWHAZ−4.49CCNF−5.94HMGXB4−9.81TRIB1−6.35ITGA3−5.38RRM2−4.23LDHA−7.24COX6B1−3.43CASP3−2.68TNFRSF10B−3.26
SLC20A1−4.16CKS2−5.89HES5−9.80ACTB−6.14LDHA−5.05DLX4−4.04EZH2−6.88NAA20−3.32E2F6−2.44SOD2−3.18
hsa-let-7cEFHD2−3.93CCNB2−5.84YWHAZ−9.79THBS1−6.11MT2A−4.81TNFSF9−3.96HMGA1−6.79SF3B4−3.17COIL−2.35RNF7−2.95
MRPL12−3.86RRM2−5.72WDR3−9.67SLC20A1−5.80HMGA2−4.77CCNB2−3.96EIF4A3−6.74KIAA0391−3.14RNFT1−2.27CEBPB−2.64
PCGF3−3.73H2AFZ−5.65LYN−9.47FHL2−5.54PSMB2−4.28FANCI−3.79MRPL12−6.67YWHAZ−3.11GABPB1−2.24ICOSLG−2.49
GLP2R−7.83CAV1−8.34SMARCA5−9.64REL−4.65CPEB3−5.23FOXC1−8.31RS1−8.88RS1−8.75SUN2−4.14CPEB1−5.76
PER1−7.67FGF2−8.05TMEM87A−9.63PIK3R1−3.59CHRM1−4.70PLAG1−8.24SH3GL2−8.86ALDH3B1−6.89ERN1−2.97CYSLTR2−3.71
hsa-mir-107CPEB1−7.62FOXO1−7.82CDC42SE2−9.52ZBTB38−3.43AMOT−4.09CPEB3−8.13TGFBR3−8.45LATS2−6.77VCAN−2.96FGF2−3.41
NFIA−7.32DST−7.71PURA−9.42CAV1−3.36TGFBR3−3.51SH3GL2−8.09CAV1−8.26PRKCE−6.36PAG1−2.85SLC28A1−2.91
DMPK−7.00KLF4−7.62UBE2Q1−9.04ADORA3−3.08NFIA−3.36CKMT1A−7.93FGF2−8.15PAG1−6.34YWHAH−2.61DAPK1−2.73
ENO1−5.57PTBP1−6.96TMEM154−9.76NWD1−6.56SERPINH1−6.16PRKD2−5.07BYSL−6.95MELK−4.10KLHL7−3.51MYOZ3−3.84
TTC4−5.48RABEP2−6.75MYADM−9.69RTN3−6.27ENAH−5.83PML−4.96ALDOA−6.91PRMT1−3.81IGDCC3−3.48IGSF3−3.45
hsa-mir-378MRPL37−5.37BBC3−6.68OPA3−9.57CYP2U1−5.99MARVELD1−5.48LDHA−4.71LDHA−6.78PTOV1−3.50ENAH−3.25NLGN2−3.35
CDK4−5.27HIST1H2BD−6.65UGT8−9.55WDR5B−5.63FBLIM1−5.46ORAI2−4.67P4HB−6.71ENO1−3.27IGSF3−3.19MSC−3.27
FEN1−5.08DCTPP1−6.34KCNN1−9.54ENPP4−5.50MYO1B−5.39STOML1−4.50PAICS−6.70DCTPP1−3.19VAMP4−3.16YPEL1−3.26
IGF2BP1−3.91MEG3−5.74TMEM59−9.63NR3C1−4.30PIGR−2.49PRDM16−7.24FERMT2−3.94PRDM16−3.79VEGFA−2.21PIGR−2.42
DEK−3.52PIGR−5.39UGT2B10−9.30ANGEL2−4.28PIAS2−2.40CMTM4−6.94ANGPT4−3.84MLEC−3.72AFTPH−1.90BCL3−2.28
hsa-mir-133aCDK5R1−3.49PER2−5.30CDC42−9.08FAM160B1−3.53RBMXL1−2.28KCNQ1−6.81ZEB1−3.75ANGPT4−3.53CIAO1−1.79MYPN−2.19
SUPT16H−3.43NGFR−4.77SEC61B−9.03MMP14−3.21CIAO1−2.23ERBB2−6.77GRID1−3.45ARHGAP31−3.50KRT7−1.72KCNQ1−2.08
TCTEX1D2−3.37NR3C1−4.65MYL12A−8.03UBA2−3.20AFTPH−2.23PNP−6.67ZFP28−3.30ZEB1−3.41NR2C2−1.64NFAM1−1.83
NCAPG−6.57PTBP1−5.38BMP7−9.88SLC8A1−5.20LHX4−3.32VEGFA−5.09RFC5−5.15EFTUD2−4.76IFI44−2.33CD63−2.85
PTBP1−6.33ATP13A1−4.69CAST−9.73C1orf27−5.03FANCI−3.32ANGPTL4−4.62KIF4A−4.92IQGAP3−4.72MYEF2−1.98PPIB−2.81
hsa-mir-1RCC2−6.23OCIAD2−4.54CEBPA−9.72SCYL3−4.94SFXN1−3.29NETO2−4.58MAD2L1−4.88DSG2−3.90TWF1−1.91MTHFS−2.72
UHRF1−6.21KIAA1522−4.54GOLGA7−9.67LZTFL1−4.80BRI3BP−3.27KAT2A−4.53MTHFD2−4.83EIF4G1−3.80FOLR1−1.90MACROD1−2.44
SFXN1−6.17RCC2−4.51OAT−9.63WDR11−4.78NCAPD3−3.01HPS4−4.44SPC24−4.76SFXN1−3.72YTHDF2−1.86CRELD2−2.24
PAM−4.93NRBP1−3.56SOCS3−9.78MARCKSL1−5.82SNAI2−3.65VIM−7.03TOMM5−4.76BIRC5−4.04ETS1−3.27PSMD7−4.26
CXCL11−4.92RUNX2−3.11ARF3−9.49VASH1−5.63SERPINE1−3.30LHFPL2−6.51RPS3−4.70MYBL2−4.03MTDH−3.15UBE2I−4.23
hsa-mir-30cSNAI2−4.22PRDM1−2.94FOXA1−9.48ETS1−5.40SLC7A5−3.22SH3GL1−6.40MRPS16−4.43SLC7A5−3.91ITGB3−3.15NDUFA12−3.64
ADAM9−3.84CASP3−2.92UBE2I−9.40DLL4−5.37CTSC−3.04IFNAR2−6.33BIRC5−4.33ECT2−3.85RRM2−2.99LLPH−3.63
MGAT2−3.79PHTF2−2.82CADPS2−8.43SOX12−4.96SLC38A7−3.03BIRC5−6.22UXT−4.29RRM2−3.81CASP3−2.99PPP2R1B−3.58
PHYHIP−8.38DMD−8.17ZYX−9.96HGF−5.72CPEB3−5.13DMRT2−9.17RS1−8.88DLC1−9.04SPRYD3−2.93PTPRT−6.30
NEGR1−8.11GNAL−8.17CLIP2−9.96CCPG1−5.68TMEM100−4.95C1orf226−8.94TGFBR3−8.70RS1−8.82ZCCHC3−2.37SLC6A4−6.24
hsa-mir-16GLP2R−8.01PLSCR4−8.15CAMSAP1−9.95IRAK3−5.49PDCD4−4.56SLC9A2−8.89WNT3A−8.65NEGR1−8.32CCND1−2.22PHYHIP−5.34
GNAL−7.85RBMS3−8.10TLL1−9.93PDXK−5.45PDK4−4.51SPTBN2−8.87SLC6A4−8.54SLC6A4−8.09BTRC−2.19PLSCR4−5.18
DIXDC1−7.73CDC14B−7.83KATNAL1−9.90PHYHIP−5.42ZBTB16−4.30SLC6A4−8.84AGER−8.53KDR−7.97BAMBI−2.11KCND3−4.74
BAZ1B−4.25CBX2−4.83NCEH1−9.64PIK3R2−5.50FSCN1−5.26SERPINE1−6.55SFXN1−6.36KIF11−6.08C8orf76−2.18PPP2R1B−4.41
LMNB1−4.22CCNE2−4.32SBF1−9.61PES1−5.43FBXO45−5.07TGFBI−6.26PHTF2−6.34KPNA2−5.87MTHFD2−2.07GRPEL2−4.39
hsa-mir-30aMYBL2−4.16RRM2−4.23SFXN1−9.31SIX4−5.41YWHAZ−4.77RUNX2−6.24PAICS−6.24CDC20−5.84CASP3−2.01DCTN4−4.16
RRM2−4.10C8orf76−4.19SOX12−9.21THOC5−5.34SLC16A1−4.62ITGA5−6.15TUBB3−6.16RRM2−5.70QRFPR−1.96IDH1−4.11
PAICS−3.98MYBL2−4.10RTP4−9.16NR2F6−5.16TUBB3−4.57CARS−6.08NUPL2−6.14PAICS−5.56CBX3−1.75STMN1−4.06
SYNJ2BP−4.59CRY2−3.32SLC20A1−9.29ONECUT2−3.74SDR42E1−1.81ANKRD46−7.49QDPR−4.07USP47−3.18HMGA2−3.31MEIS3P1−3.39
MYOCD−4.31MTUS1−2.79MSI2−9.25RABL2A−3.70CD59−1.67PAFAH2−7.28AHR−3.23GGA3−3.10ZCCHC3−3.09PBX2−3.23
hsa-let-7iACTA1−4.30DUSP1−2.64TUBB2A−9.14CDV3−3.69PRSS22−1.46TGFBR3−6.99ASCL1−3.22PIK3C2A−3.05COPS8−2.98SEMA4C−2.55
SEMA4C−4.08DNAH9−2.53SURF4−8.86NCKIPSD−3.61GLO1−1.40DNAJC28−6.96DNAH9−3.17TSC22D2−2.93PBX2−2.59RNF144B−2.53
KCNB1−3.96DNAJC28−2.31ACOT9−8.67C5orf51−3.55GRPEL2−1.17NAT8L−6.96TGFBR3−3.02IGF2BP1−2.91HMGA1−2.57DDOST−2.51
MEN1−6.99PKMYT1−7.61TTLL7−9.87FLCN−5.28TGFB1−6.65NETO2−8.35CCNB1−7.43AURKB−7.58KAZALD1−2.84PRSS8−2.95
CDK1−6.61TRIM11−7.45DCAF4−9.79CYP20A1−5.18C1QTNF6−6.46STX4−8.13ALDOA−7.41UBE2C−7.57PLAGL2−2.66AGPAT2−2.87
hsa-mir-24ADPGK−6.48CCNB1−7.28CCL4−9.73ADPGK−4.28MMP14−6.16C1QTNF6−8.08RRM2−7.39CCNB1−7.50BTBD3−2.44PKMYT1−2.67
UBE2C−6.45UBE2C−7.27SIT1−9.69SMYD4−4.27FSCN1−5.97EHD2−8.06IMP4−7.28CCNA2−7.29ZNF107−2.30TNIP2−2.62
TRIM11−6.44CDK1−7.23OLR1−9.63MDM4−4.18NETO2−5.94CDKN2A−8.03LDHA−7.25RRM2−7.04DNAJC10−2.23ATL3−2.58
DGKB−4.54ITM2B−3.83TBX18−9.74CREBL2−4.81SCD−2.30RDH11−3.52FEM1C−4.23ZNF699−3.27ZNF460−1.45ITPRIPL2−3.77
ACTC1−4.27CREBL2−3.52MRAS−8.52MCTS1−4.80RDH11−2.20MOCS2−3.41ZBTB43−4.15PER1−2.97FEM1C−1.36CRK−3.19
hsa-mir-95FOXP2−3.80INTU−3.50ACVR1−8.39B3GNT2−4.09USP8−2.07METAP2−3.16ITM2B−3.86CDKN1A−2.94B3GNT2−1.29FAM126B−3.06
PTBP2−3.79ACVR1−3.49WARS−8.15TRIP4−3.81HFE−1.98ZNF711−3.10CREBL2−3.84REL−2.87DGKB−1.26FOXJ3−3.01
NXPH3−3.65SNX1−3.35ARPC1B−8.10CEBPD−3.80B3GNT2−1.72AVPI1−3.05CELF2−3.79DUSP18−2.80CACNG8−1.26LRRC58−2.96
Figure 3

The interaction network of miRNAs and their target genes, where yellow nodes represent miRNAs and purple nodes represent genes. The edge colors red, blue, pink, dark green, light green, and black indicate that the corresponding miRNA-gene pair is associated with 6, 5, 4, 3, 2, and 1 cancer types, respectively.

Association of the 17 selected miRNAs and their top five targets in 10 cancer types. The interaction network of miRNAs and their target genes, where yellow nodes represent miRNAs and purple nodes represent genes. The edge colors red, blue, pink, dark green, light green, and black indicate that the corresponding miRNA-gene pair is associated with 6, 5, 4, 3, 2, and 1 cancer types, respectively.

3.2.3. Expression Analysis

Expression analysis was conducted using a one-way ANOVA test in order to evaluate the statistical significance of the differential expression of the 17 selected miRNAs. Alternative techniques can also be used (Conesa et al., 2016; Costa-Silva et al., 2017; Crow et al., 2019). To perform the test, the population of patients was divided into tumor and normal groups for a given miRNA in a particular cancer type. As a result of ANOVA, significant (p < 0.05) changes in expression were observed between the tumor and normal groups for the 17 selected miRNAs. For example, the p-values of hsa-mir-10a, hsa-mir-196b, hsa-mir-10b, hsa-mir-375, hsa-mir-143, hsa-let-7c, hsa-mir-107, hsa-mir-378, hsa-mir-133a, and hsa-mir-30c were 6.17E-24, 2.84E-48, 4.71E-03, 1.80E-14, 3.38E-57, 3.85E-38, 5.42E-24, 5.04E-20, 2.78E-20, and 4.48E-17 respectively. Additionally, box plots of the selected miRNAs in tumor and normal samples are provided in Supplementary Figure 3. To investigate the relationship between the expression levels of miRNAs for each cancer type, hierarchical clustering was performed on the tumor and normal samples of the 17 selected miRNAs. The results are shown in Figure 4, from which the change in expression levels of the miRNAs between tumor and normal samples is evident. In the cluster plots, red indicates high expression levels and green low expression levels; black corresponds to not significantly expressed samples.
Figure 4

Hierarchical clustering results of the differentially expressed miRNAs for the BLCA, BRCA, COAD, GBM, HNSC, KIRC, LUAD, LUSC, OV, and UCEC datasets. Red indicates high expression levels, green low expression levels, and black not significantly expressed samples.

Hierarchical clustering results of the differentially expressed miRNAs for the BLCA, BRCA, COAD, GBM, HNSC, KIRC, LUAD, LUSC, OV, and UCEC datasets. Red indicates high expression levels, green low expression levels, and black not significantly expressed samples.

3.2.4. KEGG Pathway Analysis

In order to perform KEGG pathway analysis for the 17 selected miRNAs in 10 cancer types, their targets were identified based on negative correlation values as described in section 3.2.2. Then, these target mRNAs were used in the DIANA tool (Vlachos et al., 2015) separately to identify significant KEGG pathways associated with the selected miRNAs in different cancer types. The five most significant pathways for each of the 17 miRNAs according to the FDR-corrected p-value within 5% statistical significance for the different cancer types are reported in Table 7. The detailed pathways of the 17 miRNAs with all their targets are presented in Supplementary Material. It can be seen from Table 7 that the most significantly enriched pathways are involved in various cancer types. For example, hsa-mir-205 and hsa-mir-133a are found to be enriched in pathways relating to hsa05206: MicroRNAs in cancer for nine cancer types. Similarly, hsa-mir-196b is found to be enriched in the pathway of hsa05210: Colorectal cancer for all 10 cancer types, with the FDR-corrected p-values of BLCA, BRCA, COAD, GBM, HNSC, KIRC, LUAD, LUSC, OV, and UCEC being 2.24E-07, 2.30E-06, 2.31E-07, 7.00E-04, 2.36E-07, 4.54E-06, 8.48E-05, 2.38E-06, 2.24E-07, and 2.20E-07, respectively. In addition, critical pathways such as the PI3K-Akt signaling pathway, p53 signaling pathway, Bladder cancer, Pancreatic cancer, Prostrate cancer, and Lung cancer are also found for the 17 miRNAs with FDR-corrected p-values within 5% statistical significance in 10 different cancer types. The presence of such critical pathways for the 17 selected miRNAs suggests that these miRNAs play a significant role in various cancer types, including the 10 types considered in this paper.
Table 7

Five significant KEGG pathways for each of the 17 selected miRNAs in 10 cancer types.

FDR-corrected p-value
miRNAPathwayBLCABRCACOADGBMHNSCKIRCLUADLUSCOVUCEC
hsa05206: MicroRNAs in cancer2.20E-03–1.60E-041.73E-052.23E-021.02E-062.20E-041.70E-042.00E-032.40E-03
hsa05202: Transcriptional misregulation in cancer1.33E-02–1.62E-02––1.11E-022.15E-02–9.20E-03–
hsa-mir-205hsa05219: Bladder cancer–1.06E-02–––3.54E-02–––6.30E-03
hsa05205: Proteoglycans in cancer–––2.38E-02–1.73E-02–––6.30E-03
hsa04520: Adherens junction2.13E-02–––––2.80E-02–––
hsa05016: Huntington's disease1.27E-02––––––4.64E-02––
hsa04218: Cellular senescence––––2.90E-033.01E-02––––
hsa-mir-10ahsa04714: Thermogenesis1.27E-02–––––––––
hsa04510: Focal adhesion1.27E-02–––––––––
hsa04213: Longevity regulating pathway—multiple species1.27E-02–––––––––
hsa05210: Colorectal cancer2.24E-072.30E-062.31E-077.00E-042.36E-074.54E-068.48E-052.38E-062.24E-072.20E-07
hsa01522: Endocrine resistance2.77E-075.58E-07––2.92E-075.08E-068.58E-055.75E-07––
hsa-mir-196bhsa05215: Prostate cancer2.39E-063.00E-06––2.51E-063.81E-05–3.09E-063.42E-06–
hsa05161: Hepatitis B–3.00E-06–1.52E-053.61E-07–8.58E-053.09E-065.15E-07–
hsa04915: Estrogen signaling pathway2.15E-062.88E-06––2.26E-063.27E-05–2.97E-06––
hsa05169: Epstein-Barr virus infection–3.28E-02–2.40E-032.90E-021.10E-04––1.19E-02–
hsa04550: Signaling pathways regulating pluripotency of stem cells––1.50E-03–1.48E-02–––1.19E-02–
hsa-mir-10bhsa05206: MicroRNAs in cancer2.50E-03–––––––1.19E-02–
hsa04110: Cell cycle–9.10E-03–––2.80E-04––––
hsa04914: Progesterone-mediated oocyte maturation–1.34E-02––4.95E-02–––––
hsa01521: EGFR tyrosine kinase inhibitor resistance7.00E-041.09E-02––––––7.50E-042.80E-03
hsa04550: Signaling pathways regulating pluripotency of stem cells4.02E-021.09E-02––––––––
hsa-mir-375hsa04066: HIF-1 signaling pathway––2.32E-02––––4.60E-03––
hsa05165: Human papillomavirus infection––2.32E-028.70E-03––––––
hsa05224: Breast cancer–––6.00E-03–1.90E-03––––
hsa05230: Central carbon metabolism in cancer–1.00E-03––4.09E-02–––––
hsa05213: Endometrial cancer–1.00E-03––––––––
hsa-mir-143hsa05206: MicroRNAs in cancer–1.00E-03––––––––
hsa05205: Proteoglycans in cancer–1.00E-03––––––––
hsa05161: Hepatitis B–1.00E-03––––––––
hsa05206: MicroRNAs in cancer–4.05E-021.85E-022.20E-04–1.10E-04––––
hsa04115: p53 signaling pathway–––2.90E-03–1.63E-054.43E-02–––
hsa-let-7chsa04110: Cell cycle–6.10E-03––––4.43E-02–––
hsa05222: Small cell lung cancer–4.05E-02–––4.10E-04––––
hsa04215: Apoptosis—multiple species–4.05E-02––––4.43E-02–––
hsa05200: Pathways in cancer2.50E-031.30E-03–1.31E-02––2.40E-034.70E-047.38E-05–
hsa01521: EGFR tyrosine kinase inhibitor resistance5.50E-031.30E-038.20E-03–––5.20E-035.60E-033.50E-04–
hsa-mir-107hsa05165: Human papillomavirus infection8.60E-031.30E-031.41E-02––––8.80E-032.30E-03–
hsa04151: PI3K-Akt signaling pathway5.50E-031.40E-03–––––5.60E-032.70E-03–
hsa05224: Breast cancer5.50E-03–––––5.30E-03–––
hsa03013: RNA transport1.49E-02–––––3.11E-022.04E-02––
hsa03010: Ribosome––––––6.60E-042.04E-02––
hsa-mir-378hsa01100: Metabolic pathways––––––2.56E-022.04E-02––
hsa00010: Glycolysis/Gluconeogenesis––––––2.56E-022.04E-02––
hsa05224: Breast cancer––––1.25E-02–––––
hsa05206: MicroRNAs in cancer7.80E-035.80E-031.76E-051.60E-044.70E-032.21E-021.92E-051.40E-067.20E-03–
hsa05215: Prostate cancer7.80E-031.60E-04–3.90E-037.90E-042.21E-02–8.00E-047.20E-03–
hsa-mir-133ahsa05212: Pancreatic cancer5.50E-04–4.20E-03–5.50E-042.21E-02––6.20E-03–
hsa01524: Platinum drug resistance3.30E-03–2.71E-02–1.70E-03–––6.20E-03–
hsa05205: Proteoglycans in cancer–––3.10E-03––8.34E-058.00E-04––
hsa03030: DNA replication1.34E-101.83E-07––6.25E-067.60E-036.10E-091.16E-05––
hsa03430: Mismatch repair3.80E-047.10E-03––3.10E-04–4.00E-041.75E-02––
hsa-mir-1hsa04110: Cell cycle2.50E-049.20E-03––2.10E-04–8.74E-08–––
hsa03015: mRNA surveillance pathway3.64E-02––––––1.01E-02––
hsa05166: HTLV-I infection4.42E-02–––––––––
hsa05206: MicroRNAs in cancer1.79E-05–2.90E-03––2.04E-05––1.21E-02–
hsa05200: Pathways in cancer–8.81E-05–2.40E-032.07E-023.40E-03––––
hsa-mir-30chsa05211: Renal cell carcinoma–2.30E-03–2.09E-026.40E-03–1.42E-02–––
hsa04141: Protein processing in endoplasmic reticulum–2.30E-03––1.93E-02–1.04E-022.56E-02––
hsa04380: Osteoclast differentiation2.71E-02–5.10E-03–––––9.20E-03–
hsa04510: Focal adhesion1.90E-035.80E-03––––––––
hsa04010: MAPK signaling pathway1.31E-026.30E-03––––––––
hsa-mir-16hsa05200: Pathways in cancer1.58E-02–––––7.42E-05–––
hsa04151: PI3K-Akt signaling pathway–2.60E-03––––––3.20E-03–
hsa05206: MicroRNAs in cancer–6.30E-03––––7.40E-04–––
hsa04110: Cell cycle–4.39E-06–3.90E-03–5.81E-052.30E-031.93E-06––
hsa05206: MicroRNAs in cancer–––2.10E-042.28E-026.20E-04–5.70E-03––
hsa-mir-30ahsa05203: Viral carcinogenesis–5.50E-04–––1.80E-03––2.14E-02–
hsa05130: Pathogenic Escherichia coli infection–1.24E-02––2.28E-02––5.70E-03––
hsa03013: RNA transport––––––2.00E-035.70E-03––
hsa05206: MicroRNAs in cancer––9.80E-03–1.97E-02–1.80E-04–––
hsa04550: Signaling pathways regulating pluripotency of stem cells1.20E-03–––––6.20E-04–––
hsa-let-7ihsa04066: HIF-1 signaling pathway1.20E-03–––––6.80E-04–––
hsa05130: Pathogenic Escherichia coli infection––1.60E-03–1.97E-02–––––
hsa05225: Hepatocellular carcinoma1.01E-02–––––––––
hsa04110: Cell cycle1.29E-061.57E-07––9.85E-095.43E-054.19E-104.40E-10––
hsa03030: DNA replication2.13E-073.59E-06––1.10E-04–1.20E-031.03E-07––
hsa-mir-24hsa04218: Cellular senescence–4.00E-03––1.10E-041.50E-049.90E-044.20E-04––
hsa04115: p53 signaling pathway7.49E-051.30E-03––8.40E-04–1.20E-03–––
hsa00240: Pyrimidine metabolism4.50E-04––––––4.28E-05––
hsa05211: Renal cell carcinoma4.20E-03–4.93E-02–––4.30E-034.10E-03––
hsa05220: Chronic myeloid leukemia––4.93E-02–––2.26E-02–––
hsa-mir-95hsa05219: Bladder cancer––4.93E-02–––4.90E-03–––
hsa05206: MicroRNAs in cancer––4.93E-02–––2.79E-02–––
hsa05203: Viral carcinogenesis––4.93E-02–––1.67E-02–––
Five significant KEGG pathways for each of the 17 selected miRNAs in 10 cancer types.

3.2.5. Gene Ontology Enrichment Analysis

Similar to the KEGG pathway analysis, GO enrichment analysis was also performed with the targets of the selected miRNAs using the Enrichr tool (Kuleshov et al., 2016), to assess the significance of the roles the selected miRNAs play in different biological activities. The results of different analyses for biological processes, molecular functions, and cellular components are reported in Table 8 and in Supplementary Tables 4, 5, respectively; the details of the enrichment analysis are also given in Supplementary Material. In Table 8, we see that various biological process GO terms which are related to the targets of the 17 selected miRNAs have important roles in cancer development. For example, GO:0023051 is linked with regulation of signaling, which is involved in the development of colorectal cancer. Similarly, GO:0051252 is linked with regulation of the RNA metabolic process of bladder urothelial carcinoma, with an FDR-corrected p-value of 1.30E-03, which is <0.05. Other GO terms for biological processes, such as GO:0051171, GO:0048519, and GO:0009653, are linked with the regulation of nitrogen compound metabolic process, negative regulation of biological process, and anatomical structure morphogenesis, respectively, for different cancer types.
Table 8

Five significant GO biological processes for each of the 17 selected miRNAs in 10 cancer types.

FDR-corrected p-value
miRNAGO biological processBLCABRCACOADGBMHNSCKIRCLUADLUSCOVUCEC
GO:0051239 regulation of multicellular organismal process1.30E-04–4.34E-05––––4.70E-04––
GO:0023051 regulation of signaling––6.76E-05–6.10E-04––4.70E-04––
hsa-mir-205GO:0010646 regulation of cell communication––6.76E-05–7.30E-04––4.70E-04––
GO:0060255 regulation of macromolecule metabolic process–––3.87E-07–7.21E-06–––7.59E-05
GO:0051171 regulation of nitrogen compound metabolic process–––3.87E-07–2.61E-05–––7.59E-05
GO:0006139 nucleobase-containing compound metabolic process–8.72E-05––––2.90E-04–1.10E-04–
GO:0071840 cellular component organization or biogenesis2.33E-02–––1.39E-06–––––
hsa-mir-10aGO:0016043 cellular component organization2.33E-02–––7.28E-05–––––
GO:0034641 cellular nitrogen compound metabolic process–8.72E-05–––––4.60E-03––
GO:0010467 gene expression–1.70E-04–––––8.00E-04––
GO:0048523 negative regulation of cellular process8.93E-06–1.92E-072.60E-04–6.77E-051.30E-042.62E-061.20E-04–
GO:0048519 negative regulation of biological process2.58E-05–1.92E-072.60E-04–6.77E-05–5.64E-06––
hsa-mir-196bGO:0070482 response to oxygen levels––––5.06E-06–4.43E-051.18E-051.20E-04–
GO:1901700 response to oxygen-containing compound––––1.00E-04––1.11E-053.28E-054.54E-05
GO:0051173 positive regulation of nitrogen compound metabolic process2.58E-05-1.62E-06–––––1.10E-04–
GO:0009653 anatomical structure morphogenesis2.70E-04–3.12E-05––––1.20E-04––
GO:1901564 organonitrogen compound metabolic process–––2.80E-049.30E-039.50E-03––––
hsa-mir-10bGO:0048468 cell development2.50E-04–1.10E-04–––––––
GO:1903047 mitotic cell cycle process–3.30E-03––2.65E-02–––––
GO:0044237 cellular metabolic process–––7.10E-04–7.00E-03––––
GO:0048522 positive regulation of cellular process–1.25E-055.30E-03–––5.46E-07–5.60E-045.50E-03
GO:0031325 positive regulation of cellular metabolic process–1.51E-057.60E-03–––5.82E-06–––
hsa-mir-375GO:1902533 positive regulation of intracellular signal transduction5.80E-03–––––––5.60E-04–
GO:0051173 positive regulation of nitrogen compound metabolic process–1.51E-05––3.61E-02–––––
GO:0071840 cellular component organization or biogenesis––8.40E-03–––––5.60E-04–
GO:0044237 cellular metabolic process–2.84E-02––––––2.30E-03–
GO:0008152 metabolic process–2.84E-02––––––8.40E-03–
hsa-mir-143GO:1905477 positive regulation of protein localization to membrane–2.84E-02––––––––
GO:1903829 positive regulation of cellular protein localization–2.84E-02––––––––
GO:0006807 nitrogen compound metabolic process–2.84E-02––––––––
GO:0071840 cellular component organization or biogenesis––4.50E-03–6.50E-04––7.68E-06––
GO:0016043 cellular component organization––5.70E-03–6.50E-04––1.49E-05––
hsa-let-7cGO:0010604 positive regulation of macromolecule metabolic process–––3.95E-09–––1.20E-041.50E-03–
GO:0090304 nucleic acid metabolic process4.30E-04––––––4.41E-05––
GO:0051252 regulation of RNA metabolic process1.30E-03––––––––2.16E-02
GO:0048522 positive regulation of cellular process5.54E-075.57E-065.00E-038.70E-03–––1.44E-071.40E-04–
GO:0048519 negative regulation of biological process–––8.70E-039.40E-043.38E-022.88E-06–––
hsa-mir-107GO:0051172 negative regulation of nitrogen compound metabolic process5.54E-07–––––2.31E-06-6.70E-05–
GO:0048518 positive regulation of biological process1.17E-061.07E-05–––––2.42E-07––
GO:0080090 regulation of primary metabolic process1.31E-06––––––––9.80E-04
GO:0044237 cellular metabolic process–1.07E-02–-––4.46E-052.30E-04––
GO:1901576 organic substance biosynthetic process–1.07E-02––––4.46E-05–––
hsa-mir-378GO:0044249 cellular biosynthetic process–1.07E-02––––4.46E-05–––
GO:0034645 cellular macromolecule biosynthetic process–1.07E-02–––––––1.25E-02
GO:0009058 biosynthetic process–1.07E-02––––1.45E-05–––
GO:0071495 cellular response to endogenous stimulus–4.66E-05–––2.21E-057.70E-04–8.10E-03–
GO:1901701 cellular response to oxygen-containing compound–4.40E-04–––3.10E-047.70E-04–––
hsa-mir-133aGO:0071417 cellular response to organonitrogen compound–7.30E-04––8.70E-033.10E-04––––
GO:0048518 positive regulation of biological process–––5.00E-038.70E-03––––9.23E-05
GO:0065008 regulation of biological quality––1.50E-03–––7.70E-04–––
GO:0007049 cell cycle1.76E-098.75E-06––6.45E-08–2.12E-064.40E-04–9.80E-04
GO:0051276 chromosome organization1.52E-073.59E-05––9.12E-07–5.41E-077.92E-05––
hsa-mir-1GO:0006261 DNA-dependent DNA replication4.12E-081.40E-04––––6.25E-07–––
GO:0022402 cell cycle process4.03E-07–––2.18E-06––1.70E-03––
GO:0006259 DNA metabolic process–1.70E-04––5.77E-06–5.41E-07–––
GO:0010033 response to organic substance–1.03E-025.58E-05–––––1.97E-05–
GO:0060255 regulation of macromolecule metabolic process3.20E-04––1.20E-04––––––
hsa-mir-30cGO:0070887 cellular response to chemical stimulus–1.03E-022.00E-03–––––––
GO:0016192 vesicle-mediated transport–1.03E-02–––3.60E-03––––
GO:0002376 immune system process–1.03E-02––––2.05E-05–––
GO:0051239 regulation of multicellular organismal process4.64E-055.02E-08––––––––
GO:0048731 system development4.10E-04–––––5.33E-08–––
hsa-mir-16GO:0045595 regulation of cell differentiation4.10E-04–––––2.61E-08–––
GO:0050793 regulation of developmental process–1.11E-07––––2.61E-08–––
GO:2000026 regulation of multicellular organismal development–1.29E-07––––––1.50E-03–
GO:0071840 cellular component organization or biogenesis––––1.30E-04–6.20E-031.02E-06–2.04E-02
GO:1903047 mitotic cell cycle process–1.99E-02––––6.20E-038.40E-06––
hsa-mir-30aGO:0090304 nucleic acid metabolic process–––6.20E-04–––––2.73E-02
GO:0071310 cellular response to organic substance–––––1.90E-03––3.20E-03–
GO:0006260 DNA replication–––––––6.33E-06–2.04E-02
GO:0080090 regulation of primary metabolic process––––––1.35E-022.20E-03––
GO:2000727 positive regulation of cardiac muscle cell differentiation1.05E-02–––––––––
hsa-let-7iGO:1904705 regulation of vascular smooth muscle cell proliferation1.05E-02–––––––––
GO:0071900 regulation of protein serine/threonine kinase activity1.05E-02–––––––––
GO:0061061 muscle structure development1.05E-02–––––––––
GO:0007049 cell cycle1.13E-131.08E-08––2.65E-11–1.14E-081.16E-13–2.32E-09
GO:0000278 mitotic cell cycle1.72E-131.08E-08––4.26E-137.08E-052.78E-101.17E-12––
hsa-mir-24GO:1903047 mitotic cell cycle process1.16E-114.56E-08––1.04E-107.08E-056.94E-101.17E-12––
GO:0022402 cell cycle process3.15E-11–––5.20E-119.55E-052.75E-079.89E-13––
GO:0044772 mitotic cell cycle phase transition2.15E-104.56E-08–––4.86E-101.51E-11––
GO:0060255 regulation of macromolecule metabolic process3.22E-062.03E-058.31E-076.89E-051.30E-035.67E-062.05E-051.88E-061.03E-059.40E-04
GO:0080090 regulation of primary metabolic process3.22E-062.03E-058.31E-07–––5.80E-052.99E-061.19E-059.40E-04
hsa-mir-95GO:0051171 regulation of nitrogen compound metabolic process3.22E-062.03E-058.31E-07––9.97E-065.80E-052.99E-061.19E-05–
GO:0050789 regulation of biological process9.66E-07–1.93E-076.89E-051.30E-039.97E-06––4.16E-06–
GO:0065007 biological regulation3.22E-06–8.31E-076.89E-051.30E-03–––1.19E-059.40E-04
Five significant GO biological processes for each of the 17 selected miRNAs in 10 cancer types. As with the biological processes, GO terms for molecular functions were also found to be significant in the development of various types of cancer, as reported in Supplementary Table 4. For example, GO:0008134 is linked with transcription factor binding, which plays an important role in BRCA, GBM, KIRC, LUAD, OV, and UCEC, with respective FDR-corrected p-values 8.10E-03, 8.31E-06, 6.80E-06, 4.70E-03, 3.76E-05, and 3.40E-03 all being <0.05. Other important GO terms such as GO:0044877, GO:0003723, GO:0003676, and GO:0008092 are found to be linked to protein-containing complex binding, RNA binding, nucleic acid binding, and cytoskeletal protein binding, respectively, for different cancer types. Supplementary Table 5 reports the significant GO terms for cellular components in various cancer types. GO:0070013 is linked to intracellular organelle lumen, which has FDR-corrected p-values within the 5% significance level for seven different cancer types, namely BLCA, BRCA, COAD, GBM, HNSC, KIRC, and LUSC. In addition, GO:0043227, GO:0005654, and GO:0044444 are associated with membrane-bounded organelle, nucleoplasm, and cytoplasmic part, respectively, which are critical processes in the progression of different types of cancer. The relationship between these GO terms and important activities in cancer development have also been cross-validated in other studies (Waldman et al., 1997; Dhillon et al., 2007; He et al., 2013; Reimand et al., 2013; McClurg and Robson, 2015). Taken together, all of these evidences point to the potential importance of the 17 selected miRNAs in the development of various types of cancer.

3.2.6. Protein-Protein Interaction Network Analysis

In PPI networks, a node and an edge signify the interaction of a given protein and the protein-protein association. Here, related proteins share common functions, although they do not necessarily physically interact with each other. In our study, to perform the PPI network analysis, 170 sets of targets relating to 17 miRNAs in 10 different cancer types were used to compute the PPI networks using the STRING database (Szklarczyk et al., 2019). Then, the 170 interaction networks were further analyzed to rank the proteins based on the degree of their nodes and their presence in 10 cancer types. The top 30 proteins are reported in Table 9, while the rest are given in Supplementary Material. For example, MYC has degrees 34, 28, 41, 85, 33, 133, 54, 25, 38, and 32 with respect to BLCA, BRCA, COAD, GBM, HNSC, KIRC, LUAD, LUSC, OV, and UCEC, respectively. Therefore, the total degree of MYC is 503. Similarly, other proteins have certain degrees in different cancer types. Moreover, their presence in the different cancer types is indicated by the association count; for example, MYC has an association count of 10. The proteins in Table 9 are used to construct the final consolidated PPI network shown in Figure 5, which represents the associations between the top 30 proteins and 10 different cancer types. The average node degree in this network is 13.7, with PPI enrichment p < 1.0E-16. In Figure 5, the significant proteins are MYC, PTEN, CDK1, BRCA1, AKT1, PIK3R1, etc. MYC is the most important protein, known to be oncogenic for breast cancer. Similarly, AKT1 is very significant for breast, lung, and colon cancers. PIK3R1 also plays an important role in breast cancer. The detailed list of PPI networks is provided in Supplementary Material. In summary, the PPI network analysis suggests that the 17 selected miRNAs can be considered important biomarkers with respect to diagnosis of 10 different cancer types.
Table 9

Association of top 30 proteins in 10 cancer types for the 17 selected miRNAs through their targets.

Node degree of protein in 10 cancer typesTotal degreeAssociation count
TFBLCABRCACOADGBMHNSCKIRCLUADLUSCOVUCEC
MYC34284185331335425383250310
VEGFA2310155323183636483930110
AKT11759501654320161782699
RRM217280112123303210101829
CDK12331101321021300241738
CDKN1A201917151518201781015910
UHRF118291721232514051439
CHEK124220021024280221416
H2AFX32169010016218241368
MCM10202200191120230181337
POLD133260015141630001346
IL6111272980151514161279
RHOA11159008301822121258
PCNA2026000025400171285
DTL1318031213132310161219
CCNF61851471818141901199
BRCA127120025127251201207
CDC4280221701312230191147
PTEN1510793025251801128
YWHAZ51213927411137310410
PAICS7190590232714958
PIK3R17197710016895889
UBA52110810016181608877
CTGF111210854161350849
KIF4A1715001081511100867
MTOR98909147988819
UBE2C1313001181215011837
KIF2C161500110161290796
KIF18B1111001181112011757
CHAF1B10130412012952678
Figure 5

PPI network of top 30 proteins associated with 10 cancer types, with p-value <1.0E-16 and average node degree 13.7.

Association of top 30 proteins in 10 cancer types for the 17 selected miRNAs through their targets. PPI network of top 30 proteins associated with 10 cancer types, with p-value <1.0E-16 and average node degree 13.7.

3.3. Web Application for Prediction of 10 Cancer Types

A web version of the predictor was developed for use of the scientific and medical communities to aid in cancer diagnosis. The application uses the random forest and the expression of 17 miRNAs to predict the occurrence of 10 different types of cancer. The random forest was chosen as it outperformed other classifiers in our experiments. To predict the type of cancer for a particular patient or set of patients, the expression values of 17 miRNAs in those patients need to be uploaded in a specific format, instructions are provided on the website. The application will return and display the predicted cancer types for the given set of patients.

4. Conclusions

We have proposed a statistical learning method for feature selection that integrates clustering, classification, and regression to find putative miRNAs by solving a multi-class classification problem on 10 cancer types. The first part of the method is a wrapper-based feature selection algorithm, called stochastic covariance evolutionary strategy with forward selection (SCES-FS), which involves stochastic neighbor embedding (SNE), covariance matrix adaptation evolutionary strategy (CMA-ES), forward selection (FS), and a classification technique. Features are reordered using SNE by performing clustering of highly correlated features. From the result of the clustering, a subset of features is randomly selected to perform multi-class classification on 10 cancer types. Although the features are randomly selected, the underlying classification task is treated as an optimization problem for CMA-ES in order to find the features automatically. Thereafter, a final set of features/miRNAs is obtained using forward selection. The results of the first part of SCES-FS have been compared with the results of some well-known feature selection methods, including ESVM-RFE, LASSO, NSGA-II-SE, MOGA, SVM-nRFE, SVM-RFE, CMIM, ICAP, SCAD, JMI, CIFE, mRMR, FSCOX, DISR, SNRs, and RankSum, as well as the result with all features, in terms of classification accuracy. The SCES-FS method selected 17 putative miRNAs associated with 10 cancer types and achieved higher classification accuracy than the other methods. Using the 17 selected miRNAs, a web-based multi-class cancer predictor application has been developed. These selected miRNAs are used in the second part of the proposed method, which employs Cox regression analysis to examine their importance with respect to survival of particular types of cancer. The analysis uses data on expression of the 17 selected miRNAs together with clinical data. A high Cox coefficient value signifies the importance of an miRNA for a particular cancer type. For example, it is found that hsa-mir-375 has the highest Cox coefficient, 0.9882, for the glioblastoma multiform cancer type. Similar results were obtained for other miRNAs, associated with different cancer types. The up- and down-regulations of the 17 selected miRNAs have been computed based on the ANOVA test. Furthermore, network analysis, expression analysis using hierarchical clustering, KEGG pathway analysis, GO enrichment analysis, and PPI network analysis have been performed to assess the biological significance of the selected miRNAs. The network analysis revealed the association of different cancer types with each pair of miRNA and its target mRNA. The hierarchical clustering analysis demonstrated the effective changes in expression levels of the miRNAs between tumor and normal samples. Both the KEGG and GO enrichment analyses reveals the significant pathways and biological functions in different cancer types. Moreover, using PPI networks, key cancer regulators such as MYC, VEGFA, AKT1, CDKN1A, RHOA, and PTEN are identified. All these evidences suggest that our selected miRNAs play key roles in the development of 10 different types of cancer. A future research direction is the integration of multi-omics data for finding effective regulators pan-cancer biomarkers.

Data Availability Statement

Publicly available datasets were analyzed in this study. This data can be found here: https://portal.gdc.cancer.gov/ and http://www.nitttrkol.ac.in/indrajit/projects/mirna-prediction-multicalss/.

Author Contributions

JPS, IS, and AL conceived and designed the experiments. JPS, IS, AL, NG, AD, and PL performed the experiments. JPS, IS, MW, and AD wrote the manuscript. JPS, IS, GB, and DP corrected and edited the manuscript. All authors read and approved the final manuscript.

Conflict of Interest

JPS was employed by company Larsen & Toubro Infotech Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Algorithm 1

Pseudo-code of the SCES-FS

Input:  X,Y,λ,N,A,R (dataset, class label, population size, maximum number of generations, classifier, number of runs)
Output:  Sbest (feature subset)
  1:  Initialize a NULL list, L
  2:  for i ← 1 to R do
  3:         X′ ← ConstructEmbedding (X) // With SNE
  4:         θ ← SetParamsCMAES() // μ, σ, m, C
  5:         for g ← 1 to N do
  6:               P ← DrawPopulationCMAES (X′, λ,θ)
  7:              S←ConstructFeatureSets(P)
  8:              ScoreFeatureSet(S,X′,A,Y)
  9:              SRBest ← BestFeatureSet(S,SRBest)
10:              θ ← UpdateParamsCMAES(θ)
11:         end for
12:          L ← L ∪ SRBest
13:  end for
14:  L' ← RankFrequency(L)
15:  Sbest ← ForwardSelection(L',X′,A,Y)
16:  return Sbest
  53 in total

1.  Reducing the time complexity of the derandomized evolution strategy with covariance matrix adaptation (CMA-ES).

Authors:  Nikolaus Hansen; Sibylle D Müller; Petros Koumoutsakos
Journal:  Evol Comput       Date:  2003       Impact factor: 3.277

Review 2.  Interplay between miRNAs and human diseases.

Authors:  Prosenjit Paul; Anindya Chakraborty; Debasree Sarkar; Malobika Langthasa; Musfhia Rahman; Minakshi Bari; Rk Sanamacha Singha; Arup Kumar Malakar; Supriyo Chakraborty
Journal:  J Cell Physiol       Date:  2017-04-27       Impact factor: 6.384

3.  Covariance analysis of censored survival data.

Authors:  N Breslow
Journal:  Biometrics       Date:  1974-03       Impact factor: 2.571

4.  Classification of cancers based on copy number variation landscapes.

Authors:  Ning Zhang; Meng Wang; Peiwei Zhang; Tao Huang
Journal:  Biochim Biophys Acta       Date:  2016-06-03

5.  DIANA-miRPath v3.0: deciphering microRNA function with experimental support.

Authors:  Ioannis S Vlachos; Konstantinos Zagganas; Maria D Paraskevopoulou; Georgios Georgakilas; Dimitra Karagkouni; Thanasis Vergoulis; Theodore Dalamagas; Artemis G Hatzigeorgiou
Journal:  Nucleic Acids Res       Date:  2015-05-14       Impact factor: 16.971

6.  MicroRNA-in drug resistance.

Authors:  Haoran Li; Burton B Yang
Journal:  Oncoscience       Date:  2014-01-13

7.  miRTarBase 2020: updates to the experimentally validated microRNA-target interaction database.

Authors:  Hsi-Yuan Huang; Yang-Chi-Dung Lin; Jing Li; Kai-Yao Huang; Sirjana Shrestha; Hsiao-Chin Hong; Yun Tang; Yi-Gang Chen; Chen-Nan Jin; Yuan Yu; Jia-Tong Xu; Yue-Ming Li; Xiao-Xuan Cai; Zhen-Yu Zhou; Xiao-Hang Chen; Yuan-Yuan Pei; Liang Hu; Jin-Jiang Su; Shi-Dong Cui; Fei Wang; Yue-Yang Xie; Si-Yuan Ding; Meng-Fan Luo; Chih-Hung Chou; Nai-Wen Chang; Kai-Wen Chen; Yu-Hsiang Cheng; Xin-Hong Wan; Wen-Lian Hsu; Tzong-Yi Lee; Feng-Xiang Wei; Hsien-Da Huang
Journal:  Nucleic Acids Res       Date:  2020-01-08       Impact factor: 16.971

8.  Analysis of microRNA-target interactions across diverse cancer types.

Authors:  Anders Jacobsen; Joachim Silber; Girish Harinath; Jason T Huse; Nikolaus Schultz; Chris Sander
Journal:  Nat Struct Mol Biol       Date:  2013-10-06       Impact factor: 15.369

Review 9.  Bioinformatic tools for microRNA dissection.

Authors:  Most Mauluda Akhtar; Luigina Micolucci; Md Soriful Islam; Fabiola Olivieri; Antonio Domenico Procopio
Journal:  Nucleic Acids Res       Date:  2015-11-17       Impact factor: 16.971

Review 10.  A survey of best practices for RNA-seq data analysis.

Authors:  Ana Conesa; Pedro Madrigal; Sonia Tarazona; David Gomez-Cabrero; Alejandra Cervera; Andrew McPherson; Michał Wojciech Szcześniak; Daniel J Gaffney; Laura L Elo; Xuegong Zhang; Ali Mortazavi
Journal:  Genome Biol       Date:  2016-01-26       Impact factor: 13.583
View more
  1 in total

Review 1.  Ontologies and Knowledge Graphs in Oncology Research.

Authors:  Marta Contreiras Silva; Patrícia Eugénio; Daniel Faria; Catia Pesquita
Journal:  Cancers (Basel)       Date:  2022-04-10       Impact factor: 6.575
  1 in total

北京卡尤迪生物科技股份有限公司 © 2022-2023.