Metaheuristic Model of Gene Selection for Deep Learning Early Prediction of Cancer Disease Using Gene Expression Data

Abstract

Cancer prediction in the early stage is a topic of major interest in medicine since it allows accurate and efficient actions for successful medical treatments of cancer. Mostly cancer datasets contain various gene expression levels as features with less samples, so firstly there is a need to eliminate similar features to permit faster convergence rate of prediction algorithms. These features (genes) enable us to identify cancer disease, choose the best prescription to prevent cancer and discover deviations amid different techniques. To resolve this problem, proposed a hybrid novel technique CSSMO, based feature selection for cancer prediction. First, combine the use of Spider Monkey Optimization (SMO) along with Cuckoo search (CS) algorithm viz. CSSMO for feature selection, which helps to combine the benefit of both metaheuristic algorithms to discover a subset of genes which helps to predict a cancer disease in early stage. Further, to enhance the accuracy of the CSSMO algorithm, choose a cleaning process, minimum redundancy maximum relevance (mRMR) to lessen the gene expression of cancer datasets. Next, these subsets of genes are classified using Deep Learning (DL) to identify different groups or classes related to a particular cancer disease. Six different datasets have utilized to analyze the performance of the proposed approach in terms of cancer sample classification and prediction with Recall, Precision, F1-Score, and confusion matrix. Proposed gene selection method with DL achieves much better prediction accuracy than other existing Deep Learning (DL) and Machine learning models with a large gene expression datasets.

1. INTRODUCTION

Successful cancer therapy has remained a significant issue despite enormous improvements in healthcare over the past century, and it is the second leading cause of mortality globally, after cardiovascular disease [1]. According to data from the World Health Organization (WHO), cancer is the leading cause of death worldwide. Of the estimated 18.1 million cancer cases worldwide, 9.3 million cases involved males and 8.8 million involved women. The most common types of cancer are lung, liver, prostate, colon, breast, and rectum. Figure 1 depicts the anticipated global number of new instances in 2022, broken down by age groups and gender. Clinical research and the treatment of many diseases are significantly influenced by the gene expression levels in an organism. Gene expression data is also known as gene-chip is a scientific advanced tool used by many researchers to study the magnitudes of several genes expressed in the abnormal sample. It serves as a tool that reflects the possible spectrum of the genome to analyze and investigate the root cause of the diseases. Problems related to gene expression profile could be solved using DNA microarray and Rna-Sequence based platform [2]. The use of gene expression profile in genetic research is a potent strategy that presents the data scientist with several analytical difficulties [3]. In order to locate the relevant gene that is conveyed, advanced biomarker machine learning approaches help by using gene expression data. The development of trustworthy cancer biomarkers is crucial for the field of clinical diagnostics [4]. Microarray gene expression technology, the most popular instrument used in illness prediction, makes it possible to categories and analyze a variety of genetically linked disorders. More than two hundred forms of cancers and their subtypes caused due to abnormal mutation of the genes in DNA that leads to uncontrolled growth of cells. To take preventive measures, to improve the early prognosis methods and to initiate clinical treatment, it is important to identify the changes in the complete set of DNAs or its genome. Gene expression profiles like microarray technology and Rna-Sequence based platforms with machine learning and deep learning are useful in managing and isolating the genes responsible for inherited diseases [5, 6]. It helps to design suitable treatment in suppressing the magnitude of expressed genes linked with inherited diseases during the early development of the organism. Gene expression profiles helps to strengthen the genes responsible for developing healthy crops in bulk, in the management of effective use of pesticides, fertilizers, and in designing nationwide agriculture policies. The gene expression profiles generate high dimensional data, which is a major issue to deal with before creating the actual classifier. The accuracy and cost of computation affect the performance of the classifier. The specific methods to decrease the dimensionality of the gene expression and to conquer the related problems are the Feature selection technique & method of Feature extraction [7]. The latter provides new fewer size features, condensing the properties of high dimensional features as far as possible and the previous, filters irrelevant and reductant features and includes critical informative features. The datasets of standardized microarray consist of thousands of gene expression and samples of a few hundred. Every individual expression of a gene quantifies the magnitude of interest shown in the given tissue of the sample. It enables comparing the abnormally expressed gene in the tissue with those in common tissues, which provides a good insight and intuition into the diagnosis as well as treatment and further clinical management, predictions of the future samples. Feature selection or Feature extraction methods are the few scientific approaches to the curse of dimensionality problem [7]. The optimization techniques of linear algebra and the core part of statistics are the fundamental tools of most of the algorithms developed for gene expression data analysis [8].

Gene expression profiles have been analyzed in bulk by systematic implementation of machine learning and deep learning techniques [9, 10]. Results connected to the information recorded by gene expression determined the prognosis of the majority of illnesses and different forms of cancer. It is possible to examine the data of expressed genes using a variety of machine learning approaches, but current findings connected to deep learning algorithms are more precise and accommodating since they are good at identifying and categorizing useful genes. The isolated genes may be useful for further clinical trials and diagnosis of the disease.

Malfunctioning of the apoptosis process, uncontrolled growth, and division of the body's cells is the primary cause of almost all types of cancers [11]. Spread of abnormal cells through basic organ systems of the human body, scientifically termed as Metastasizing, is the major cause of death from cancers. The scientific community is looking for the specific reasons for a person to develop cancer. The known unavoidable cancer developing factors are family history, age, climate change, and avoidable factors that could be exposed to radioactive substances or certain harmful chemicals. Consumption of tobacco and alcohol, unhealthy diet, physical inactivity is the prime risk factor worldwide, and are also leading contributors to no communicable diseases [12]. Worldwide sixteen percent deaths are because of some types of cancer, which will reach twenty-five million cases by the end of 2030. Hence, early screening for cancer is important before they damage vital organs, as it is very difficult to treat once it invades. Most cancers have a moderately high chance of being cured if diagnosed and treated at early stages [13].

The important findings of this work defined as:

• Hybrid metaheuristic learning-based approach has been designed with DL classifier for gene selection that explains about infection accurately using SMO and CS as CSSMO for classifies cancer more accurately even if the patients are in an early stage.

• Enhance the CSSMO results by adopting filtering method mRMR, to reduce the dimensionality of gene expression data.

• Identify different classes with an associated disease using deep learning with various evaluation criteria.

• The result of deep learning model with proposed hybrid approach achieves much better accuracy than other existing DL models. Figure 2, shown the complete framework of proposed model.

2. LITERATURE REVIEW

Deep learning methods can be used for the prediction of cancer data, though it has several challenges, Joseph et al. explored various problems related to the implementation of the deep learning model in solving the classification of cancer in gene expression data. Transformation of the 1D gene expression values into a 2D image takes into account the overall features of the genes and evaluates features that are required for the classification purpose using a convolution neural network (CNN). The proposed model achieved a steady result with an accuracy of 95.65% when operated on the data related to 33 different cancer types [14].

Mostavi et al. successfully implanted the different CNN models such as 1D-CNN, 2D-Vanilla-CNN, and 2D-Hybrid-CNN which are based on unconstructed gene expression input, which is not only able to classify the sample as tumor or non-tumor but also into their respective types of cancer or normal. All the created classifier operated on 10,340 samples of 33 distinct cancer types, the Cancer Genome Atlas was used as the source which resulted in an average accuracy of 94.45% [15].

The backpropagation method is the universal technique for training artificial neural networks. In the estimation of the feature selection in gene expression data is carried out with the hybrid models of backpropagation neural networks along with fast Genetic Algorithms (GA), which examine those algorithms which are sensitive towards restricted minima in search space. Most of the gene selection algorithms are based on selecting the genes with higher ranking, however many low-rank genes if selected properly, are capable of increasing the performance of the classification. The Hybrid Fast GA-BPN algorithm encircles an informative small gene subset capable of improving the overall accuracy of the classification. Using microarray technology to do genetic research is very effective but, the curse of dimension creates many analytical challenges for the data scientist [16].

Zeebaree et al. tackled the main challenges of the classification of cancer microarray data with the help of deep learning algorithms based on convolutional neural network (CNN), which show improvement in accuracy and extraction of informative genes as compared to mSVM-REF-iRF and varSeIRF [17]. Feature selection is a crucial task in any classification problem, as the overall accuracy of the classifier depends upon the relevance between the properties of the selected gene and the properties of each class.

Mao et al. used the Randomization test (RT) and the partial least square discriminant analysis (PLSDA) to measure the relevance of the significance of the selected genes with the objective class. The suggested model provided satisfactory result upon classifying the data into four classes, using principal component analysis (PCA) & multiple linear regression (MLR) [18].

Cancer microarray data is represented in a non-squared matrix as the number of features is very large compared to the size of the sample (Curse of dimensionality), which needs to resolve by employing a suitable feature extraction method. Generally, distance-based feature selection techniques are based on Euclidean distance (PLSDA). Zhong et al. incorporated Bhattacharya distance to determine the class of the gene in a binary classification model. Genes were selected based upon the lowest misclassification rates obtained through a support vector machine. The highlighted method improved the accuracy and processing time as compared to SVM-REF and SWKC/SVM [19].

International communities stand united to fight against cancer. One out of six death worldwide is because of cancer. Thus, to facilitate researchers across the globe, different data sets related to types of cancers are made available publicly to expedite the process of finding suitable, affordable treatment. Comparative studies of accuracies of available algorithms with known data sets may give some insights to the researcher community. Tabares et al. have shown the comparative studies on the 11 tumor database and recorded accuracies of 90.6% & 94.43% respectively on logistic regression and convolutional neural networks. The proposed algorithm based on deep learning methods shows more promising results on microarray data analysis [20].

Medical treatment of patent diagnosed with cancer depends not only on the gene that caused it but also on the type of the cancer. Hence, classification of types of cancers plays vital role in clinical management of the disease. Despite many available microarray data analysis algorithms that have already shown remarkable results, they possess their limitations too. Hence, improved techniques based on machine learning algorithms are required for efficient implementation. Salem et al. implemented Information Gain & Standard Genetic Algorithm to classify human cancer disease depending on gene expression profiles. The Information Gain algorithm serves the purpose for feature selection followed by feature reduction and cancer type classification is achieved through Genetic algorithm and Genetic programming algorithm respectively. The hybrid technique based on IG and SGA improved the accuracy of the classifier [21].

Deep learning methods are getting popularized day by day as researchers started to use them in a wide range of areas however its use in the classification of cancer microarray data is limited and many connected domains remained untouched due to the Scarcity of the training datasets. The hybrid approach with deep learning techniques is the newly emerging area that will try to overcome limitations with available algorithms. Malignant and benign are the two types of tumors. Early detection and classification of tumors are important in preventing its harmful effects. Liu et al. suggested Sample Expansion Based SAE and 1DCNN are two deep learning approaches used for categorization of the microarray data. The authors claimed improvement in the accuracy of the classifier after testing the data with proposed algorithms [22].

Elastic net is very effective in variable selection and its regularization. The improved performance of Elastic net, under weak regularization, is the unique property of it which enabled the implementation in multidisciplinary fields. Wang et al. classified microarray data of leukaemia and colon cancer, using the hybrid technique with Adaptive Elastic Net with Conditional Mutual Information (AEN-CMI) & conditional mutual information. The proposed hybrid algorithm dominates traditional methods not only by improving the accuracy but also by using the minimum number of genes [23].

Microarray data analysis is one of the advanced and reliable tools in medical science, which strives to link the cause of the disease like cancer with an informative gene. Medjahed et al. developed a unique two steps algorithm. It is based on Support Vector Machine Recursive Feature Elimination to extract the genes and the latest Binary Dragonfly Algorithm (BDA) to improve performance of the previous. Author/s, for the first time, incorporated the application of the metaheuristics BDA with microarray data analysis that enhanced the accuracy of the classifier with a minimum number of genes [24].

Microarray data analysis of tumour requires a sufficient number of training models to build a classifier with better accuracy. Liao et al. tackled this problem by incorporating the different datasets of the multiple types of cancer and applied the Multi-Task Deep Learning (MTDL) algorithm to analyse the microarray data. MTDL resolved the issues related to the scarcity of the data and significantly enhanced the accuracy of a classifier, towards identifying the type of cancer when tested on multiple cancer datasets [25].

Prostate cancer in males causes seven percentage of deaths worldwide out of total cancer deaths. Research community is still debating on the reliability of the Prostate-specific antigen (PSA). Hence, there is a need of the biomarker that receives acceptance form the community. Hou Q et al. created a diagnostic and prognostic forecast model for prostate cancer and discussed the effectiveness of the biomarker named C1QTNF3. Author also expects the application of the proposed biomarker in other parallel domains to map oncogenes [26].

Mapping the type of disease with the best informative gene is the central idea behind using microarray data for the classification. Jansi et al. implemented two-stage algorithms based on Mutual Information Genetic Algorithm (MI-GA). Screening of potential genes with high mutual values is followed by creating an optimal set of genes through Genetic Algorithm and SVM. The proposed method shows improvement in accuracy when applied on datasets of different type of cancers [27].

Filtering only useful features capable of affecting the overall classifier is the prime thought behind selection of features through appropriate algorithms from available search space. Chen et al. implemented binary particle swarm optimization and CCFS to design cost effective and confidence-based features election method. Initially, author has filtered the most promising features by utilizing the available statically parameters within the features as well as within different categories. Secondly, discriminant statistical parameters designed in to the model to develop best possible classifier [28].

Reductant and irrelevant features reduce the classifier's accuracy and raise the computational cost, which leads to the mismatches of the training set with the desired class. Rouhi et al. proposed a hybrid approach which initially reduces the dimension of the features followed by implementation of Advanced Binary Ant Colony (ABACOh) meta-heuristic algorithm. The constructed hybrid approach enhanced the accuracy of the classifier when compared with available methods [29].

The study of the Genetic mutations inside the cell, is one of the primary applications of the microarray data, which provides the spectrum of expressed genes for further analysis. Jain et al. proposed a two-phase algorithm that combines Correlation-based Feature Selection (CFS) with improved-Binary Particle Swarm Optimization (iBPSO). At later stages, it implements Naive–Bayes classifier with stratified 10-fold cross-validation. It accelerates the convergence and provides highly accurate classifier on seven data sets out of ten, based on different types of cancers [30].

Advancements in computing facilities and vectorization of the iterative steps involved in algorithms enable data scientist to explore more useful insights from the data. It also helps in experimentation with a large number of possible subsets of features related to microarray data. Several parallel computing packages like Hadoop, Spark and Mahout provides a wide number of tools like extracting/selecting features to tune the classifier up to the desired level of accuracy. Venkataramana et al. implemented Parallelized hybrid feature selection (HFS) method. It not only incorporates the statistics related to subsets of features but also ranks them to set the selection of most effective, informative genes. The proposed method established the accuracy of 97% on the data sets related to gastric cancer and improved the accuracy to some extent when compared with available methods [31].

Alzaqebah et al. presented a study demonstrating use of cuckoo search methods for feature selection. This study involved use of cuckoo search alongside a memory-based mechanism to save optimal solutions (feature vectors) to find features that enhanced the classification accuracy. Additionally, the proposed algorithm was contrasted with the original algorithm using a microarray dataset. This study concluded that the proposed algorithm produced outcomes that are superior than those of the original and contemporary algorithms [32].

Swathypriyadharsini et al. have put out a methodology for identifying co-expressed genes that combines triclustering methods with a hybridized CS (cuckoo search) algorithm and clonal selection. After that, to ascertain the biological importance of the genes in the generated clusters, this technique makes use of gene ontology, functional annotation, and transcription factor binding site analysis. In comparison to both conventional cuckoo search techniques and other current triclustering algorithms, the experimental results of this approach were shown to be superior. [33].

Zhao et al.proposed a new search algorithm namely, the Elite Hybrid Binary Cuckoo Search (EHBCS) algorithm which employed feature weighting and elite strategy to improve over Cuckoo Search. The proposed algorithm showed results outperforming binary genetic algorithm and binary particle swarm optimization algorithm in terms of standard deviation, sensitivity, specificity, precision, and F-measure [34].

Othman et al. use of innovative operators for genomic selection is included in a hybrid multi-objective CS (cuckoo search) algorithm that has been developed. To do this, this study employed single crossover and double mutation operators. Using seven high dimensional cancer microarray datasets that are freely available, the suggested method was assessed. According to the experimental findings, the suggested technique selected fewer relevant genes while outperforming multi-objective cuckoo search and classic cuckoo search algorithms in terms of performance [35].

Pandey et al. presented a novel feature selection technique based on the binomial cuckoo search metaheuristic was described in the study. To increase stability, a hybrid data transformation procedure was developed that combines Fast Independent Component Analysis (FICA) and Principal Component Analysis (PCA). On 14 benchmark datasets from the UCI repository, the suggested method was evaluated and contrasted with other methods like the binary cuckoo search, the binary bat algorithm, the binary gravitational search algorithm, the binary Whale Optimization with Simulated Annealing, and the binary Grey Wolf Optimization [36].

Swathypriyadharsini et al. conducted a comparative study exploring two commonly used bio-inspired optimization algorithms namely, Cuckoo Search (CS) and Particle Swarm Optimization (PSO) for triclustering the microarray gene expression data. Both the algorithms were applied to three real-life three-dimensional datasets with mean square residue as fitness function. The experimental results in this study showed CS having better computational efficiency compared to PSO [37].

Scaria et al. proposed a user-friendly rule-based classification model for processing microarray gene data. Here, cuckoo search optimization algorithm was used to form classification rules and pruned by associative rule mining. This study concluded that the performance of the proposed approach was adequate enough in terms of accuracy, sensitivity, specificity and time consumption [38].

Balamurugan et al. presented a study showcasing use of shuffled cuckoo search with Nelder-Mead (SCS-NM) for finding significant biclusters in large expression data. This study used shuffling and simplex NM to diversify and intensify the search space to gain an edge over compared algorithms. The proposed work was benchmarked on four benchmark datasets where it showed significant improvement in fitness value when compared with the swarm intelligence technique and various bi-clustering algorithms [39].

Boushaki et al. performed a study involving hybridization of parallel cuckoo search optimization (PCSO) algorithm and Naïve Bayes (NB). In order to locate feature subsets that optimize accuracy, this led to the development of the parallel cuckoo search with naive bayes (PCSNB) wrapper technique, which coupled the exploration capacity of PCSO with the speed of naive bayes. On seven distinct datasets, the suggested method was evaluated and compared against existing metaheuristic methods. The experimental findings demonstrated more accurate prediction than previous algorithms and were relatively efficient [40].

Pandey et al. presented a study that demonstrated a cutting-edge metaheuristic technique that combined k-means clustering with enhanced cuckoo search to expand the potential of conventional clustering techniques. Three microarray datasets were used to assess the effectiveness of the suggested approach. This study found that the suggested algorithm beat the benchmarked state-of-the-art approaches. [41].

Kulhari et al. explored a novel metaheuristic gauss-based cuckoo search clustering method to broaden traditional algorithm capabilities for unsupervised data classification. The proposed algorithm was tested on three microarray datasets. This study concluded via. result-drive experimentation that the proposed algorithm was more performant than existing methods [42].

3. METHODS USED