Genetic variations analysis for complex brain disease diagnosis using machine learning techniques: opportunities and hurdles

Search keywords

Keywords have been carefully chosen for the initial search. After the initial search, new words found in several relevant articles have formulated multiple keywords. The study’s keywords included “genetic analysis for brain disease using machine learning”, “genetic variations based on single nucleotide polymorphism (SNP)”, “Gene analysis for complex brain disease based on machine learning”.

Data sources

Our survey used different academic database sources to obtain articles, as shown in Table 2.

Article inclusion/exclusion criteria

Inclusion/exclusion criteria have been developed to determine which papers are eligible for the next review phase. Relevance for research has been considered for articles that meet inclusion criteria and excluded articles that do not fulfill inclusion criteria. The following criteria are defined for inclusion/exclusion.

Inclusion

1. Our paper only focuses on genetic variation for complex diseases related to genes.

2. Only articles that performed on genetic variation for complex diseases related to genes using machine learning.

3. For inclusion, only articles in English were taken into account.

Exclusion

1. Papers that do not focus on genetic variation for complex diseases related to genes were excluded.

2. Articles that do not concentrate on genetic variation for complex diseases related to genes using machine learning were excluded.

3. Articles that are not in English were excluded.

Article selection

Research has established criteria for inclusion and exclusion to choose which articles are eligible for the next review phase. Articles that satisfy the inclusion criteria were considered research-related, and items that did not meet the inclusion criteria were excluded. In the previous section, we present the list of inclusion/exclusion criteria. Three phases were followed to select an article for this research. The first step was to consider extracting only the titles and abstracts of the articles. The second stage was to analyze the abstract, introduction, and conclusion to refine the first stage’s choice. At the end of the process, the articles were carefully read and the quality of the papers was then measured according to their relevance to the research.

Types of genetic variation

Candidate genes tend to decrease the number of SNPs in the study to some mutations where genes are intended to provide the genetic basis for the disease under study. Although overall candidate genes’ data sets are considered to be, the various tests of hundreds or even thousands of polymorphisms in our situation make link detection troublesome. A conceivable method to overcome testing overpowering quantities of SNPs, mainly based on the studies of a candidate gene, is to order SNPs as per their priority of functional significance (Clare & King, 2001). Previous natural information can be used in current databases to reduce the number of SNPs by focusing on specific genomic regions, and computational methodologies and aptitude are used to separate the neutral polymorphisms from poly-morphisms of potential functional interest (Mikhail et al., 2020).

The critical biological determinant is a genetic variation that supports evolution and determines the phenotype’s genetic basis. A researcher may wish to deal with the genetic data depending on the researcher’s viewpoint (Barnes, 2010). Researchers of the biomedical view tend to focus on genetic or phenotypic directions. The function of a gene cannot be understood fully without realizing the possible variation within the gene. This means that the biomedical researchers, who study genetics, must know what variants present and what effect these variants are on the gene’s function and thus the phenotype. The genetic variation comes in many forms, but every form arises from only two types of mutation events. The simplest type of different genetic variation results from a simple primary substitution that comes from mutation. This type of mutation event explains the most common form of variation (Barnes, 2010), SNP. However, it also represents rare mutations that may manifest Mendelian inheritance in families. Most other kinds of variation arise directly or indirectly from inserting or deleting a portion of DNA.

Microarrays

Microarray is a technique in which 1,000 nucleic acids are attached to a surface. They are used to measure DNA sequences’ concentration in a mixture by hybridization and the subsequent detection of hybridization events (Bumgarner, 2013). Millions of sequences in one reaction are simultaneously analyzed using microarrays. Microarrays come in three basic types: (A) spotted arrays on glass, (B) self-assembled arrays and (C) insitu synthesized arrays. Many types of microarrays depend on the type of data (Berry et al., 2019; Mao, Young & Lu, 2007; Prince, 2017; Smith et al., 2007; Isik, 2010; Yin et al., 2017). These types are DNA microarray, SNP microarray, cDNA microarrays, SNP microarrays, protein microarrays, MM chips, peptide microarrays, etc.

All studies on genome association aim to determine the genetic basis for traits and disease sensitivities using SNP microarrays. They carry the genetic variants, which are the most common in the human. Variants were identified for families at risk for several diseases (Coelho et al., 2009; Chen et al., 2012). However, with only a few notable exceptions, such as the related age of macular degeneration, risk variants usually describe only a small portion of the genetic risks known for their existence. Many factors are found that favor is contributing to this observation. Common variants may have minor effects on the phenotype or have variable penetration due to cognitive or epigenetic effects. Two other factors are CNVs and rare variants (Barnes, 2010). Genomic variation types have an essential effect. However, examine these variants on disease phenotypes should perform.

SNP microarrays

SNP microarray is a type of DNA microarrays, which is used to identify polymorphisms within a population. SNP is a variation of a single site in DNA, and it is the most common kind of variation in the genome. About 335 million polymorphisms have been specified in the human genome (Berry et al., 2019; Mao, Young & Lu, 2007). Microarrays of SNP have grown as a robust tool for the large-scale detection of epigenetic changes in genomes of predictive and/or predictive value. For genotyping in 1998; SNP matrix technology was developed (Association, 2019). Since then, this technique has been extensively developed and has become one of the most robust genetic analyses (Latkowski & Osowski, 2015; González & Belanche, 2013; Xue, Zhang & Browne, 2012; Teng, Dong & Zhou, 2017; Kong, Mou & Yang, 2009; Saeys, Inza & Larranaga, 2007). In this study, the very long execution times result in an enormous volume of data. Some preliminary tests were performed to select the three following constraints for the primary sequence of experiments (Mao, Young & Lu, 2007). First, it depends on the used data, which we could not generalize all data analysis. The second issue is the selection of SNP. Finally, the third issue is the used ML method.

Available datasets

In this section, we provided several benchmark databases that can be used to test the researchers’ proposed techniques and methods (Guyon et al., 2002) as shown in Table 3. Research in biology and medicine may benefit from higher-order gene screening to confirm recent genetic disease research discoveries or new ways to explore them. Therefore, the ADNI (Rémi et al., 2011) dataset’s primary goal is to support researchers with the chance to combine genetics, imaging, and clinical data with supporting the investigation of disease mechanisms. Genotyping and sequencing data were generated for ADNI 1, ADNI GO, and ADNI 2 subjects and are available to ADNI investigators. ADNI was released in 2003 as a public-private partnership. ADNI phase 1 data were collected from 757 subjects (214 controls, 366 MCI and 177 AD cases).

Preprocessing

There is a lot of information in a raw dataset prepared by researchers. It depends entirely on the requirements and purposes if the information is valuable or not. Data preprocessing is the first step in ML. It is necessary to ensure that the dataset is fully adapted to the needs. First, we removed unwanted attributes, missing values, and correct arrangement, and the expression value is normalized. The Bioconductor package completes the two steps. This raw file must be processed to extract the correct attributes for the following study. R programming language is chosen as a preferred language for analyzing data. Therefore, R programming language packages were applied for the preprocessing of the dataset. The Bioconductor analyzed the value of the expression and set the data set further, using the data normalization that reduces the data range to be studied as shown in Eq. (1). Unwanted attributes and the samples that have missing values must be removed. Data rearrangement according to format requirements will be completed before the next phases proceed. For classification purposes, only desired attributes are kept.

Gene removal with no expression value through all samples is one of the most simple and straightforward preprocessing techniques used in Karthik & Sudha (2018) and Daoud & Mayo (2019). Researchers applied this step by removing samples with 20% of the deleted features. After then, the data are normalized. Normalization is to remove the unimportant technical differences in data and to filter matrices with the p-value. Preprocessing is a step towards facilitating the use of data. The normalization was performed in the preprocessing stage by changing the scale or range of data from 0 to 1. Data normalization is necessary because microarray data has a significant range difference. The function of data normalization is presented by Eq. (1).

(1) $$y^{\mathrm{\prime }}={\displaystyle \frac{y-y_{min}}{y_{max}-y_{min}}}$$

where y′ is the value of features in the normalization domain, while the data’s value before the normalization process is y. At the same time, y_min and y_max refer to the smallest and largest values of all data in an attribute to be normalized, respectively. Algorithms of ML tend to be affected by noisy data. To avoid unnecessary complexity, noise should be decreased as much as possible in the inferred models (Karthik & Sudha, 2018; Daoud & Mayo, 2019; Adiwijaya et al., 2018). Two common types of noise can be known: (1) class noise and (2) attribute noise. Class noise is affected by samples classified as belonging to more than one class, which leads to wrong classifications. Simultaneously, the attribute causes attribute noise value errors, such as variables with wrongly measured and missing values.

The classification of patterns with missing data generally affects two issues: handling missing values and classifying patterns. The ability to handle missing data has become an essential requirement for pattern classification because inadequate data processing can lead to significant errors or false classification results. Most literature approaches can be grouped into four different types according to how both problems are resolved, as shown in Fig. 5.

1. Removal of incomplete cases and design of classifier using only the full data portion.

2. Imputation or estimation by editing set of missing data and the classification problem learning, i.e., with imputed values for complete data portion and incomplete patterns.

3. Use of model-based processes where the distribution of data is modelled using certain procedures, e.g. by expectation–maximization (EM) algorithm.

4. Use of ML methods in which missing values are included in the classification system.

The two first types of approaches are solved separately: the handling of missing value (data deletion, imputation) and the pattern classification, but the third type of approaches model the probability density function (PDF), used by the theory of decision in Bayes to classify the input data (complete and incomplete). Finally, the classifier is designed to process incomplete input data without previous estimates of missing data in the last types of approaches.

Dimensionality reduction

As for increasing dimensions, the cost of computations also increases, usually exponentially. To get around this problem, it is essential to find a way to decrease the number of features in mind. Often two methods are used, which are feature selection and extraction (Daoud & Mayo, 2019; Adiwijaya et al., 2018).

Classification

After reducing the dimensional complexity of the data, the next step is the classification process. Classification is not only the primary purpose of this research, but also it is critical to detect biomarkers of complex diseases. At this stage, the data was diagnosed (classified) based on whether they are affected by a specific disease or not. Biomarkers that can detect and diagnose brain diseases accurately are urgently required. So, we introduce a comprehensive review of genome sequencing analysis for discovering complex genes related to genetic brain diseases. Most of the genetic variation is contributed by SNPs about the human genome. Many complex and common diseases are associated with SNPs like AD. Early diagnosis and treatment can improve by investigating SNP biomarkers at different loci of these diseases. The investigation of genetic variants in the human genome related to complex diseases is one of the most important study objectives (Sherif, Zayed & Fakhr, 2015; Saeys, Inza & Larranaga, 2007). So, we divided the review into three sections. Each section discusses a disease, which is summarized in Tables 8–10 and 12. Table 11 provides some of the available datasets that the researcher can use.

Performance evaluation metrics

Performance evaluation metrics are used to measure the classification model’s performance and investigate how a model works well to achieve the goal. On the test dataset, performance evaluation metrics are used to estimate the classification model’s performance and effectiveness, chosen correct metrics. It is essential, such as the confusion matrix in Table 6 to evaluate the model performance. Some commonly used performance metrics include accuracy, precision, and Matthews’s correlation coefficient (MCC). which are listed in Table 7 (Jain, Jain & Jain, 2018; Singh & Sivabalakrishnan, 2015; Raj & Masood, 2020; Le et al., 2020; Do & Le, 2019).

Autism spectrum disorder (ASD)

Spencer et al. (2018) presented a heritable genotype system. Their system consisted of five stages, as shown in Fig. 6. The preprocessing is the first step, which contained the missing value imputations for the SNPs, and then they selected the most SNPs significantly. The second stage is population division. Prioritization is used as a primary association for each genome-wide subgroup procedure. The third stage is the prioritization of genome-wide by turning to the question: what sets of SNPs are they to test? The answer is to use the minimum threshold support in the frequent models (FPM) algorithm. The filter is performed for SNP sets, according to its prevalence among the affected population. The fourth stage is FPM, one of the most DM techniques that excellently identifies the most common occurrence in feature combinations. The data must be interpreted in binary form with both reference states in one person. The presence or absence of the item in FPM shows potential interactions among variants. Finally, the stage is the UICsup, which is a contrast mining utilizing. In the end, the comparison with another application is carried out, and its system for all was superior.

Al-Diabat (2018) presented an autism screening process which comprises presenting some questions to parents, family members, and caregivers to answer the child’s behalf to identify potential autism traits. The data mining and fuzzy rule are adapted to improve the screen efficiency and accuracy, which is considered one potential solution. To discover the features of ASD, the fuzzy unordered rule induction (FURIA) algorithm was evaluated. The automated method, FURIA, is designed to examine patterns from controls and historical cases. After then, they used the models to determine the possibility of autism traits in new individuals. The higher performance of fuzzy data mining models is evidenced by experimental results regarding predictive accuracy and sensitivity rates.

Alzubi et al. (2017) developed an accurate method based on hybrid feature selection for identifying the most SNPs that are informative and chosen an optimal SNP subset. Their method was based on the fusion of the wrapper and filter method. The system performance was evaluated against four feature selection methods and four classifiers on SNPs’ five different data sets. The experiment results show the adopted feature selection approach’s efficiency and achieve classification accuracy up to 96% for the used data set. Overall, they concluded that the whole genome could efficiently differentiate between people with complex diseases and healthy individuals. Their method has been validated in an independent and large case and evidence study.

Krishnan et al. (2016) introduced a complementary ML approach based on a network for the human brain’s specific genes to present a wide prediction genome of autism risk genes, comprising candidates who may be hundreds is minimal or no previous genetic evidence. Leveraging these genome- and network-specific predictions of the brain, we showed that a large group of ASD genes converge in fewer major pathways and developmental stages of the brain. Finally, they identified potentially pathogenic genes within autism-related CNVs and suggested genes and pathways that are likely mediators of ASD across multiple CNVs.

The model decision function output ranks the gene lists comprising an ASD risk gene and adjacent genes. Cogill & Wang (2016) developed SVM with training on brain developmental gene expression data for classification and prioritization of ASD risk genes. The pre-reduction process has been shown to improve the accuracy of the SVM classification. The results showed that each filter procedure identifies different gene sets with a specific gene repetition. Due to the high variance in ASD of several genes, further selection steps are necessary.

Jiao et al. (2012) presented an SNP-based predictive model which could predict the severity of ASD’s symptoms. They divided 118 ASD children into a moderate group of autisms (n = 65) and a severe group of autisms (n = 53). 29 SNPs of 9 ASD-related genes were obtained for every child. They applied three ML techniques to create predictive models: Decision stumps (DSs), ADTrees, and FlexTrees. With an accuracy = 67%, DS and FlexTree produced modestly better classifiers. All models of the SNP rs878960 in GABRB3 were selected and linked to the CARS evaluation.

Cancer & brain diseases

Ismaeel & Ablahad (2013) developed a method that can predict the disease by mutations. Bioinformatics techniques were used to train and used back-propagation algorithms to test whether the patient has the disease or not on the collected data, using all expected mutations for genes of some diseases (e.g., BRCA1 and BRCA2). They Implemented their method as the first way of predicting the disease based on mutations in the gene sequence causing this disease which showed two decisions were achieved successfully, as shown in Fig. 7. The first way was to diagnose whether a patient had cancer mutations or not by using bioinformatics techniques. Back-propagation is the second classification of these mutations.

Jain, Jain & Jain (2018) introduced a two-phase hybrid model for cancer classification. A low-dimensional set is selected for this model of prognostic genes. For biological samples classification into binary and multi-class cancers using naïve Bayes (NB) classifier with a cross-validation stratified 10-fold. Different cancer types have been evaluated in their system by using 11 microarray datasets. The experiment results are compared with seven methods. Their system achieved better results, showing the classification accuracy and the selected number of genes in most cases. It is obtained classification accuracy for seven datasets up to 100% with prognostic gene subset with very small sizes (up to <1.5%) for all datasets.

Liu et al. (2008) developed a supervised recursive feature addition (SRFA) method for feature selection. This approach chose candidate features/SNPs. Supervised learning and statistical measures are combined to control the redundancy information and enhance classification performance in association studies. Additionally, they have developed a support vector-based recursive feature addition (SVRFA) scheme in the SNP association analysis of disease.

Hussain et al. (2019) analyzed the DNA sequences of cancer patients using various classification techniques. Based on this study, it can be concluded that people’s cancer diseases can be diagnosed easily based on their DNA sequences. Different classifiers have been used to analyze purposes, such as an artificial neural network (ANN), KNN, decision trees (DTs), and fuzzy classifiers, NB, random forest (RF), and SVM. Based on the analysis of results, it has been found that these classifiers provide sufficient performance in terms of accuracy, recall, specificity, and other parameters. By using the gene dataset, the paper used different classifiers to analyze the data. They found the RF provides the best performance for the classification of cancer patients.

Alzheimer’s disease (AD)

The most common forms of dementia that degenerate neurons are AD and PD found in the brain cells. Tejeswinee, Shomona & Athilakshmi (2017) described a computational framework to investigate neurodegenerative disorders, as shown in Fig. 5. The collection of data that they performed has not been the computational explorations of earlier utilization. In the first stage, the genes related to AD and PD are collected as the dataset generation. Genes uniquely of AD are 74 and 38 genes uniquely of PD, representing 112 of genes collected. Both diseases have 95 common genes. Feature selection is the second step. Three feature selection methodologies are used in this study to select the optimal feature set. The result revealed that it is the best-optimized feature subset. This subset is used to feed the six classification algorithms individually to each of them in the classification phase, as shown in Fig. 8. The accuracy of all these algorithms was measured when predicting the correct diagnostic class.

Park et al. (2018) introduced genetic interactions based on ML from heterogeneous gene expression profiles, as presented in Fig. 9. Related gene determination of disease and dis-ease mechanisms is a research goal and necessary process; this problem has been approached with many studies problems by analyzing gene expression profiles and datasets interaction for genetic networks. Among pairs of genes, correlations or associations must be determined to construct a gene network. However, when heterogeneous data for gene expression is noisy with high levels for assigned samples to the same condition, it is difficult to specify whether a gene’s pair represents a significant (GGI). To find a solution for this challenge, they introduced an RF-based method to classify data of gene expression of significant GGIs. The model is trained by defining sets of novel features and utilizing various datasets with high confidence interactome.

Abd El Hamid, Mabrouk & Omar (2019) developed a method for detecting biomarkers SNPs associated with the disease with high accuracy classification, leading to early disease prediction and diagnosis. For discovering new biomarkers, ML techniques are utilized for the dis-ease. Many common diseases have SNPs related to AD. SNPs are recognized for this disease as significant biomarkers. In its early stages, they allow in understanding and detecting the disease. They apply many techniques that are performed on all genetic data of AD. They achieved the highest accuracy of classification. The results detected that K2 learning and NB algorithms accomplished an accuracy of 98.40% and 98%, respectively.

Shahbaz et al. (2019) introduced an AD classification using ML techniques to classify AD. In their paper, they performed six different ML and DM algorithms. The first step in their work is a data processing and feature selection. The second step is to partition data partitioning using 70% of training data and the reset for testing. Finally, classification techniques are used. Five stages of AD’s can efficiently classify using the generalized linear model (GLM) described in the study results on the test dataset with 88.24% accuracy. In medicine and healthcare, these techniques can be successful for early disease detection and diagnosis.

Zafeiris, Rutella & Ball (2018) discussed, explored, and evaluated an integrated ANN pipeline in AD for biomarker discovery and validation. Over time, the most popular form of dementia and no specific cause and no treatment available. The developed system consists of analyzing data that are public with a categorical and predicting gene interactions. A continuous stepwise algorithm is used and further examined through network inference. Novel markers can be generated by their methodology and another well-known study that could be used to guide future research in AD.

Bringas et al. (2020) implemented deep learning model AD stage identification. Data from a daycare center collected for one week from 35 AD patients on smartphones were used. The data sequences for each patient recorded changes in the accelerometer during daily operations and labeled with the disease stage (early, middle or late). Their methodology uses the CNN model to identify the patterns which identify each step. Their methodology is used to process these time series. As a result, the 90.91% accuracy and an F1-score of 0.897 were achieved through the CNN-based method, which greatly improves the results achieved by classifying features.

Different applications

Narayanan et al. (2019) provided a computer-aided detection (CAD), and it is considered as a hot research area that is attracting significant interest in the past ten years. ML algorithms have been widely used for this application because they provide clinicians with valuable opinions to make the right decision. Regardless of the many ML models available for medical imaging applications, not much has been implemented in the real world due to the inexplicable nature of the network’s decisions. Their research paper investigated the results supported by deep neural networks to detect malaria, brain tumors, etc., in various imaging methods. They visualized category activation mappings for all applications to improve understanding of these networks. This work will help data science experts understand their models and assist clinicians in the decision-making process. The main drawback of their work should extend by studying the region of interest (RoI).

Carter, Dubchak & Holbrook (2001) developed an ML approach using ANN and SVM to extract standard features among known Ribonucleic acids (RNAs) to predict new RNA genes in regions not shown on them. Results showed RNA genetic differences in composition and structural parameters in known RNAs compared to non-coding sequences could be recognized with high confidence using M L-Approach in both bacterial and archaeal genomes. Besides cross-validation testing, highly predicted sequences in E. coil, M.geneitalium, etc., not involved in the training datasets, were experimentally distinguished and reported in the literature as expressed fRNAs.

Narayanan, Hardie & Kebede (2018) presented a study of lung cancer and explored SVM’s performance based on a wide range of features. The SVM performance was studied according to the number of features. Their results demonstrated greater robustness, faster computation with a wide range of features, and less prone to overtraining than conventional classifiers. Besides, they also offered a computationally efficient approach to select SVM features. Results are presented to the publicly available 2016 lung nodule analysis dataset. Their results show that the SVM is superior to the Fisher linear discrimination classifier based on 10-fold cross-validation by 14.8%.

Yang et al. (2020) provided a review that introduced sequencing technology development and explains the structure of DNA sequence data and sequence similarity. Second, they analyzed the necessary DM process, summarized several of the significant ML algorithms, and highlighted the future challenges faced by ML algorithms in extracting biological sequence data and possible future solutions. Third, they reviewed four typical applications of ML in DNA sequencing data. Fourth, they analyzed the background and relevance of the corresponding biological application. Fifth, they have systematically summarized the evolution and potential problems in the field of DM for DNA sequencing in recent years. Finally, they summarized the content of the review and looked forward to the future for some research directions.

Bellinger et al. (2017) performed a systematic review of the application of DM and ML methods in the epidemiology of air pollution. A systematic review identified current trends and challenges and explored recent trends for applying data extraction methods to air pollution epidemiology. Their work presented that DM is increasingly performed in the epidemiology of air pollution. The potential is to enhance air pollution epidemiology continues to grow with DM’s advances for temporal and geospatial mining and deep learning. New sensors and storage media, which allow greater and better information, also support this. This shows that a large number of successful applications can in the future be expected.

Bracher-Smith, Crawford & Escott-Price (2021) introduced a review for psychiatric disorders based on ML for genetic prediction. ML methods have been performed to make predictions in psychiatry from genotypes, with the potential to enhance the process of prediction of outcomes in psychiatric genetics; however, their current performance is unclear. This paper aimed to systematically review ML methods for psychiatric disorders prediction from genetics alone and estimate their discrimination, bias, and implementation.

Romero-Rosales et al. (2020) compared three ML methods that have been shown to build robust predictive models (genetic algorithms, LASSO, and progressive wisdom) and provided the involvement of markers from misclassified samples to enhance the accuracy of overall prediction. Their results described that adding markers from a prototype plus model markers fitted to misclassified samples enhances the AUC about 5%, to ≈ 0.84, which is very competitive using only the genetic information. The computational strategy applied here can support good ways to enhance classification models for AD. Their work could have a positive effect on early AD diagnosis.

De Velasco Oriol et al. (2019) performed systematic comparisons of representative ML models to predict the late onset of AD (LOAD) from the ADNI cohort’s genetic variance data. Their experimental results demonstrated that classification performance is the best-tested model resulting in 72% of the area under the receiver operating characteristic (ROC) curve.

Mishra & Li (2020) introduced the first intensive reviews for applying AI technology in medicine and the current genetic research state in AD. Next, the extensive review concentrated on applying AI in genetic research of AD, including diagnosis and prediction of AD based on genetic data, analysis of genetic variance, gene expression profile, and gene expression. AD analysis based on the knowledge base. Although several studies have yielded some meaningful results, it is still in the preliminary stage. Fundamental shortcomings include database limitations, failure to leverage AI to conduct a systematic biological analysis of multi-level databases, and a theoretical framework for analysis results. Finally, the future direction of development is to aspire to high quality, comprehensive sample size, and data sharing resources should be developed. An AI analysis strategy for multi-level system biology is a development trend. Computational innovation may play a role in constructing and validating the theoretical model and designing new intervention protocols for AD.

Ahn et al. (2019) introduced fuzzy logic dependent on ML tools for enhancing business big data business analytics in complex AI environments. Specifically, they have provided a suitable and extended the traditional C4.5 algorithm by incorporating fuzzy logic for business processes. To classify and predict the individual wage, government staff, help take actions appropriate to those in need and/or make better use of their available resources. The resulting tool supported analysts in this process. Additionally, they have also released an enhanced version, and this base copy has been reinforced by branch extension. The enhanced version speeded up the mining process and the business analytics and led to higher accuracy.

Gao & Tembine (2016) presented networked systems for the effective way of extraction and utilization and the availability of accurate prediction, proactive tools of mean-field type dynamical systems. They introduced large-scale networked systems based on the distributed mean-field filter (DMF). The filter exploits the network’s topology and breaks it down into highly independent components concerning marginal mean-field correlations. The experiments showed for two object tracking scenarios are performed to illustrate the performance of their algorithm. Results of the evaluation show that DMF superior to the existing filtering algorithms.

Yazdani et al. (2020) introduced simultaneously complex and common features that arise from multiple genes interacted and regulated. Therefore, to unravel the fundamental biological networks, it is vital to characterize genes’ interconnectivity. They have systematically combined transcription, genotyping, and Hi-C data to determine the interconnections between individual genes as a causal network. They used various ML techniques to extract information from the network, determine differential regulatory patterns between cases, and control schizophrenia.

Yazdani et al. (2018) introduced a method which is known as Bounded Fuzzy Possibilistic Method (BFPM) that takes into account preserving the flexibility of the search space for higher accuracy cluster sampling (Yazdani, 2020). The method assessed samples for their movement from one cluster to another. This technique lets us finding important samples in advance of those with the potential to belong to other clusters in the near future. BFPM was performed on the metabolism of individuals in a lung cancer case-control study. Metabolism as close molecular signals of actual disease processes may be powerful biomarkers of the current disease process. They aimed to find out whether healthy human serum metabolites can be distinguished from those with lung cancer. With BFPM, some differences were noticed, pathology data were evaluated, and essential samples were identified.

Xu et al. (2020) used the methods of SVM and RF for developing smoking prediction models. For a model building of 10- fold cross-validation, they first used 1,431 smokers and 1.503 non-smokers and tested the model prediction models on independent datasets of 213 smokers and 224 non-smokers. AUC of 0.691, 0.721 and 0.720 for training, tests, and independent test samples were obtained by SVM with the 500 top SNPs, selected using LR methods (p < 0.01). AUCs 0.671.665 and 0.667 were obtained for the training, the tests, and independent test samples of 500 top SNPs selected using LR (p < 0.01).

Le & Nguyen (2019) applied DL in the prediction of SNARE proteins, which is one of life science’s most essential molecular functions. Several human illnesses involve a functional loss of SNARE proteins (e.g., neurodegenerative, mental illness, cancer, and so on). There is thus a critical problem in understanding these diseases and designing the drug targets by establishing a precise model to determine their functions. With 2D convolutional neural network (CNN) and position-specific scoring matrix profiles, their SNARE-CNN model could identify the accuracy reached with the SNARE proteins of 89.7%.

In many aspects of cellular life activities, antioxidant proteins are essential. Cell and DNA are protected against oxidative agents. In this study, Ho Thanh Lam et al. (2020) developed an ML model from a benchmark set of sequence data that was applied for this prediction purpose. The experiments have been performed through 10-fold cross-validation during the training and validated by three separate datasets. On the optimum set of sequence features, various ML and DL algorithms were evaluated. Among them, RF was identified as the best model for identifying antioxidant proteins with the highest performance. Their optimal model has achieved 84.6% high accuracy.