A Study of Bio-inspired Computing in Bioinformatics: A State-of-the-art Literature Survey

1.2.1. Bioinspired ML in Bioinformatics and Applications

The exponential growth in the size and rate of capture of biomedical data during the “big data” age is posing a challenge to traditional analysis methods. Deep learning, a subset of machine learning methods with biological origins, promises to use massive data sets to find hidden patterns and make precise predictions. Machine learning has a lot of potential for the analysis of biological data sets, as is known. Building complex models that reveal their underlying structure theoretically enables greater exploitation of the accessibility of increasingly large and high-dimensional data sets. The learned models include advanced properties, enhanced interpretability, and a better knowledge of the structure of biological data [13]. In recent days, researchers have attempted to apply bioinspired machine learning to a variety of bioinformatics domains, including DNA sequence optimization, sequence alignment, feature selection, training, neural networks, genome mining, omics, biomedical imaging, electronic health records, optimizing support vector machines, clustering application, and many others. The Basic aims and application areas of Bioinformatics are depicted below in Table 1.

1.2.2. DNA calculation

The length of a sequence of individual genes, larger genetic areas (i.e ., clusters of genes or operons), full chromosomes, or entire genomes can be determined by calculating the precise order of nucleotides within a DNA molecule. Calculation of the length of individual nucleotides in DNA or RNA (usually represented as A, C, G, T, and U) isolated from cells of animals, plants, or almost any other source of genetic information is performed. Quick DNA computation techniques have significantly increased biological and medical research and discovery. Choose an encoding method [16] that translates the original alphabet into strings of A, C, G, and T and then synthesizes the resulting information-encoding strings as DNA single strands to encode information with DNA. For storing biological information, DNA is an excellent medium.

1.2.3. Feature Selection

In data mining and machine learning, feature selection is a crucial procedure. The purposes of feature selection are to reduce the number of features, select those that are the most representative, and remove redundant, noisy, and unimportant features. To assess the potential feature subsets, Emary et al. [17] created a wrapper technique that combines a binary version of GWO with the fitness function k-nearest neighbor (k-NN). They compared their method with GA and PSO across three benchmark datasets. They showed improved class accuracy and faster convergence speed using their GWO-based feature selection technique.

1.2.4. Training Neural Networks

Information processing models called “artificial neural networks” (ANNs) were developed as a result of studying biological nervous systems. ANNs are often used in research and applications because of how well they can capture nonlinearity and dynamicity. However, the structure and connection weights of ANNs have a significant impact on their performance. Every new metaheuristic technique has a lengthy history of being evaluated for success by rapidly optimizing neural network connection weights. One of the most popular varieties of neural networks, MLPs, was trained by Mirjalili et al. [18] using GWO. In his research, a single hidden multilayer perceptron (MLP) network's weights and biases were maximized using GWO. The proposed training approach was contrasted with some well-known evolutionary trainers, including PSO, GA, ACO, ES, and GWO. GWO's advantage in training MLP networks was shown by comparative results based on five classifying datasets and three function approximation datasets.

1.2.5. Support Vector Machines (SVM)

SVM is a classifier and regression model that is very efficient. V. Vapnik created SVM, which has a strong mathematical foundation [19]. To enhance SVM efficiency, both the kernel parameters and the error penalty parameter C must be modified. The standard approach to solving the issue is to employ a basic or thorough grid search. However, this approach is ineffective because it takes too long to consider every possible combination. To maximize these hyperparameters, numerous researchers have looked at the application of metaheuristic algorithms [19, 20]. Recently, GWO has been used to optimize the SVM's hyperparameters. Gamma and sigma parameters in SVM were modified by Eswaramoorthy et al. [21] for categorizing intracranial electroencephalogram signals. In comparison to another classifier, the results revealed greater accuracy rates.

1.2.6. Clustering Applications

Clustering is a standard machine learning and data mining technique that divides data instances into those that share specific features [22]. For clustering problems, metaheuristic algorithms are widely used and implemented. One of the best-known metaheuristic clustering algorithms is the k-means method, and most of the metaheuristic clustering algorithms presented in the literature are alternatives to this approach. The starting centroids used for the K-means approach heavily influence its performance, and a local minimum is a highly likely trap for it. To overcome the limitations of the k-means method, based on GWO, Kumar et al. [23] designed a clustering algorithm. The fundamental idea is that each member of GWO corresponds to a set number of centroids in a dataset. They calculated the fitness by adding the squared Euclidean distance between each data point and the cluster centroids. Their tests on eight datasets showed that GWO performed better than both their metaheuristic techniques and the k-means algorithm and the results revealed greater accuracy rates [21].

1.2.7. Genome Mining

Genome mining is the technique of examining genetic information to discover naturally occurring proteins, metabolic pathways, and potential interactions [24]. It requires tools such as Computational biology and bioinformatics. A large amount of data, represented by DNA sequences, is required for the mining process and is available in genomic databases. These models are applicable in several areas of medicinal chemistry, such as the identification of natural compounds [25], where new information may be generated using data mining techniques [26, 27].

1.2.8. Omics

The growth of omics applications has been driven by technological advancements. Biological systems with complex interactions are frequently measured and studied using these methods. Numerous fields of study are covered by the term “omics,” including genomes, transcriptomics, proteomics, interactomics, metabolomics, phenomics, and pharmaco- genomics, to mention a few. Each of these areas may also have numerous subdomains, each of which calls for more skill in analytical and computational methods [28].

1.2.9. Biomedical Imaging

Machine learning has had a major influence on many naturally inspired approaches in recent years. Significant advancements in voice and image recognition have resulted from it, and it can train computer programs to outperform humans while generating artistically pleasing new pictures and images. Numerous of these tasks were thought to be insurmountable for computers to do [29] and this technology is essential for biomedical imaging.

1.2.10. Electronic Health Record

One of the reasons hospitals have adopted electronic health record (EHR) systems at such a rapid rate over the past 20 years is the Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009, which provided hospitals and medical practices with $30 billion in incentives to adopt EHR systems [30]. The number of hospitals employing the fundamental EHR system has grown since 2008, according to the Office of the National Coordinator for Health Information Technology. To carry out clinical informatics activities using machine learning, several studies use EHR data. In terms of preprocessing time and feature extraction, the new strategy performs better than the existing approaches. This is brought on by both the vast amount of patient data and the growing use of machine-learning techniques.

3.1.1. Genetic Algorithm (GA)

Step 1: Initializes the population randomly.

Step 2: Find the fitness of the population.

Step 3: Repeat until the termination criteria are met:

• Choose parents from the general population.
• Crossover results in a new population.
• Create a new population and experiment with mutations.
• Determine a new population's fitness.

Step 4: Return to step 3 until the halting criteria are met.

3.1.2. Literature Survey on GA and its Application in Bioinformatics

This section examines the current innovative genetic algorithms and their applications over the last several decades of computational biology research and development. Iqbal et al. [32] report employing genetic algorithms to orient contentious graph edges to uncover paths in protein-protein interaction networks. Their work reconstructs biologically significant routes in yeast species' weighted networks of protein interactions. The proposed method arranges the edges of the weighted network, and PGMOGA reconstructs paths. They also compare their model with four cutting-edge techniques, and the experimental results perform better. Bader et al. [33] used intelligent algorithms to perform automatic clustering of DNA sequences.

The basic bat algorithm and the popular GA were combined to create the innovative hybrid GABAT algorithm, which was proposed as a solution to the automated data clustering issue. The test findings show that the hybrid algorithm was superior to both the bat algorithm and the fundamental genetic algorithm. To statistically validate the built clusters, the Mann-Whitney-Wilcoxon rank-sum test was performed, and it generated a p-value of less than 5%. Suri et al. [34] provide a framework for objective genetic algorithm optimization using a chaos-DNA hybrid method. In the second step, the suggested model employs two fitness functions with an average changing intensity and entropy with the number of pixel change rates, as well as optimizing the encrypted data. It improves single-objective function optimization significantly. Chaudhary et al. [35] offer a genetic strategy for biological sequence alignment that includes controlled crossover and directed mutation. Their methods are a revolutionary genetic algorithm-based alignment technique that is presented for determining the best alignment of a sequence pair in an efficient manner. They also compared the suggested method's analytical and statistical performance to that of existing comparable sequence alignment approaches. The experimental findings improve the performance of the proposed model by allowing for faster convergence to the optima. Ahmed et al. [36] propose automated dermatological image segmentation by combining a new hybrid intelligent ACO-GA algorithm with a TSVM classifier for disease detection. In dermatological image segmentation for disease diagnosis and their suggested algorithm is used to segment various types of skin lesions. Following image segmentation, their proposed algorithm classified pixels in the image into different classes to investigate precise skin diseases. The hybrid ACO-GA can identify 24 different types of skin diseases. An EPGA-SC is proposed by Lio et al. [37] for the assembly of single-cell sequencing reads. Their contributions include an error-correcting method for nonuniform sequencing, data, reading categorization to a lower percentage of false reads, and the adoption of several sets of high-precision paired-end reads produced from high-precision assemblies. Their experimental findings demonstrate better assembly than the majority of modern tools. Zhang et al. [38] present a model for heart failure disease prediction based on GA to optimize an extreme learning machine. Their model possesses characteristics such as test-training speed and strong generalization ability. Their simulation results have better predictive performance and can better predict the occurrence of heart failure. This model could help doctors predict who will develop heart failure. Finally, Table 2 represents the literature survey on DNA sequencing by using GA application on bioinformatics with advantages and disadvantages.

3.2.1. Particle Swarm Optimization Algorithm (PSO)

Step 1: Construct a population of particles that are distributed uniformly along a vector X.

Step 2: Use the objective function to determine each particle's position.

Step 3: If a particle's current location exceeds its previous top position, its velocity should be updated.

Step 4: Select the best particle (according to the particle's last best position).

Step 5: The velocity of the particle is updated:

v_i^t+1₌ w_. v_i^t + c₁r₁ (p_b^t- p_i^t) + c₂r₂ (g_b^t- p_i^t)

(p_b^t -Local best solution, g_b^t-Global best solution)

Step 6: Distribute the particles to their new position:

p_i^t+1₌ p_i^t₊ v_i^t+1

Step 7: Continue with step 2 until the termination criteria are met.

3.2.2. PSO Literature Survey and its Application in Bioinformatics

This section examines the current innovative PSOs and their applications over the last several decades of computational biology research and development. Papetti et al. [42] apply particle swarm optimization techniques to simplify global optimization. For a multidimensional search space, they propose a dilation function. The search space can improve the fitness distribution by creating a random population and improving the overall optimization process. This automated approach will allow us to investigate the performance of local bubble dilation functions on challenging optimization issues in real-life scenarios. Zhu et al. [43] provide a novel approach for better PSO based on elastic collisions for DNA coding design. Using the sparrow search algorithm, the suggested EC-PSO enhances the optimal and worst positions within the population. The harmony search technique is then introduced to increase the quality of DNA sequencing. The algorithm's effectiveness is validated by eight test functions. In the DNA sequence design, their model was more rational, and the produced sequences were of greater quality. Som-In et al. [44] present an improved particle swarm optimization-based approach for detecting different motifs in DNA sequence collection. They proposed PSO-HD, which improves the efficiency of finding transcription factor binding sites (TFBS) by using particle swarm optimization with Hamming distance (PSO-HD). They compare the related algorithms using the F1-score as a measurement unit. The experimental results increase the detection of TFBS. Khan et al. [45] proposed a modified PSO technique for solving DNA problems. To evaluate multiobjective issues in DNA sequencing, the proposed technique makes use of four objective functions. To determine how effective the proposed method is, they use average and standard deviation data. Results of the experiment show that the recommended technique outperforms others. Vijayalakshmi et al. [46] applied particle swarm optimization with non-dominant sorting to give a multimodal prediction of breast cancer. They proposed a multimodel prediction strategy for breast cancer. The model makes use of publicly available WBCD and WDBC ultrasonic computed tomography breast cancer data sets (UCI). The results of the findings showed that kernel density estimation performed well (98.8% and 98.3%), which is extremely promising. Indumathy et al. [47] proposed a model using PSO with inertia weight and constriction factor to solve DNA sequence assembly problems. The proposed strategy optimizes the DNA fragments' overlapping scores. To find the optimal answer, they use computational tools. According to the experimental data, the proposed method beats earlier techniques in terms of overlap score. Elsayed et al. present a new hybridization technique for recreating DNA sequences based on probabilistic cellular automata with PSO [48]. The proposed model is split into two parts. The first phase involves analyzing the evolution of DNA sequences, and the second involves reproducing the sequences. Using cellular automata models, the DNA sequence is then analyzed and predicted. In this article, they assess a large number of DNA sequences to anticipate the changes that occur in DNA mutations. Wang et al. [49] present a PSO-based approach for generating DNA barcode sets. In this work, a PSO technique was adapted and used to build DNA bar code sets that use GC content as ideal self -complementary and continuity constraints. The GC content ranges between 40% and 60%. The proposed PSO method is used to generate new DNA barcode sets to improve lower limits, and the results are essentially equivalent to or better than the modified genetic algorithm that improves the lower bounds of DNA barcode sets. Using PSO and an ensemble learning approach, Wang et al. [50] created a feature fusion predictor for RNA pseudouridine sites. They presented PsoEL-PseU, a new feature fusion predictor for pseudouridine site prediction. They use binary PSO to find the best feature subset for each of the six feature subsets. Using the six best feature subsets, they trained six individual predictors. They also used a sliding window technique to get around the predictors' inability to identify RNA sequences inside nonterminal lengths. In addition, powerful deep-learning techniques are used to determine RNA sequence information. Finally, they demonstrate that the improved predictor performed better in the independent dataset evaluation. Using coupled BA-PSO and infection propagation, Issa et al. [51] propose a technique for detecting biological subsequences in COVID-19. This article proposes a modified version of FLAT that uses the hybridization of the PSO method with the BA algorithm to improve the performance of the BA algorithm. By using BA operators to find the optimum solution, the recommended technique, BPINF, identifies the best-investigated solution. The proposed model outperforms all existing state-of-the-art models in detecting LCCSS between a set of real biological sequences with substantial lengths, including COVID-19 and other well-known viruses, with an accuracy of 88 percent. Goncalves et al. [23] present CNN architecture optimization using bioinspired cancer detection in infrared images. The goal of the suggested paper is to demonstrate that early identification of breast cancer is critical to improving patient cure and recovery rates. In this paper, they apply bioinspired optimization techniques to determine suitable hyper-parameters and the architecture of completely connected layers. They were also able to improve the layer's ResNet-50 outcome from 0.83% to 0.90% of the F1 scores. Liu et al. [53] present a BPSON algorithm that has been applied to the construction of DNA codes. The bat algorithm and the PSO algorithm are combined in this BPSON algorithm. To defeat PSO and improve the global search ability, a bat algorithm is utilized in this article. The experimental results reveal that the suggested approach may avoid the formation of secondary structures as well as hybridize them. The results demonstrate that their technique may generate an optimal code for DNA computing, and the performance is compared with the existing model. Table 3 depicts the literature survey on DNA sequencing by using PSO application in bioinformatics with advantages and disadvantages.

3.3.1. ACO Algorithm

Step 1: Each ant's journey represents a potential solution to a particular problem.

Step 2: The amount of pheromones left behind on a path by an ant corresponds to the problem's quality.

Step 3: The path with the highest pheromone has a higher probability of being selected when an ant must choose between two or more paths.

Step 3.1 Once all ants have completed their journey to meet the requirement; we continue to apply the pheromone updating rule.

Step 4: Produce the best worldwide solution.

3.3.2. Literature Survey on ACO and its Application in Bioinformatics

This section examines the current innovative Ant Colony Optimization algorithms and their applications over the last several decades of computational biology research and development. A novel hybrid ant colony optimization method was developed by Bir-Jmel et al. [56] for the classification of cancer in high dimensional data. This algorithm uses a model of gene selection. Based on adopting a new graph-based technique for gene selection, they propose a hybrid strategy (MWIS-ACO-LS) for the gene selection problem in their work, to minimize gene redundancy. They put their approach to the test on ten high-dimensional microarray datasets ranging from 2308 to 12600 genes. Their suggested technique is effective, as shown by the experimental findings. To solve the DNA sequence design problem, Misinem et al. [57] developed a new hierarchical pheromone update system and population-based ant colony optimization. They describe an improved multiobjective optimization approach called PACO. To construct the DNA sequence, they employ a finite state machine model. Each level uses a 20-level pheromone matrix. Each node can connect to the next node in the level via four different paths. Each pheromone matrix has a mechanism for exchanging its information with another pheromone matrix, allowing them to access the best results from other populations. The findings indicate that this new approach can produce a more exact DNA sequence. To optimize DNA sequences, Kurniawn et al. [58] created an ant colony system. To deal with the DNA sequence optimization challenge, in this work, the ant system (AS) was introduced. The method used to find solutions based on pheromone data involves using certain ants. Four nodes—which stand for the DNA bases—make up its structure. The results of the proposed strategies are contrasted with those of alternative techniques, such as the evolutionary algorithm. Wang et al. [59] proposed an intelligent privacy-preserving technique for discovering associations in genomic association studies. They devise a multiobjective search strategy to find prospective single-nucleotide polymorphism (SNP) sets linked to an illness phenotype to develop a logical epistemic privacy protection technique. They employ four widely used techniques to find K-order SNPs. The experimental results indicate that detecting numerous models improves search accuracy and makes the framework more stable. For DNA sequencing by hybridization, Blum et al. [60] propose an ant colony optimization technique. To determine the average global similarity scores for their research, they provide a multilevel framework that uses the ant colony optimization technique for DNA sequencing. The results suggest that the ML-ACO techniques have performed better than conventional techniques. Sun et al. [61] improve the ant colony optimization approach for locating epistasis with heuristic data, which is then incorporated into ant decision rules in the linear time. Their recommended techniques have an O(L_j + nm²) temporal complexity, where L is the number of ants, j is the number of iterations, n is the number of SNPs, and m is the number of samples. It requires less time to implement the proposed method in practical applications. To find evolutionary relationships between species that are closely related based on their distance, Perera et al. [62] proposed a modified ant colony-based algorithm for phylogenetic tree construction. In regard to creating distance matrices and calculating Kolmogorov complexity as a measure of distance between species from an entire mitochondrial genome alignment, their recommended method performs better than conventional techniques. The summary of the literature survey on DNA sequencing by using ACO application on bioinformatics with advantages and disadvantages is represented in Table 4.

3.4.1. Bat Search Algorithm (BS)

Step 1: Initialize the bat population, including all bats' positions, velocities, and frequencies.

Step 2: Set all bats' echolocation parameters.

Step 3: After evaluating the bats in their initial position, the determined answer is saved.

Step 4: In descending order, sort the current population.

Step 5: Produce candidate bat solutions.

Step 6: Fly at random by changing your position, velocity, and direction.

Step 7: Select a bat at random from the population.

Step 8: Assess the bat based on its fitness function.

Step 9: Update the echolocation parameters for a more accurate next position.

Step 10: Choose the best bat for the next iteration.

Step 11: Repeat Steps 4 through 10 until the termination criteria are met.

3.4.2. Literature Survey on BS and its Application in Bioinformatics

This section examines the current innovative Bat Search Optimization algorithms and their applications over the last several decades of computational biology research and development. Zemli et al. [64] describe a model that solves numerous sequence alignment challenges by using a bioinspired approach. They introduce a new bioinspired approach to solving problems by employing BA-MSA (Bat Algorithm-Multiple Sequence Alignment). This strategy offers a new mechanism for generating the initial population.

It entails creating a guide tree for each solution using a progressive method and altering some parameters. For experiments, they employ the Balibse 2.0 dataset. They used a statistical non-parametric test to determine whether BA-MSA increases the effectiveness of the proposed strategy. Al-Betar et al. offer a bat algorithm inspired by Triz [65] for gene selection in cancer categorization. A hybrid rMRMR-MBA for the gene selection problem is presented in this paper. To find the most promising genes in this model, the robust minimum redundancy maximum relevance (rMRMR) is used as a filter and the modified bat algorithm (MBA) is used as a search engine to select a small number of beneficial genes. Using ten gene expression datasets, the effectiveness of the suggested technique was evaluated. Their proposed strategy is capable of producing highly promising results for gene selection with high classification accuracy. Kaur et al. [66] introduced the automated diagnosis of COVID-19 by using deep features and parameter-free bat optimization. In this work, deep learning models are trained to assess the severity of COVID-19 using CT scans. Researchers have developed a parameter-free bat (PF-Bat) optimized fuzzy k-nearest neighbor (PF-FKNN) classifier to identify the coronavirus. Mobile-fully Net2's connected layer of transfer was used to extract features, and FKNN was used to train the features. The experimental results show 99.38 percent accuracy, which is higher than the previous model and is, based on Covid CT scan benchmark data. Professionals using this model will be able to effectively manage patients' healthcare needs, resulting to a faster recovery. Based on the bat algorithm, Badr et al. [33] introduced an intelligent method for the automatic grouping of DNA sequences. For problems involving automatic clustering, they propose a soft computing based metaheuristic framework. The results of the tests show that their hybrid technique outperformed the bat algorithm as well as the standard genetic algorithm. To statistically validate the development clusters, a Mann-Whitney-Wilcoxon rank-sum test was used. The Wilcoxon test has a p-value of less than 5%. Table 5 depicts the summary of the literature survey on DNA sequencing by using BS application on bioinformatics with advantages and disadvantages.

3.5.1. Initialization

Creating a fitness function can be tricky, so it's critical to select the one that appropriately represents the facts in the problem. Then, after calculating the fitness of all members, pick a subset of the highest-scoring members.

3.5.2. Genetic Operator

Crossover has a higher chance of producing new solutions by swapping genes. First, we must generate a population of solutions. The population will have an arbitrary number of members, which are viable solutions to the problem.

3.5.3. Selection

After a population has been generated, members of the population will be evaluated based on their chromosomes. When a mutation is performed by flipping some of a string's digits, new solutions are produced, and the chance of mutation is quite low.

3.5.4. Termination

When an algorithm reaches its finish, one of two things normally happens: either the program has reached its maximum runtime or it has reached a performance threshold. After that, a final solution is chosen and returned.

3.5.1.1. Evolutionary Algorithm (EA)

Step 1: Create a population of potentially good solutions by selecting them at random.

Step 2: Calculate each person's fitness in the group.

Step 3: Use a selection technique to pick the parents.

Step 4: Crossover and mutation operators can be used to generate offspring.

Step 5: Assess the new offspring's fitness.

Step 6: Using a selection technique, pick members of the population to die.

Step 7: Repeat Steps 2–7 until all termination conditions have been met.

3.6.1. Cuckoo Search Algorithm (CS)

Step 1: Create an N-host nest population, with each h₁ and c₁ being a candidate for optimal parameters.

Step 2: Every time a cuckoo lays an egg, they choose a nest at random to place it in.

Step 3: Compare the cuckoo egg's fitness to the host egg's fitness.

Step 4: If the fitness of the cuckoo's egg exceeds that of the host egg, the cuckoo's egg will be replaced with another host

nest.

Step 5: If the host bird notices it, it can either discard the egg outside the nest or abandon it and establish a new one (nest).

Step 6: Repetition of Steps 2–5 until the termination criteria are met.

3.8.1. Grey Wolf Optimization Algorithm (GWO)

Step 1: Set the population size and maximum iteration settings.

Step 2: Evaluate all search agent's fitness.

Xα = most effective first searching parameter of the agent.

x_β = second searching agent’s parameter.

x_δ = most effective third searching agent’s parameter.

Step 3: Once the given number of iterations has been achieved, evaluate each agent's fitness.

Step 4: Next, the position of the current search agent is updated, and the fitness level of each search agent is evaluated.

Step 5: Repeat Steps 2–5 until the termination requirements have been met.

Step 6: Generates the best solution x_α.

3.8.2. Literature Survey on GWO and its Application in Bioinformatics

This section examines the current innovative GWO algorithms and their applications over the last several decades of computational biology research and development. Nematzadeh et al. [78] describe how to tune the hyperparameters of machine learning algorithms and deep neural networks using metaheuristic applications in biomedical and biological cases.

They introduce the tuning hyperparameters of machine learning algorithms using Grey Wolf optimization and the evolutionary algorithm. Their proposed method employs 11 distinct algorithms and datasets in biological and biomedical domains to demonstrate and experimental results deliver greater performance with faster convergence, and the optimizer aims to decrease the algorithm cost (space/time complexity). Using the Grey Wolf optimizer and operators with TRIZ-inspired design, Alomari et al. [79] provided a model for gene selection for microarray data classification. The production of a single chip with 1,000 genetic instructions using DNA microarray technology is well recognized. By using nine well-known microarray datasets and their proposed modified Gray-wolf optimizer, which evaluates the quality of a selected group of genes using support vector machines for classification tasks, they can identify the top-ranked genes. Additionally, they contrasted the results of their study with seven cutting-edge gene selection techniques. Six of the nine datasets were able to pinpoint the ideal gene combination, and their results are the best in four of the dataset. Chen et al. [80] provide a model for grey wolf optimizer adaptation for the disassembly sequencing problem. They investigate a grey wolf optimizer (GWO) for continuous optimization issues in their study; however, it cannot be immediately applied to a disassembly sequencing problem. The experimental results reveal that SGWO outperforms previous methods. Hasim et al. [81] presented methods for grey wolf optimization for motif discovery (GWOMF). They presented and assessed six synthetic datasets in their research. The experimental findings demonstrate an outstanding success rate with a less processing time that is comparable to other available approaches. With an average success rate of 86.6%, these novel models indicate their effectiveness as a motif discovery tool. Table 8 shows the summary of the literature survey on DNA sequencing by using GWO application on bioinformatics with advantages and disadvantages.

3.9.1. Whale Optimization Algorithm (WOA)

Step 1: Whales’ population is initialized randomly.

Step 2: Then initialize the iteration counter t.

Step 3: Next, let’s initialize the value of a=2.

Step 4: Afterward evaluate each search agent _i by calculating its fitness function.

Step 5: Update the position vector according to and .

Step 6: Repeat steps 4 and 5 until the termination requirements have been met.

Step 7: Finally return the best solution.

3.9.2. Literature Survey on WOA and its Application in Bioinformatics

This section examines the current innovative WOA algorithms and their applications over the last several decades of computational biology research and development. A triplet-base unpaired constraint is a novel constraint proposed by Li et al. [83], which can form a new combination constraint by combining it with existing constraints. The Harmony Search (HS) method and the Whale Optimization Algorithm (WOA) are combined to form a novel method known as HSWOA, which they used to produce DNA sequences that adhere to the new combination requirement. Finally, when compared to past findings, their findings show that their approach not only increases the efficiency of hybridization reactions but also generates a greater fitness value. A novel evaluation function, FCC, is presented by Abdel-Basset et al. [84] to assess the quality of various assemblers. In 30 benchmark instances, DWOA-LS was validated in two tests, and it was contrasted against several modern, highly resilient state-of-the-art DFAP algorithms. To show DWOA's superiority in converting continuous whale behaviors to discrete, it was contrasted against the WOA, DE, and SCA versions in the first trial. To show the relevance of the DWOA over those algorithms, the Wilcoxon rank sum test was also run. The purpose of the second experiment was to show that DWOA-LS outperformed some contemporary, reliable, state-of-the-art assemblers that were suggested for the DFAP.

A drawback of the suggested method, in addition to its temporal complexity, is that it occasionally fails to produce better overlap scores than CSA-P2M*Fit. An improved method for selective opposition whale optimization is provided by Wang et al. [85]. The first step is to update the predator's position selectively using improved quasi-oppositional learning (EQOBL), compute the population's fitness before and after, and maintain the positions of the best people as food sources. The position update of the food supply is finished by the second technique, which is an improved time-varying update technique for predator inertia weight position. 23 benchmark functions from CEC 2005 and 15 benchmark functions from CEC 2015 are used to compare the algorithm's performance in several dimensions. The nonparametric rank test by Friedman and the rank sum test by Wilcoxon both provide better outcomes. Applications in the field of biological computing, in conclusion, demonstrate its relevance. Building high-quantity DNA code sets is one of ISOWOA's objectives in their study. The experimental results demonstrate that the lower bounds of the multi-constraint storage coding sets used in this study are on par with or better than those of earlier optimal implementations. The results show that ISOWOA enhanced the quantity of DNA storage codes filtered by 2–18%, proving the algorithm's dependability in real-world optimization tasks. Stephan et al.'s [86] algorithm for a hybrid artificial bee colony (HAW) suggests an employee bee’s attacking phase by combining ABC's exploitative employee bee phase with the whale optimization's bubble net attacking technique. During the attack phase, worker bees employ humpback whales to find better food sources. The proposed mutative initiation step, which is part of the HAW algorithm's exploratory portion, improves on regular ABC's mediocre exploration. The simultaneous feature selection and parameter optimization of an ANN model use the HAW method. Back-propagation learning is used to implement HAW, which consists of momentum-based gradient descent, Levenberg-Marquart, and robust back-propagation (HAW-RP and HAW-GD). Several breast cancer datasets are used to evaluate the accuracy, complexity, and processing time of these hybrid versions. The HAW-RP variant achieved good accuracy using a low-complexity ANN model in comparison to HAW-LM and HAW-GD. Table 9 shows the summary of the literature survey on DNA sequencing by using the WOA application on bioinformatics with advantages and disadvantages.

3.10.1. Slime Mold Algorithm (SMA)

Step 1: First slime molds positions are initialized.

Step 2: Calculate the fitness of all slime molds W by Eq. (17)

Step 3: Calculate each search position (Xi).

Step 4: Update parameters P, ub and vc.

Step 5: Update positions of slime mould by Eq. (18).

Step 6: Repeat steps 4 and 5 until the termination requirements have been met.

Step 7: Finally return the best Fitness and X_t

3.10.2. Literature Survey on SMA and its Application in Bioinformatics

This section examines the current innovative SMA algorithms and their applications over computational biology research and development. Qiu et al. [88] introduced an enhanced SMA (GLSMA) algorithm based on Levy flight and Gaussian mutation. According to experimental results, the two tactics are critical in expanding SMA's global search and preventing it from sliding into the local optimum. Beginning with comparisons to DE, PSO, GWO, and other well-known algorithms, the value of the GLSMA approach is shown.

Second, when compared to other advanced swarm intelligence algorithms such as MPEDE, LSHADE, ALCPSO, and CLPSO, GLSMA finds the optimal solution faster. After that, BGLSMA is created by translating GLSMA into binary space using a transformation function and is then used to highlight selection problems in 14 frequently used UCI high-dimensional gene datasets to show how well GLSMA performs in real-world applications. It is clear that GLSMA, in the application of feature selection, still has good global search ability and can select fewer features with higher classification accuracy when compared to a great metaheuristic optimizer in terms of general average characteristics selected, average error rate, average fitness, and average cost. To solve this issue, Qiu et al. [89] introduced a wrapper gene selection technique based on the slime mold algorithm. SMA is a revolutionary approach to feature selection with numerous applications. Their study expands on the original SMA by merging the Cauchy mutation mechanism with the crossover mutation technique based on differential evolution (DE). The transfer function then solves the gene selection problem by converting the binary continuous optimizer. ISMA, the continuous form of the approach, is initially used to study 33 classical continuous optimization tasks. The effect of the discrete version, or BISMA, was then explored further by comparing it to other gene selection methodologies using 14 gene expression datasets. The experimental results reveal that the continuous version of the algorithm achieves the optimal balance of local and global search capabilities, whereas the discrete version of the algorithm gets the greatest accuracy when selecting the fewest genes. Ewees et al. [90] proposed a modified slime mold algorithm based on the Marine Predators Algorithm (MPA) operators as a local search technique. This strategy enhances the SMAMPA technique's convergence rate while minimizing the attraction to local optima. The effectiveness of SMAMPA is tested using 20 datasets, and the findings are compared to cutting-edge FS approaches. It is also feasible to assess how adaptable SMAMPA is to real-world situations by employing it as a quantitative structure-activity relationship (QSAR) model. The results demonstrate how the developed SMAMPA method outperforms other approaches in terms of lowering the size of the datasets under test. Table 10 shows the summary of the literature survey on DNA sequencing by using SMA application on bioinformatics with advantages and disadvantages.

According to Fig. (2), the number of papers published in Scopus exceeds that of other publications. Furthermore, Figs. (2a and 2b) show the distribution of presented articles relevant to the bio-inspired computing technique from 2003 to the present year 2022. According to the findings, the number of published papers has been steadily increasing from 2003 to the present. However, the number of articles published in 2022 is noticeably higher than in previous years. According to the findings, the use of bioinspired computing techniques in various research and studies is expanding significantly year after year.

4.5.1. Improved Diagnostic Models for Rare Genetic Disorders

Genetic diseases can be classified as either monogenic or polygenic. Monogenic diseases are extremely rare and are caused by mutations in a single gene. They are handed down from generation to generation. In contrast, polygenic disorders are more common and are brought on by mutations in many genes. Such inherited diseases have recently become excessively common. Accurate disease diagnosis will benefit from the application of bio-inspired approaches in ML classification and prediction models.

4.5.2. Cancer Diagnosis and Prediction

As cancer is a complicated illness with several subtypes, it is essential to identify patients as soon as possible to provide them with the best therapeutic care. Many studies have been done on bioinspired approaches used in bioinformatics and machine-learning applications since it is crucial to discriminate between high and low-risk patients. To develop classification and prediction models, machine learning models such as support vector machines (SVM), artificial neural networks (ANN), and Bayesian networks (BN) are essential.

4.5.3. Comparison of Multiple Sequences Alignment

When applied to gene expression data, individual machine-learning models will produce better outcomes. The effectiveness of hybrid techniques, on the other hand, has been demonstrated in many applications. More research is required on the topic of combining various clinical reports and hybrid approaches. Although challenging and time-consuming, it will result in superior outcomes.

4.5.4. Gene sequence data analysis for pharmacogenomics

It is essential to predict how medicine will respond to a genetic anomaly or disease. Outstanding efforts have been made recently to detect cancer sensitivity and response. However, the main obstacle to developing a drug analysis technique is the high dimension and small sampling size. Machine learning with feature selection techniques aids in minimizing dimensions and improving the accuracy of predicting drug reactions.