Rise of Deep Learning Clinical Applications and Challenges in Omics Data: A Systematic Review

This research aims to review and evaluate the most relevant scientific studies about deep learning (DL) models in the omics field. It also aims to realize the potential of DL techniques in omics data analysis fully by demonstrating this potential and identifying the key challenges that must be addressed. Numerous elements are essential for comprehending numerous studies by surveying the existing literature. For example, the clinical applications and datasets from the literature are essential elements. The published literature highlights the difficulties encountered by other researchers. In addition to looking for other studies, such as guidelines, comparative studies, and review papers, a systematic approach is used to search all relevant publications on omics and DL using different keyword variants. From 2018 to 2022, the search procedure was conducted on four Internet search engines: IEEE Xplore, Web of Science, ScienceDirect, and PubMed. These indexes were chosen because they offer enough coverage and linkages to numerous papers in the biological field. A total of 65 articles were added to the final list. The inclusion and exclusion criteria were specified. Of the 65 publications, 42 are clinical applications of DL in omics data. Furthermore, 16 out of 65 articles comprised the review publications based on single- and multi-omics data from the proposed taxonomy. Finally, only a small number of articles (7/65) were included in papers focusing on comparative analysis and guidelines. The use of DL in studying omics data presented several obstacles related to DL itself, preprocessing procedures, datasets, model validation, and testbed applications. Numerous relevant investigations were performed to address these issues. Unlike other review papers, our study distinctly reflects different observations on omics with DL model areas. We believe that the result of this study can be a useful guideline for practitioners who look for a comprehensive view of the role of DL in omics data analysis.


Introduction
Omics refers to the branch of study in biological sciences that end with "-omics," such as genomics, transcriptomics, proteomics, or metabolomics. The suffix "-ome" refers to the research subjects in domains such as the genome, proteome, transcriptome, or metabolome. Omics studies mainly aim to identify, describe, and quantify all biological molecules involved in the structure, function, and dynamics of a cell, tissue, or organism.
Multi-omics research flourishes in the areas of genomes, proteomics, transcriptomics, microbiome, metabolomics, pathomics, and radiomics [1]. Researchers have paid close attention to the connection between multi-omics data, medications, and illnesses. Moreover, multi-omics may accurately predict the diagnosis, prognosis, and course of treatment for illnesses. At various levels, a network may represent study entities, including genes, RNAs, proteins, microorganisms, metabolites, pathways, and pathological and medical imaging data. Additionally, some computer science and biology researchers have attempted to employ computational techniques to investigate potential connections between biological elements [2]. Instead of single-omics data, multi-omics data can improve the performance of these techniques because they obtain a vast landscape for understanding biological systems and mechanisms. However, heterogeneous data from multiple sources result in high complexity and different kinds of noise, which are detrimental to information extraction [3].
For instance, the median overall survival for tumor patients receiving the current medical standard of care is three years [4]. This outcome is due to the high level of heterogeneity in any form of tumor and the complicated biological molecular markers. Therefore, the community urgently needs the development of computational techniques for identifying novel treatment targets. Numerous attempts have been made to determine the molecular subtypes of gliomas and understand the heterogeneity across tumor types [4]. Numerous new molecular subtypes have been discovered based on several data sources, including histology, gene expression patterns, and driver genes [5][6][7][8]. However, applying the existing subtypes to clinical practice is challenging because the survival prognosis in the study of molecular subtypes is barely considered and has poor performance. The consideration of survival time is crucial in classifying tumor patients into useful subcategories with varying prognoses. Additionally, identifying community-wide high-risk tumor subtypes as appropriate treatment targets remains difficult. Consequently, an urgent need arises for computational methods to categorize patients accurately into subgroups linked with survival and identify prospective treatment targets.
Many computational methods have been developed with the rapid development of computer technology. Many research methods on intermolecular associations of humans related to different diseases, such as heart disease [9], Parkinson's disease [10], hearing loss [11], and COVID-19 [12][13][14], were developed. Big data and artificial intelligence technologies are expected to drive the rapid development and modernization of traditional Chinese medicine [15]. Machine learning (particularly deep learning [DL]) technologies play an active role in computational biology; they also gain remarkable success in biological fields with the advancement of artificial intelligence and high-performance computing [16][17][18][19]. For instance, DL has been used in identifying genetic variations, DNA methylation, and image analysis, thereby leading to remarkable advancements in these areas [4]. DL is a promising subject when used with data from various omics. However, the following factors often result in limitations of current techniques. The majority of multiomics approaches concentrate on two to three data kinds with a narrow scope for a specific sample or patient because of the growing number of accessible data types. Unsupervised learning has poor interpretability in anticipated outcomes or classifier labels [4].
On the basis of the above concepts, we introduced a comprehensive study covering all aspects of omics data that comprise the DL approaches. All research articles were scanned from two perspectives: technical and medical perceptions. The healthcare research field has enormous potential and problems because of the widespread availability of complex and unstructured omics data. Thus, the literature on omics data must be investigated thoroughly. The review paper's contributions and uniqueness are as follows: • The development studies are highlighted to improve the medical processes based on DL models. • Notable achievements by other researchers in response to omics needs are summarized.

•
The real benefits of using DL models in omics data analysis are emphasized.

•
The current challenges in DL models when used with omics data are clarified. • Using DL, we provide a taxonomy that organizes the corpus of already-published materials and specifies several omics research trajectories. We believe that the findings are useful to other researchers.
• The challenges associated with using omics data in DL-based clinical applications, such as the high dimensionality and sparsity of omics data, the lack of labeled data, and the need for robust evaluation methods, are identified.

•
The importance of interpretability, generalizability, and robustness in DL models applied in the omics field is highlighted.

•
The ongoing research on using DL algorithms for omics data analysis and the integration of multi-omics data to improve disease outcome prediction are analyzed. This paper is organized into different sections. Section 2 introduces the details of the systematic review procedure. Section 3 provides the results of the adopted systematic review protocol. Section 4 focuses on the existing datasets used for omics data analysis. Section 5 highlights the challenges linked to omics studies when DL approaches are used. Section 6 concludes the contributions of the present study and maps the addressed challenges with achieved outcomes.

Information Sources
We selected four of the most popular online search engine databases for a systematic search: Web of Science (WOS), ScienceDirect (SD), PubMed, and IEEE Xplore Digital Library. The selection was made by following an index that simplifies the formulation of complex search queries and keeps track of numerous scientific journals and conference papers in computer science, biology, medicine, biomedical engineering, and biomedical computing. We aimed to include as much material as feasible and as many articles on Omics data and DL as possible in the selection. We also aimed to provide a holistic view of researchers' achievements in a wide but pertinent variety of disciplines.

Study Selection
The research selection method required a thorough two-step search of linked literature. Reading the titles and abstracts enabled the initial removal of duplicated and irrelevant items. Second, full-text reading was used to filter the articles scanned in the preceding stage. The same procedures were used for both levels. Figure 1 shows the study selection protocol used in the present work.

Searching Settings
The article search procedure was conducted on 18 April 2022. The search query was entered into the search boxes of the IEEE, WOS, PubMed, and SD databases. The searches were conducted in all of the aforementioned databases using terminology-related keywords ("Omics"). Afterward, these keywords were paired with the keywords "Deep Learning" or "Reinforcement Learning" using the "AND" operator, as shown in Figure 1. All authors agreed that the selected query and search engines were suitable for the search process. The most relevant articles were obtained during this process. Each search engine's advanced search settings were used to limit the search to relevant journals and conference papers and exclude book chapters and other sorts of publications, such as white papers and workshops. We also looked at papers that were clearly based on the most recent and appropriate scientific research relevant to DL in omics data analyses.

Eligibility Criteria
All articles that match the criteria shown in Figure 1 were included. We set the primary goal as mapping the compass of research on omics into a wide-range and coarse-grained taxonomy of three groups. After a thorough preview of the available literature, the groups were chosen without limitation. We eliminated the articles that did not meet the qualifying requirements to remove duplicate articles. The following points comprised the exclusion criteria: (1) The article is not in English. (2) Omics and DL are briefly discussed. Figure 2 shows the proposed taxonomy.

Searching Settings
The article search procedure was conducted on 18 April 2022. The search query was entered into the search boxes of the IEEE, WOS, PubMed, and SD databases. The searches were conducted in all of the aforementioned databases using terminology-related keywords ("Omics"). Afterward, these keywords were paired with the keywords "Deep  the groups were chosen without limitation. We eliminated the articles that did not meet the qualifying requirements to remove duplicate articles. The following points comprised the exclusion criteria: (1) The article is not in English. (2) Omics and DL are briefly discussed. Figure 2 shows the proposed taxonomy.

Results
The details of the first run of the search query filtering 325 articles are as follows: 29 articles from the IEEE Xplore search engine, 142 articles from SD, 104 articles from Pub-Med, and 50 articles from WOS over a period of five years (2018-2022). The number of duplicate articles was 22. A total of 130 articles were left after 173 were excluded for being unrelated via title and abstract scanning. After the review of all the manuscripts, 66 articles were eliminated. The final group of articles eventually included 64 articles. These publications were carefully read to obtain a general research overview of this new field. However, several studies concentrated on the same topic. The articles were sorted into groups according to the study goals and used in the taxonomy procedure. Figure 2 shows a potential taxonomy for evaluating studies applying DL in omics. The acquired taxonomy may be broken down into three distinct article categories. First, only 35 of the 64 publications focused on the research on DL as a clinical application in the omics field. These publications include studies on disease subtyping, biomarker discovery, pathway analysis, and omics data prioritization. Second, 22 studies conducted a review and survey. They reviewed two aspects: single omics and multi-omics. Third, seven articles focused on other aspects of omics data, such as comparative analysis and guideline studies. We identified the general types of articles via summaries. Then, we changed the categorization into a literature taxonomy, as shown in Figure 2. We also illustrated numerous distinct subcategories of the primary classes in this diagram without any overlap. The following sections describe the recognized categories and condensed data related to them.
As illustrated in Figure 2, we rebuilt the categorization into a literary taxonomy indicating the broad types of articles through outlines. Multiple nonoverlapping subclasses of

Results
The details of the first run of the search query filtering 325 articles are as follows: 29 articles from the IEEE Xplore search engine, 142 articles from SD, 104 articles from PubMed, and 50 articles from WOS over a period of five years (2018-2022). The number of duplicate articles was 22. A total of 130 articles were left after 173 were excluded for being unrelated via title and abstract scanning. After the review of all the manuscripts, 66 articles were eliminated. The final group of articles eventually included 64 articles. These publications were carefully read to obtain a general research overview of this new field. However, several studies concentrated on the same topic. The articles were sorted into groups according to the study goals and used in the taxonomy procedure. Figure 2 shows a potential taxonomy for evaluating studies applying DL in omics. The acquired taxonomy may be broken down into three distinct article categories. First, only 35 of the 64 publications focused on the research on DL as a clinical application in the omics field. These publications include studies on disease subtyping, biomarker discovery, pathway analysis, and omics data prioritization. Second, 22 studies conducted a review and survey. They reviewed two aspects: single omics and multi-omics. Third, seven articles focused on other aspects of omics data, such as comparative analysis and guideline studies. We identified the general types of articles via summaries. Then, we changed the categorization into a literature taxonomy, as shown in Figure 2. We also illustrated numerous distinct subcategories of the primary classes in this diagram without any overlap. The following sections describe the recognized categories and condensed data related to them.
As illustrated in Figure 2, we rebuilt the categorization into a literary taxonomy indicating the broad types of articles through outlines. Multiple nonoverlapping subclasses of the primary classes are also depicted below. The sections that follow detail various perceptions and provide concise data for support.

Clinical Application Studies
Clinical application (development) papers (35/64) focused on enhancing the process of diagnosis, prognostication, therapy, and knowledge discovery. They comprised the majority of research works on omics data using DL. We categorized development articles according to the case study information offered in the literature, particularly in the parts of datasets and findings.
In general, precision medicine seeks to replace generic treatments for a large population with customized, targeted medications and treatment plans based on each patient's molecular profile [20]. Alternatively, it seeks to develop preventative medicine plans using disease susceptibility assessment [21]. Omics data are crucial to this shift because they allow for the simultaneous analysis of illnesses at several levels (such as the DNA sequence, gene expression, and medical imaging). Moreover, the specific elements of the affected complex biological activities can be identified. Several ML-based technologies have been used in this new context to apply medicine [22].

Disease Subtyping
Disease subtyping is relevant to studies aimed at finding groups of patients who exhibit different therapeutic/prognostic outcomes [23]. Moreover, disease subtyping concerns diseases comprising multiple subtypes implicated in prognosis [24]. Disease subtyping is vital for diagnosis and individualized patient therapy [25]. In particular, 5 out of 35 articles on the disease subtyping based on DL models were found, as shown in Table 1.
Computational multi-omics approaches are based on DL techniques; they typically aim to classify patients according to disease subtypes [26]. These methods are usually exploited to detect regularities and patterns revealing different disease molecular subtypes or disease classifications that share a common pattern of pathway perturbation [27]. The most common disease is cancer; it is divided into several categories according to various criteria, such as phenotype, molecular portraits, and histopathology [28]. Several works in this direction have been proposed using direct DL classification models or clustering methods. Cancer is the most common disease that requires solution subtyping. According to the authors in [26], next-generation sequencing (NGS) methods accelerate human genome mapping. Sophisticated NGS methods indicate that several genetic molecules are involved in developing breast cancer and its subtypes. Consequently, a DL model based on a stacked autoencoder (SAE) is needed to diagnose different forms of breast cancer accurately. In the present study, the authors used multi-omics data, such as miRNA, mRNA, and DNA methylation, which are pivotal in breast cancer subtype classification. They also play a crucial role in breast cancer formation. Moreover, the authors concluded that SAE could reduce the dimension of high-dimensional multi-omics data. In [29], the authors highlighted that most of the existing methods for breast cancer subtyping used only gene expression to identify cancer subtypes; another major issue is that most of the existing clustering methods completely ignore the results from prior knowledge. Thus, the authors designed a new DL fusion clustering framework to integrate multi-omics data (mRNA expression, miRNA expression, and DNA methylation) with The Cancer Genome Atlas (TCGA) BRAC dataset for breast cancer subtype identification. Table 1 shows that subtyping can be done via a classification task, clustering process, or both. The most frequent DL model for such a process is the autoencoder (AE) model and its improved versions. However, the data type most frequently used for classification or clustering type is miRNA and mRNA. However, data subtyping is used mainly for cancer disease subtyping, particularly breast cancer.

Biomarker Discovery
Biomarker discovery involves studies that detect omics characteristics that indicate a disease state [23]. Biomarkers, sometimes referred to as biological markers or biomarkers, are measurable and evaluable indicators of specific biological states in healthy and pathological processes and potential pharmacologic reactions to therapeutics [32]. As one of the most prevalent categories, molecular biomarkers are extensively researched across many disciplines. Examples include important genes, RNAs, proteins, and metabolite molecules from tissues, blood, and other bodily fluids [33].
The discovery of novel biomarkers for early illness diagnosis, therapy response, and categorization is one of the most popular uses of omics technology in biomedical research. TCGA [34] and other publicly available omics datasets focusing on cancer allow the identification of novel biomarkers via DL [22]. Molecular biomarkers are found by analyzing the information supplied by various omics [35]. For instance, a precise and quantitative risk assessment for cardiovascular disease is provided by a high-sensitivity C-reactive protein test [36]. The planning of preventative actions and patient decisions is mainly influenced by biomarkers, which may be categorized as either diagnostic, prognostic, or predictive [37]. Prognostic biomarkers offer information on the overall prognosis with or without the standard course of therapy. In comparison, diagnostic biomarkers are used to determine the presence of illness in a patient. Before the onset of their physiological symptoms, biomarkers may help identify high-risk patients. They also help determine how a condition progresses [38,39]. In the present study, we considered that prognostic, predictive, survival analysis, and disease recurrence have the same meaning.
Drug discovery/repurposing research aims to find new medications or potent ones that are already on the market and created for other ailments [23]. The traditional drug discovery approach is dominated by target-based high-throughput screening. For decades, it has been the focus of computer-aided drug discovery, including the recent DL applications [40]. Precision medicine depends critically on the accuracy of prediction models that forecast the response of medications to treatment based on the molecular profiles of patients. Prediction models provide options for choosing acceptable individuals for clinical trials; they also assist physicians in choosing the most successful therapy option [41]. The objective of biomarker discovery research was emphasized in our study based on many factors, including data type, illness type, analytic techniques, and the presence of medical validation. The final number of articles in the indicated category is 22 out of 35, as shown in Table 2. Random projection [30] and PCA [31] Liver cirrhosis (CIR), colorectal cancer (COL), obesity (OBE), inflammatory bowel disease (IBD) and type 2 diabetes(T2D) [26] p-value and t-test --No  According to the above table, most biomarker discovery studies were conducted for prognostic purposes. Moreover, many studies on diagnostics of different case studies exist. However, studies on solving the issues with patients' treatments are few. A GNN model is often used for biomarker identification because GCNs use a different methodology by attempting to categorize network nodes based on node characteristics and network design. Similar to CNNs, GCNs gather information from adjacent nodes in a hierarchical manner; they may also function in a semi-supervised setting when labeled nodes are in short supply [63].
Given the data issues, a statistical test based on p-value and t-test has been widely used to measure the validity of used data and obtained results. Finally, we observed that few studies validated the results from a medical viewpoint. This observation may direct us to a question: how meaningful are these results for the medical sector?

Pathway Analysis
Pathways analysis or knowledge discovery is essential in studies aimed at discovering the relation among -omics terms, such as genes or proteins, and measuring the influences of these terms on DL performance. Thus, this category looks for relationships of used omics data rather than outcomes of DL models for diagnostic or prognostic purposes. In particular, only 5 out of 35 studies were found to belong to the pathway analysis category, as shown in Table 3. According to the above table, most data used for omics analysis are usually text or numeric. However, image data, such as CT images, can also be used for this purpose. In [66], the authors used CT images to map the features between medical images and gene expression profiles and quantify their correlations. For improvement, different studies took full advantage of DL methods and characterized lung cancer clinically at both genome and image levels. However, medical validation can be directly from doctors or based on a validation method that applies only to the medical field. Adding clinical information, such as gender, age, stage, and smoking history, can also be used to validate the results of DL from a medical point of view. However, only a few studies conducted medical tests to show how vital the obtained results were for the medical sector.

Omics Data Prioritization
The enormous amount of omics data offers a previously unheard-of potential for computational algorithms to prioritize illness candidates, thereby illuminating the course of human diseases and considerably facilitating cancer prevention, detection, and therapy. Thus, we discuss this trend in this section even though it demands some work. The three methods of prioritization are as follows: prevention, diagnosis, and treatment prioritization or gen prioritization.
The authors of [70] introduced a computational technique called MRSLA to find disease-associated miRNAs. The disease miRNA prioritization task was designed as a recommender system suggesting the miRNAs that tend to be responsible for a given disease based on a low-rank approximation framework. This robust AI algorithm can successfully incorporate multimodal features into the prediction model and produce a high-performing result [70]. In the same way, [71] utilized DL approaches for prioritizing complex disease loci by investigating the landscape of ML applications in three parts: selected models, input features, and output model performance.
On the contrary, the data analysis from several omics has indicated potential genes for measurable features. However, these data are not integrated well, particularly in nonmodel species. Thus, choosing candidate genes for further experimental confirmation becomes difficult. The authors of [72] utilized the DL method (CNN) that integrates multi-omics information to prioritize the candidate genes of objective traits.

Single Omics
In this subsection, our study has focused on review papers that considered only a single type of common omics data. Mainly four types of omics directions have been investigated as follows:

Genomics
The study of the structure, function, evolution, and mapping of genomes is known as genomics. Its goal is to characterize and quantify the genes that control the creation of proteins with the help of enzymes and messenger molecules. Genomics uses techniques to analyze the DNA sequences for studying the structure and function of genomes, gene regulation, and genetic alterations that can be associated with several diseases [22].
We found further in the literature that several works have investigated this topic. In [40], the authors presented a mechanism-driven neural network-based approach, called DeepCE, for predicting the differential gene expression profile perturbed by de novo chemicals. DeepCE uses a GNN and multi-head attention mechanism to model chemical substructure-gene and gene-gene associations. Additionally, the authors suggested a unique data augmentation method that uses faulty trials in the L1000 dataset to extract relevant information. Extensive work was introduced in [73]. First, that work defined the objectives of learning perturbation response in single-cell omics (genomics). It also surveyed the existing approaches, resources, and cited datasets in GitHub and discussed how a perturbation atlas could enable DL models to construct an informative perturbation latent space. In the same work, the authors examined future avenues toward robust and explainable modeling using DNNs, thereby enabling the integration of disparate information sources and an understanding of heterogeneous, complex, and unseen systems.
The authors of [74] investigated how DL has contributed to and been utilized in the mining of biological data. Well-liked open-source DL tools relevant to these data and accessible open-access data sources related to the three data kinds were studied by concentrating on the application of DL to analyze data patterns from several biological areas. Additionally, comparative analyses of these tools from the qualitative, quantitative, and benchmarking viewpoints were presented.
Many ML-and DL-based metastatic prediction techniques created to date are summarized in the present article [75]. Various molecular data types utilized for the creation of models and crucial signatures obtained using various techniques were also described. Additionally, the authors emphasized the difficulties in using DL and ML techniques and provided recommendations to enhance their ability to anticipate outcomes.
In the utilization of genomics in Alzheimer's disease (AD), [76] focused on the latest developments for AD prediction using DL techniques with the principles of neuroimaging and genomics. The authors first described several experiments that employ DL algorithms to develop AD prediction using genomes or neuroimaging data. They particularly outlined pertinent integrated genomics and neuroimaging studies that use DL techniques to predict AD by combining both genomic and neuroimaging data.
Finally, the authors in [68] aimed to help fill the gaps in data preprocessing and genome-wide association study (GWAS) methodologies by reviewing novel techniques and methodologies. The data preprocessing performed prior to a GWAS presents challenges in Hardy-Weinberg estimation, genotyping, and accounting for factors such as sample structure.

Transcriptomics
The collection of all messenger RNA molecules in one cell, tissue, or organism is known as the transcriptome. In addition to the chemical identities, it contains the quantity or concentration of each RNA molecule. Transcriptomics measures the degree of expression for each RNA transcript generated in a cell. Transcriptomics raw data are often processed to produce expression matrices, which are frequently the input of DL techniques. These matrices contain an estimate of each gene's or transcript's level of expression across many samples and situations. Transcriptomics applications span a broad spectrum, and DL has been effectively used in some of these applications [22].
In the theoretical literature review covering the transcriptomics keyword, we found only two articles during the search period. First, the authors presented the most recent methods for determining the link between learned latent and observed variables and the external phenotypes in [77]. Thus, the scholars improved the comprehensibility of the DL methods. They exhibited the usefulness of the proposed techniques in an application with single-cell gene expression data (transcriptomics). The relationships between observed gene expressions and experimental variables or phenotypes can be understood well because of the work presented here. Deep generative models may also provide synthetic observations by offering a generative model for the latent and observed variables, thereby enabling us to evaluate the uncertainty in the learned representations.
Second, the authors analyzed the most recent data modeling techniques used with transcriptomics data I TGx in [78]. The researchers demonstrated how the TGx data might be used for the benchmark dose study. Then, the approaches for reading across and modeling adverse outcome pathways were also reviewed here. Additionally, the authors addressed how network-based strategies may effectively define the mechanism of action or particular exposure biomarkers. Moreover, they explained the primary AI techniques used for developing predictive classification and regression models using TGx data while addressing the existing challenges.

Metabolomics
The metabolome represents the collection of all metabolites in a biological cell, tissue, organ, or organism. These metabolites are the end products of cellular processes. Metabolomics is the science that studies all chemical processes involving metabolites. In particular, metabolomics is the study of chemical fingerprints that specific cellular processes establish during their activity. It is the study of all small-molecule metabolite profiles.
We found several state-of-the-art works in this type of omics. A historical perspective work [79] introduced the readers to the fundamental computational concepts of ANNs. It also provided a brief historical context of their use in metabolomics, discussed the advantages and disadvantages of this method when applied to omics data, and finally looked toward future applications and challenges.
In [80], the authors explored breakthrough approaches for natural product discovery from plant microbiomes, emphasizing the promise of DL as a tool for endophyte bioprospecting, endophyte biochemical novelty prediction, and endophyte regulatory control. It concludes with a proposed pipeline for harnessing a global database to uncover and unsilence desirable natural products.
Finally, [81] presented the applications of DL that have recently emerged in metabolomics research. Then, the scholars presented the utilization of DL in the data preprocessing step. They also reviewed the use of the CNN model in developing DL for metabolomics data.

Proteomics
The term proteome refers to the sum of all the proteins in a cell, tissue, or organism. Proteomics is the science that studies these proteins in terms of their biochemical properties and functional roles. How the quantities, modifications, and structures of these proteins change during growth in response to internal and external stimuli was also investigated.
In this area, we found only one article that covered proteomics data. This scenario is related to the natural behavior of researchers because protein is usually used with other omics data to detect many tumors and cancers; moreover, protein is not used in single omics except in a few cases, including [82]. In the same work, the authors demonstrated the effectiveness of ensemble learning for reliable protein abundance prediction using single-cell multimodal omics data. Their study paved the way for knowledge discovery by mining single-cell multi-omics data on a large scale. Their work was accomplished by contrasting numerous tree-based ensemble learning techniques with neural network models. According to their work, ensemble learning frequently outperformed neural networks. In particular, random forest (RF) exhibited the greatest overall performance. Moreover, they interpreted the biological mechanisms driving the prediction using the feature significance ratings from RF.

Multi-Omics
As we mentioned above, multi-omics currently has promise for filling the gaps in the understanding of human health and disease. Many researchers are working on ways of generating and analyzing disease-related data. They also describe the relations between omics data to predict the diagnosis, prognosis, and treatment of diseases. In this section, we reviewed the survey articles working on all previously mentioned aspects, as shown in Table 4.
As shown in Table 4, different types of literature papers have been mentioned in the second column as review, survey, systematic review, comprehensive review, methodological review, and system biology review. Sequentially, The Survey paper gives information about the amount of research done so far in a particular domain or on topic. In contrast, the Review paper analyzes published works and provides insights with technical evidence. A review article or review paper is based on other published articles. It does not report original research. Where and how one searches for evidence is an important difference. While literature reviews require only one database or source, systematic reviews require more comprehensive efforts to locate evidence. Multiple databases are searched, each with a specifically tailored search strategy (PRISMA). A systematic review that is less rigorous is often called a "Comprehensive review". On the same side, a methodological review is a type of systematic review that focuses on summarizing the state-of-the-art methodological practices of research on a substantive or essential topic. Finally, a system biology review is a new type of paper that aims to provide specialists with a unique and educational platform to keep up-to-date with the expanding volume of information published in the field of Systems Biology, which belongs to a systems biology journal [83]. We found four types of reviewing and survey articles in the literature about multiomics areas with DL approaches. Some of the omics data take up all four omics data. Some of them only take two of the omics data. As shown in Table 4, the DL models included in the articles varied from not mentioning any model and confining to mentioning DL in general to mentioning six DL methods and their relationship with multi-omics. The target disease type must be mentioned in the review articles. Some of the works dealt with cancers and tumors in general. Some of them identified a specific type of disease and the works that dealt with this disease. From the above, we can summarize that review and survey works are vital for researchers to understand the relationship of multi-omics with DL from two perspectives: medical and technical points of view.

Others
Some works were not classified under development and could not be considered a review of previous research. Therefore, we decided to classify them as follows.

Guidelines
Guideline papers aim to simplify the processes of utilizing DL in multi-omics according to a routine or sound practice. The three types of guideline papers are proposal, perspective, and narrative papers. They serve the same purpose in terms of presenting ideas. In addition, the articles' main objectives differ. Table 5 shows all types of guideline papers.

Comparative Studies
Comparative studies answer the question of which DL model is accurate in omics data. This objective was attained by comparing two or more DL models with single-or multi-omics data based on the metrics shown in Table 6.

Current Omics Datasets
The cornerstone of conducting an omics analysis is the availability of a dataset. In general, a dataset can support omics or multi-omics information. Moreover, this dataset can be associated with a single disease or multiple diseases. However, most omics datasets are not injected directly into a computational model unless a little preprocessing is performed for data enhancement or integration. In the present review, we tried to cover the existing datasets with details links to data type, number of samples, disease type, and implementation links, as shown in Table 7. In particular, all existing datasets are different in terms of the number of samples and the frequency of usage. For instance, we found that the cases in TCGA are the most frequently used data by omics studies. Moreover, the number of samples in this dataset is the largest. Moreover, TCGA is preferred by researchers because of the availability of tools for the required analysis and different types of omics data.  A comprehensive description of the structural and evolutionary relationships between all proteins whose structure is known

Challenges
In general, DL models link various challenges that affect the outcomes of classification or regression processes. In the DL models employed for omics analysis, we discovered many DL-related issues. The preprocessing procedure, datasets, model validation, and testbed applications are discussed in the following subsections (see Figure 3). on 23 December 2022)

Challenges
In general, DL models link various challenges that affect the outcomes of classification or regression processes. In the DL models employed for omics analysis, we discovered many DL-related issues. The preprocessing procedure, datasets, model validation, and testbed applications are discussed in the following subsections (see Figure 3).

Low Noise Reduction Efficiency
The most popular experimental method in proteomics is separating proteins into short amino acid chains (peptides) and then using an MS to evaluate these peptides. The MS output signals are matched with the peptide profiles kept in open or private databases to identify the peptides. These databases are still unreliable. They also lack information. Additionally, DL is used to infer protein secondary structures from their amino acid sequences [108]. The creation of data for proteomics and metabolomics depends on NMR technologies. Given its technological limitations, it often returns noise signals that must be filtered to increase accuracy. Using DL approaches to metabolomics data is particularly difficult because identifying unique components contributing to individual samples in these kinds of research is crucial [22,109].

Traditional Feature Selection Methods
The authors in [51] suggested using techniques based on DMPs and DEG to train models. These authors did not employ traditional feature selection or dimension reduction methods, such as LASSO, Relief-F, or PCA, because they cannot capture the bifurcated processes logically. This finding implies that using these techniques may result in few characteristics or dimensions; however, the occurrence of a few features with biological significance is not ensured. Moreover, these methods are unsuitable for application to the multi-omics dataset. The interaction between two various omics datasets cannot be considered by these approaches. One biological factor that may regulate and govern the level of expression of genes adjacent to the CpG site is DNA methylation, which is particularly well-known. Conventional feature selection or dimension reduction methods cannot handle this trait [51].

High Computational Cost
In the early days, GNNs learn a node's features by iteratively propagating information from the neighboring nodes until convergence [110]. High computing costs and the learning filters' lack of the localization attribute are two notable drawbacks of these models. Chebyshev polynomials were proposed by Defferrard et al. in 2016 as localized learning filters in a spectral-based GCN (ChebNet) for converting computing costs into linear complexity [31,111].

Selection of Best Prediction Model
Different omics applications have generated several prediction algorithms. These applications are founded on various computational techniques and diverse omics data sources. However, several studies [112,113] have demonstrated that various factors affect how well prediction algorithms function. We must first discover the important criteria affecting the performance of these algorithms to choose an appropriate prediction algorithm for a certain application. Then, we can utilize the identified factors to suggest a prediction algorithm for the application [52].
However, extensive training, testing, and cross-validation are required to determine the most appropriate weights for each parameter and its influence on algorithm performance. Nonetheless, the efficacy of unstable techniques varied based on the combined omics profiles, drugs, and performance indicators. The evaluations of these methods depend on (1) performance measurements and (2) supported data (e.g., omics characterization integrated into the model and drug structure/pathways). Thus, a difficult issue is to integrate different existing prediction models and determine which model is most effective in each scenario [47]. However, the application of CNNs to omics data poses several obstacles, such as the processing of complicated network architectures and its integration with transcriptome data [49].
However, the kind of model architecture limits how the information may be integrated. For instance, in the bioinformatics literature, a feedforward NN called a visible neural net-work can only represent directed acyclic graphs, which do not match with some knowledge networks (KEGG and STRING). Additionally, only the first hidden layer is connected to the input genes; the subsequent hidden layers are unrelated. The connections between failing levels are also removed. Consequently, a piece of information must be shortened to be absorbed into the neural network [56].

Ignoring Prior Knowledge
Most of the existing clustering methods ignore established results from prior knowledge. Prior biological knowledge (e.g., Pam50 breast cancer subtypes) can guide representation learning. According to the gene expression profile analysis [114], breast cancer was first classified into five intrinsic subtypes according to the Pam50 breast cancer subtyping: Luminal A, Luminal B, basal-like, normal-like, and HER2 [29].

DL Model Explainability
The lack of interpretability of such approaches is one of the most difficult issues impeding the growth of ML in healthcare. Most ML techniques, including DL techniques, are regarded as black boxes because of their complexity. The choices of these models cannot be explained. One of the most crucial current concerns is making ML algorithms interpretable. In the medical field, final users (e.g., researchers, clinicians, and patients) must understand why a phenotype has been predicted to ensure that it is based on reliable medical features rather than on irrelevant artifacts. This understanding has a considerable impact on their decisions and confidence in the model, regardless of the model's efficacy. Finally, the model examination may aid biological discovery by exposing intriguing signs. The application of sophisticated machine learning models, including DL, on omics data, enables the emergence of precision medicine. However, their use in clinics is limited and not explainable [56].
Although DL approaches allow for the uncovering of hidden patterns in opaque and complex data, the models themselves are complex and opaque; thus, researchers cannot easily infer how learned latent representations relate to the observed variables [69]. Enhancing the interpretability of machine learning from a "black box" from genotype to phenotype is also challenging, particularly for uncovering causal mechanisms from risk variants to diseases [115].

Omics Interaction Identification
Identifying the interactions and crosstalk between heterogeneous RNA classes is essential for uncovering the functional role of individual RNA transcripts, particularly for unannotated and sparsely discovered RNA sequences with no known interactions. As high-performing and adaptable techniques that can either predict RNA-RNA interactions from sequence or infer missing interactions from patterns that may exist in the network topology, sequence-based DL and network embedding methods have been gaining popularity. However, the majority of the approaches used today have several drawbacks, such as the inability to make inductive predictions, determine the directionality of interactions, or combine different sequences, interactions, expressions, and genomic annotation datasets [116].

Time-Consuming
Matrix factorization (MF) methods are general models that use shared features across diseases and genes to predict genetic links between diseases [117]. However, MF-based algorithms often take too long to converge and can only handle a restricted variety of data during actual performance [55].

Single-Omics Dimension
Most of the existing methods use only gene expression to identify cancer subtypes [29]. The individual type of omics data only represents a single view that suffers from data noise and bias [60,61]. Huge and various types of genetic data have been produced with the advancement of high-throughput sequencing techniques. The integration of multi-omics data contributes to cancer subtype identification.

Reproducibility and Generalization
DL success depends on the random numbers produced at the beginning of training; moreover, practitioners may adjust hyperparameters, including learning rates, batch sizes, weight decay, momentum, and dropout probabilities [118]. Setting the same random seeds across many stages ensures the same experimental outcome. Although the hyperparameters and random seeds were often not provided in the present study, keeping the same code bases was crucial [119]. Reproducing the study and achieving the same results are challenging because of the uncertainty of the configuration and the randomness involved in the training [120].

Selection of Suitable Evaluation Metric
Classification accuracy is often used for evaluating algorithm efficacy. However, it may be deceptive when used for data imbalance because an unbalanced training set may result in a classifier that favors frequent classes. In binary classification, the classifier may assign each test point to the dominant class and provide an optimistic accuracy estimate. BACC [121] was created to handle unbalanced training data (true negative rate) by averaging sensitivity (true positive rate) and specificity. In particular, BACC balances the contributions of different classes by giving each class a weight. Thus, the classifier produced can learn the same amount from each class. With binary predictions, BACC is also equal to the area under the ROC curve. This metric is the same as the regular accuracy if the dataset is balanced. Thus, the disease and control samples were not the same size in this application [115].

Classification Type
The GCN is blind to different subpopulations within a cancer tissue because of the averaging of characteristics across patients for a specific cancer type. A possible drawback of the present situation is the vague idea of cancer genes. A multiclass, multilabel scenario becomes increasingly appropriate and enables the prediction of the cancer genes that are particularly unique to a single cancer type, thereby opening the door for precision oncology. In comparison, certain genes substantially affect numerous cancers [54].

Data Integration
Identifying disease-related lncRNAs is crucial for the diagnosis, prevention, and treatment of diseases. Many computational approaches have been developed; however, effectively integrating multi-omics data and accurately predicting potential lncRNA-disease associations remain challenging, particularly regarding new lncRNAs and new diseases [64]. In omics, several datasets comprise diverse data in terms of variable count, scale, distribution, and modality. The analytical goal in healthcare is challenging because of the heterogeneity of the data. Genetic instability is the primary contributor to heterogeneity [122], thereby complicating the determination of particular prognostic genes using only gene expression patterns [43]. Many machine learning techniques, including clustering, graph and network analysis, kernel learning, and DL, may be used in addressing the issue of heterogeneity [26,30].
The process of extracting heterogeneity from GRNs is another challenge. Previous works used GNN algorithms to combine homogeneous GRN with single-omics data. However, this approach is primarily ineffective in extracting the complex information present in heterogeneous GRN, particularly its rich variety of interconnection [44].

Data Scarcity
In the field of brain-related diseases, the omics data obtained using advanced sequencing technology typically have few patient samples (tens to hundreds of samples) [123]. Reliably training complicated model architectures with limited labeled data remains a challenge. The scarcity of labeled multi-omics data may result in the overfitting/underfitting of complicated models, e.g., DL methods [44].

Data Selection
Millions of genes have been sequenced in the data-driven genomics era; however, their causal relationships with disease phenotypes remain limited because of the difficulty of elucidating underlying causal genes using laboratory-based strategies [67].

Temporal Data
The illnesses in healthcare evolve and change over time in a non-deterministic way. Machine learning algorithms have trouble dealing with the time component because they assume that inputs are static [124,125].

Missing Data
Missing data refers to observations that are absent for various reasons, such as a protein's decreased sensitivity or inability to process NGS data. When the majority of the samples in a dataset have missing data, the missingness becomes problematic. This issue is solved by list-wise deletion, which removes the whole sample of missing variables from the dataset.
However, the approach can result in loss of information if the percentage of missing data is large. Another option is imputation, which assigns the variable with missing data some random values, often the mean or median. The k-nearest neighbors method is mainly used to impute missing data [91,124].
Furthermore, a study defining the specific ratio for excluded features with missing data remains lacking. For instance, the authors in [61] eliminated the features that were absent from each cancer dataset in more than 20% of the patients, in addition to excluding patient samples if these data missed more than 20% of the remaining multi-omics characteristics. Then, the authors eliminated the cancer datasets that included less than 50 unfiltered samples.

Dimensionality Curse
Most healthcare datasets include few samples but many characteristics or variables. Overfitting becomes an issue for many types of machine learning algorithms as the number of variables increases. Imaging, clinical, and omics data are just a few examples of the many kinds of information that may be found in healthcare databases. Various data reduction methods must be performed to make high-dimensional datasets suitable for learning. Feature extraction and feature selection methods are commonly used [126].
PCA and non-negative matrix factorization are two examples of the types of analyses utilized in feature extraction. Feature extraction plays a central role in unsupervised learning, in which answer labels are unknown. Machine learning with feature extraction is commonly utilized in multi-omics data to identify disease groupings. Selected features are used in feature selection. One can pick features using a filter, a wrapper, or an embedded method. Unlike other feature selection techniques, the filter chooses a subset of characteristics independently of any model. Pearson correlation, ANOVA, and information gain are just a few filtering techniques that may be applied. The wrapper technique uses a prediction model to select the most useful traits iteratively while retaining the least useful ones. Boruta, Jackstraw, and Recursive Feature Elimination are examples of the many wrapper techniques that may be applied. Wrapper approaches perform computation on massive datasets inexpensively. The embedded technique falls between the filter method and the wrapper method in terms of computational complexity. LASSO is an embedded technique. Supervised learning is mainly used for feature selection in classification and regression. Not all features are meaningful in high-dimensional datasets; thus, feature selection methodology is utilized prior to dataset integration [43,48].

Imbalance Data
In healthcare, an imbalance of data occurs when the samples with positive values in the target class in the dataset are less than those with negative values.
Several techniques can address the imbalanced data in machine learning, including data sampling, algorithm modification, and ensemble learning. The dataset is balanced before the machine learning for classification in sampling is utilized. Undersampling and oversampling are typically used together to address the problem of an unbalanced dataset. When algorithms are changed, the same dataset with imbalance is used to make the changes [59]. Giving samples with minority types of high weights modifies the algorithm. In ensemble learning, the aggregate of all classifier judgments is obtained by applying each model subset of the majority of examples and considering all minority data [91,124,127].

Sample Group
Most of the integrated molecular studies do not use the multi-omics dataset from the same sample group. Unlike the TCGA database, no database has produced considerable omics data from one sample [51].

Lack of Outcome Validation
Various DL-based single-or multi-omics publications have emerged in the past five years. The evaluation process has usually been conducted based on well-known measurements in the data science field, such as AUPRC, AUC, precision, recall, F-score, and ROC. These metrics can reveal how meaningful the obtained results are from the theoretical perspective. Most of the reviewed articles in this study have this kind of evaluation but lack supportive validation based on the medical perspective. However, the question of how remarkable and satisfying these results are for the medical sector ensues. Therefore, defining the measurements and procedure is essential for achieving the purpose and providing reliable results.

Shortage of Testbed Application
Many proposed works have reviewed omics data analysis using DL. However, no study has been deployed into medical infrastructure and translated into clinical practice.
One of the key issues hindering the clinical deployment of DL methods is transparency. A transparent and explainable ML algorithm seems essential to building trust for clinical decision-making [39,128]. Moreover, this deployment is missing because of the gap in consistent result validation between theory and clinical perspectives.

Conclusions
In the past decade, researchers have drawn considerable attention to the involvement of DL in omics data analysis, thereby making the DL and omics data one of the major research topics. In this context, no clear boundaries have been observed in the development of this field. Thus, further study is necessary to provide a holistic view and track this research line. Our study attempts to provide an extensive view and deep understanding by reviewing and classifying the highly pertinent literature. Consequently, this study maps the final set of relevant articles in three main categories: clinical applications, reviews, and other studies conducted using DL and omics data. In addition to offering an extensive investigation of the body of literature, the three main classes of articles are divided further into subcategories: such as disease subtyping, biomarker discovery, pathway analysis, data prioritization, review articles conducted to single-or multi-omics data, comparative analysis, and guideline studies. Other information is provided, including the difficulties and limitations of DL models and the information on the data and validation procedure. In the present review study, we found that the clinical applications that use DL in addressing various medical problems have received considerable attention from researchers. Additionally, the existing literature shows minimal concern for DL characteristics, feature data, and definition for a straightforward medical validation approach. Furthermore, testbed DL applications used in real-world settings or accepted by healthcare organizations are uncommon. Thus, the omics community should make considerable efforts to close the gap between theoretical and application viewpoints by creating practical applications. Finally, this systematic review assists researchers in tracking the crucial concerns about DL in omics by expanding and outlining new research avenues. In the future, the authors intend to extend this review by focusing on data integration methods and issues relevant to the existing methods.