Harnessing Deep Learning for Omics in an Era of COVID-19

Abstract

Omics data are multidimensional, heterogeneous, and high throughput. Robust computational methods and machine learning (ML)-based models offer new prospects to accelerate the data-to-knowledge trajectory. Deep learning (DL) is a powerful subset of ML inspired by brain structure and has created unprecedented momentum in bioinformatics and computational biology research. This article provides an overview of the current DL models applied to multi-omics data for both the beginner and the expert user. Additionally, COVID-19 will continue to impact planetary health as a pandemic and an endemic disease, with genomic and multi-omic pathophysiology. DL offers, therefore, new ways of harnessing systems biology research on COVID-19 diagnostics and therapeutics. Herein, we discuss, first, the statistical ML algorithms and essential deep architectures. Then, we review DL applications in multi-omics data analysis and their intersection with COVID-19. Finally, challenges and several promising directions are highlighted going forward in the current era of COVID-19.

Introduction

Multi-omics studies are vital to unpack the systems biology of COVID-19 (Soko et al., 2022). Because omics science is composed of big data, dimension reduction methods have been widely applied in data analysis (Meng et al., 2016). Machine learning (ML) and artificial intelligence (AI) offer new prospects at the intersection of COVID-19 and multi-omics systems science. Deep learning (DL) is a subset of ML that is poised to accelerate bioinformatics and computational biology research concerning COVID-19 and life sciences broadly.

DL is currently a rapidly developing frontier of ML and AI applications in the field of omics systems science. DL brings about an end-to-end mechanism through a multilayer network structure without human intervention (Su et al., 2020). DL techniques display robust performance in many real-world applications ranging from criminology to health care (Zhang et al., 2019).

Because COVID-19 is here to stay as an endemic disease after the pandemic course, unraveling its multi-omics basis through robust computational approaches is timely and bears significance for planetary health. This article provides an overview of the current DL models applied to multi-omics data for both the beginner and the expert user. We discuss, first, the commonly used ML algorithms and DL architectures, such as support vector machine (SVM), logistic regression (LR), random forest (RF), deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), autoencoder (AE), and generative adversarial network (GANs). Next, we highlight the recent DL models applied to omics data and the relevant literature at the intersection of multi-omics and COVID-19. We conclude with an overview of the challenges and promising directions going forward in the current era of COVID-19.

Common ML and DL Models

The data analysis algorithms in ML aim to exploit meaningful correlations between data and perform predictions using historical data (Abiodun et al., 2018). As shown in Figure 1, ML models are divided into the following four major types (Pugliese et al., 2021): (i) supervised learning, in which new input data are predicted based on the prelabeled input data; (ii) unsupervised models are applied to unlabeled and unclassified data, and the primary purpose is to discover and analyze patterns in unsorted data (Pugliese et al., 2021); (iii) semi-supervised is a hybrid approach for labeled and unlabeled samples, which eliminates the need for numerous labels; and (iv) reinforcement learning is suitable for sequential decision-making in which the agent attempts to fulfill a target in an uncertain and potentially complex situation (Mahmud et al., 2018).

FIG. 1.

The approaches in machine learning. The boxes where there are instances of real-world health care have their references inserted: Disease Prediction (Bowler et al., 2022), DNA methylation regression (Tian et al., 2019), Biomarker Gene Clustering (Pal et al., 2007), Pathway's pattern discovery (Huang et al., 2014), Microarray dimension reduction (Aziz et al., 2017), Human activity recognition (Zahin et al., 2019), Privacy-preserving clustering (Huang et al., 2016), Diagnostic clinical concept (Moeinizade et al., 2022), Robotic-assisted surgery (Kassahun et al., 2016).

Contrary to statistical ML methods, where features are manually imported into the model, deep models automatically extract features without the intervention of a human. Consequently, large data sets are required for effective feature extraction and meaningful correlation (Pirmoradi et al., 2020). Briefly, common models were discussed in summary.

Random forest

RF (Schonlau and Zou, 2020) is a supervised ML algorithm that applies multiple decision trees to handle regression and classification problems by majority vote. This algorithm employs ensembles of trees where random variables extend each tree from the original set of attributes.

Logistic regression

LR is a derived version of linear regression with binomial variables. It could be an effective model for binary class issues to estimate the relationship between whole independent variables, ranging from categorical to continuous (Shipe et al., 2019).

Support vector machine

SVM is a robust algorithm capable of recognizing subtle patterns in complex data sets. It comes in two types, namely support vector classifier (SVC) and support vector regression (SVR) (Nayak et al., 2015). The SVC is divided into three categories. First, SVC, with maximum margin, was used to establish a linear decision boundary between two classes to investigate hyperplanes and achieve generalization capability and the consequent high-performance classification and accurate prediction. In real life, many problems do not have linear variables, and then, they are unable to be solved linearly. In this respect, two approaches are introduced, soft margin SVC using curved decision boundaries and kernel trick allowing nonlinear boundaries to be formed (Suzumura et al., 2017). Kernel-based models are common for handling noisy data (Yu et al., 2021). However, SVR was performed for regression problems ranging from linear to nonlinear, using an alternative loss function (Parbat and Chakraborty, 2020).

Deep neural network

The DNN model is a deep artificial neural network model that contains input and output layers and hidden layers, which enable resolving complicated and nonlinear model relationships. Each layer consists of neurons that receive inputs from the preceding layer to prepare an outputs map and transform them into the next layer (Sze et al., 2017).

Convolutional neural network

CNN was inspired by a neurobiological model of the visual cortex of the human brain. The visual cortex comprises a map of local receptive fields that decrease granularity in the anterior of the cortex. It contains two types of cells: simple and complex cells. Simple ones respond to primitive patterns, whereas complex cells receive information from simple cells and recognize the complex form (Min et al., 2017). CNNs operate on high-dimensional inputs such as images, videos, and three-dimensional (3D) medical images. The basic model consists of three layers: convolution, pooling, and fully connected (Albaradei et al., 2021), in which the output of each layer could be an input for the next one (Min et al., 2017). A convolution activation layer, the first critical component, performs most of the calculations and is responsible for identifying features of input data through several small filters (Ravì et al., 2016).

A pooling layer is located between convolution layers, which receives feature maps and then reduces input size to conserve critical characteristics. Finally, a multilayer perceptron (MLP) network makes up the fully connected layer. This layer takes the input vector from the previous section to explore the global feature interactions (Su et al., 2020) and predicts the class of data.

Recurrent neural network

RNN analyzes sequential or time series data with a cyclic connection structure (Min et al., 2017). RNNs evaluate the current state's output depending on the previous steps, whereas traditional DNNs consider independent inputs and output data. RNNs evaluate the current state's output depending on the previous steps and have a memory-like property. Short-term memory allows RNNs to remember information from prior states (Mahmud et al., 2018). There are two inputs to this network, the present and the recent past, which combine to determine the state of the inputs. RNNs have usages in bioinformatics, generating medical image descriptions and deoxyribonucleic acid (DNA) sequences (Su et al., 2020).

Long short-term memory

The LSTM network is the advanced version of RNN (Sze et al., 2017), which overcome the RNN weakness. The major impediment to RNN is the vanishing gradient issue, lack of handling long sequences of data, and long-term dependency (Yu et al., 2019). This dependency makes the training process difficult and causes gradient explosion problems, which LSTM could overcome. As a result of these algorithms, it is possible to determine long-term lags between significant events of unknown size (Ravì et al., 2016).

Deep belief network

Deep belief network is a powerful probabilistic generative model combining stacked latent variable layers with restricted Boltzmann machines (RBM) networks. The training involves two steps: layer-by-layer training is referred the unsupervised training of each RBM; fine-tuning refers to the backpropagation technique to perform classification and other tasks on a small labeled data set after the unsupervised training (Sohn, 2021).

Autoencoder

AE is one of the most influential unsupervised neural networks (Yu et al., 2019) that attempt to reduce dimensions and reconstruct the most crucial input features (Sarker, 2021). In the network structure, the bottleneck controls the flow of information and permits only highly ranked information to pass. In addition, the decoder is a part of the network that decompresses the represented knowledge. It reconstructs the data passed through the bottleneck as a mirror image of the encoder and generates similar data with the same input data distribution. In this regard, the small bottleneck decreases the risk of overfitting, but an extremely small size impedes sending beneficial information (Pirmoradi et al., 2020). Denoising, detecting anomalies in time series, detecting network penetration, and generating new data are some applications of AEs (Su et al., 2020).

Generative adversarial networks

GAN is a contemporary architecture introduced by Goodfellow et al. It has been used for semi-supervised and unsupervised learning. GAN consists of two competing networks: a generator and a discriminator (Goodfellow et al., 2020). As an unsupervised technique, the generator model tries to generate new examples using random noise Z. The generator has no accessibility to original data, thus generating fake data via interaction with the discriminator (Alqahtani et al., 2021). However, the discriminator receives fake and real data to predict original or false data. The networks are trained until the discriminator is fooled approximately half the time.

Single Omics Data Analysis Using DL

Over the past few years, DL has gained a strong foothold in many precision medicine domains and addressed different challenges with high-throughput omics data. The omics study initiates with genomics, which assesses genome and the genetic variants related to complex diseases, DNA sequencing, and microarray Data (Fig. 2a and Table 1) (Kaur et al., 2021). Notwithstanding rapid advancements in sequencing technologies, precisely calling variants in the genome is difficult. In this regard, Poplin et al. (2018) applied the CNN method, DeepVariant, which can learn statistical likelihoods to call genetic variation in aligned next-generation sequencing (NGS). It allows nonhuman sequencing projects to benefit from a wealth of human data through the learned model, which generalizes across genome builds and mammalian species.

FIG. 2.

The Application of DL techniques for omics data analyses. (a) In the first box, the omics data, termed in the article, are illustrated as inputs of the second box. (b) Next, several DNNs that have been employed for omics data are listed. (c) The last box shows some applications of DL models in omics data analysis. AE, autoencoder; CNN, convolutional neural network; DL, deep learning; DNA, deoxyribonucleic acid; DNN, deep neural network; GAN, generative adversarial network; LSTM, long short-term memory; MS, mass spectrometry; ncRNA, noncoding RNA; NMR, nuclear magnetic resonance; RNA, ribonucleic acid; RNN, recurrent neural network; TF-DNA, transcription factor DNA.

Table 1.

Leading Deep Learning Models for Multi-Omics Data Analyses

References	Omics type	Aims	Data type	Deep model	Contribution
Agarwal and Chen (2023)	Genomics	Chromatin interaction prediction	pcHi-C	CNN (DenseNet)	DeepPHiC is a supervised multitask model based on transfer learning for promoter-centered 3D chromatin interaction prediction.
Singh et al. (2019)		Chromatin interaction prediction	EPI data sets/somatic mutation	LSTM	SPEID signifies the EPIs through sequence-based features.
Hong et al. (2020)		Data augmentation	Hi-C	GAN	DeepHiC predicts high-resolution data where the sample resolution is low.
Hazra et al. (2022)		Data augmentation	DNA sequences	WGAN-GP	The framework generates artificial synthetic nucleic acid sequences. A predictor between the generator and discriminator networks is employed to detect samples that contain promoter sequences.
Bahado-Singh et al. (2022)	Epigenomics	Disease detection	DNA methylation	DNN	The model predicts nonsyndromic CoA. The study presented reasonable proof of epigenetic dysfunction in related gene pathways.
Levy et al. (2020)		Disease subtype classification	DNA methylation	VAE	MethylNet is a multipurpose API base model that can swap between modules concerning embeddings, prediction, and generation tasks. It can survey unknown diseases and cellular heterogeneity and predict age.
Levy et al. (2020)		Disease subtype classification	DNA methylation	MLP
Lopez-Garcia et al. (2020)	Transcriptomics	Classification	RNA-seq	CNN	The DL approach combines the CNN model to derive correlated features from structured input data with the transfer learning strategy to prevent overfitting for lung cancer prediction.
Majji et al. (2021)		Classification	Gene expression	DeepRNN	JayaALO-based DeepRNN calculates the optimal weights and improves the results of cancer prediction. JayaALO, a combination of Ant lion and Jaya algorithms, plays a role as an optimizer.
Marouf et al. (2020)		Data augmentation	Single-cellRNA-seq	CGAN	cscGAN and realistic synthetic data enhance downstream analyses such as marker genes prediction, revealing gene–gene dependencies, classification, and reproducibility rate. RF algorithm is trained on generated data for model evaluation.
Marouf et al. (2020)			Single-cellRNA-seq	RF
Han et al. (2022)			Gene expression profile	WGAN	CWGAN improves the diversity and quality of generated data by removing the output layer's activation function and adding a constraint penalty.
Han et al. (2022)			Gene expression profile	CGAN
Wen et al. (2020)	Proteomics	Retention time prediction	CPTAC data sets	CNN	AutoRT accurately predicts RT to assess identification strategies for mutated peptides. The model employs a genetic algorithm for NAS and transfer learning for developing accuracy with small-size training data.
Wen et al. (2020)		Retention time prediction	CPTAC data sets	LSTM
Tiwary et al. (2019)		Spectrum prediction	PRIDE database	RNN	DeepMass: Prism, a DL regression method using one-hot encoding, prediction peptide sequences. It benefits from DDA and DIA computational workflows to enhance peptide prediction rates.
Tiwary et al. (2019)		Spectrum prediction	PRIDE database	LSTM
Dührkop et al. (2021)	Metabolomics	Multilabel classification	Mass	SVM	CANOPUS is conducted based on class assignment and ontology prediction, which accurately predict compound classes from high-resolution fragmentation spectra data.
Dührkop et al. (2021)		Multilabel classification	Spectral data/molecular structures	DNN
Cakmakci et al. (2020)		Prediction	NMR spectra	RF/DNN	The real-time model detects informative regions in the HRMAS NMR spectrum and malignancy biomarkers of intraoperatively resected tissue to use this information for tumor margin assessment during surgery.
Xu et al. (2019)	Multi-omics	Subtype classification	Gene expression/miRNA/DNA methylation	SAE	HI-DFNForest applies SAE for feature representation and a hierarchical integration of deep flexible neural forest as a classifier.
Lin et al. (2020)			mRNA/DNA methylation/CNV	DNN	DeepMO integrates features based on the chi-squared test and uses DNN for multiclassification.
Islam et al. (2020)			CNA/gene expression	CNN	DCNN_Concat learns two data sets independently, concatenates deep features, and classifies tumor subtypes by CNN. A major benefit is a significant reducing the dimension of inputs through CNN.
Zhang et al. (2018)		Subgroup clustering	CNA/gene expression	AE	The bottleneck layer of AE identifies 100 features, and then, Cox-PH selects 35 highly correlated features. Finally, K-means clustering employs to infer the survival risk of subgroups.
Yang et al. (2021)		Subgroup clustering	miRNA/mRNA/DNA methylation/CNV	GAN	The subtype-GAN model derives latent features. The cluster ensembles with the Gaussian Mixture model are employed to examine the cluster number K and specify the subgroups labels.

3D, three-dimensional; AE, autoencoder; CGAN, conditional GAN; CNN, convolutional neural network; CNV, copy number variation; CoA, coarctation of the aorta; Cox-PH, Cox proportional hazard; DDA, data-dependent acquisition; DIA, data-independent acquisition; DL, deep learning; DNA, deoxyribonucleic acid; DNN, deep neural network; EPIs, enhancer–promoter interactions; GAN, generative adversarial network; HRMAS, high-resolution magic angle spinning; LSTM, long short-term memory; MLP, multilayer perceptron; NAS, network architecture search; NMR, nuclear magnetic resonance; RF, random forest; RNA, ribonucleic acid; RNN, recurrent neural network; SVM, support vector machine; VAE, Variational Auto-Encoder; WGAN-GP, Wasserstein GAN gradient penalty.

In the same direction, a new design termed missense variant pathogenicity prediction (MVP) diagnoses the diseases of missense variants with supervised practice. MVP used the ResNet, a deep residual neural network model introduced for computer vision, to enhance the performance of prioritizing pathogenic missense variants over former models (Qi et al., 2021).

An approach that identifies genetic variants associated with specific traits and diseases is genome-wide association studies (GWAS) (Uffelmann et al., 2021). GWAS investigate the genome of a huge crowd to assess millions of single nucleotide polymorphisms (SNPs). Demontis et al. (2019) employed GWAS for attention deficit hyperactivity disorder (ADHD) prediction. GWAS is a practical method, but its effectiveness is reduced by insignificant SNPs and is sensitive to significant p values (Sun et al., 2020). To this end, Liu et al. (2021) have introduced a CNN-based binary classifier to predict ADHD diseases based on SNPs. Additionally, researchers nominated a number of potential risk genes for ADHD. The authors used three SNP sets with different p values, and the final results indicated that a lower p value achieves the most accurate classification evaluation.

With the growth of massive genetic data, such as SPNs and high-quality time-to-event phenotypes, the applications of DL approaches in survival prediction need more attention. Due to this, a survival framework compares three prediction models applied to GWAS data. The authors compared the proposed DNN with random survival forest and least absolute shrinkage and selection operator (LASSO).

A major component of the disease is the disruption of transcription factor (TF) binding. TF is a set of proteins that control the transcription rate of genetic information by binding to a specific DNA sequence (Zeng et al., 2020b). In this regard, MTTFsite offers a practical framework for predicting transcription factor binding sites (TFBSs) without labeled data using a multitasking algorithm (Zhou et al., 2019). Considering that some cell types have more labeled data than others, this framework employed a shared-private model to leverage the labeled data available. A shared CNN learns common features of all cell types, and a private one learns private features of each cell type. This article also proposed a new gene expression prediction technique termed TFChrome that incorporates TFBSs diagnosed by MTTFsite with histone modification features. It can predict gene expression for labeled and unlabeled cell types. TFChrome outperformed the method employing only histone modification features.

Multi-Omics Integration Using DL

A complex disease can be caused by cascades of modifications at different stages of omics, such as gene expression dysregulation, DNA methylation, and proteomics. In this context, multi-omics integration integrated the multiple omics techniques into the biological aspects. High-throughput NGS technologies have made unprecedented progress in the growth of multi-omics data, which allow analyzing the molecular mechanisms in multiple diseases simultaneously (Nicora et al., 2020).

Chaudhary et al. (2018) proposed a DL model that successfully used ribonucleic acid (RNA) and miRNA sequencing data, methylation data, and clinical information to predict subgroups in liver cancer. The authors have claimed that this is the first attempt that uses DL methods for the multi-omics hepatocellular carcinoma (HCC) data set integration as the most common type of liver cancer. Researchers selected AE architecture to integrate heterogeneous transcriptomics and epigenomics data. The bottleneck layer in AE generates new features from the omics data. Besides, the supervised learning classifier, SVM, has been used to diagnose healthy and patient individuals.

Francescatto et al. (2018) combined integrative network fusion, a bioinformatics technique incorporating similarity network fusion (SNF), with ML. The proposed framework identifies integrated multi-omics biomarkers based on predictive profiling and considers a new strategy to integrate genomics and transcriptomics data. RF and linear support vector machine (LSVM) classifiers were applied at three levels to achieve final biomarkers. First, they trained on direct concatenation multi-omics data. Then, the model learned the high-ranked feature rendered by SNF. Eventually, it learned the intersection outcomes of two previous steps and ranked features again to integrate multi-omics biomarkers.

Integrated multi-omics studies can significantly assess the complex molecular underpinnings of diseases, improving the biological interpretability of results and characterizing disease mechanisms and pathways at a detailed molecular level. Park et al. (2020) designed a DL model that merges DNN and Bayesian models to predict Alzheimer's disease. The authors used heterogeneous omics data sets, large-scale gene expression, and DNA methylation data. Since both genetic and epigenetic alterations have an impact on gene expression regulation, Seal et al. (2020) have established a system based on multi-omics integration. The authors employed deep denoising autoencoder to extract features and MLP to estimate gene expression from genetic and epigenetics. The effectiveness of the DL method has been demonstrated in data from The Cancer Genome Atlas (TCGA) liver HCC study.

A research team reported multi-omics late integration (Sharifi-Noghabi et al., 2019), which benefits from the RBM, to integrate the epigenomics and transcriptomics data and predict drug response. This model includes three feed-forward encoding subnetworks to the number of omics data types, and each one takes their related data to encode them into the learned feature space. These subnetworks are developed by a fully connected layer that outputs the trained features for corresponding omics. Then, the framework using the late integration technique concatenated the learned features to demonstrate a multi-omics representation.

Another deep framework has suggested survival analysis learning with multi-omics neural networks, SALMON, that integrates gene expression data (mRNA and miRNA) and cancer biomarkers to explore breast cancer corresponding co-expression modules (Huang et al., 2019). Original data before inputting to the DNN converted to an eigengene matrix inferred from co-expression network analysis that significantly decreases the feature space dimension. The SALMON located Cox proportional hazard (Cox-PH) regression after neural networks layers. Thus, high-ranked features selected by neural networks, extra information such as age, tumor mutation burden, and copy number burden are sent to Cox-PH to estimate survival probabilities.

Omics-Based COVID-19 Diagnosis Using AI

The coronavirus is turning into an endemic, so understanding its genetic characteristics and developing accurate identification techniques are imperative. AI has afforded new insights into the evolution of COVID-19. Since multi-omics analysis provided a comprehensive perspective of COVID-19, a two-level disease prediction model was implemented. The framework initially diagnoses the patient or control individuals and then predicts the disease's severity. For negative and positive classification, several ML models such as SVM, RF, LR, KNN, RUS, and MLP were applied to single omics data type, DNA methylation, and RNA-Seq. Then, the severity prediction was conducted in two strategies. At first, they used single omics, including DNA methylation, proteomics data, metabolomics data, and the mentioned ML models in the article. For the second approach, they applied an AE neural network on multi-omics data containing transcriptomics, proteomics, metabolomics, and lipidomics. The authors proved that multi-omics analysis outperformed single-type omics data (Liu et al., 2023).

Randhawa et al. (2020a) identified an intrinsic genomic signature of the COVID-19 virus by a combined ML with digital signal processing (MLDSP) model, which was improved as MLDSP-GUI applying decision tree algorithm as a supervised approach, to classify a large data set of viral genomic sequences within a few minutes. Consequently, Spearman's rank correlation coefficient analysis was used for the final evaluation (Randhawa et al., 2020b).

Studies have indicated that the patient reaction to the coronavirus differs between individuals ranging from asymptomatic to critical. Additionally, some confusing factors such as age, sex, and comorbidities cause difficulty in identifying mediators of COVID-19. Moreover, molecular mechanisms of patients without comorbidities were unexplored. Therefore, researchers performed a range of multi-omics profiling on this cohort to detect a critical gene for the COVID-19 severity using an ensemble of ML, DL, quantum annealing, and structural causal modeling techniques (Carapito et al., 2022).

As genomic sequencing is a leading component in monitoring COVID-19, several studies report the usage of AI for genome sequence analysis. Ahmed and Jeon (2022) also aimed to comparatively study the basic patterns of the genome sequence of COVID-19 and other viruses, namely SARS, MERS, and Ebola. Further, they performed an AI-based model for prediction. Firstly, they analyzed DNA sequences and then, using visualization, discovered the nucleotide information or length of the genome sequence of four diseases. Finally, SVM classified COVID-19 genome sequences and other viruses.

Another hybrid algorithm incorporated CNN with bidirectional LSTM for two purposes. The first aim is to correctly classify SARS-CoV-2 sequences among coronaviruses. The second is to classify genome sequences to indicate candidate regulatory motifs. The model has been evaluated significantly using several comparison parameters, such as various hyperparameters, different neural network models, and data sets, where the proposed model outperformed (Whata and Chimedza, 2021).

Drug repositioning, also termed drug repurposing, provides new remedial prospects for known drugs and detects cures for untreated conditions (Jarada et al., 2020). This approach delivers a promising path for accelerating drug discovery and treatment strategies for COVID-19. To this aim, a network-based DL methodology named CoV-KGE (Zeng et al., 2020a) integrated Amazon's AWS supercomputing resources and a DL framework. In the proposed model, the knowledge graph embedding model (KGE) rotatE trains vectors for entities to infer various relations from a global network of biomedical relationships (Percha and Altman, 2018) and DrunBanks, a comprehensive and free online resource including information about drugs (Wishart et al., 2018). After thoroughly investigating drug connections, gene expressions, biological pathways, and proteins, the researchers identified repurposable drugs for COVID-19.

Alvarsson et al. (2016) developed a large-scale ligand-based virtual screening named ChemAI for drug discovery to discover new therapies for COVID-19. The authors applied the DNN model on the ZINC database, a collection of chemical compounds that aim to illustrate the biologically relevant, and DrugBank databases to suggest a screening library of potential inhibitors of molecules SARS-CoV-2. A similar study has been introduced by Karki et al. (2021), which employs ZINC and DrugBank databases. It was a rapid screening method based on neural networks named SSnet and merged with the Smina docking algorithm (Koes et al., 2013) to specify protein–ligand interaction. The outcomes of this method expand the libraries that identify probable high-affinity compounds and are possible clues for new drug discovery.

A group of researchers analyzed immune system protein interactome networks, scRNA sequencing, and AI networks to demonstrate potential therapeutic marks for drug repurposing against COVID-19. They assessed immune system proteins to discover probable therapeutic targets significantly overexpressed in individuals with several immunopathologies. Then, a fully connected DNN, optimized with grid search, performed the best classification to diagnose drug activity and discover several approved drugs with the highest characteristics (Lopez-Cortes et al., 2021).

Another leading field in COVID-19 is protein structure prediction since nonsynonymous mutations may change the function of the resulting protein (Tiwari and Mishra, 2020). In this regard, Senior et al. (2020) developed a model that trained a CNN named AlphaFold to accurately predict the difference between pairs of residues to reveal extra details about the structure compared with contact predictions. The same authors developed AlphaFold2, a modern evolutionary ML approach quite different from AlphaFold. This bioinformatic tool indicates the 3D coordinates of whole atoms for a protein through the amino acids and multiple sequence alignments. The proposed network has extraordinary impacts on the biological ability for protein structure prediction, which extensively surpasses other methods (Jumper et al., 2021).

Challenges Faced by the Omics Data Analysis and DL

Even though omics technologies have collected massive amounts of biological data and have demonstrated many characteristics of primitive biological processes, analyzing these complex data remains a challenging task. A critical challenge is finding a suitable method that simultaneously investigates and extracts biological interactions among data (Grapov et al., 2018).

Omics data

According to some natural characteristics of omics data, their analysis and interpretation have faced some problems in extracting meaningful insights, which are listed below.

Curse of dimensionality

The mentioned problem occurred when the number of features was more extensive than the samples (Zhang et al., 2019). Thus, developing a model to discover significant patterns is challenging due to an insufficient number of observations to train correctly. Various dimensional reduction methods exist to overcome the limitation, including feature selection, feature extraction, regularization (Kaur et al., 2021), and data augmentation (Wei et al., 2022).

Heterogeneous data

Such circumstances, populations, samples, or results are statistically different. Omics data are usually obtained from various samples and thus have different data types and variable formats (Mirza et al., 2019). For example, two persons with the same cancer may have different biomarkers that cause their cancer. Due to this, identifying and prioritizing cancer-related genes is challenging. In this regard, ML approaches such as clustering attempt to tackle the problem.

Missing value

Missing data problems occur when some variables (observations or genes value) are not stored and are lost for reasons such as lower sensitivity or noise. As collecting omics data is intrinsically a noisy technique, missing data problem is common (Momeni et al., 2020). Because missing data is a typical issue, varying solutions exist that categorize into two main methods, imputation and removing data (Mirza et al., 2019).

Deep learning

Although DL technologies have been applied successfully in a wide range of areas, several challenges exist, as mentioned subsequently.

Large volumes of data

Deep models require big data to avoid overfitting. If the data set is too small, deep models harm more than statistical ML algorithms. Like the human brain, the artificial neural network needs much data and experience to learn more precisely and infer knowledge well. For instance, the small number of samples in gene expression microarray data sets can cause problems in data analysis through deep models. An efficient solution for addressing this issue is generating new data in two manners. The first is employing GAN models to synthesize new fake data based on existing data distribution. The second solution is various slight changes in data, such as image rotation or flip (Albaradei et al., 2021).

Overfitting

When a model learns more details over a long time or the design is very complex, it can fit training data and the learned noises in omics data. Therefore, this model will not be able to predict new data, and its performance will decrease significantly (Pirmoradi et al., 2020). Data augmentation, dropout, feature selection, early stopping, and a well-designed model are used to address overfitting.

Hyperparameter tuning

Unlike the parameters that automatically determine during learning, hyperparameters should be predefined manually before starting the learning. Although all hyperparameters have no equal effects on the learning process, finding appropriate values for all of them is of paramount importance. Noteworthy, since they are sensitive values and small changes resulted in obtaining significant results, thus optimizing algorithms is highly recommended to set them accurately (Min et al., 2017).

The black box nature

Mathematical methods were unable to prove DL algorithms. DL models are usually based on a trial-and-error process, and the logic is hidden. Hence, basic principles are imperceptible even in optimal and best solutions (Kaur et al., 2021).

Computational complexity

Deep models are more complex than statistical ML and traditional models. As a result, they have a longer training time. The deep models have numerous parameters and hyperparameters for tuning, and adjusting them should be performed in complex processes. Thus, they require heavy mathematical computations and robust systems (Kaur et al., 2021).

Conclusions

AI provided opportunities for bioinformatics scientists to improve machines' abilities. In this context, DL is a promising approach for analyzing complex and heterogeneous data. Moreover, omics comprehensively describe biological molecules and have the benefit of applications. It can be applied to clarify the relationship between multi-omics data and COVID-19 consequences. Therefore, we focused on various ML-based techniques in the diagnosis of COVID-19 based on omics data. In the inclusive survey, first, widely used ML algorithms and DL architectures were discussed. Then, diverse leading DL techniques have been applied to omics data and multi-omics data, and COVID-19 analysis has been summarized. In the end, several potential challenges have been highlighted. The survey intends for the beginner and the expert users interested in incorporating AI with omics data and can expand computer researchers' insights into omics data analysis.

Future works are required to establish comprehensive ensemble models for multi-omics integration and employ diverse methods in combination and simultaneous manner to obtain better predictive performance. Since multi-omics data suffer from the curse of dimensionality, developing dimension reduction approaches can be considered promising future directions. For instance, hybrid feature selection as a novel technique can attempt to pursue the advantages of multiple models. Moreover, data augmentation, one of the hottest paradigms in AI, can control the shortage of sample amounts next to too large a feature space by generating realistic synthetic samples. The pathophysiology of coronavirus and its clinical phenotypes is still an open issue. Therefore, more emerging and robust computational models for investigating the intersection of COVID-19 and multi-omics data are essential to unravel this complicated network for future direction.

Footnotes

Acknowledgments

The authors thank the anonymous reviewers and the editor for their valuable and constructive suggestions to improve the article.

Authors' Contributions

B.J.: Conceptualization (lead), writing—original draft (lead), formal analysis (lead), writing—review and editing (equal). H.T.: Conceptualization (supporting), writing—review and editing (equal). A.R.: Methodology (lead), original draft (supporting), writing—review and editing (equal).

Author Disclosure Statement

The authors declare they have no conflicting financial interests.

Funding Information

No funding was received for this article.

Abbreviations Used

References

Abiodun

, Jantan

, Omolara

, et al. State-of-the-art in artificial neural network applications: A survey. Heliyon, 2018; 4(11):e00938; doi: 10.1016/j.heliyon.2018.e00938

Agarwal

, Chen

. DeepPHiC: Predicting promoter-centered chromatin interactions using a novel deep learning approach. Bioinformatics, 2023; 39(1):btac801; doi: 0.1093/bioinformatics/btac801

Ahmed

, Jeon

. Enabling artificial intelligence for genome sequence analysis of COVID-19 and alike viruses. Interdiscip Sci, 2022; 14(2):504–519; doi: 10.1007/s12539-021-00465-0

Albaradei

, Thafar

, Alsaedi

, et al. Machine learning and deep learning methods that use omics data for metastasis prediction. Computa Struct Biotechnol J, 2021; 19:5008–5018; doi: 10.1007/s12539-021-00465-0

Alqahtani

, Kavakli-Thorne

, Kumar

. Applications of Generative Adversarial Networks (Gans): An updated review. Arch Comput Methods Eng, 2021; 28:525–552; doi: 10.1007/s11831-019-09388-y

Alvarsson

, Lampa

, Schaal

, et al. Large-scale ligand-based predictive modelling using support vector machines. J Cheminform, 2016; 8:39; doi: 10.1186/s13321-016-0151-5

Aziz

, Verma

, Srivastava

. A novel approach for dimension reduction of microarray. Comput Biol Chem, 2017; 71:161–169; doi: 10.1016/j.compbiolchem.2017.10.009

Bahado-Singh

, Vishweswaraiah

, Aydas

, et al. Precision cardiovascular medicine: Artificial intelligence and epigenetics for the pathogenesis and prediction of coarctation in neonates. J Matern Fetal Neonatal Med, 2022; 35(3):457–464; doi: 10.1080/14767058.2020.1722995

Bowler

, Papoutsoglou

, Karanikas

, et al. A machine learning approach utilizing DNA methylation as an accurate classifier of COVID-19 disease severity. Sci Rep, 2022; 12(1):17480; doi: 10.1038/s41598-022-22201-4

10.

Cakmakci

, Karakaslar

, Ruhland

, et al. Machine learning assisted intraoperative assessment of brain tumor margins using HRMAS NMR spectroscopy. PLoS Comput Biol, 2020; 16(11):e1008184; doi: 10.1371/journal.pcbi.1008184

11.

Carapito

, Li

, Helms

, et al. Identification of driver genes for critical forms of COVID-19 in a deeply phenotyped young patient cohort. Sci Transl Med, 2022; 14(628):eabj7521; doi: 10.1126/scitranslmed.abj7521

12.

Chaudhary

, Poirion

, Lu

, et al. Deep learning–based multi-omics integration robustly predicts survival in liver cancerusing deep learning to predict liver cancer prognosis. Clin Cancer Res, 2018; 24(6):1248–1259; doi: 10.1158/1078-0432.CCR-17-0853

13.

Demontis

, Walters

, Martin

, et al. Discovery of the first genome-wide significant risk loci for attention deficit/hyperactivity disorder. Nat Genet, 2019; 51(1):63–75; doi: 10.1038/s41588-018-0269-7

14.

Dührkop

, Nothias

L-F

, Fleischauer

, et al. Systematic classification of unknown metabolites using high-resolution fragmentation mass spectra. Nat Biotechnol, 2021; 39(4):462–471; doi: 10.1038/s41587-020-0740-8

15.

Francescatto

, Chierici

, Rezvan Dezfooli

, et al. Multi-omics integration for neuroblastoma clinical endpoint prediction. Biol Direct, 2018; 13(1):1–12; doi: 10.1186/s13062-018-0207-8

16.

Goodfellow

, Pouget-Abadie

, Mirza

, et al. Generative adversarial networks. Commun ACM, 2020; 63(11):139–144; doi: 10.1145/3422622

17.

Grapov

, Fahrmann

, Wanichthanarak

, et al. Rise of deep learning for genomic, proteomic, and metabolomic data integration in precision medicine. OMICS, 2018; 22(10):630–636; doi: 10.1089/omi.2018.0097

18.

Han

, Zhu

, Ling

, et al. Gene-CWGAN: A data enhancement method for gene expression profile based on improved CWGAN-GP. Neural Comput Appl, 2022; 34(19):16325–16339; doi: 10.1007/s00521-022-07417-9

19.

Hazra

, Kim

M-R

, Byun

Y-C

. Generative adversarial networks for creating synthetic nucleic acid sequences of cat genome. Int J Mol Sci, 2022; 23(7):3701; doi: 10.3390/ijms23073701

20.

Hong

, Jiang

, Li

, et al. DeepHiC: A generative adversarial network for enhancing Hi-C data resolution. PLoS Comput Biol, 2020; 16(2):e1007287; doi: 10.1371/journal.pcbi.1007287

21.

Huang

, Chen

B-W

, et al. A semi-supervised privacy-preserving clustering algorithm for healthcare. Peer-to-Peer Netw Appl, 2016; 9(5):864–875; doi: 10.1007/s12083-015-0356-9

22.

Huang

, Dong

, Ji

, et al. Discovery of clinical pathway patterns from event logs using probabilistic topic models. J Biomed Inform, 2014; 47:39–57; doi: 0.1016/j.jbi.2013.09.003

23.

Huang

, Zhan

, Xiang

, et al. SALMON: Survival analysis learning with multi-omics neural networks on breast cancer. Front Genet, 2019; 10:166; doi: 10.3389/fgene.2019.00166

24.

Islam

, Huang

, Ajwad

, et al. An integrative deep learning framework for classifying molecular subtypes of breast cancer. Comput Struct Biotechnol J, 2020; 18:2185–2199; doi: 10.1016/j.csbj.2020.08.005

25.

Jarada

, Rokne

, Alhajj

. A review of computational drug repositioning: Strategies, approaches, opportunities, challenges, and directions. J Cheminform, 2020; 12(1):1–23; doi: 10.1186/s13321-020-00450-7

26.

Jumper

, Evans

, Pritzel

, et al. Highly accurate protein structure prediction with AlphaFold. Nature, 2021; 596(7873):583–589; doi: 10.1038/s41586-021-03819-2

27.

Karki

, Verma

, Trozzi

, et al. Predicting potential SARS-COV-2 drugs—In depth drug database screening using deep neural network framework SSnet, classical virtual screening and docking. Int J Mol Sci, 2021; 22(4):1573; doi: 10.3390/ijms22041573

28.

Kassahun

, Yu

, Tibebu

, et al. Surgical robotics beyond enhanced dexterity instrumentation: A survey of machine learning techniques and their role in intelligent and autonomous surgical actions. Int J Comput Assist Radiol Surg, 2016; 11:553–568; doi: 10.1007/s11548-015-1305-z

29.

Kaur

, Singh

, Chana

. Computational techniques and tools for omics data analysis: state-of-the-art, challenges, and future directions. Arch Comput Methods Eng, 2021; 28:4595–4631; doi: 10.1007/s11831-021-09547-0

30.

Koes

, Baumgartner

, Camacho

. Lessons learned in empirical scoring with Smina from the CSAR 2011 benchmarking exercise. J Chem Inf Model, 2013; 53(8):1893–1904; doi: 10.1021/ci300604z

31.

Levy

, Titus

, Petersen

, et al. MethylNet: An automated and modular deep learning approach for DNA methylation analysis. BMC Bioinformatics, 2020; 21(1):108; doi: 10.1186/s12859-020-3443-8

32.

Lin

, Zhang

, Cao

, et al. Classifying breast cancer subtypes using deep neural networks based on multi-omics data. Genes, 2020; 11(8):888; doi: 10.3390/genes11080888

33.

Liu

, Feng

, Li

, et al. Deep learning model reveals potential risk genes for ADHD, especially ephrin receptor gene EPHA5. Brief Bioinform, 2021; 22(6):bbab207; doi: 10.1093/bib/bbab207

34.

Liu

, Hasan

, Ahmed

, et al. Machine learning to analyse omic-data for COVID-19 diagnosis and prognosis. BMC Bioinformatics, 2023; 24(1):7; doi: 10.1186/s12859-022-05127-6

35.

Lopez-Cortes

, Guevara-Ramirez

, Kyriakidis

, et al. In silico analyses of immune system protein interactome network, single-cell RNA sequencing of human tissues, and artificial neural networks reveal potential therapeutic targets for drug repurposing against COVID-19. Front Pharmacol, 2021; 12:598925; doi: 10.3389/fphar.2021.598925

36.

Lopez-Garcia

, Jerez

, Franco

, et al. Transfer learning with convolutional neural networks for cancer survival prediction using gene-expression data. PLoS One, 2020; 15(3):e0230536; doi: 10.1371/journal.pone.0230536

37.

Mahmud

, Kaiser

, Hussain

, et al. Applications of deep learning and reinforcement learning to biological data. IEEE Trans Neural Netw Learn Syst, 2018; 29(6):2063–2079; doi: 10.1109/TNNLS.2018.2790388

38.

Majji

, Nalinipriya

, Vidyadhari

, et al. Jaya Ant Lion optimization-driven deep recurrent neural network for cancer classification using gene expression data. Med Biol Eng Comput, 2021; 59:1005–1021; doi: 10.1007/s11517-021-02350-w

39.

Marouf

, Machart

, Bansal

, et al. Realistic in silico generation and augmentation of single-cell RNA-Seq data using generative adversarial networks. Nat Commun, 2020; 11(1):166; doi: 10.1038/s41467-019-14018-z

40.

Meng

, Zeleznik

, Thallinger

, et al. Dimension reduction techniques for the integrative analysis of multi-omics data. Brief Bioinform, 2016; 17(4):628–641; doi: 10.1093/bib/bbv108

41.

Min

, Lee

, Yoon

. Deep learning in bioinformatics. Brief Bioinform, 2017; 18(5):851–869; doi: 10.1093/bib/bbw068

42.

Mirza

, Wang

, et al. Machine learning and integrative analysis of biomedical big data. Genes, 2019; 10(2):87; doi: 10.3390/genes10020087

43.

Moeinizade

, Hu

, Wang

. A reinforcement learning approach to resource allocation in genomic selection. Intell Syst Appl, 2022; 14:200076; doi: 10.1080/23307706.2022.2136273

44.

Momeni

, Hassanzadeh

, Abadeh

, et al. A survey on single and multi omics data mining methods in cancer data classification. J Biomed Inform, 2020; 107:103466; doi: 10.1016/j.jbi.2020.103466

45.

Nayak

, Naik

, Behera

. A comprehensive survey on support vector machine in data mining tasks: Applications & challenges. Int J Database Theory Appl, 2015; 8(1):169–186; doi: 10.14257/ijdta.2015.8.1.18

46.

Nicora

, Vitali

, Dagliati

, et al. Integrated multi-omics analyses in oncology: A Review of machine learning methods and tools. Front Oncol, 2020; 10:1030; doi: 10.3389/fonc.2020.01030

47.

Pal

, Aguan

, Sharma

, et al. Discovering biomarkers from gene expression data for predicting cancer subgroups using neural networks and relational fuzzy clustering. BMC Bioinformatics, 2007; 8:5; doi: 10.1186/1471-2105-8-5

48.

Parbat

, Chakraborty

. A Python based support vector regression model for prediction of COVID19 cases in India. Chaos Solitons Fractals, 2020; 138:109942; doi: 10.1016/j.chaos.2020.109942

49.

Park

, Ha

, Park

. Prediction of Alzheimer's disease based on deep neural network by integrating gene expression and DNA methylation dataset. Expert Syst Appl, 2020; 140:112873; doi: 10.1016/j.eswa.2019.112873

50.

Percha

, Altman

. A global network of biomedical relationships derived from text. Bioinformatics, 2018; 34(15):2614–2624; doi: 10.1093/bioinformatics/bty114

51.

Pirmoradi

, Teshnehlab

, Zarghami

, et al. A self-organizing deep auto-encoder approach for classification of complex diseases using SNP Genomics Data. Appl Soft Comput, 2020; 97:106718; doi: 10.1016/j.asoc.2020.106718

52.

Poplin

, Chang

P-C

, Alexander

, et al. A universal SNP and small-indel variant caller using deep neural networks. Nat Biotechnol, 2018; 36(10):983–987; doi: 10.1038/nbt.4235

53.

Pugliese

, Regondi

, Marini

. Machine learning-based approach: Global trends, research directions, and regulatory standpoints. Data Sci Manag, 2021; 4:19–29; doi: 10.1016/j.dsm.2021.12.002

54.

, Zhang

, Zhao

, et al. MVP predicts the pathogenicity of missense variants by deep learning. Nat Commun, 2021; 12(1):510; doi: 10.1038/s41467-020-20847-0

55.

Randhawa

, Hill

, Kari

. MLDSP-GUI: An alignment-free standalone tool with an interactive graphical user interface for DNA sequence comparison and analysis. Bioinformatics, 2020a;36(7):2258–2259; doi: 10.1093/bioinformatics/btz918

56.

Randhawa

, Soltysiak

, El Roz

, et al. Machine learning using intrinsic genomic signatures for rapid classification of novel pathogens: COVID-19 case study. PLoS One, 2020b;15(4):e0232391; doi: 10.1371/journal.pone.0232391

57.

Ravì

, Wong

, Deligianni

, et al. Deep learning for health informatics. IEEE J Biomed Health Inform, 2016; 21(1):4–21; doi: 10.1109/JBHI.2016.2636665

58.

Sarker

. Deep learning: A comprehensive overview on techniques, taxonomy, applications and research directions. SN Comput Sci, 2021; 2(6):420; doi: 10.1007/s42979-021-00815-1

59.

Schonlau

, Zou

. The random forest algorithm for statistical learning. Stata J, 2020; 20(1):3–29; doi: 10.1177/1536867X20909688

60.

Seal

, Das

, Goswami

, et al. Estimating gene expression from DNA methylation and copy number variation: A deep learning regression model for multi-omics integration. Genomics, 2020; 112(4):2833–2841; doi: 10.1016/j.ygeno.2020.03.021

61.

Senior

, Evans

, Jumper

, et al. Improved protein structure prediction using potentials from deep learning. Nature, 2020; 577(7792):706–710; doi: 10.1038/s41586-019-1923-7

62.

Sharifi-Noghabi

, Zolotareva

, Collins

, et al. MOLI: Multi-omics late integration with deep neural networks for drug response prediction. Bioinformatics, 2019; 35(14):i501–i509; doi: 10.1093/bioinformatics/btz318

63.

Shipe

, Deppen

, Farjah

, et al. Developing prediction models for clinical use using logistic regression: An overview. J Thorac Dis, 2019; 11(Suppl 4):S574; doi: 10.21037/jtd.2019.01.25

64.

Singh

, Yang

, Póczos

, et al. Predicting enhancer-promoter interaction from genomic sequence with deep neural networks. Quant Biol, 2019; 7:122–137; doi: 10.1007/s40484-019-0154-0

65.

Sohn

. Deep belief network based intrusion detection techniques: A survey. Expert Syst Appl, 2021; 167:114170; doi: 10.1016/j.eswa.2020.114170

66.

Soko

, Dlamini

, Ntsekhe

, et al. The COVID-19 pandemic and explaining outcomes in Africa: Could genomic variation add to the debate?. OMICS, 2022; 26(11):594–607; doi: 10.1089/omi.2022.0108

67.

, Xu

, Pathak

, et al. Deep learning in mental health outcome research: A scoping review. Transl Psychiatry, 2020; 10(1):116; doi: 10.1038/s41398-020-0780-3

68.

Sun

, Wei

, Chen

, et al. Genome-wide association study-based deep learning for survival prediction. Stat Med, 2020; 39(30):4605–4620; doi: 10.1002/sim.8743

69.

Suzumura

, Ogawa

, Sugiyama

, et al. Homotopy continuation approaches for robust SV classification and regression. Mach Learn, 2017; 106:1009–1038; doi: 10.1007/s10994-017-5627-7

70.

Sze

, Chen

Y-H

, Yang

T-J

, et al. Efficient processing of deep neural networks: A tutorial and survey. Proc IEEE, 2017; 105(12):2295–2329; doi: 10.1109/JPROC.2017.2761740

71.

Tian

, Zou

, Tang

, et al. MRCNN: A deep learning model for regression of genome-wide DNA methylation. BMC Genomics, 2019; 20(2):1–10; doi: 10.1186/s12864-019-5488-5

72.

Tiwari

, Mishra

. Investigating the genomic landscape of novel coronavirus (2019-NCoV) to identify non-synonymous mutations for use in diagnosis and drug design. J Clin Virol, 2020; 128:104441; doi: 10.1016/j.jcv.2020.104441

73.

Tiwary

, Levy

, Gutenbrunner

, et al. High-quality MS/MS spectrum prediction for data-dependent and data-independent acquisition data analysis. Nat Methods, 2019; 16(6):519–525; doi: 10.1038/s41592-019-0427-6

74.

Uffelmann

, Huang

, Munung

, et al. Genome-wide association studies. Nat Rev Methods Primers, 2021; 1(1):59; doi: 10.1038/s43586-021-00056-9

75.

Wei

, Li

, Huang

, et al. Cancer classification with data augmentation based on generative adversarial networks. Front Comput Sci, 2022; 16:1–11; doi: 10.1007/s11704-020-0025-x

76.

Wen

, Li

, Zhang

, et al. Cancer neoantigen prioritization through sensitive and reliable proteogenomics analysis. Nat Commun, 2020; 11(1):1759; doi: 10.1038/s41467-020-15456-w

77.

Whata

, Chimedza

. Deep learning for SARS COV-2 genome sequences. IEEE Access, 2021; 9:59597–59611; doi: 10.1109/ACCESS.2021.3073728

78.

Wishart

, Feunang

, Guo

, et al. DrugBank 5.0: A major update to the DrugBank Database for 2018. Nucleic Acids Res, 2018; 46(D1):D1074–D1082; doi: 10.1093/nar/gkx1037

79.

, Wu

, Chen

, et al. A hierarchical integration deep flexible neural forest framework for cancer subtype classification by integrating multi-omics data. BMC Bioinformatics, 2019; 20(1):527; doi: 10.1186/s12859-019-3116-7

80.

Yang

, Chen

, Li

, et al. Subtype-GAN: A deep learning approach for integrative cancer subtyping of multi-omics data. Bioinformatics, 2021; 37(16):2231–2237; doi: 10.1093/bioinformatics/btab109

81.

, Kim

, Mechefske

. Remaining useful life estimation using a bidirectional recurrent neural network based autoencoder scheme. Mech Syst Signal Process, 2019; 129:764–780; doi: 10.1016/j.ymssp.2019.05.005

82.

, Lan

, Liu

, et al. Progressive Ensemble Kernel-based broad learning system for noisy data classification. IEEE Trans Cybern, 2021; 52(9):9656–9669; doi: 10.1109/TCYB.2021.3064821

83.

Zahin

, Tan

, Hu

. Sensor-based human activity recognition for smart healthcare: A semi-supervised machine learning. Artif Intell Commun Netw, 2019; 2019:450–472; doi: 10.1007/978-3-030-22971-9_39

84.

Zeng

, Song

, Ma

, et al. Repurpose open data to discover therapeutics for COVID-19 using deep learning. J Proteome Res, 2020a;19(11):4624–4636; doi: 10.1021/acs.jproteome.0c00316

85.

Zeng

, Gong

, Lin

, et al. A review about transcription factor binding sites prediction based on deep learning. IEEE Access, 2020b;8:219256–219274; doi: 10.1109/ACCESS.2020.3042903

86.

Zhang

, Lv

, Jin

, et al. Deep learning-based multi-omics data integration reveals two prognostic subtypes in high-risk neuroblastoma. Front Genet, 2018; 9:477; doi: 10.3389/fgene.2018.00477

87.

Zhang

, Zhao

, Liao

, et al. Deep learning in omics: A survey and guideline. Brief Funct Genomics, 2019; 18(1):41–57; doi: 10.1093/bfgp/ely030

88.

Zhou

, Lu

, Gui

, et al. MTTFsite: Cross-cell type TF binding site prediction by using multi-task learning. Bioinformatics, 2019; 35(24):5067–5077; doi: 10.1093/bioinformatics/btz451