To understand the roles of ncRNAs to diseases, determining their interactions is the key. Several disease related ncRNAs databases have been developed. The databases are composed of a collection and integration of resources focusing on circRNAs as disease biomarkers, and lncRNAs/mRNAs/miRNAs interactions with diseases. The ncRNA related databases, their description and URL links are listed in Table 1.

Multi-omics data gives multiple views of a problem that are aggregated into context (Maghsoudi et al., 2022). Performing representation learning to explore information from multiple views is a challenging problem. The machine and deep learning models rely on feature information extracted from unlabeled data. Several tools have been proposed to facilitate interpretation of molecular features derived from multi-omics datasets that contain the biology of diseases. The tools incorporate concepts such as loss functions meant for obtaining puissant feature learning and prediction ability. For instance, a meta-learning deep learning method termed DeepLIFT was recently proposed that implements cox hazard loss to improve performance, intelligibility and interpretability of the model (Cho et al., 2023). Meta-learning, a learning-to-learn method, based on back propagation and cox hazard loss trained on transcriptomics, proteomics, and clinical datasets showed better performance than direct and transfer learning-based models.

Contrastive learning via data augmentation and other strategies and frameworks have attracted attention vastly in multi-omics data analysis. Contrastive learning is a self-supervised instance-level discriminative method that learns latent information with the aim of pulling the different views closer (Cui et al., 2023). Adapting algorithms that incorporate contrastive loss to evaluate cross-modal associations between datasets is paramount in understanding genetic modalities in the multi-omics data. Additionally, deep neural network algorithms improve the learned representations through contrastive masking to model nonlinear relationships. Many researchers have been contrasting high-dimensional features in the attempt to explore discriminative information from multiple views. However, low-dimensional representation is ignored albeit the significance of the information in the learned representation on downstream tasks. Li et al. proposed a method, regularized and hybrid Multiview coding (RHMC) a variant of contrastive learning method for comprehensive modeling of consistent information between multiple views (Li X et al., 2022). The method obtrudes global alignment on the learned representations in the latent space by computing the probability distribution of the views using the Wasserstein distance-based view alignment regularization. Therefore, contrastive learning learns high-quality representations and effectively supports multi-omics integration (Yang et al., 2022).

Comprehensive understanding of human complex diseases requires the analysis of multi-omics data reciprocally with clinical information. Based on the analysis, useful insights into the cellular functions are derived. The integration helps to understand the interplay of the biomolecules and in the assessment of the information from the omics data. Further, it improves prognostic and predictive accuracies hence better treatment and prevention of the diseases (Subramanian, et al., 2020). The gap between genotype and phenotype is traversed through the flow of information from one omics level to the other. However, challenges associated with data integration arise from complexity of the data and difficulty in the interpretation of the analysis results. There are platforms such as GraphOmics for exploring and integrating multiple omics datasets and also used for hypothesis generation (Wandy and Daly, 2021). The goal of the platforms is to uncover associations between unknown entities not captured in the knowledge base and methods such as correlation analysis and other analysis methods. A hybrid multi-omics network from longitudinal multi-omics data was proposed to facilitate the interpretation of the data. The method provides interpretation guidelines to explore network generated from multi-omics data to highlight inter and intra omics mechanisms and interactions (Bodein et al., 2021). Figure 1 below is an illustration of the data multimodalities and how the machine learning prediction models are used to predict diseases.

Disease diagnosis is primarily based on the patient history and physical examination done by health experts. It is often a difficult task due to the complexity of the disease mechanism and ambiguity of the symptoms that require proper diagnostic procedures. Moreover, a series of medical tests are needed which are expensive and because of human error patients may be misdiagnosed. Therefore, artificial intelligence develops algorithms and techniques for correct disease diagnosis, prognosis, and treatment. Machine learning (ML) is a subset of artificial intelligence that is based on mathematical and statistical approaches. The ML based diagnosis methods are developed based on healthcare data such as X-ray, MRI and tabular data with patients’ conditions, age, gender, body mass index (BMI) and blood pressure (BP). Other features used for prediction include genetic features and interaction network-based features. The features are fed into the algorithm which is able to generalize the knowledge based on a pattern observed from a group of samples. Generally, the immense growth of ML over the years is attributed to the advancement of technology and availability of data generated by academics and practitioners (Ahsan et al., 2022).

Deep learning models are effective for decoding pathological images, interaction and prognosis prediction. Several machine and deep learning algorithms have been used for breast cancer detection namely support vector machine (SVM), CNN, bidirectional RNN (BiRNN) and naïve bayes (Vaka et al., 2020). It has been proven that the expression levels of ncRNAs are altered in cancer cells or tumor tissues. Therefore, research on the expression of ncRNAs under pathological conditions is valuable for identification of novel biomarkers and target therapeutics. A biomarker is a molecule that is relatively easy to detect and offers credible information on diagnosis, prognosis and other disease parameters (Volovat et al., 2020). The implementation of the models is done through platforms such as Tensorflow, PyTorch, and Caffe (Shoaib et al., 2023). Tensorflow and PyTorch are open-source libraries for computation via creation of dataflow graphs and distributed training respectively. Caffe is an open-source deep learning framework for video and image classification.

Yu Z et al. (2022) published a review paper on popular deep learning algorithms for disease prediction. One of the popularly used deep learning model is CNN which is mostly suitable for learning image features. The model combines local receptive field, shared weight and down sampling. Receptive field in the convolutional kernel extracts visual features such as edges and corners. Shared weight feature is realized through scanning of the images by the convolution kernel using the same weight. One limitation of CNN is large amount of labeled data required to train the model. In a study conducted by Khan et al. deep CNN model was trained and used to classify normal and abnormal breast tissue (Heenaye-Mamode Khan et al., 2021). Digital images and health records of women were used to train and test the proposed model. The model predicted biopsy malignancy and differentiated normal from abnormal screening examinations. The features extracted include binding site, morphological and genetic features. Other deep learning models include LSTM for learning sequence-based features, stacked autoencoder for dimensionality reduction and classification and deep belief network. Factorization machine deep learning (FMDNN) was proposed to solve the problems of DNN (Yu et al., 2021). This model learns low and high order feature interactions. Factorization machine (FM) eliminates pre-training and facilitates an end-to-end training of the neural network. An FM based neural network termed DeepFM was proposed to predict the presence or absence of hepatitis (Yu et al., 2021).

Graph neural networks (GNNs) operate on graph-structured data and have successfully been used in network biology applications. Hernández-Lorenzo et al. (2022) proposed an Alzheimer’s disease prediction model through Graph Neural Networks. The study presented a genotype-to-phenotype prediction pipeline that uses GNNs in combination with protein-protein Interaction (PPI) and functional biological networks. Huang and Chung (2022) proposed edge-variational graph convolutional networks (EV-GCN) for prediction of Autism Spectrum Disorder, Alzheimer’s disease and ocular disease. Further, Monte-Carlo edge dropout uncertainty estimation was implemented to estimate the predictive uncertainty related to the constructed graph. EV-GCN, a population-based disease analysis method uses multi-modal medical data to evaluate the proposed method. In another recent study, a weighted-link GNN algorithm that combined graph auto-encoder and graph convolutional network was put forward (Cheng et al., 2023). The algorithm produced the best classification performance in the lung cancer knowledge classification compared to other state-of-the-art methods. From the highlighted GNN based methods, it is observed that the algorithms extract meaningful features which enable them to achieve superior performance. The methods are knowledge guided such that they inject knowledge from a graph structure medical ontology into deep models via attention mechanisms. Despite the advantages of GNNs, their limitations include scalability such that it is difficult to scale the edges of graphs based on the type and relations. Table 2 consists of recently proposed prediction models, the datasets used and the task performed by the models.

The association between multi-omics data, and diseases has been extensively explored by researchers. It has been established that multi-omics data can effectively predict the diagnosis, prognosis, and treatment of diseases (Pan et al., 2022). The application of computational biology tools to integrate omics data to investigate disease pathogenesis is known as network medicine. The analytical methods in network medicine based on molecular networks exploit protein-protein interaction, correlation-based networks, gene regulatory networks and Bayesian networks (Silverman et al., 2020). The network-based approaches exploit graph-theoretic (random walks, network propagation and path search), machine and deep learning. Analyzing topology of the network nodes uncovers specificities and similarities in how genes play regulatory roles and draws insights on diseases similarities.

Disease associated genes prediction can be accomplished through graph-theoretic algorithms, machine learning algorithms and their integration. This task is based on the assumption that diseases linked to the same genes are closely located in a molecular network such as PPI, co-expression networks and gene regulatory networks. PPI is the regularly used network among the three, this is attributed to the factuality that interacting proteins perform common biological functions. While gene expression refers to the process of converting genetic information into functional RNA or protein, gene regulation is the process of controlling the expression of genes. In a gene regulatory network, the edges represent not only interaction but also other biological processes such as reaction, activation or inhibition. A study by Hasankhani et al. (2021) revealed that through the integration of co-expression networks based on the hub genes and PPI networks, key hub-high traffic genes were identified as potential therapeutic targets for COVID-19 pandemic. In 2022, a study was done to predict gastric cancer diagnosis, prognosis, and drug repurposing based on gene expression signatures. The study used gene expression datasets to predict novel diagnostic candidates. Recently, co-expression network analysis of down syndrome was conducted to explore cell types associated with abnormal brain development (Seol et al., 2023). Through cell-type enrichment analysis on gene expression modules, gene modules associated with specific brain types were identified and functional annotation provided insights into the role of specific cell types in biological processes.

Wu et al. (2013) proposed qNABpredict, a taxonomy-agnostic model that predicts content of the nucleic acid-binding residues. The tool is designed to predict details of protein-NA (nucleic acid) interactions for large protein families and proteomes. Interactions between proteins and nucleic acids from protein sequences are critical in a wide range of cellular functions such as gene expression and regulation. Discovery of biomarkers through differential expression and molecular associations is a focal point of research. Technological advancements in molecular analysis have enabled identification of a large number of candidate biomarkers for complex diseases. Biomarkers can be used to determine disease stage in disease diagnosis. Additionally, they are used to assess the efficacy and monitor the response to new drugs or therapeutic intervention.

The profound advancements in technology assist in the development of decision support systems that provide accurate and reliable evidence-based solutions in different domains such as finance and medicine. In this section, we investigate prediction methods based on different types of diseases including breast cancer, brain cancer and hybrid disease detection. We examine emerging technologies such as blockchain, internet of things (IoTs), their evolution and integration with deep learning. Blockchain is a technology designed to offer high-level security, transparency and tamper proof data management for applications. It uses cryptographic signature that links blocks in the chain and generate unbroken chain of records. On the other hand, IoT based platforms are based on intelligent hardware, deep learning and mobile terminals to develop applications. By integrating blockchain technology and machine/deep learning, developed applications are able to extract valuable insights from data while preserving privacy. Figure 2 below shows the data sets, integration of deep learning with emerging technologies such as blockchain and IoT and their application in computational biology.

Breast cancer is a major cause of mortality worldwide. Studies have shown that it emerges from abnormally replicated breast cells. Detection methods include mammography, CT, MRI, ultrasound and biopsy. Machine and deep learning models have been proposed to aid in the detection of the malignant breast cancer such as inflammatory, invasive among others. Internet medical of things (IoMT) is a recently proposed method for detection and management of breast cancer. The method was implemented through gated recurrent units (GRU) a recurrent neural network model (Aldhyani et al., 2023). The method employs blockchain technology using advanced encryption standard (AES) cryptosystem. A cloud health resource-sharing model based on consensus blockchain technology is a platform developed to perform breast tumor diagnosis (Zhu et al., 2019).

Imaging technology provides an interior anatomy of patients that assist in the detection of abnormal tissues. Brain tumor, causes impairment and death in both men and women. Diagnosis of brain cancer is usually based on studying MRI scans which is laborious, error-prone and time-consuming. Therefore, deep learning models in conjunction with IoT and blockchain technology offer fast, secure, and precise prediction mechanism. Several authors have proposed classifiers that utilize the two technologies. An adaptive neuro-fuzzy system classifier for detecting brain tumor implemented with IoT through simulations was developed by (Sandya et al., 2023). A deep learning method for brain tumor detection based on blockchain technology was proposed to predict using MRI images (Mohammad et al., 2023).

Hybrid disease detection systems are models that diagnose multiple ailments. A blockchain-based multi-diagnosis deep learning model for disease classification was recently proposed to provide security of data shared in the healthcare sector (Rahal et al., 2023). The approach combines data from multiple sources for disease diagnosis by training and testing deep learning models on breast cancer, lung cancer and diabetes datasets. Another study based on federated learning, blockchain technology and deep learning models was proposed for classification of four respiratory diseases, COVID-19, Pneumonia, Tuberculosis, and Lung Opacity (Noman et al., 2023). The model is a web-based real-time classification tool.

This article is a qualitative research comprising of recent studies on the prediction of complex disease diagnosis, prognosis and treatment using multi-omics data and emerging technologies. To effectively measure the prediction models performance, cross validation system which circumvents cross-section prejudice is applied. Five-fold and ten-fold cross-validation techniques are most commonly used to assess classifiers performance. In the cross-validation techniques, the dataset is randomly fragmented into training and testing sets. To quantitatively appraise the efficacy of the classifiers, the evaluation metrics include accuracy, Mathew’s correlation coefficient (MCC), precision, recall, specificity and area under the curve (AUC). Accuracy measures the ratio of correct predictions over all the samples. Precision indicates the ratio of correctly predicted positive samples over all the predicted positive samples. Recall indicates the ratio of correctly predicted positive samples over all the positive samples.

The number of publications on the computational prediction of diseases has steadily increased over time. The authors main aim has been to improve the performance through different approached including dimensionality reduction and feature selection mechanisms. Figure 3 illustrates a summary of the number of publications between 2013 and 2023, the information is obtained from Scopus data. Table 3 represents the studies’ findings and the respective references. The accuracies obtained by the methods is notably high ranging between 76% and 99%. Particularly, the performance of DRAE (Menagadevi et al., 2023) recorded an accuracy of 98%. The high performance is attributed to the pre-processing techniques implemented, modified optimal curvelet thresholding and Octagon histogram equalization. The two techniques removed noise from the datasets hence Then residual autoencoder architecture is used for extracting features and SVM implemented for classification. Moreover, it can be observed in Table 3 that the deep learning models performance were positively influenced by the optimization and feature normalization mechanisms they implemented. A checkmark in the third column indicates whether the model was designed to address the specific challenge. In the fourth column, the checkmark indicates implementation of heterogeneous or hybrid machine learning for disease prediction.

Genome, transcriptome, proteome and other omics data collaborate to perform complex cellular processes. Some researchers have proven that multi-omics datasets help to unravel the molecular mechanisms. The cancer genome atlas (TCGA) data identify distinct molecular subtypes of cancer with the aim of improving diagnostic methods, treatment standards, and finally to prevent cancer (Tomczak, et al., 2015). The data combines DNA methylation, mRNA/microRNA expression and proteomics data (reverse-phase protein arrays). The analysis of the datasets output a comprehensive catalog of genetic and epigenetic drivers of cancer e.g., breast cancer subtypes (Koboldt, et al., 2012).

Aside from disease diagnosis, multi-omics data has been effectively used for precise cancer treatment. Chai et al. (2021) integrated multi-omics data for accurate cancer prognosis prediction. The method implemented denoising autoencoder for robust representation of multi-omics data to estimate patient’s risk of cancer through the Cox proportional hazard model. Another model named MSDLM was recently proposed to predict prognosis and therapy response in colorectal cancer (Foersch, et al., 2023). The model was developed based on cellular patterns of anti-tumor immunity and was determined to outperform clinical, molecular and immune cell-based parameters.

The biggest challenge in training models is data quality. Particularly in the application of deep learning models in disease related predictions, high quality medical data is of great significance. However, the quality of medical data is of low quality despite the availability of the data in large quantity. The problems include, the need for medical expertise to label samples, abnormal features and mismatch between the training and actual data samples. To circumvent the challenges of low amount and quality of image, speech and text data, techniques like up sampling, Fourier transform and augmentation are used to improve the quality. Moreover, attention mechanisms are implemented to improve representation ability and interpretability.

Feature selection entails the selection of the most relevant features during model construction. Some algorithms have built-in feature selection methods and penalization mechanisms for reducing overfitting while others rely on the features fed into them. An exhaustive search strategy is usually implemented on the feature subsets to identify the most optimal features suitable for the task. The search strategies include filter methods such as Chi-square and linear discriminant, wrapper methods such as genetic algorithms and embedded methods such as least absolute shrinkage and selection operator (LASSO) and ridge regression. The benefits of feature selection include reduced model training time, increased performance and Feature selection is also a means of achieving dimensionality reduction which has advantages such as complexity reduction. Models that rely on knowledge graph information with historical information and domain knowledge have been found to have high accuracy on diagnosis prediction tasks. To leverage the knowledge graphs, an end-to-end robust solution can be achieved through the features of GNN algorithm. Also, Huang et al. proposed combining graph structure learning and multi-task representation learning through multi-task node cluster to address the challenge of model interpretability (Huang et al., 2023).

The main challenge in disease diagnosis, prognosis, and treatment is medical uncertainty which affects both human (physicians) and deep learning models implementation in healthcare. This is a scenario whereby, there is no definite solution to a symptom presented by a patient for diagnosis. In this regard, mathematical and statistical models have been used to describe the mechanisms and dynamics of biological experimental findings and the degree of uncertainty quantified. Several deep learning models that quantify uncertainty in the classification results have been proposed including (Arco et al., 2023; Cifci 2023; Ren et al., 2023). Therefore, collaboration between computational biology experts who develop the prediction models with medical professionals to test the proposed models in real clinical scenarios is highly recommended. Thus, findings from the computational prediction models should be verified using wet-lab experiments and extensive pathway analysis. To overcome the shortcomings of the computational and biological experiments, firstly, the quantitative experimental methods can be used to measure dynamics of the diseases in vitro. Then, the wet-lab studies and data generated from them can be integrated with the mathematical modeling approaches for more accurate and context-based interpretation of the results.

Deep learning algorithms have been the most promising computational models for multi-omics data integration analysis. The models have achieved great success due to their superior feature representation capability and the end-to-end training paradigm. Generating correlations between the omics data is of utmost significant (Gong et al., 2023). Other challenges encountered in the algorithm implementation include overfitting, inequality, poor interpretability, privacy protection, and lack of reliable validation. To address model overfitting, feature selection and deep learning based multimodal feature fusion has successfully been implemented. Moreover, legal and ethical challenges in the implementation of deep learning models in healthcare is a major problem due to the use of personal information which can cause harm to the people involved through exposing them to discrimination etc. The different ways of ensuring protection of data are ethical guidelines, transparency and explainability, robustness, privacy, and accountability. Practically, blockchain technologies have been proposed to deal with the privacy protection as a means of boosting trust in technology. To solve reliable validation techniques challenge, methods such as cross-validation a resampling technique is used. To further enhance the performance of deep learning algorithms, deeper and wider neural networks with many layers and channels is crucial to cater for the depth of representation.

Attention is an algorithm that suppresses irrelevant information and accentuates relevant information. Adding attention mechanisms to neural network algorithms such as CNN enhances network performance and it is also used for dimensionality reduction. It is implemented during feature extraction process. Attention mechanisms dynamically assigns weights to the features to minimize the effect of less important features. The variants of attention mechanisms are self-attention (Arooj et al., 2022), graph attention (Wekesa et al., 2020a), coordinate attention (Xie C et al., 2022), dimensionality reduction attention (Wang and Wang, 2022), residual attention (Zhao et al., 2022), and spatial attention. Self-attention enhances information content by focusing on a single sequence to compute the sequence representation. Coordinate attention mechanism embeds position/location information in the feature map to enable the network to focus on important regions. Dimensionality reduction attention mechanism aims at limiting skewness error during feature extraction. Combining the attention mechanisms such as spatial and residual significantly improves the algorithm performance through broadening the representation power of the baseline network in a classification problem.

Researchers are developing learning paradigms such as meta-learning categorized as metric (similarity based on distance metrics), model (internal and external memory) and optimization (optimizing model parameters for fast learning). Meta-learning helps to solve data scarcity problems in disease diagnosis (Liu H et al., 2023). Multi-diagnosis methods are based on the three dimensions of meta-learning. Since deep learning algorithms are applicable to data-intensive learning tasks, integration of knowledge representation and reasoning in the development of complex systems is an area that is yet to be widely explored. Knowledge representation and reasoning is applicable in medical robotics and other domains. For instance, Ontology for Robotic Orthopedic Surgery (OROSU) is a robot used to perform surgical procedures (Gonçalves and Torres, 2015). Such systems require ethical procedures to be adhered to for their adoption to be acceptable in practice.