To date, genomics approach has highly sustained the finding and investigation of variations at both the germline and somatic levels thanks to many progresses in genome-exome sequencing techniques, for instance from the Sanger sequencing-based approaches to the NGS-based sequencing. Bioinformatics has always had a central role in the analysis of downstream genetic data. For example, in the multiscale scale project “The Cancer Genome Atlas” (TCGA), researchers used NGS sequencing associated to bioinformatics tools with the aim to discover somatic mutational landscape across thousands of tumor samples and to understand the complexity underlying different cancer types [56,57]. For the analysis of NGS data a sequence aligner tool is used on the sequence data (stored in FASTQ format). Some popular aligners are the stand-alone BWA [58], Bowtie [59], Bowtie2 [60], and SNAP [61], with aligned sequences being stored in SAM (Sequence Alignment Map, text-based) or BAM (Binary Alignment Map) files.

Epigenomics is concerned with the genome-wide identification of chemical modifications (i.e., methylation and acetylation of DNA) which are involved in regulatory mechanisms controlling gene expression and cellular phenotypes [62]. Chromatin immunoprecipitation (ChIP) assays-coupled NGS (ChIP-seq) and methylation analysis through whole-genome bisulfite sequencing (WGBS) or bisulfite sequencing (BSSeq) are the most widely used methods in epigenomics analysis [6]. By exploiting the advances in NGS field, it is now possible to analyze genome-wide methylome patterns at a single nucleotide resolution and to detect the methylated cytosine bases in genomic DNA. Data from array-based techniques can be analyzed using dedicated packages such as methylationArrayAnalysis [63], whereas for ChIP-seq data processing tools like SICER2 [64], PeakRanger [65], GEM [66], MUSIC [67], PePr [68], DFilter [69], and MACS [70] are used.

The detection and quantification of RNA transcripts (mRNA, noncoding RNA and microRNAs) is possible owing to the employment of several transcriptomics techniques. Differently from the static nature of genome, transcriptome dynamically changes as consequence of temporal cellular and extracellular stimuli. Microarray was the technique of choice to detect alterations in cellular mRNA levels in a high-throughput manner owing to its ability to quantify the relative abundance of mRNAs for thousands of genes at the same time. Microarrays are widely used to facilitate the identification of genes with differential expression between normal and cancer conditions. With the advent of NGS, the identification of the presence and the abundance of RNA transcripts in genome-wide manner became possible. In contrast to microarrays technique, RNA-seq does not depend on the transcript-specific probes and thus can effectively perform an unbiased detection of novel transcripts, also the less abundant, with high specificity and sensitivity. Starting points for RNA-seq bioinformatics analysis include alignment-based methods, such as Bowtie [59], and STAR [71], or alignment-free methods, such as kallisto [72] and Salmon [73]. Cancer-related omics experiments often rely on specific, tailor-made analytic pipeline. TCGA and other repositories give the great opportunity to analyze the omics data by a pan-cancer approach where different types of cancers can be compared in terms of genomic and transcriptomic landscapes [74].

Given the high complexity and dynamic range of proteins, their identification and quantification in large scale are significantly challenging. Proteomic analyses are applied to identify and quantify the set of proteins present within a biological system of interest. Progressions of the tandem mass-spectrometry (LC-MS/MS) techniques in terms of resolution, accuracy, quantitation, and data analysis have made it a solid instrument for both the identification and quantification of cells proteome [75]. Recently, the advent of cutting edge high-resolution “Orbitrap” mass-spectrometer instruments associated with powerful computational tools (i.e., MaxQuant [76] and Perseus [77]) simplified the genome-wide detection of all expressed proteins in human cells and tissues paving the way for a first draft of the human proteome [78,79]. MS-based proteomics techniques have been extensively applied also to investigate the proteome alteration in several human cancer tissues [80]. In particular, the study of cancer proteomes is a promising path for biomarkers and therapeutic targets identification because proteins are the molecular unit from which cellular structure and function arise [81].

The application of MS techniques is not restricted to proteomics but rather can be extended to smaller molecules such as metabolites. Metabolomics is characterized by the quantifications of metabolites that are synthesized as products of cellular metabolic activities, such as amino acids, fatty acids, carbohydrates, and lipids. Their levels can be dynamically altered in disease states reflecting aberrant metabolic functions in complex disorders like cancer. Indeed, metabolic variations are significant contributors to cancer development [82]. This is the reason why cancer metabolomics has become an important research topic in oncology [83], with the aim to get new insights on cancer progression and potential therapeutic targets. Lipidomics is a subset of metabolomics [84], specifically cancer lipidomics has recently led to the identification of novel biomarkers in cancer progression and diagnosis [85]. Metabolomics is still an ongoing field with the potential to be highly effective in the discovery of biomarkers, especially in cancer. This is possible due to the support of bioinformatics tools like metab package [86], which provides an analysis pipeline for metabolomics derived from gas chromatography-MS data, or metaRbolomics package [87], which is a general toolbox that goes from data processing to functional analysis. Similarly, the lipidr package [88] is an analogous framework focused on lipidomics data processing.

Selecting an appropriate dataset that allows for easy manipulation and data calculations could affect the performance of a computational model and reduce the main obstacles to multi-omics data analysis by improving data science applications of multiple omics datasets:

• The MultiAssayExperiment Bioconductor database [95] contains the information of different multi-omics experiments, linking features, patients, and experiments;
• The STATegRa dataset [96] has the advantage of allowing the sharing of design principles, increasing their interoperability;
• MOSim tool [97] provides methods for the generation of synthetic multi-omics datasets.

Integrating large amounts of heterogeneous data is currently one of the major challenges in systems biology, due to the increase in available data information [98]. The problem of missing and mislabeled samples, is a common problem in large-scale multi-omics studies [99]. It is common for datasets to have missing data related to some individuals. This often happens in clinical studies, where patients can forget to fill out a form. In other cases, it is possible that the acquisition of data reveals to be too expensive, need much time to be obtained or it is difficult to measure. Missing row values for a table are difficult to manage because most statistical methods cannot be applied directly to incomplete datasets. In recent years, several approaches have already been proposed to address missing row values [100]. The missRow package combines multiple imputation with multiple factor analysis to deal with missing data [99]. The omicsPrint method detects data linkage errors and family relations in large-scale multiple omics studies [101].

Understanding the nature of the data is a critical step in omics analysis [102]. For this purpose, it is possible to use exploratory data analysis (EDA) techniques which allow better assessments at a further modeling step. The main techniques for EDA include cluster analysis and dimension reduction, both widely applied to transcriptomics data analysis [103]. While cluster analysis consists of a set of methods for grouping objects into homogeneous classes, based on measures related to the similarity between the elements, dimension reduction is the process of reducing the number of variables, obtaining a set of variables called “principal.” Both cluster analysis [104] and dimension reduction [105] are applied to cancer studies, as shown in Table 2.

Together with dimensionality reduction and data clustering, data visualization is also an important part of EDA [2]. The combinations of these three factors make it possible to identify complex patterns, subpopulations within a dataset, and understand the variability within a phenomenon. Even if the scatter plot is the most common method for data visualization, there are other visualization tools available. Hexbins [111] can be used to explore sc-RNAseq data, while Circos diagram [112] can be used for the detailed representation of multi-omic data and their position in specific genomic regions.

Recently it is stated that mapping omics data to pathway networks could provide an opportunity to biologically contextualize the data. A network representation of multi-omics data can enhance every aspect of the multi-omics analysis because the functional level of biological description is fundamentally composed of molecular interactions [2]. The main tools for a network representation of multi-omics data are Pathview [113] and Graphite [114].

In recent years, machine learning has been proved to be capable of solving many biomedical problems. These mathematical models can represent the relationships between observed variables and provide a useful description of biological phenomena. A ML tool can perform several tasks, including classification task in which the input data are divided into two or more classes and the learning system produces a model capable of assigning one class among those available to each input. These models have important biomedical applications [94], because they are capable of discriminating between health and disease, or between different diseases outcomes [2]. In a regression task instead, the output belongs to a continuous rather than discrete domain. These models provide insights into the molecular mechanisms driving physiological states, reveal interactions between different omics, and have been used in prognostic tools [115]. In this context, due to the large amounts of heterogeneous data, the removal of non-informative characteristics which simplifies the model, increases its performance, and makes it less expensive to measure, reveals to be a crucial process [2]. Feature selection algorithm is a process which selects the variables that contribute most to the prediction, removing the irrelevant or less important features that can negatively contribute to the performance of the model. Both classification and regression ML techniques combined with feature selection algorithms have been widely used for cancer prognosis and prediction [2]. Moreover, many packages, which combine exploratory, supervised, and unsupervised tools, have been recently implemented in oncology. Table 3 provides a list of some of these new tools.

The interpretation of a ML model results could be a difficult task. A strategy that can provide readily interpretable results consist in mapping omic data on functional characteristics, in order to make them more informative and to associate them with a wider body of biomedical knowledge [2]. Some functional enrichment approaches are listed below:

• Over-Representation Analysis (ORA) [124];
• Gene-Set Enrichment Analysis (GSEA) [125];
• Multi-Omics Gene-Set Analysis (MOGSA) [126];
• Massive Integrative Gene Set Analysis (MIGSA) [127];
• Exploratory Data Analysis (PCA) [128];
• Divergence Analysis [129].

The first two enrichment approaches, ORA and GSEA, are feature extraction methods generally employed as dimensionality reduction methods. The output of these methods could be the starting points for more complex models such as interactions among functions. In particular, ORA method is based on a statistical evaluation of the fraction of pathway components found among a user-selected list of biological components. This input list fulfils the specific criteria (i.e., log fold change, statistical significance, and cutting-off the majority of components from the input list such as all the genes of a microarray experiment). GoMiner [130] is one of the most popular examples of ORA method. It was developed for gene-expression analysis of microarray data. It takes as input a set of over-/under-expressed genes plus the complete set list of the microarray, then it calculates over-/under-representation for Gene Ontology categories by means of Fisher’s exact test. Similarly, GSEA was developed for gene expression analysis from microarray data. The input is a list of ranked genes in accordance with their differential gene expression between two phenotypic classes. For each set of genes, an enrichment score (ES) is calculated based on a Kolmogorov–Smirnov pathway-level statistic. Multiple hypothesis testing is applied for the evaluation of ES significance. In the study of [131], the GSEA methodology was used to validate the proliferative role of growth-supporting genes involved in cancer treatment [132]. Multi-omics gene-set analysis (MOGSA) is an enrichment approach that uses multivariate analysis, which consists in integrating multiple experimental and molecular data types measured on the same data set. The method projects the features across multiple omics data sets to reduce dimensional spaces and calculates a gene set score with the most significant features. MOGSA’s multi-omics approach compensates for missing information in each single data type to find sets of genes not obtainable from the analysis of single omics data. A different approach is the massive integrative gene set analysis (MIGSA). It allows to compare large collections of datasets from different sources and create independent functional associations for each omic layer. The utility of MIGSA was demonstrated in [133] by applying the multi-omics perspective method to functionally characterize the molecular subtypes of breast cancer. There are enrichment approaches, such as pathwayPCA and divergence analysis methods, which use functional aggregation as support for other data analysis studies. In pathwayPCA, exploratory data analysis is performed using statistical methodologies to analyze the functional enrichment of each omics set and aggregating them via consensus. pathwayPCA overcomes alternative methods for identifying disease-associated pathways in integrative analysis. Among various case studies, the model was applied for the identification of sex-specific pathway effects in kidney cancer for the construction of integrative models for the prediction of the patient’s prognosis and for the study of heterogeneity in an ovarian cancer dataset. Divergence analysis method instead, is an enrichment approach that uses functional aggregation to classify large amounts of omics data. The omic profile is reduced to a digital representation based on that of a set of samples taken from a baseline population. The state of a subprofile that is not within the basic distribution is interpreted as “divergent.” In [134] an application of the divergence analysis within the study of metabolic differences among the interpersonal heterogeneous cancer phenotypes has been described.