High-dimensionality Data Analysis of Pharmacological Systems Associated with Complex Diseases

II. High-dimensionality Data in Pharmacological Paradigms

The routine use of H-D data in pharmacological settings to facilitate appropriate matching of drug mechanisms to disease signatures allows scientists to begin to effectively exploit the multidimensional nature of biologic factors (e.g., transcript, metabolite, and especially protein) activity. The complex and dynamic interactions between these multiple entities during therapeutic interventions further increases the complexity of drug efficacy profiles (Maudsley et al., 2015, 2016; Bradley and Tobin, 2016; Besserer-Offroy et al., 2017; Wooller et al., 2017).

Considering the importance of pharmacologically associated H-D data, it is important to define how we consider the description of a data set as possessing high-dimensionality. In the simplest of the terms, a data set (of any specific size) that possesses a large number of attributes or features is considered to possess a “high-dimensionality” nature. In our current context, the use of the word “dimension” indicates a “feature” of the data set that can be visualized or annotated, for example, a specific drug-regulated mRNA transcript or protein. The designation of high-dimensionality relates to the relationship between data point number (n) (e.g., a specific protein) and extractable dataset features (p). An H-D data set is defined as one that possesses more features (p) than data points (n), i.e., p > n. This simple dominance within the dataset of features over data points is independent of the size of either p or n. For example, a data set of only four data points that possesses six features each represents a small, but actionable, high-dimensional set.

In recent years the concept of inspecting drug signaling functionality (Appleton et al., 2013; Gesty-Palmer et al., 2013; Maudsley et al., 2015; Williams et al., 2016) as well as disease signatures (Mattison et al., 2014; Bakula et al., 2018; Melouane et al., 2018) has gained traction as a vital component of therapeutic intervention against aging-associated disease. With specific reference to the enormous diversity of individual transcripts and especially protein functionality at individual and multiprotein complex levels, it is evident how even a very small data set can be considered as high-dimensional, since a single protein can interact (in)directly with between 20 and 100 other proteins. These protein complexes then create essentially de novo signaling entities of incredibly diverse compositional nature (Martin et al., 2009; Maudsley et al., 2012; Westermarck et al., 2013; Alanis-Lobato et al., 2018). For seemingly simple data sets, such as differential drug effects on the magnitude of a single effector response (e.g., intracellular calcium mobilization), which may only involve one or two extracted features per data point, simple linear separation techniques can be effectively used to visualize the overall data set. These factor discrimination techniques are easily performed by humans within a scale setting of up to a three-dimensional feature space. However, once the number of dimensions increases, e.g., when the drug response or disease profile is measured using transcriptomic or proteomic profiling, it becomes more difficult to simply apply linear separation. To assist this, formalized data separation and/or machine-learning approaches have been developed to contend with higher dimensionality data. One of the earliest adopted mechanisms for this was principal component analysis (PCA). In a multidimensional data point space, PCA is employed to identify the orthogonal principal drivers, also known as eigenvectors, of the global dataset in a feature classification (e.g., control vs. test) task. From a machine-learning-based approach, linear separations of H-D data sets are now widely refined using algorithms such as support vector machines (SVMs) that aim to identify discriminatory hyperplanes between differential groups of data points. Such processes attempt to reduce the dimensionality of the data set, resulting in the definition of specific subgroups without destroying the “essential” discovery and discriminatory information in the data. In addition to data separation, impurity measures like entropy and information gain are also used for dimensionality reduction (Sakhanenko and Galas, 2015). In the following sections, we shall address the diverse methodologies that are currently being applied in pharmacological science to challenge the issue of complex drug response and disease-based data analysis.

III. Management of High-dimensionality Experimental Data

On a daily basis, terabytes of biomedical data are generated through the now widespread implementation of mRNA transcriptomics, RNA-sequencing, quantitative proteomics, and metabolomics. Each created data set may indicate the quantitative variation of between 5,000 and 10,000 experimental indices (transcripts, proteins, or metabolites), with perhaps 100s–1000s of these being statistically significant. In this context the ability of an individual scientist to appreciate the connectivity between these factors, which likely represents the true biomedical and pharmacological meaning of the data, is profoundly limited without the assistance of machine-based clustering and annotation (Chen et al., 2018; Dai et al., 2018; Lee et al., 2018; Lin et al., 2018; Lim and Xie, 2019). While the intrinsic depth of such data streams is a tremendous analytical advance for the study of complex drug activities, a major hurdle for the clinical translation of such data are the pace of advanced data management and investigational platform development. Not only is the collection and processing time for H-D data becoming problematic, it is often difficult to rapidly and efficiently generate a quantitative and therapeutically meaningful interrogation of such H-D data.

While computational methodologies for handling complex multivariate data have reached a highly advanced level in computer science, a similar level of advance in pharmacological science is still lacking. The future goal for this expansion phase will likely center up our ability to condense data vectors that exist beyond the realm of physical space (i.e., nth dimensional) into easily interpretable, aesthetic, and translationally relevant forms (Gong et al., 2018; Sharma and Rani, 2018; Vogt, 2018). In this regard, the creation of diverse informatic platforms is vital for future biomedical science and translational medicine, as these disciplines are associated with distinct and disparate forms of input data streams. Here we propose that an effective “complexity science” approach (Kenakin, 2017; van Gastel et al., 2018a) should be adopted to best appreciate aging-related disorders. Such an approach necessitates the generation of informatic systems that allow cross-platform correlation between H-D data sets of distinct types, e.g., transcriptomic, proteomic, and metabolomics (Topol, 2014; Gojobori et al., 2016; Satagopam et al., 2016; van Zimmeren et al., 2016).

H-D data analytics encompasses a broad computational space ranging from bottom-up dynamical systems modeling to top-down probabilistic causal approaches (Chang et al., 2015). Across many scientific fields, modeling and simulation have come to complement theory and experiment as a key component of the scientific method (Geerts et al., 2016; Xie et al., 2017). Therefore, a variety of methodological frameworks have been developed for modeling and analyzing complex multivariate data that can be adapted to the field of therapeutic gerontology. Here we will specifically focus on several new approaches, e.g., latent semantic analytics and topological data investigation, that have shown great promise for the study of aging-associated complex disease mechanisms.

IV. Classic Informatic Interrogation and Combinatorial Integration

The most common mechanism to interrogate H-D data involves literature searches from the primary data upon the highest and lowest regulated factors (gene/protein). This approach, while yielding actionable data, is often criticized for ignoring the correlated biologic relevance of the multiple factors in the rest of the data set that do not individually demonstrate either large or significant differential regulation between test and control samples (Mootha et al., 2003). If we consider the posit that complete knowledge of all gene or protein interactions currently does not exist (and is unlikely to be ever realized), it is prudent to assume that all gene and/or protein interactions could occur and may exert important pharmacological effects (Boyle et al., 2017; Wray et al., 2018). Hence at the present time, the relationships between a given factor (gene/protein) and the biologic/signaling property are considered to be “one-to-many.” Given our ever-expanding global data corpus, it is likely that in the near future we should consider this to be a “one-to-all” relationship, i.e., no drug-driven protein-protein interaction may be completely impossible or noncontributory to a highly complex pathologic mechanism. In this context, the scientist should only disregard analytical data for factors that one can empirically identify as “non-meaningful.”

Currently employed functional gene/protein clustering structures only reflect the status quo of pharmacological signaling knowledge and therefore they must be implemented as highly plastic entities. The next generation of drug discovery tools should be designed with a potentially nth degree level of dimensionality so that prejudicial elimination of combinatorial low-significance gene/protein effects does no longer occur. In this context of potential nth dimensional functionality of drug-driven gene/protein interactions coupled to the fact that genomes/proteomes can be simply spanned by just a small series of interacting connections, it is clear to see how considerable patient pharmacological variability could be engendered. The challenge for future H-D analytics is to rationally cluster these variable but specific molecular phenotypes into physiologic/pharmacological strata that are effective for both disease prognostication and/or remediation (Laifenfeld et al., 2012; McMahon et al., 2016; Balbas-Martinez et al., 2018). Considering that functional signaling cascades or disease states are likely the resultant composite of multiple, linked gene-protein networks, an effective gestalt appreciation of the entire data set is needed to draw accurate inferences at a biomedical and clinical level (Tabei et al., 2019).

Simple literature-based searches primarily enrich data output for biologic similarities between multiple components (genes/proteins) of an H-D data set. Biologic similarity, functional activity and predicted downstream functional sequelae can then be explored effectively with the use of informatic term annotation clustering. This process can be simply performed using a wide variety of “supervised” approaches, e.g., Gene Ontology (GO), signal transduction pathway analysis (KEGG pathways, Canonical Signaling Pathways (IPA), Pathway Commons, WikiPathways, REACTOME pathways, biophysical parameters (Pfam, PIR_Superfamilies, SMART), interactomic profiles (BioGRID, BIND, MINT, STRING), and a functional overlap with experimental molecular signatures (MSigDB-GSEA, L1000CDS2) (Maudsley et al., 2011). While individual applications of such platforms can still generate actionable data, a synergistic combination of such first-line tools underpins the future value of these canonical platforms (Chadwick et al., 2012a). Li et al. (2015a) recently demonstrated the utility of combinatorial informatics in an integrated genomics framework to elucidate pathologic AD mechanisms from large-scale H-D data sets. This research team extracted differentially expressed genes (DEGs) in published data sets comprising 450 late onset AD (LOAD) brains and 212 controls. Employing canonical signaling pathway analysis (IPA based), GSEA, and protein-protein interaction investigation via a protein-protein interaction network generated from the HPRD (Human Protein Reference Database: http://www.hprd.org/), the researchers created a large-scale LOAD-related data set of 3124 DEGs. Pathway analysis of this data identified several crucial LOAD-driving processes including nitric oxide and reactive oxygen species generation in macrophages, nuclear factor-κB modulation, and mitochondrial dysfunction. These predictive data outputs further align with our existing knowledge concerning the converging nature of the pathophysiologies between neurodegenerative, cardiovascular disease (nitric oxide related), and mitochondrially associated metabolic dysfunction (Ninomiya, 2014; Desikan et al., 2015). Demonstrating the efficacy of such combinatorial data extraction processes, discrete melatonin-associated signaling networks associated with cardiovascular disease were recently deconvoluted from publicly available H-D data using weighted gene coexpression network analysis coupled to differential gene expression analysis (Li et al., 2019).

Functional integration of standard informatics approaches using large-scale genomic data sets has also demonstrated its utility in identifying potential new remedial agents that reinforce the integrated nature of major age-related pathophysiologies. For example, the prioritization of the angiotensin receptor II antagonist candesartan as a potential novel antineurodegenerative therapeutic was demonstrated with the use of an employed combination of KEGG, IPA, GSEA, and Gene Expression Omnibus (GEO) data analytics (Elkahloun et al., 2016). This analysis was performed using data gathered from both wet and curated data sources, but was still able to identify a molecular therapeutic target that strongly reinforces the role of proaging cardiovascular mechanisms in neurodegenerative etiologies (Maudsley and Mattson, 2006). It is clear from these examples that while novel informatic pipelines can be beneficial to H-D data analysis, there is still a strong need for basic informatics analysis as this can often create a highly standardized form of data interpretation. This standard interpretation can then be used to integrate diverse H-D data streams for further synergistic exploratory analysis.

V. Metadata Extraction and Crowd-Sourced Curation and Investigation

The effective exploitation of existing, yet relatively cryptic H-D data repositories has been augmented through the use of intelligent, user-based functional clustering of publicly available metadata. H-D data depositories are now becoming commonplace, e.g., Gene Expression Omnibus (www.ncbi.nlm.nih.gov/geo/), Human Genome (www.genome.gov/) and Human Proteome (https://hupo.org/human-proteome-project) Organizations, Human Cell Atlas (https://www.humancellatlas.org/), Broad Institute Connectivity Map (cmap) (https://www.broadinstitute.org/connectivity-map-cmap), Allen Institute (https://alleninstitute.org/), interactomic data at BioGrid (https://thebiogrid.org), as well as the diverse forms of data associated with the LINCS program (http://www.lincsproject.org/). Many of these data repositories are now in the process of developing and refining advanced informatic platforms specifically tailored to assist biomedical scientists in accessing and efficiently analyzing the highly complex material contained within.

The Gene Expression Omnibus (GEO), based at the National Center for Biotechnology Information, represents a crucial data repository of validated transcriptomic data. In addition to standard search options available at GEO, the introduction of alternative search mechanisms including GEOBLAST and GEO2R have significantly improved advanced functional data retrieval (Barrett et al., 2013). In addition, multiple external GEO-using platforms are now commonly in use such as GEO2Enrichr, GEN3VA, and shinyGEO (Gundersen et al., 2015, 2016; Dumas et al., 2016). As the GEO database is constantly growing alternative data, clustering/structuring strategies have been developed to continually exploit and interrogate this rich resource. Demonstrating the utility of curated expression data-driven approaches, GEO, cmap, and the LINCS1000 signature database were explored to discover novel therapeutic repurposing functionalities for central nervous system antineoplastic interventions (Zador et al., 2018). Specific GEO data set extraction procedures also demonstrated the ability to facilitate novel receptor target identification as well. For example, the G protein-coupled receptor Gpr125 was identified as a potent determinant of the beneficial prognostic outcomes of patients presenting with colorectal cancer (Wu et al., 2018). In addition to the identification of antineoplastic interventions, large-scale computational biology initiatives using GEO interrogation have also shown promise in prioritizing novel GPCR-targeting therapeutic agents, e.g., glucagon-like peptide 2, relaxin 3, and follicle-stimulating hormone subunit beta, for the effective treatment of diabetes-associated retinopathy (Platania et al., 2018).

Perhaps the simplest but most effective method may be the use of crowd-sourced data analytics. This analysis strategy is built upon the intelligent user-based creation of biomedical domain-specific data clusters curated from GEO, e.g., Crowd Extracted Expression of Differential Signatures (CREEDS) (Wang et al., 2016). These advanced platforms for crowd-sourced H-D data analysis, i.e., CREEDS, GEN3VA, and others [Gene Wiki (Good et al., 2012); BioGPS (Wu et al., 2013)] have already demonstrated their efficacy with respect to assisting biomedical research in multiple aspects of neurodegenerative diseases (Fu and Fu, 2015; Allen et al., 2016), cardiovascular disease (Gottlieb et al., 2015; Scruggs et al., 2015; Rumsfeld et al., 2016), and metabolic dysfunction disorders such as type 1 and type 2 diabetes mellitus (Lau et al., 2016; Fadini et al., 2017; Mudie et al., 2017).

In addition to the employment of crowd-sourced informatics based on GEO expression data, this alternative form of data curation and analysis has recently been employed for drug-based investigations. Li et al. (2016) developed a crowd-sourcing workflow for extracting chemical-induced disease relations from publicly available texts from PubMed. Crowd-sourced–based analytics are also being employed for drug discovery investigations (Talikka et al., 2017). Using a combination of large-scale crowd-based interpretation with expert knowledge input will potentially expedite the ability and accuracy of extracting actionable inference from large H-D data sets. In this vein, Tang et al. (2018) recently developed the Drug Target Commons (DTC). The DTC possesses tools for crowd-sourced compound-target bioactivity data annotation, standardization, curation, and intraresource integration. The DTC platform was demonstrated to possess the capacity to significantly advance drug discovery and drug repurposing applications. In addition to the use of crowd-sourced data for repurposing, this novel human-machine workflow has also been effectively applied to natural product library screening to discover novel pharmacological agents (Tang et al., 2018). Kearney et al. (2018) recently developed the Canvass platform that uses human-curated public H-D data to assist in the computational evaluation of potential pharmacologically active naturally occurring agents. Using this database platform, this research group was able to specifically identify the activity of (−)-2(S)-cathafoline, which was found to stabilize calcium levels in the endoplasmic reticulum, both processes critical to the development of neurodegeneration and age-related DNA damage (Chan et al., 2002; Chadwick et al., 2010, 2012b; Zhou et al., 2014; Kearney et al., 2018). The workflow described here illustrates a pilot effort to broadly survey the biologic potential of natural products by utilizing the power of automation and high-throughput screening.

While predicting drug based signaling mechanisms is essential for therapeutic discovery, crowd-based informatic approaches have furthermore shown tremendous promise for revealing the potential safety issues of prescription drug combination adverse effects (Zhao et al., 2013). This research team demonstrated that implementation of a crowd-sourced informatic approach via drug response databases, such as FDA’s Adverse Event Reporting System, can help to identify drugs that could potentially be repurposed for mitigation of serious adverse events. Crowd-based H-D analytics have also shown utility with respect to assistance in the effective patient selection and ultimate generation of clinical drug trials (Leiter et al., 2014; Nayak et al., 2018) as well as the pharmacovigilant monitoring of postapproved drug effects on a population level (Tricco et al., 2017). Integration of both clinical trial and clinical use drug data using crowd-sourced H-D will also help facilitate the generation of effective “reverse drug translation” (Gibbs et al., 2018; Heatherington et al., 2018).

Quantitative proteomic data currently represents perhaps one of the most common and important forms of H-D data for both disease and drug therapy investigation (Yoshikawa and Kanazawa, 2013). An important facet of large proteomic data sets is the latent ability, through intelligent informatic interrogation of such data sets, to reveal the true ramifications of the pluridimensional signaling capacity of individual proteins through their myriad interactions. Crowd-based interpretation has been employed to refine signaling pathway analysis strategies (Martin et al., 2013b), generate protein-protein interaction data (Tastan et al., 2015), assist in the more efficient planning of initial experimental design (Barsnes and Martens, 2013), and interpret signal transduction cascades across diverse species (Bilal et al., 2015). Receptor-based signaling paradigms likely mediate their effects through a series of interactions between distinct multiprotein complexes. Thus, specific signaling cascades can be “encrypted” by the qualitative and quantitative stoichiometry of proteins that are formed into discrete interactomes (Martin et al., 2009). These functional protein complexes can then be interconnected and coherently regulate larger signaling networks (Martin et al., 2009). These specific signaling networks may then be simultaneously employed across different tissues, generating a coherent somatic signal transduction activity. Therefore, the acquisition and analogy of complex signaling data from multiple tissues simultaneously will likely create a more comprehensive view of systemic signaling paradigms. To this end, crowd-based platforms were recently developed to assist in network signaling prediction (Prill et al., 2011) and systems biology interpretation (Guryanova and Guryanova, 2017).

VI. Latent Semantic Analysis-Based Data Interpretation

Vast amounts of important and detailed information concerning individual age-related disease genes/proteins is incorporated in the text of published scientific literature. One of the best-curated examples perhaps is the Human Aging Genomic Resource [HAGR: http://genomics.senescence.info/ (Tacutu et al., 2018)]. The development of novel informatic platforms to interrogate efficiently such literature has been significantly augmented in recent years with the adoption of natural language processing (NLP) techniques (Jensen et al., 2006; Chen et al., 2013a; Plaza, 2014; Jimeno Yepes et al., 2015; Duque et al., 2018). An important aspect of large volume H-D data set interpretation is the transformation of a multidimensional corpus into an interpretable and manageable output. Bridging the gap between biomedical domain-specific datasets and naturalistic English language is crucial for a variety of applications linked to aging-related research, including the discovery of previously unknown biologic connections using the process of “Swanson Linking” (Swanson, 1988). NLP informatics also facilitates the identification of potential research topics that connect disparate tissues, visualization of biologic themes, and the effective discrimination between specific data sets and validation of existing data in a clinically actionable manner. To extract biomedically relevant information from diverse public literature sources, the reproducible generation of a mathematical association between scientific language and meaningful words or sentences is imperative. Analysis platforms for the interpretation of H-D data presently place a reliance on controlled-languages, i.e., Gene Ontology (GO), Medical Subject Headings (MeSH), BioCarta, or Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. Moreover, most platforms use standard Boolean and co-occurrence models to identify enriched categories (Ashburner et al., 2000; Coletti and Bleich, 2001; Kanehisa, 2002). To contend with these issues, algorithmic interpretation mechanisms for natural language sources have been developed. One such NLP data extraction process is latent semantic analysis (LSA). In a short period of time, this approach has now gained significant interest in the field of biomedical science (Klie et al., 2008; Xu et al., 2011; Roy et al., 2016). LSA employs singular value decomposition to elucidate patterns of correlation between information terms/concepts (i.e., text from PubMed Central Abstracts) and user-defined input interrogator items (words or Gene Symbols) within an unstructured collection of text (Han et al., 2011; Chen et al., 2013a). This investigation technique relies upon the elucidation of potentially meaningful association of words/items within a body of text and thereafter across multiple other texts. With the creation of a matrix of scientific (e.g., PubMed abstracts) “text-to-word” associations, it is possible to identify and quantitate even non-curated/experimentally derived links between input concepts and a specific gene or protein (Chadwick et al., 2010, 2011a; Xu et al., 2011). In recent years, further improvement of existing web-based informatic tools and novel platforms specialized for the machine-based analysis of natural language corpora have been developed, such as LigerCat (Sarkar et al., 2009), AmiGO (Carbon et al., 2009), Textrous! (Chen et al., 2013b), GeneIndexer (Cashion et al., 2013), Textpresso (Muller et al., 2004, 2008), and Genes2WordCloud (Baroukh et al., 2011).

VII. Electronic Medical Data File Analytics

Meaningful exploitation of well-curated empirical H-D data describing murine phenotype-genotype connectivity, e.g., from Jackson Laboratories Mouse Genome Informatics (http://www.informatics.jax.org/), has significantly supported the understanding of how simple genomic perturbations are physically connected to whole-organism phenotypes. With the future transition from paper-based patient medical data files to electronic medical data (EMD) files, it will be likely that a similar connectivity of biomedical/drug history data to complex human phenotypic analysis will be possible. EMD health records and their interrogation with machine-learning/deep learning platforms are currently being used to develop both augmented diagnostic aids as well as assisting in the discovery of novel “off-target” therapeutic effects or contraindications of prescription agents (Zhao et al., 2013; Chen et al., 2015).

However, the efficacy of these approaches is currently limited by data quality, quantity, and computational structure. Existing patient-based diagnostic aids commonly use machine-learning methods that employ data presented in a low-dimensional vector space. Current algorithms often exclude unstructured human-generated data, such as text, which can however contain critical predictive information (Perry et al., 2014). Clinical written data, generated from the patient-physician interaction is critical to assess the pathology and to recommend appropriate pharmacological interventions. An effective exploitation of this textual data is vital to improve the effectiveness of machine-learning-based diagnostic and potential drug discovery aids. Machine-based semantic examination of EMD data may facilitate novel clinical decision support (Prokosch and Ganslandt, 2009), conduct biomedical association studies (Tusch et al., 2000; Lyman et al., 2008; Melamed et al., 2014), and assess the cost effectiveness of pharmacotherapeutic treatments (Muranaga et al., 2007). While EMD analytics potentially represent a huge breakthrough in connecting medicinal compound effects to complex human phenotypic responses, there are significant future challenges to these goals. These potential pitfalls include a lack of common standards to merge clinical data and translate clinical concepts between disparate healthcare systems (Weiskopf and Weng, 2013). Furthermore, it will be critical to develop scalable methods to learn clinical concepts that can be translated across disparate healthcare systems. Extraction of actionable EMD insights has already been demonstrated for neurologic disorders, such as psychoses and severe mental illness (Kadra et al., 2015) as well as diabetic (McAdam-Marx et al., 2011) and cardiovascular (Tanaskovic et al., 2018) pathologies, all of which possess potential aging pathology causes (Kennedy et al., 2014; Franceschi et al., 2018; van Gastel et al., 2018a).

Along with these EMD analytics, automated domain-specific knowledge condensation for clinicians is currently a productive field of informatic development. The ever-increasing body of freely accessible biomedical data, while containing vital information, is becoming increasingly difficult to interpret for clinicians. To alleviate this issue, a platform for clinicians containing an automatic generation of AD-specific knowledge summaries has been developed, composed of relevant sentences extracted from Medline citations (Jonnalagadda et al., 2013). Their applied methodology combined information retrieval and semantic information extraction techniques to identify relevant sentences from Medline abstracts. In this given example of the Alzheimer’s domain, over 90% of the semantically retrieved sentences demonstrated a strong pertinence for clinical importance. Similar ventures have been made for cardiovascular disease knowledge (Torii et al., 2015; Sharafoddini et al., 2017; Hemingway et al., 2018). In this manner, over 2000 patient records were recently used from a large single-center pediatric cardiology practice to devise algorithms to predict if patients could be automatically diagnosed with cardiac disease (Perry et al., 2014). This research group then employed a supervised method using Laplacian Eigenmaps to enable existing machine-learning methods to estimate low-dimensional representations of textual data and at the same time accurate predictors based on these low-dimensional representations. This methodology allowed existing machine-learning predictors to effectively and efficiently capture the potential of textual predictors for cardiac disease, especially those based on short texts. Unsurprisingly, given the global incidence of glycometabolic disorders, the application of EMD analytics to diabetes diagnosis and treatment now represents a major computational tool in the endocrinological field (Chen et al., 2016; Zheng et al., 2016; Capobianco, 2017). The implementation of multiple forms of informatic interrogation (e.g., artificial neural networks, semantic analyses and machine learning) of EMD sources was shown recently to enhance phenotype description (Anderson et al., 2016; Gabert et al., 2016; Hall et al., 2018), disease trajectory progression (Jensen et al., 2014; Oh et al., 2016), diabetic comorbidities (Petrasek, 2008; Sancho-Mestre et al., 2016; Li et al., 2018), and eventual therapeutic efficacies (Ozery-Flato et al., 2016; Vashisht et al., 2016; Kang, 2018).

Given the expanding body of interrogable EMD files it is now becoming easier to assess postexposure drug response idiosyncracies in patients themselves. Given the considerable volume of typical EMD data, it is not surprising that machine-learning approaches have helped in this process. For example, machine-based analysis of depression treatment efficacy across patient EMD files at a meta-analysis level was recently used (Lee et al., 2018). EMD analysis has also recently been deployed to assist in the dose-response modeling of high-dimensionality drug interaction effects upon myopathies (Zhang et al., 2015). Intelligent EMD-based analytics have proven crucial for the creation of phenome-wide associations between pathologic gene variants and conditions such as obesity that can drive multiple age-related pathologies (Cronin et al., 2014). Furthermore, such EMD-based phenome-wide association networks have also proven effective for the development of precision medicine applications (Denny et al., 2016) as well as rational drug repurposing (Pulley et al., 2017).

As patient data gathered on a longitudinal basis provides an optimal internal data control it is not surprising that highly nuanced and actionable pharmacological data can be elucidated using EMD analytics. For example, the efficacy profiles of multitherapy data sets was recently simultaneously measured for a broad range of chronic age-related diseases (Khotimah et al., 2018). Perhaps one of the costliest age-related conditions, from a public health pharmacological intervention standpoint, is T2DM. While effective pharmacotherapies currently exist, their optimal implementation at a population level will be significantly improved with EMD analytical capacity. Advanced EMD analytical pipelines were implemented recently to monitor treatment quality measurements associated with treatment intensification (Arnold et al., 2018). In a similar vein, EMD analytic patient stratification, based on insulin efficacy measurements of glycated hemoglobin, has allowed the identification of distinctly responding subgroups of patients (Sidorenkov et al., 2018). In addition to monitoring insulin-based therapy profiles, EMD analysis of gestational diabetic states has also shown an ability to uncover the potential use of calcium channel blocking agents (nifedipine) as candidate repurposed agents extracted rationally from a broad panel of test agents for this poorly understood and important condition (Goldstein et al., 2018) that is actually linked to multiple aging-related conditions (Pinheiro et al., 2019).

VIII. Graph Theory Implementation for Signaling Network Deconvolution

Age-related disorders comprise a daunting number of molecular alterations due to a complex interplay between genetic, proteomic, and environmental factors. In a similar manner, the advent of commonplace transcriptomic/proteomic data analysis of drug signaling functions coupled to an appreciation of receptor signaling complexity (Martin et al., 2009; Maudsley et al., 2012, 2015, 2016; Luttrell et al., 2018) also significantly enhances the intricacy of drug signaling networks. Classic reductionist approaches to disease and its remediation previously have focused on a limited number of involved functional elements. This classic approach provides only a narrow overview of both the etiopathogenic complexity of multifactorial diseases (Hay et al., 2014; Topol, 2014; Boyle et al., 2017) and the subtlety of therapeutic signaling systems. This status has led to a severe hindrance for ultimate therapeutic development, e.g., in the realms of age-related central nervous system degenerative disorders (Alzheimer’s disease), CVDs, major psychiatric conditions, and cancer (Schadt et al., 2014; Kesselheim et al., 2015; Papassotiropoulos and de Quervain, 2015; Ramsay et al., 2018; Teneggi et al., 2018; van Gastel et al., 2019a). However, H-D data investigations into disease pathomechanisms/drug signaling allow for the simultaneous evaluation of multiple components of these biologic systems and their behaviors. While these data are vital to fully appreciate such intricate signaling systems, it represents a significant barrier for data presentation and connection-based investigation. To contend with this, the employment of network “graphs” has proven to be an effective mechanism for efficient data inference from H-D data sets (Cirillo et al., 2018; Hampel et al., 2018; Ning et al., 2018; Pirkle et al., 2018; Pita-Juarez et al., 2018). Network-based graphs are mathematical structures used to model pairwise/multiple relations between objects. Graphs in this context are made up of nodes, also called vertices, which are connected by edges. Graphs can be “undirected,” meaning that there is no distinction between the two vertices associated with each edge, or it may be “directed” from one node to another, i.e., a functional effect of one node on another can be shown.

Graph theory-based investigations have proven to be especially effective in enhancing the appreciation of the hypercomplex nature of both age-related dysglycemic (Barreda-Pérez et al., 2013; Khan et al., 2018) and cardiovascular (Huang et al., 2016) pathologies. Graph theory-based investigations are also especially pertinent to neurodegenerative diseases such as AD, as the connectivity between neuronal circuits in the central nervous system represents a biologic transposition of the “connectomic” relationships between graph/network components. Noninvasive magnetic resonance imaging (MRI) of brain functional connectivity has played a fundamental role in cognitive neuroscience. MRI-based imaging techniques possess the capacity to map neuronal activity/cerebral perfusion to the intricate connective structure of the brain. Independent component analysis, closely related to the classic PCA technique, allows for a network-based functional exploration of the brain when combined with applied graph theory (de Haan et al., 2009; Sanz-Arigita et al., 2010; Xie and He, 2012; Dipasquale and Cercignani, 2016). Graph-based analysis of CNS structure-function relationships has revealed multiple subtleties of the disruption of brain-wide communication issues in age-related neurodegenerative disease (Tijms et al., 2013; John et al., 2017). Recently, structural networks were constructed out of 87 brain regions, using data from 135 healthy elders and 100 early-onset AD patients selected from the Open Access Series of Imaging Studies (OASIS) data base (John et al., 2017), demonstrating the importance of access to shared H-D data. It is likely that neurodegenerative processes take place at different rates in different brain areas due to discontinuities in local age-related cellular damage events. Hence in this context, simply focusing analyses upon singular subnetworks may leave global interconnected molecular changes undetected. This graph-network–based investigation therefore suggests that neurodegenerative processes impact volumetric networks in a non-global fashion. Thus reinforcing the future importance of a more integrative-based approach to “connectomics” in diseased tissue analysis (Rubinov and Sporns, 2010; LaPlante et al., 2014). Resting state MRI has demonstrated that brain networks degrade during symptomatic AD. Recently, the graph theory metrics of functional integration (path length), functional segregation (clustering coefficient), and functional distinctness (modularity) were investigated as a function of AD severity (Brier et al., 2014). Clustering coefficient and modularity, but not path length, were reduced in AD. Cognitively normal participants who harbored AD biomarker pathology also showed reduced values in these graph measures, demonstrating brain changes similar to, but smaller than, symptomatic AD. It was furthermore demonstrated that AD has a particular effect on “hub”-like regions in the brain, underlining the ability of connectivity analysis via graph theory to unravel the complexity of disease mechanisms across the human brain (Brier et al., 2014).

While not considered a discrete “disease” process itself, the age-related presence of chronic inflammation (now codified as “inflammaging”) is a major driver of multiple severe pathologies and, in itself, causes significant collateral tissue damage both in the gut and lungs. The use of networks has already been greatly exploited to investigate the molecular interactome of irritable bowel disease (IBD) (Moco et al., 2014), which has been useful in the identification of specific master regulator proteins, termed “keystone” or “hub” proteins (Palmieri et al., 2016). Hence classification through drug indications (Suzuki et al., 2012) has suggested the presence of distinct disease states, i.e., mild and severe. In this manner, these two extracted data sets lie at the basis of a molecular fingerprint of IBD. Multiomics use involving genomics (Franke et al., 2010; D’Addabbo et al., 2011), transcriptomics (Fisher and Lin, 2015; Kalla et al., 2015), proteomic (Alex et al., 2009; Gazouli et al., 2013; Bennike et al., 2014), metabolomic (Bjerrum et al., 2015; De Preter, 2015), and epigenomic (Harris et al., 2012; Karatzas et al., 2014) variations together with environmental contributions, underpin an impaired immune system in patients with IBD, which is now one of the most investigated common complex human disorders. H-D data-based research into the molecular nature of this disorder, now described as the “IBD integrome,” is a strong example of the importance of “omic” integration between data platforms (Fiocchi, 2015; Palmieri et al., 2016).

The ability accurately to appreciate and predict a global somatic impact of pharmacological signaling will likely create a greater understanding of disease etiology and eventual network-level precision drug-based disease remediation (Prill et al., 2011; Hasan et al., 2012; Janssens et al., 2014; Muhammad et al., 2018). The appreciation of a network hypothesis for biologic activity presents many important new avenues for pharmacological research, especially in the aspect of prioritizing the most crucial factors/tissues within a hypercomplex signaling system. For example, the ability to identify molecular “keystone” or “hub” factors that exert the most profound actions upon the state of a given pathologic network may facilitate the creation of collateral efficacy pharmacological strategies (Gesty-Palmer and Luttrell, 2011b; Chadwick et al., 2012a; Maudsley et al., 2012). Combinatorial bioinformatic platforms are currently being developed to assist and expedite the discovery of novel network regulating factors that use advanced machine-based retrieval processes such as LSA-based GeneIndexer (Cashion et al., 2013) or Textrous! (Chen et al., 2013b). Such network-regulating factors may represent the most efficient means by which complex biologic processes can be controlled therapeutically. As a specific example, the G protein-coupled receptor (GPCR) kinase interacting protein 2 was identified as a molecular “hub” or “keystone” in the physiologic aging process using a coherent combination of standard data annotation (Gene Ontology/KEGG pathway analysis) followed by a successive application of “cross-pathway” gene/protein prioritization using the LSA-based GeneIndexer platform (Chadwick et al., 2012a). The potential keystone nature of kinase interacting protein 2 in the aging process has subsequently been reinforced by findings that this protein appears to regulate the intercommunication between metabolic function pathways and those responsible for protecting DNA integrity (Lu et al., 2015; Martin et al., 2016; Leysen et al., 2018). This level of pathologic connectivity is crucial to control the extent of age-related damage incurred during times of age-dependent stress in the brain and in peripheral metabolic tissues. Extending this aging network controlling locus to therapeutic exploitation, it was demonstrated recently that indeed LSA-based H-D interrogation can help elucidate and prioritize receptor-based targets for anti-aging therapies (van Gastel et al., 2016, 2018a,b, 2019b).

In addition to the application of data deconvolution to identify key factors within complex networks, graph-based pipelines have been used to define drug-signaling pathway association analytics. For example, Dai et al. (2018) defined a computational process, integrative graph regularized matrix factorization, to enhance the drug-induced signaling cascade classification and prioritization. Integrative graph regularized matrix factorization employs graph regularization to encode data geometrical information and prevent possible overfitting in the prediction of the association of specific therapeutic agents with the strongest associated signaling paradigm. Given the ability of drugs to affect multiple cellular transduction cascades, it is vital to demonstrate that graph/network dynamic analysis can actually generate therapeutically relevant information regarding in vivo drug activities. Recently de Anda-Jáuregui et al. (2019) demonstrated their ability to elucidate the nuanced effects of pioglitazone in the context of diabetic neuropathies using advanced pathway analytics. This research team generated the graph-theory–based pathway crosstalk perturbation network (PXPN) model. This model was created to aggregate H-D data using a pathway-based approach to associate molecular results with complex functional features related to the studied disorder/signaling system. Within a disease or drug response process, transduction pathways communicate with one another through the crosstalk phenomenon, forming large networks of interacting processes. With the PXPN model, changes in activity and communication between signaling pathways, observed in transitions between physiologic states, were represented as networks. Such graph-based models possess an agnostic nature with regard to the type of biologic data and pathway definition employed and can thus be implemented to analyze any type of high-throughput perturbation experiments. To demonstrate the efficacy of such an approach, de Anda-Jáuregui et al. (2019) analyzed the interactions between transcriptomic data from experiments in a BKS-db/db mouse model of T2DM-associated neuropathy and the effects of the thiazolidinedione class agent, pioglitazone, in this paradigm. Using their PXPN network approach, this group was able to identify changes in the specific connectivity of perturbed signaling pathways associated to each biologic transition, such as reorganizations between extracellular matrix, neuronal system, and GPCR signaling pathways (de Anda-Jáuregui et al., 2019).

Graph-based analytics can also be used effectively to coordinate meta-level HD data to define linkages between molecular drug signatures and patient-based sequence analysis. Recently Musa et al. (2017) outlined how advanced cmap (connectivity map: https://www.broadinstitute.org/connectivity-map-cmap) analytics, combined with an integration of signature data from LINCS (Library of Integrated Network-based Cellular Signatures: http://www.lincsproject.org/) could be aligned with the latest pharmacogenomics data. This approach will likely enhance and refine drug repurposing/discovery for a wide variety of aging-associated conditions (Musa et al., 2017). While graph-based analytic workflows are effective for molecular deconvolution of complex disease/drug-based data, this approach is also now proving its worth with respect to the prediction of adverse drug-drug interactions (Kamdar and Musen, 2017). To most effectively mine the current data concerning long-term patient drug-drug interactions integration and analysis of patient biomedical data as well as inferred knowledge from diverse and distinct sources with varying schemas, entity notations and formats is required. To contend with this, the Semantic Web community recently published and linked several datasets in the Life Sciences Linked Open Data (LSLOD: http://srvgal78.deri.ie/roadmapViz/) cloud using established W3C standards. While informative in themselves, such complex graph structures require a nuanced interrogation tool to most efficiently extract pertinent pharmacological data rapidly, hence Kamdar and Musen (2017) developed the PhLeGrA (Linked Graph Analytics in Pharmacology) platform. Using advanced query federation, this research team was able to effectively generate functional drug-reaction networks from integrated sources within the LSLOD cloud data base. This drug response graph could then be represented as a hidden conditional random field (HCRF), a discriminative latent variable model that is used for structured output predictions. Using this, Kamdar and Musen (2017) were able to calculate the underlying probability distributions in the drug-reaction hidden conditional random field, using the data sets from the FDA Adverse Event Reporting System. The PhLeGrA platform was also able to incorporate other sources published using Semantic Web technologies, also enabling it to facilitate the discovery other types of pharmacological associations.

IX. Data Visualization and Topological Methods

Given the prodigious size of many disease/drug response H-D data sets, the use of advanced mathematical modes of data management investigation is becoming increasingly commonplace. One such approach is the use of topological data analysis (TDA), which aims to interpret complex data through the identification/classification of its essential, scale-independent “shape.” TDA analyses data sets by the use of techniques drawn from classic surface topology investigations typically applied to geometrical morphology fields. H-D biomedical data sets present several difficulties with respect to TDA-based applications, i.e., they are often 1) generated using diverse experimental platforms, 2) relatively incomplete, and 3) “noisy.” Data “noise” has two main sources, implicit errors introduced by measurement variation, such as different types of apparatus and random errors introduced by batch processes or experts when the data are gathered. In the field of aging-related diseases such as AD, metabolic syndrome, and advanced CVDs, real patient data in high dimensions are nearly always sparse (incomplete as a translational set across diverse patients) but still tend to have relevant low dimensional features, i.e., accurate and predictive individual biometric data such as pulse-wave velocity or glycated hemoglobin levels. TDA provides a versatile framework to analyze such data in a manner that is insensitive to the particular metric and provides dimensional reduction while also being robust to “noise.” This analysis approach has combined algebraic topology and other pure mathematical tools to give a mathematically rigorous and quantitative study of data “shape.”

The primary effective facet of TDA-based structure classification is the elucidation of “persistent homology” patterns within the data. This process seeks to reveal consistently connected components of the data irrespective of any attached numerical scale. In general, the assumption is that features, e.g., drug-induced signaling protein expression, that persist for a wide range of scale parameters are “true” features. Features persisting for only a short period are presumed to be noise. Recently, a TDA-inspired platform [Plurigon (Martin et al., 2013a)] was developed to facilitate nth-dimensional analysis “data texturization,” i.e., creation of near-limitless levels of “feature” extraction from mass analytical data in a simple visual/textural format that is readily appreciable to users. This open access platform has already been used to identify and classify neurodegenerative aging mechanisms, experimental animal model phenotypes, and context-specific proneurotrophic drug efficacy signatures from complex H-D data (Martin et al., 2013a). Analysis of transcriptomic data using Plurigon revealed specific texturized data features that were responsible for the maintenance of the proneurotrophic/procognitive activity of amitriptyline (Chadwick et al., 2011b; Janssens et al., 2017) even in the presence of age-related cellular stressors. The future application of Plurigon data texturization may also help identify key molecular features in data sets that could reveal causative molecular interactions from quantitative plasma proteomic data that allow more accurate patient grouping for hypercomplex syndromic disorders such as AD (Fig. 2).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Plurigon-based mechanism for three-dimensional data “texturization.” (A) The input of a simple .txt file allows the creation of a highly nuanced and interrogatable three-dimensional structure overlaid onto a spherical base (1 and 2). The three-dimensional rendering can be freely rotated (3); increased or decreased in viewer screen (4); rendered in red/cyan three-dimensional viewing mode (5); coordinated with applied fixed axes (6); interrogated for center of mass (COM) calculation (7); interrogated for general structural features (8). (B) Implementation of Plurigon to detect gestalt signature distinction between quantitative differential expression plasma proteomic datasets from human patients either with clinically diagnosed Alzheimer’s disease (AD), with a genomically predicted low risk of generating AD (LR) or with a genomically predicted high risk of generating AD (HR) (unpublished data). Patients classified as either LR or HR were not diagnosed with AD. Differential plasma proteomic expression data sets (created using isobaric mass-tag labeling) were created using either AD: LR (1) or HR:LR ratiometric comparisons (2). For these differential protein expression data set comparisons (1 and 2), the resultant Plurigon representations were oriented in an identical fashion with the yellow z-axis extending from the page to the reader in a perpendicular manner. Analogous structural components found in both data set comparisons are indicated by the orange arrays in both Plurigon display modes. It is evident from these simple representations that functional commonalities are observable between fully diagnosed AD patients and those that are disease-free but possess a high genomic risk score.

Topological analysis of paired pathophysiological gene expression networks (extracted from the ArrayExpress data base) was applied recently to hippocampal data from AD patients (Yue et al., 2016). Yue and coworkers identified 144 DEGs across multiple studies. Five groups of coexpression gene pairs and five gene networks were identified and constructed using four existing bioinformatic methods: 1) weighted gene coexpression network analysis; 2) empirical Bayesian (EB) analysis; 3) differentially coexpressed genes and links (DCGL), a search tool for the retrieval of interacting genes/proteins data base; and 4) a novel rank-based algorithm with combined score. Using this combinatorial approach Yue et al. (2016) found that the resultant co-expression network was scale-free (an important feature of TDA) and had the tendency to exhibit small-world characteristics (Watts and Strogatz, 1998), i.e., possessing a variety of nodes possessing a diverse range of degree scores. This small-world form of network suggests that it is likely that signaling networks are coordinated by so-called keystone factors (Chadwick et al., 2012a).

Furthermore, advanced data topology investigations recently demonstrated their efficacy with respect to molecular patient stratification in the field of age-related diabetes. For example, Li et al. (2015b) aimed to investigate patient diversity in the scope of T2DM. This research group employed EMD files coupled to genotype data from over 11,000 individuals. In this manner Li et al. (2015b) were able to identify three distinct diabetic molecular subtypes using Ayasdi–Iris (Lum et al., 2013) based topological data classification from this large cohort of patients. These three topologically separate diabetic types were typified by their comorbidity linkages to 1) nephropathy and retinopathy (subtype 1); 2) cancer and CVDs (subtype 2); and 3) neurologic disorders, allergic responses, and CVDs (subtype 3). Coupling this structural approach with the highly nuanced TDA, (Li et al., 2015b) were then able to correlate specific gene single nucleotide polymorphisms with these diverse T2DM subtypes and therefore generate considerable functional insights with respect to the generation of specific disease pathomechanisms. In addition to the work of Li et al. (2015b), imaged informed precision medicine analysis of T2DM EMD files recently was performed by Perer et al. (2015). This research group employed the Care Pathway Workbench (Yu et al., 2014) to stratify over 11,000 diabetic patients into therapeutically informative subtypes based on distinct data visualization patterns (Perer et al., 2015).

With respect to drug discovery workflows, drug target interactions can be predicted based on observed topological features of a semantic network across a chemical and biologic space using TDA-based approaches. In a semantic network, the types of nodes and links are different. To take the heterogeneity of the semantic network into account, metapath-based topological patterns can be investigated for link prediction. Fu et al. (2016) recently constructed supervised machine-learning models based on meta-path topological features of an enriched semantic network derived from Chem2Bio2RDF. This structure was then further built upon by adding compound and protein similarity neighboring links obtained from the PubChem data bases (Fu et al., 2016). The additional semantic links significantly improved the predictive performance of the supervised learning models. Topological drug-discovery approaches, with respect to signal transduction monitoring, were applied recently to one of the more recently developed HD data pipelines, i.e., single cell transcriptomic data (Gong et al., 2018). In addition to this topological appreciation of complex drug-induced signaling “landscapes” (Vogt, 2018) as well as simpler data set structures [columns within multirelational data sets (Partl et al., 2014)], TDA has also demonstrated promise with respect to effective definition of novel therapeutic agents (Vogt, 2018).

In addition to TDA implementation for drug target interactions, H-D data topological studies have also shown promise with respect to a capacity for prediction and/or discovery of novel and potentially deleterious drug-drug interactions. Recently topological community cluster generation, from an initial crude drug-drug interaction network using ForceAtlas2 (Jacomy et al., 2014; Udrescu et al., 2016), was used to help inform drug interaction investigation. These clusters were then combined with inferences gained from modularity-based detected communities within the initial drug-drug network to create a color-coded community-based drug-drug interaction network. By using this community drug network approach, this research group was not only capable of identifying potentially deleterious drug-drug interactions, but they suggested that with such a nuanced network-based view of drug functionality these H-D data structures could also assist in effective drug repurposing.

X. Machine Learning and Pattern Recognition

XI. Conclusions

The presence of high quality and informative H-D datasets related to age-related disease research is now a widespread and commonly employed aspect of biomedical research. Advances in molecular profiling technologies and the development of sophisticated computational approaches for analyzing these data are providing a new systems-oriented approach toward drug discovery, e.g., quantitative systems pharmacology (McQuade et al., 2017; Maudsley et al., 2018), personalized diagnoses, and patient stratification (Topol, 2014). Multiple H-D data acquisition and interrogation (Fig. 3) mechanisms have now been firmly placed into the realms of disease stratification and pharmacologically targeted H-D drug signaling definition. Systems-oriented approaches to drug discovery effectively use the parallelism and H-D of molecular data to construct more inclusive molecular models that aim to enhance our appreciation of pathophysiological and pharmacological mechanistic systems. This H-D data model of molecular biology offers a means to explore complex molecular states (e.g., age-related disease) where thousands to millions of molecular entities comprising multiple molecular data types (e.g., proteomics and transcriptomics) can be evaluated simultaneously as components of a coherent network.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Diverse platforms for effective acquisition and analysis of high-dimensionality data. (A) There is currently a wide array of efficient, high-throughput mechanisms by which to acquire high-dimensionality (H-D) biologic data. These data can be derived from cell, tissue, experimental animal, or even from direct human electronic medical records (EMR). Nucleotide sequencing platforms can easily generate data pertaining to genomic or transcriptomic activities. Protein-based analysis can generate highly nuanced data from either classic tissue-based quantitative proteomics or from affinity purification mass spectrometry (AP-MS)-based interactomics. The use of additional diverse forms of mass spectrometric analyses can also generate in-depth metabolomic data as well. (B) Once the H-D data has been acquired, efficiently stored, and curated using expert domain knowledge, multiple forms of analytics can be applied to extract the most meaningful interpretation of these hypercomplex data sets.

Interconnected age-related disorders are likely caused by a subtle blend of genetic, environmental, and individualized lifestyle factors. This makes it highly challenging for specific differential diagnoses and thus effective “precision” therapeutic design. However, as already indicated in this review, by integrating different “omics” approaches and through H-D data analysis, this therapeutic goal may actually be achievable. While this field is still under construction, certain studies have already been performed to facilitate diagnosis, biomarker discovery, elucidation disease progression, and design of targeted therapeutics, driving a new era of precision medicine. Several studies have concentrated on the expression alterations of biochemical pathways as opposed to single markers of disease. This whole network approach underpins the causation of several multicausal disorders. Here the use of protein chips could be future tools for better analysis of proteomic disease signatures or networks (Palmieri et al., 2016; Hellström et al., 2017).

Here we present the potential diverse interrogation platforms that can be used by researchers to investigate the intricate relationship between H-D disease phenotypes and a similarly detailed H-D pharmacological signature. This multidimensional investigation approach can represent an effective model for a gestalt appreciation and remediation of age-related conditions. An interesting example of this H-D data system integration lies in the deconvolution and potential treatment of amyotrophic lateral sclerosis (ALS). ALS is a rare but potentially fatal heterogeneous neurodegenerative disorder that is strongly linked to the condition of frontotemporal dementia. Interestingly, both of these conditions are now strongly considered to be age-related diseases (Niccoli et al., 2017). This age-associated nature of ALS or frontotemporal dementia is strongly linked to the presence of profound oxidative stress mechanisms linked to alterations of superoxide dismutase in these conditions. Two forms of ALS are known, familial (5%–10% of cases) and sporadic, which accounts for the rest. The use of H-D data analysis has already been shown to allow progress in the identification of possible biomarkers for ALS; however, as emphasized before, the integration of the different H-D data acquisition modalities is necessary (Mitropoulos et al., 2018). From classic genomics, which have been the recent staple of ALS research, it appears that a transition to a multiomics approach has been increasingly suggested. Through single omics approaches, several biomarkers have already been identified; e.g., superoxide dismutase 1 through transcriptomics (Freischmidt et al., 2015), TDP43 through proteomics (Sreedharan et al., 2008), and neurofilament proteins (NF), especially phosphorylated neurofilament heavy chain, also through proteomics (McCombe et al., 2015). However, we hypothesize that one single protein will not be enough to adequately design a therapeutic and that by combining the separate omics, a possible network of drug target proteins can be discovered that could be much more efficacious (Mitropoulos et al., 2018).

Many issues regarding the acquisition and storage of H-D data have been perfected. However, the generation of simple yet intelligent data interpretation systems is still a future challenge. The usefulness of a specific platform depends on the nature of questions to be addressed, i.e., identifying actionable risk factors for disease prevention, is a very different problem from identifying new targets for drug discovery that is furthermore different from predicting who will benefit from a specific therapeutic intervention or what might be prognostic markers for disease progression. Simplicity and mechanistic diversity of new bioinformatic platforms is needed. The ultimate goal of the integrative creation of a coherent analytical pipeline for gerontological research represents a vital goal for novel drug development.

We envisage that an effective quantitative system would first possess the ability to identify molecular causes of disease (at both pre- and postsymptomatic phases) using advanced patient molecular stratification (e.g., with topological data analysis and EMD semantic analytics). Secondly, this ideal informatics system would enable and inform the prediction of the signaling patterns entrained by these alterations, e.g., using classic pathway analysis and natural language processing-based de novo pathway construction and theoretical data set comparisons for biased receptor-based signaling definition (Maudsley et al., 2016). Thirdly, such an effective pipeline should facilitate the generation of whole somatic appreciation of systemic pathologies entrained by the age-related disease. This appreciation could be constructed using complex deep learning machine analytics applied to standard biomedical images, as well as “connectomic” and proteomics-based “diseaseome” maps (Lau et al., 2018). Fourthly and finally, this intelligent informatics system should engender the ability to inform and eventually test the efficacy of investigational new therapeutics. For example, machine-learning assisted TDA as well as semantically created theoretical data set comparisons for biased signaling specificity may help to develop therapeutic agents with enhanced specific efficacies and minimized side-effects/contraindications.

The number of research groups applying H-D data analytics to complex gerontological therapeutic paradigms is only likely to increase due to the reduction in unit costs of assays and further advances in multiplexing genomic, transcriptomic, and proteomic technologies. It is therefore imperative to continue the rational development of advanced algorithms and analytical platforms to rapidly and meaningfully extract therapeutically actionable data from this ever expanding data corpora. It is likely that in the future, the development of novel data generating (hardware systems such as mass spectrometers) and analytical platforms will eventually feed into each other’s development to provide synergistic data gathering and processing units. The generation of more nuanced and intuitive informatic systems will potentially accelerate novel drug development and simultaneously aid rational drug repurposing for age-related disease conditions.