Systems Biology in Biomarker Development for Cancer Signaling Therapy

C H A P T E R

3 Systems Biology in Biomarker Development for Cancer Signaling Therapy Evgenii Generalov1,2, Tara Clarke3, Lahiru Iddamalgoda4, Vijayaraghava Seshadri Sundararajan4, Prashanth Suravajhala4,5 and Alexey Goltsov3 1

Cell Proliferation Laboratory, Engelhardt Institute of Molecular Biology, Moscow, Russia Faculty of Physics, Lomonosov Moscow State University, Moscow, Russia 3School of Applied Sciences, Abertay University, Dundee, United Kingdom 4Bioclues.org, India 5Department of Biotechnology and Bioinformatics, Birla Institute of Scientific Research, Jaipur, India

2

O U T L I N E 3.1 Introduction

28

3.2 Systems Biomarker in Personalized Cancer Therapy

29

3.6 Systems Biomarker Identification by Exploring Contextual Genomic Features 40

3.3 Toward Systems Biomarker Development Through In Silico CDx Assay Model 33

3.7 Machine-Learning Approaches in Systems Genomics and Pharmacology: Biomarker Identification Based on Omics Data 44

3.4 Systems Biomarker Identification in Signaling Control Interface 36

3.8 Conclusion

46

Acknowledgments

47

References

47

3.5 Reprogramming of Signaling Networks as a Challenge in Systems Biomarker Identification in Cancer Therapy 37

Companion and Complementary Diagnostics DOI: https://doi.org/10.1016/B978-0-12-813539-6.00003-1

27

© 2019 Elsevier Inc. All rights reserved.

28

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

3.1 INTRODUCTION The paradigm of predictive biomarkers of therapeutic responses to drug intervention has dramatically transformed cancer drug development, disease diagnosis, prognosis, and therapy. Successful development of the molecular marker driven cancer therapy laid a foundation of personalized therapy and medicine when biomarker design and biomarker screening form the basis of stratification of patients who benefit from the drug therapy. Over the past three decades, we have witnessed a rapid translation of cancer research in the field of biomarker development to the US Food and Drug Administration (FDA)-approved biomarker screening in clinical practice. More than 80 molecularly targeted drugs have been approved for the treatment of various cancers and most of these drugs have genetic biomarkers associated with therapeutic targets of the drugs [1]. Most of these biomarkers are associated with the well-established gain- and/or loss-offunction mutations in oncogenes or oncosuppressors such as ErbB2, TP53, VEGF, KRAS, BRAF, and others. The biomarkers indicate the abnormalities in the oncogenic signaling pathways which are frequently activated in different cancers such as the PI3K/PTEN/ AKT/mTOR, mitogen-activated protein kinases (MAPK), WNT pathways, and others and cause carcinogenesis, uncontrolled cell proliferation, and drug resistance [2]. Although these key pathways in tumorigenic-transformed cells are well known, detailed mechanisms of cancer cell addiction to these prosurvival and antiapoptotic pathways remain elusive. Despite fast progress in molecular marker driven cancer therapy targeting oncogenic pathways in cancer patients, clinical outcomes showed that some biomarker-positive patient cohorts exhibit a lack of positive response to the therapy. For example, in anti-HER2 (human epidermal growth factor receptor 2) therapy, specific HER2-positive breast cancer patient populations treated by Herceptin demonstrate either de novo (primary) resistance or acquired resistance when initial therapeutic response weakens following a prolong treatment and leads to progression of the disease [3]. This problem in personalized therapy highlights unmet necessity to develop the next generation of predictive biomarkers. Progress in cancer translational research and a better understanding of tumorigenesis at the molecular and genetic levels lead to the discovery of the novel molecular mechanisms of targeted therapy resistance [3 5]. In part, new therapeutic strategies in anti-HER2 treatment were developed to suppress activation of the downstream and cross talk pathways as a result of drug intervention [1,6,7]. This research was translated into the development of novel drug and drug combination therapy targeting therapeutic resistance and showed high efficiency regarding the different cancer types unresponsive to anti-HER2 treatment. So, further progress in the development of truly predictive biomarkers for exact stratification of cancer patients requires the development of the next generation of molecular biomarkers which would take into consideration the new mechanisms of de novo and acquired resistance. The development of reliable next-generation biomarkers to predict efficacy and toxicity of anticancer drugs for cancer patients is one of the key challenges in cancer therapy [8]. In this way, many strategies of the biomarker development have been suggested [9,10], but, in general, standard strategy and methodology in this field are absent nowadays. For this reason, many promising anticancer drugs cannot exhibit their potential thanks to the absence of proper biomarkers for right patient cohorts showing clear therapeutic benefits.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.2 SYSTEMS BIOMARKER IN PERSONALIZED CANCER THERAPY

29

In addition, some drug candidates failed registration because of the problem of identification of exactly targeted patients for these drugs at the clinical stage of drug development [11]. In addition to traditional models of the clinical design for biomarker identification [9], one of the promising technologies in this direction is drug-diagnostic codevelopment combined with companion diagnostic (CDx) [1,12]. According to this strategy, the drugs are developed based on a well-established mechanism of action (MoA) of the drug candidates that leads to establishing tangible biomarkers of the compound efficacy and toxicity. So, the drug-diagnostic codevelopment strategy suggests biomarkerguided drug development, where the predictive biomarker CDx assay is developed jointly with the targeted drug to guide the use of targeted cancer drugs [12]. Thus the convergence of the molecular biomarker design with drug development significantly increases the predictive power of biomarkers and enhances the role of CDx assay in clinical diagnostics. However, the challenge remains how to integrate molecular diagnostics to drug development based on the therapeutic response of tumor to drug action at the systems level. In this chapter, we discuss a systems biology approach to the biomarker development based on the systems-level MoA of anticancer drugs and the mechanisms of the resistance emergence to targeted therapy. We show that the integrative systems biology approach is consistent to the methods and technologies used in drug discovery which are based on the integration of multiomics data generated at the molecular, cellular, and organism levels at the different stages of drug development. Building-up of the systems biology methods and tools to analyze and model heterogeneous high-throughput data makes it a promising approach to incorporation of multilevel omics data into computational models of MoA of drugs. We suggest that response characteristics of the cellular signaling networks to drug action obtained in the framework of the developed computational models can be used complementary to the development of CDx assay. In Sections 3.2 and 3.3, we discuss the concept of an in silico CDx assay model which may be developed in parallel to the development of in vitro CDx assay in order to support the design of systems biomarkers. In Section 3.4 we focus on the complexity of the signaling networks embracing the drug targets and discuss a concept of the cellular signaling control interface (SCI) useful for identification of the systems biomarkers. Section 3.5 is devoted to the analysis of one of the challenges in systems biomarker development related to the reprogramming of signaling networks under drug treatment. Finally, we discuss exploration of genomic features for systems biomarkers identification with the advances of availability of high-throughput technologies generating a huge amount of sequences. Various bioinformatics-based techniques, such as machine learning, network-based approach, and text mining for identification and characterization of systems biomarkers, are discussed.

3.2 SYSTEMS BIOMARKER IN PERSONALIZED CANCER THERAPY Rapid development of CDx and its successful introduction into the clinical practice have enabled a shift to molecular biomarker driven precision medicine. In 2016, the FDA has approved 26 CDx assays to screen HER2, EGFR, and KRAS overexpression and

COMPANION AND COMPLEMENTARY DIAGNOSTICS

30

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

mutations which are recommended to use as companion tests for the administration of anti-HER-targeted cancer therapy [1]. Commonly, CDx assay is designed in parallel with the drug marketing approval. For example, in case of Herceptin, its CDx assay HercepTest (Dako) to screen HER2-positive breast cancer patients was introduced at the same time as Herceptin was approved by the FDA in 1998. After this, several successful in vitro CDx devices have been approved using different assay techniques. In spite of the success of anti-HER2 therapy, not all patients with biomarker-positive test respond to Herceptin treatment [1,4,6], that is, some cancer patient populations demonstrate either de novo (primary) or acquired resistance following prolonged treatment and leading to the patients’ relapse. These problems of the personalized therapy revealed an unmet need to identify new predictive biomarkers to assess the effectiveness of cancer therapy, in part, anti-HER2 therapy. Currently, there is no definite biomarker of the resistance to HER2-targeted therapy and a set of biomarkers has been proposed based on the different mechanisms of resistance revealed in the joint experimental research and clinical trials [4,6]. The marker identification underling Herceptin resistance is impaired by the fact that Herceptin MoA is not entirely clear. To tackle this problem, different experimental and computational systems approaches have been developed [13,14]. In the framework of systems biology, the traditional approach to drug discovery “one drug one target one biomarker one disease” is considered as insufficient for the treatment of complex diseases such as cancer. Here, we discuss the challenge in the development of systems biomarkers for personalized therapy targeting oncogenic signaling pathways. Systems biomarkers are suggested to be developed based on the detailed information on a drug target and consideration of its network environment which is enriched by feedback loops and cross talk connecting a targeting pathway with compensatory pathways to enable cancer to bypass a direct drug action [15]. Next, we focus on several systems biology approaches successfully developed and applied to the identification of the next-generation biomarkers to support the drug and combination therapy, in part, anti-HER2 therapy which enables to overcome de novo and acquired resistance to wellestablished drugs. One of the promising approaches developed in systems biology is computational modeling of complex metabolic and signaling pathways and its application to the identification of molecular mechanisms of a malfunction in cellular response to drug action at the different genetic alterations [16 18]. This systems approach was successfully applied, in part, to the investigation of downstream pathways of HER2 signaling and analysis of the effectiveness of the resistance biomarkers to Herceptin such as loss-of- and gain-function mutations of tumor suppressor PTEN and PIK3CA oncogene causing insensitivity of the PI3K/PTEN/AKT pathway to HER2 inhibition [14,19,20]. However, mutations in these genes were identified before high-throughput era; further investigation is needed to understand the results of mutual mutations in this pathway to support the development of the effective combination therapy to overcome drug resistance. Other sets of biomarkers were identified as a result of the joint computational and experimental investigation of complex receptors cytokines drugs interaction networks, in part, the network of the family of ErbB receptors [21 23]. As reported, receptor tyrosine kinase (RTK)-targeted cancer therapies including anti-HER2 therapy failed when cancer

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.2 SYSTEMS BIOMARKER IN PERSONALIZED CANCER THERAPY

31

cells bypass the action of a single agent due to coexpression of different ErbB receptors in various cancer types that results in the resistance development [19]. In a series of works devoted to the complexity of the ErbB signaling network, it was shown that ErbB receptors and their ligands play a transforming role in the generation of malignant phenotypes, and different coexpression profiles of ErbB receptors and epidermal growth factors define different cancer types and subtypes [24 27]. This suggests that the progression, growth, migration, and survival of carcinoma cells are sustained, at least in part, due to a network of RTK of the ErbB family and their ligands. Coexpression profiles of different ErbB receptors and their ligands in human carcinomas define different tumor response to drugs targeting ErbB receptor system, ranging from drug sensitivity to resistance, and the clinical data on the ErbB receptor expression correlate to some degree with disease-free survival and anti-RTK treatment outcome [28]. In part, it has been shown that the binding of trastuzumab (Herceptin) to ErbB2 receptor does not prevent ligand-induced HER2 heterodimerization with ErbB1/EGFR and ErbB3/HER3 that leads to activation of the RAS MAPK and PI3K/PTEN/AKT pathways and continued proliferation in the presence of Herceptin [29,30]. It was concluded that coexpression of EGFR and ErbB3 in HER2 1 cancer cells can be biomarkers of Herceptin resistance [31]. Another mechanism of the resistance development is related to the phenomena of reprogramming of cellular signaling networks induced by drug intervention. Currently, this mechanism of resistance emergence is intensively investigated by systems biology methods using different proteomic and genomic profiling techniques [32]. Drug-induced reprogramming of signaling networks is a systems mechanism of acquired resistance to single drug therapy [5]. This mechanism lies in the adaptive rewiring of targeted networks and activation of the compensatory networks to enable cancer cells to escape the inhibitory effects and uphold the tumorigenic phenotype. Reprogramming of RTK signaling networks following inhibition of one of the members of RTK family was reported to lead to a compensatory response to single drug intervention [5,33]. Transcriptional and post-transcriptional upregulation of ErbB3 receptors were reported to compensate for the RTK-targeted therapy [34,35]. As a result of the increasing understanding of the mechanism of network dynamic reprogramming, combination therapies targeting multiple RTKs and successfully suppressing de novo and acquired resistance have been developed [19]. The coexpression profile of different ErbB receptors is used as a basis for rational drug codevelopment and drug-diagnostics combination strategies [36,37]. Following this strategy, new anticancer drugs have been approved and recommended for use in combination therapy. For example, pertuzumab, a humanized anti-HER2 monoclonal antibody, has been approved only in combination with trastuzumab [38]. In contrast to trastuzumab, pertuzumab mainly blocks ligand-dependent receptor heterodimerization of ErbB2 and ErbB3 receptors [39] while trastuzumab is more effective at ligand-independent tumor growth [40]. In Section 3.5, we discuss the drug-induced network reprogramming in detail. The systems approach is also useful for biomarker identification in complex oncogenic signaling networks which are controlled by multiple positive/negative feedback loops and cross talk connecting a targeting pathway with the compensatory pathways enabling cancer cells to bypass a drug action [41]. The complex regulatory and compensatory negative feedback can be illustrated by mTORC1/2 pathway which is a metabolic and signaling hub controlling many cellular processes such as nutrient uptake, energy metabolism,

COMPANION AND COMPLEMENTARY DIAGNOSTICS

32

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

translation and involved in autoimmune disorders, cancer, obesity, and aging [42]. Thus drugs that selectively target mTORC1, such as rapamycin, are expected to impair cancer metabolism and considered as a promising anticancer therapy. However, complex regulation of mTORC1 pathway is activated in response to rapamycin treatment that causes activation of the upstream AKT signal as a result of the release of negative feedback in the PI3K/AKT/mTORC1 pathway. This activation of antiapoptotic AKT signal was suggested to lead to rapamycin resistance [41]. Investigation of correlation of rapamycin-induced AKT activation with sensitivity and resistance to rapamycin in a large panel of cancer cell lines was reported to show that rapamycin-induced AKT activation is, on the contrary, associated with rapamycin sensitivity in vitro [43,44]. Cells with PI3KCA/PTEN mutations were reported more likely to be rapamycin sensitive [43,45,46]. Thus on the one hand, PI3KCA/PTEN mutations are biomarkers of trastuzumab resistance, and on the other hand, they can be biomarkers of sensitivity to rapamycin. So, the development of systems biomarkers closely links to the investigation of different compensatory mechanisms and negative feedback loops which are differentially regulated in different cancer cell lines. Unwinding these control networks is a challenge of systems biology of biomarker identification that can lead to the development of systems biomarkers, in part, for the efficacy of monotherapy or combination rapalogs therapy. A key problem in this area is a need of further systems investigation on logic circuits of the control interface which integrates multiple cellular signals (metabolic status, nutrition, cytokine levels, and others). In Section 3.4 we discuss a concept of the signaling hub/interface in cellular networks as promising targets for new drug development and biomarker identification. Development of the systems methods of multiscale genomic, transcriptomic, and proteomic analysis combined with the network-based analysis of cancer cell responses to drug treatment significantly enhanced systems biology approaches to the identification of novel biomarker [7,32,47 49]. For example, computational network based analysis of microarray data on gene expression was carried out for cell treatment by anti-HER2 drug, lapatinib targeting RTKs that allowed determining a set of genes characterizing lapatinib resistance ErbB2-positive breast cancer cell line (NDRG1, HSPA5, HK2, IRE1) [47]. Application of a functional enrichment analysis of global expression data revealed compensatory pathways, mainly the glucose-deprivation and estrogen receptor stress response networks, which are chronically activated in response to targeted therapy and counteract lapatinib-induced toxicity in resistant ErbB2-positive breast cancer cells. Application of a systems-based approach to the construction of a gene gene interaction network helped authors to identify genes responsible for the chronic activation of compensatory networks in response to the therapy targeting RTK pathways and map the genes highly correlated with relapse in ErbB2 1 tumor patients. Translation of this systems research to the clinical data analysis allowed validation of the identified genetic biomarkers and, moreover, to suggest novel drug combination therapy targeting signaling and metabolic pathways in tumors with acquired resistance to anti-HER2 therapy. Thus the network-based approach of systems biology combined with the large-scale clinical analysis of gene expression profiles of cancer patients can synergize the development and validation of systems biomarkers of targeted therapy. Now we can see an expansion of systems approach to the identification of biomarkers which link oncogenic pathways with the signaling pathways related to inflammation,

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.3 TOWARD SYSTEMS BIOMARKER DEVELOPMENT THROUGH IN SILICO CDX ASSAY MODEL

33

oxidative stress, metabolic reprogramming, and immunity response [32,47,50]. The link of oncogenic and immunity pathways is well illustrated by the immune function of trastuzumab investigated in clinical trials [4]. It was reported that additionally to the wellestablished MoA of trastuzumab as an inhibitor of cell proliferation, trastuzumab induces cell death due to an immune response activation of an antibody-mediated cellular cytotoxicity. It was found that besides trastuzumab binding to extracellular epitopes of ErbB2 receptors of cancer cells, it also binds to Fc receptors expressed on natural killer cells and macrophages in tumor microenvironment. This causes their activation and leads to trastuzumab-dependent cellular phagocytosis and lysis of the HER2 overexpressed tumor cells. The Fc-receptor polymorphism was studied as the predictive markers of trastuzumab response that showed that FcgRIV-expressing immune cells in tumor tissue were increased in response to trastuzumab treatment in the HER2 1 breast cancer xenograft tumor model and demonstrated that FcgRIV expression on the macrophage can be a biomarker of trastuzumab-associated antibody-dependent cellular phagocytosis and lysis of HER2 1 tumor [51]. This study of a new MoA for trastuzumab and systems biomarkers related to this mechanism suggests that the activation of tumorassociated macrophages can improve the anticancer efficacy of trastuzumab and cancer immunotherapy as a whole. This mechanism also suggests the benefits of the combination of anti-HER2 therapy with checkpoint inhibitors such as anti-PD-1 and anticytotoxic T-lymphocyte-associated protein 4 [52]. In the following section, we discuss applicable systems biology approaches to the development of systems biomarkers to support CDx.

3.3 TOWARD SYSTEMS BIOMARKER DEVELOPMENT THROUGH IN SILICO CDX ASSAY MODEL A CDx assay, according to the FDA guidance, is defined as an in vitro diagnostic device that provides information essential for the safe and effective use of a corresponding therapeutic product [1,12]. The main purpose of developing of the CD assays in conjunction with a drug is to have a test that can predict whether a patient is likely to benefit from it. Hence, for many drugs, the CDx assays will take up a central role as a decisive stratification tool both during the clinical development phases and, later, after approval, when the drug is in the clinic. As the CDx assay is developed in the progress of drug design, vast amounts of multilevel experimental data are produced at the molecular, cellular, and organism levels in in vitro and in vivo experiments and clinical trials on patients’ response to the treatment. So, in vitro CDx assay can be considered as an in vitro predictive model of a specific cancer type and its response to targeted therapy. Moreover, the process of integration of multilevel experimental data in in vitro CDx assay is consistent with the systems biology approach to generating and integrating of multilevel omics data into an in silico model to predict the response of targeted signaling pathways to a drug inhibition action. Thus systems biology approach can be used complementary to the development of in vitro CDx assay and supports it by the development of an in silico CDx assay model. In silico CDx assay model represents a computational framework which includes a computational

COMPANION AND COMPLEMENTARY DIAGNOSTICS

34

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

model or a set of interlinked models which are trained and validated on a basis of experimental data obtained in the progress of drug development. In Fig. 3.1, we show a scheme of the in silico CDx assay model development complementary to in vitro CDx assay based on the multiomics data stream generated at the different stages of drug development. Incorporation of systems biology modeling into the drug development process represents a mutual exchange of experimental data and information generated by model-based data analysis. Among computational systems biology approaches, we highlight next the following modeling pipelines which are currently intensively developed and can be directly applied to the in silico CDx assay model development. First, the computational systems biology of metabolic and signaling network response to drug intervention [14,17,22] is now rapidly expanding to an emerging area of Quantitative Systems Pharmacology (QSP) [53]. In the framework of QSP, computational systems biology approaches are applied to the development of model-based pipeline which combines computational models of drug-targeted cellular networks with pharmacokinetics/pharmacodynamics (PK/PD) models [54,55]. As a result of the integration of experimental and clinical data with computational models, a PK/enhanced PD (PK/ePD) modeling approach has been developed and applied to preclinical and clinical trials. For example, the PK/ePD model of the VEGFR signaling pathway response to the antiangiogenesis drug, sunitinib, was developed and applied to the optimization-based control analysis and design of multidrug regiment treatment protocol [56]. Other computational systems biology approach extensively implemented in systems pharmacology is a logic modeling-based technique and software tools which are well suited to the development of large-scale signaling networks, where detailed biological knowledge is often incomplete [57]. This modeling approach focuses on the regulatory logic in the drug-targeted signaling networks embedded in a complex environment of interconnection and cross talk activated by different cellular stimuli. In particular, this

FIGURE 3.1 Life cycle of in silico CDx assay model development complementary to in vitro CDx assay based on the experimental data stream generated at the different stages of drug development.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.3 TOWARD SYSTEMS BIOMARKER DEVELOPMENT THROUGH IN SILICO CDX ASSAY MODEL

35

systems approach was successfully applied to the elucidation of drug resistance mechanisms in colorectal cancer [58] and was expanded to patient-specific logic models of signaling pathways based on screenings of cancer biopsies to inform personalized combination therapy [59]. The current challenge in computational systems biology is in the adaptation of wellestablished computational techniques as well as the development of new computational tools to model the large-scale multidimensional experimental data on in vitro 3D organoid models and, in part, 3D tumor spheroid heterocellular models which replace 2D cellular population models in the preclinical stage of drug development [60]. A transition from 2D to 3D tumor models in screening antitumor drug pharmaco-activity is defined by the benefits of taking into account 3D morphological complexity and heterogeneity of tumors in 3D computational tumor models that are close to in vivo tumor structure and processes [61,62]. As suggested, the complexity of heterocellular tumor structure allows tumor to successfully bypass the therapeutic effects of drugs that causes drug resistance in 3D models when 2D models show drug sensitivity. Such 3D heterocellular tumor models mimic the tumor microenvironment including various growth factors and patient’s immune system action. The main advantage of 3D heterocellular tumor models is the reproduction of the tumor microenvironment processes, when a tumor sample is cultivated in the patient’s microenvironment, including a set of cells of the immune system, stroma, and vascular network [60]. One of the promising computational technologies capable to cope with modeling and analysis of large-scale multidimensional data generated in a 3D experimental assay is machine-learning modeling and classification of phenotypic responses at drug screening. Machine-learning and deep learning modeling are computational technologies, which rapidly enter into biomedical and pharmaceutical research [63 65]. By using network analysis approach of systems biology, machine-learning models are adjusted to infer signaling and metabolic networks based on the proteomic and genomic data and large databases of like-drug chemical compounds to predict new drug targets and reveal drug MoA [65]. A powerful classification potential of machine-learning models is widely adopted for cancer patient classification into high or low survival groups and development of cancer prognosis biomarkers [66,67]. The development of a machine-learning technology for the analysis of massive datasets generated at the preclinical stage of drug development is an excellent ground to apply machine learning to pave a way to phenotype-driven drug discovery technology complement to the target-based strategy and their combination [68]. Application of deep learning to the joint analysis of high-throughput omics data together with data on phenotypic screening of large drug-like compound databases is likely to facilitate phenotypic-based biomarker development. In Section 3.6, we discuss in detail machine-learning application to the development of biomarkers based on the genetic data. The main strategy in the development of in silico CDx assay model aims at the design of a computational framework to identify systems biomarkers of the response and outcome of cancer patients to drug therapy depending on the genetic profiles of the disease. At the cellular and tissue levels, the model should accurately describe a short- and long-time dynamic response of the heterogenous tumor to drug therapy and predict sensitive and resistant cancer phenotypes at the early and late stages of drug treatment [49].

COMPANION AND COMPLEMENTARY DIAGNOSTICS

36

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

The predictive power of systems biomarkers will be attained by training and validation of in silico CDx assay model based on large-scale omics data aggregated during all stages of drug development (Fig. 3.1).

3.4 SYSTEMS BIOMARKER IDENTIFICATION IN SIGNALING CONTROL INTERFACE The development of in silico CDx assay model in the framework of the systems biology paradigm suggests an integration of hierarchical levels of the cellular organization into multilevel computational models (MLMs) in order to predict response to drug treatment. Prediction of both short- and long-term behavior of the heterocellular cellular systems such as organoid models, tissue, and organs requires development of multiscale approach to multilevel modeling. The progress of complex diseases such as cancer, diabetes mellitus, neurovegetative and cardiovascular diseases, and others are strongly multilevel and multiscale processes which embrace different cell types, tissues, and organs and progress at different time scales, days, months, and years. To develop the hallmarks of these heterogeneous diseases, it is necessary to integrate several risks factors, such as genetic, epigenetic, and environmental factors, into a systems biomarker predicting the onset of the diseases. We suggest that a key mechanism regulating time multiscale behavior of cellular systems and their response to drug therapy is defined by signaling control interface (SCIs) in multilayer cellular organization. The SCI is a regulatory gate between adjacent layers of a cellular organization where multiple external/internal cues converge to promote a coordinated response of the whole system to drug intervention at a wide range of time scales. So, identification and modeling of the SCI molecular structure and function in healthy and pathology are critical for the biomarker development. Next, we discuss the results of the computational modeling of the SCIs regulating time multiscale responses of the signaling networks to activation and drug intervention. Computation modeling of SCI between the receptor protein signaling network and gene regulatory network in the framework of the multilevel computational approach was successfully undertaken in a series of works devoted to RTK signaling networks [69 72]. The developed models allowed authors to describe the short- and long-time scale signaling in the RTK networks and its link to cell fate. Consideration of the SCI between receptor/protein and genetic networks revealed that the SCI ensures different types of output AKT/ERK signaling: sustained, transient, and oscillatory signaling dynamics that are observed in different cancer cells [70]. One of the features of the SCIs is their enrichment by multiple regulatory feedback loops which form logic circuits. Particularly, the emergence of the logic circuits in SCIs of gene regulation can result from specific activation mechanisms of some transcription factors requiring several temporally correlated signals for their activation which can be arranged as AND or/and OR gates [72]. Investigation of these logic circuits by the coupled signaling-transcription models revealed the mechanism of signal discrimination between transient and sustained ERK1/2 signaling that defines cell fate at the long time scale [72].

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.5 REPROGRAMMING OF SIGNALING NETWORKS

37

Other feature of the SCI, related to a complex response of the systems at a wide time range, is the hierarchical time scales in gene expression response to receptor/protein signaling [71,73]. Early and delayed expressing genes were observed at the ERK1/2 signaling activation [73]. Consideration of the coupled protein signaling-genetic network in ERK1/2 pathway and the detailed functional structure of the SCI revealed that positive/negative feedback loops between receptor/protein and genetic networks lead to stabilization of a prolonged signal, transition from graded to biphasic one, and switch to an irreversible response to receptor activation [71]. Early- and long-time response genes are examples of the realization of the up- and downscaling in signaling network regulation through the SCI. This property of the SCI defines the mechanism of multiscale hierarchical response in the coupled receptor/protein and genetic networks. Introduction of up- and downscaling approach into MLMs can help to bridge the gap between different temporal scales in MLMs of cellular response to receptor stimuli and drug treatment. One of the methods to study the SCI properties consists of the experimental and computational investigation of the input/output (I/O) characteristics of the cellular signaling networks [18]. The I/O curves were obtained for the SCIs of the ErbB receptor network and PI3K/PTEN/AKT pathway as a result of computational modeling of the response of this pathway to ErbB receptor activation and inhibition [18]. The I/O curves corresponding to the SCIs were also defined for the antioxidant regulatory system in ovarian cancer cells [74] and the regulatory systems of epithelial-to-mesenchymal transition at cancer progression [75]. We suggest that the complexity of nonmonotonic and multibranch I/O curves obtained in these works may define a wide spectrum of cellular responses to drug intervention. Searching and modeling of the SCIs is a challenge in the MLM development which is complicated by the inherent complexity of coupled and distributed receptor protein signaling and genetic networks. Application of statistical methods combined with pathway analysis of gene expression data can be useful for identification of the SCI patterns in cellular signaling networks [76]. We suggest that specific molecular components and subnetworks of the SCIs can be selected as systems biomarkers of therapeutic responses and considered as new targets for drug development. Thus identification of the SCIs and understanding of their regulatory mechanisms in healthy and pathology can assist in systems biomarker design and new drug target identification in drug therapy of cancer as well as other heterogeneous diseases showing multilevel and multiscale properties. In the following section, we discuss the role of SCIs in the emergence of therapeutic resistance due to reprogramming of signaling networks induced by prolonged drug therapy and the development of systems biomarkers of this process.

3.5 REPROGRAMMING OF SIGNALING NETWORKS AS A CHALLENGE IN SYSTEMS BIOMARKER IDENTIFICATION IN CANCER THERAPY The further development of the systems approach to MLM of time multiscale responses to drug treatment is required to predict drug efficacy for overcoming resistance acquired due to reprogramming of cellular signaling networks induced by drug intervention

COMPANION AND COMPLEMENTARY DIAGNOSTICS

38

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

(see also Section 3.2). Drug resistance due to the reprogramming mechanisms is developed during prolonged drug treatment as a result of the adaptive rewiring of a drug-targeted network and activation of the compensatory networks to enable cancer cells to escape drug inhibitory action [5,33,48]. Substantial investigation into the genetic and epigenetic mechanisms of signaling network reprogramming was stimulated by clinical observation of adaptive resistance to monotherapy regimes. As a result, key biomarkers for some of these transformations have been obtained and applied to cancer diagnostics and prediction of therapeutic responses to drug therapy [77,78], but, in general, the control mechanisms of the reprogramming and SCI behind these mechanisms are still elusive and need further investigation. The systems analysis of the multiomics profiling of multicellular systems treated by drugs can provide deep insight into these resistance mechanisms such as upregulation of distinct genetic subnetworks responsible for activation of redundant signaling pathways that leads cancer cells to circumvent the action of a single drug action. For example, reprogramming of the kinome in response to a MEK inhibitor (AZD6244) in SUM159 cells of triple-negative breast cancer was observed when upregulation of a set of kinases and RTKs occurs [33]. Reconstruction of the kinome map revealed a compensatory kinome response through the activation of multiple kinases and RTKs: PDGFRβ, VEGFR2, HER2/ 3, and others that allows cancer cells to bypass MEK inhibition. These data and other results of drug-induced changes in gene expression have shown that some inhibitors of specific kinases/receptors induce reprogramming of signaling networks in some cancer lines that leads to the upregulation of unique networks of kinases and receptors responsible for the activation of signaling pathways not targeted by drug therapy. This reprogramming thus leads to new phenotypes of cancer cells which are resistant to primary drug therapies, yet may be sensitive to other drugs targeting the newly activated networks [3,5]. Scenarios of dynamic reprogramming can be considered in terms of a hierarchical structure of signaling networks, in part, kinase and RTK networks consisting of dominant and secondary signaling networks. After inhibition of a dominant network, signaling is rewired to activate a secondary network, and upregulation of the secondary kinases/ receptors occurs [79]. It is also plausible that dominant and secondary RTK/kinase subnetworks cooperate with each other and drive cancer progression together. In this case, inhibition of the dominant kinase unmasks the rest of the signaling network and sensitizes the cell to signaling through this network and so sensitizes cell to a second drug targeting the secondary network [5]. For example, trastuzumab sensitizes HER2 1 cancer cells to EGFR and HER3-targeted therapy [3,30]. An acquired sensitivity to a second inhibitor targeting a drug-activated network suggests that the first drug action may significantly broaden the signaling network activity. This increase in the breadth of activity following drug-induced perturbation is manifested through the activation of various cross talks, feedback loops, and gene regulation circuits occurring due to the SCIs. In general, inhibition of a module of signaling network can cause an increase in sensitivity to perturbation, for example, through mutations or drug action, in the rest of the network modules [18]. Thus systems-level studies and modeling of signaling network response to drug intervention should consider much wider networks than perturbed networks alone, and consideration of the experimental data on gene

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.5 REPROGRAMMING OF SIGNALING NETWORKS

39

expression following drug treatment can unmask these expanded networks and unravel the mechanisms of the acquired sensitivity of the signaling network. This concept of drug-induced reprogramming of signaling network activity significantly expands the conventional view on oncogene addiction in targeted therapy. Typically, cancer therapy involves targeting single key signaling pathways, for example, cancer drivers such as the PI3K/PTEN/AKT/mTOR, RAS/RAF/ERK pathways, and others, and inhibition of these oncoprotein pathways to abrogate tumor growth. With drug-induced reprogramming of signaling networks, disruption of oncogenes by drugs fails, in many cases, to cause cell death and leads to cancer sensitivity to other kinase/receptor inhibitors. For example, not all HER2 1 breast cancer cell lines are sensitive to anti-HER2 therapy, that is, these lines do not exhibit addiction to the HER2 oncogene [5]. Following targeted anticancer therapy, cancer cell addiction to one oncogene can indeed be abolished, yet the dependence on another gene may emerge due to dynamic reprogramming. Thus the development of systems biomarkers should take into account signaling network rewiring mechanisms that enable cancer cells to change the dependence of tumor growth from one subnetwork to another. A key deliverable from the systems biomarkers would be the mapping of cell line specific drug-induced reprogramming in protein protein association networks for drugs targeting different signaling networks. Such mapped systems biomarkers of the targeted network reprogramming could significantly impact on the development of combination cancer drug therapy. First, unpicking the drug-activated subnetworks can help to identify new drug targets in designing combination therapy to overcome acquired resistance to monotherapy. As discussed earlier, a first drug may well broaden signaling network activity and thus expand the range of possible targets for a secondary drug in combination therapy. Second, it will help to define those signaling networks responsible for resistance to drug action and inform the design of the optimal strategy of combination therapy to overcome de novo and adaptation resistance [77]. Further research should be carried out to elucidate the role of the first drug as a primer to initiate network reprogramming. It has been established that the drug can retain its activity even at de novo resistance and activate network reprogramming in some cancer lines. For example, in the case of trastuzumab, it is assumed that this drug is active in primary trastuzumab resistance tumors and causes gene expression and reprogramming of signaling networks [3,5]. Moreover, a clinical trial of HER2 1 breast cancer progression following HER2 inhibition by trastuzumab showed that chemotherapy in combination with trastuzumab was more effective than chemotherapy alone. This finding suggests that trastuzumab sensitizes cellular response to a second drug despite the fact that cancer cells are insensitive de novo to trastuzumab and that there is a therapeutic benefit to continue trastuzumab therapy in combination with other drugs beyond progression [80]. In order to develop drug therapy targeting dynamic reprogramming signaling networks in cancer cells and systems biomarkers of this process, it requires proper combination therapy. Further, it is necessary to bind tightly drug development, diagnostics, and therapy, following CDx and drug codevelopment strategies in combination therapy design [81]. Novel compounds and their combinations should be developed in a combinatorial context with in vitro CDx assays, where codevelopment of in silico CDx assays can give an integrative benefit with respect to the design of systems biomarkers for suppression of adaptation drug resistance.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

40

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

3.6 SYSTEMS BIOMARKER IDENTIFICATION BY EXPLORING CONTEXTUAL GENOMIC FEATURES A complex cellular system of higher organisms consists of a huge number of processes running simultaneously. Maintaining a particular cellular state is one of the major tasks regulated in a synchronous way in a cell. Transition of cellular state from one state to another requires the alterations in regulatory pathways as well, which consists of a large number of genes functioning to get a particular state. Such genes with functional relatedness are called contextual genes [82]. Genome context methods were used to predict this functional relatedness between genes using the patterns of existence and relative locations of the homologs. While annotating functional and structural information of genes in newly sequenced genomes, homology is used as reliable tools, we explore contextual genes in targeted genomes to find out their functional relatedness. Widely used context methods in literature are phylogenetic profiles, gene neighbor, gene cluster, and gene fusion (Fig. 3.2) [82]. The important criteria for these methods are the presence of homologous sequences for the genes in the target genome where in this case, the degree of sequence similarity is checked. The contextual gene methods are classified by and large into (1) full coverage and (2) restricted coverage with the former generating scores for all possible gene pairs from a genome, while the other essentially generates scores only for some pairs. Widely employed full-coverage methods are the phylogenetic profile (Fig. 3.3) and gene neighbor methods (Fig. 3.2) which are used to compare sequences based on the degree of full coverage, whereas the gene cluster and the gene fusion are a part of restricted coverage methods (Fig. 3.2). However, the contexts used to infer functional relationships between genes with or without sequence similarity based on evolutionary relationships can be biased in some cases [83]. To check this, crude statistics and normalization are done to enhance the performance and genome dependence [84]. In addition, combining scores from two different methods, machine learning and classification with parameter selection and tuning, would help to achieve better results. Among these methods, the gene neighbor methods are known to outperform the phylogenetic profile method by as much as 40% in sensitivity when compared with the gene cluster method at low sensitivities. Methods used for biomarker identification: Identifying biomarkers is one of the greatest challenges as different genotype phenotype correlation from large-scale biological data is assessed. There are a good number of text-mining, knowledge-based, and network-based methods to discover biomarkers from the literature [85]. These can be better made by constructing a resource based on indexed text or dictionary used from a finite state machine. One of the key components of identifying contextual genes is condition-specific interactions in biological networks. Valuable mechanistic insights can be derived from functional analyses of genomic data. A gist of commonly used data mining methods is given next: • Unpaired null hypothesis testing is a univariate filter method in which p-value serves as the evaluation measure for the discriminatory ability of variables [84]. • Principal component analysis (PCA) is an unsupervised projection method which calculates linear combinations of variables based on the variance of the original data space [86].

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.6 SYSTEMS BIOMARKER IDENTIFICATION BY EXPLORING CONTEXTUAL GENOMIC FEATURES

41

FIGURE 3.2 (A) Contextual gene methods and their classification, (B) an overview of phylogenetic profiles methods, (C) gene neighbor methods, and (D) gene fusion and gene cluster methods.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

42

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

FIGURE 3.3 An overview of context mining process flow. The experimental data measured from contextual information before annotations are drawn from various methods.

• Information gain is also a univariate filter method which calculates how a given feature separates data by pursuing reduction of entropy [85]. • ReliefF (RF) is a multivariate filter method in which RF score relies on the concept that values of a significant feature are correlated with the feature values of an instance of the same class and uncorrelated with the feature values of an instance of the other class [87]. • Associative voting is a multivariate filter method and uses a rule-based evaluation criterion by a special form of association rules and considers interaction among features [88]. • Unpaired Biomarker Identifier is a univariate filter method, a statistical approach in which score can be determined by combining a discriminant measure with a biological effect term appropriate for two class problems only [89].

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.6 SYSTEMS BIOMARKER IDENTIFICATION BY EXPLORING CONTEXTUAL GENOMIC FEATURES

43

• Support vector machine recursive feature elimination (SVM-REF) is an embedded selection method SVM-REF which optimizes the weights of SVM classifier to rank features [90]. • Random forest (RF) model is an embedded selection method and uses bagging and random subspace methods to construct a collection of decision trees aiming at identifying a complete set of significant features [91]. • Aggregating feature selection is an ensemble selection method which aggregates multiple feature selection results to a consensus ranking [92]. • Stacked feature ranking is an ensemble selection method which has stacked learning architecture to construct a consensus feature ranking by combining multiple feature selection methods [93]. • Wrapper approach evaluates the merit of a feature subset by accuracy estimates using a classifier and produces a subset of very few features that are dominated by stronger and uncorrelated attributes [85]. There are a few guilt-by-association studies to check these features [94]. • Paired Biomarker Identifier is a univariate filter method which uses a statistical evaluation score by combining biological effects [89]. Next-generation sequencing (NGS) has given us tremendous impetus in identifying the biomarkers and with the help of feature selection or classification methods, there arose an interest to discern deep learning approaches [95]. Several combinatorial based methods are known for the identification of biomarkers, for example, heuristic searches, hybrid methods, and exhaustive methods which have been reviewed [96]. There is still a paucity of resources employing these features. For example, a database of mutation gene drug relations for such classification features could prove to be a valuable resource for researchers interested in personalized medicine and genealogy [97]. For example, Wu et al. have developed a large-scale text mining system to generate molecular profiles of thyroid cancer using literature-based classification scoring schemes [97]. In this process, they are able to identify key genes and pathways for easy prioritizing diagnostic biomarkers for therapeutic use. Similarly, Li et al. prioritized cancer-related miRNAs using network-based approaches by employing the random-walk algorithm to elucidate miRNA-disease networks [98]. In the recent-past, knowledge-driven text mining, approaches have shown tremendous applications for extracting biomarker information covering all therapeutic areas [99]. These are preferentially used by taking MEDLINE publications containing the MeSH terms and other bibliometric analysis. Automated extraction of biomarkers using semantic graphs has also been attempted recently [100]. With the last few years, systems biology approaches clubbed with NGS more amenable in extracting the useful information [101], biomarker identification, and computationally expensive and challenging task methods such as knowledge graphs prove to be very useful in bringing exponential growth of biomedical literature and databases [99]. Structured knowledge is better represented in the form of knowledge graph where unique relationships are defined using the unified medical language system (UMLS) which integrates relationships extracted from MEDLINE abstracts [102]. The knowledge graph runs on a 1.8.3 Neo4j graph database, the 3,527,423 biomedical concepts are represented as vertices, with 68,413,238 relationships between them. The individual concepts

COMPANION AND COMPLEMENTARY DIAGNOSTICS

44

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

represent units of thought, which are atomic and unique. Two concepts can be connected to each other with one or more semantic relationships (also referred to as predicates), such as “causes” or “inhibits,” thereby forming a triple. In addition, ranking and filtering methods are applied by checking recall, precision, and cumulative gains. The list of potential biomarker compounds generated by the method could be retrieved by the aforementioned processes and later assigned to the error categories. This will also allow us to recognize and apply normalization methods for the disease and biomarker entities in biomedical publications by means of the biomedical named entity recognition (BioNER) system. A gene dictionary constituting a huge collection of terms referring to human genes and proteins is integrated with data from three biological databases. This is complemented by a disease Dictionary with UMLS Metathesaurus [103], a large, multipurpose, and multilingual thesaurus that contains millions of biomedical and health-related concepts, their synonymous names, and their known relationships. Both dictionaries are curated and extended semiautomatically using different rules to facilitate the matching task. Methodology focused on the extraction of disease-biomarker associations reported in the literature is not just limited to the methods as mentioned earlier but knowledge-driven approach takes advantage of the annotation of MEDLINE publications pertaining to biomarkers with MeSH terms, narrowing the search for specific publications and therefore minimizing the false-positive ratio. The application of this methodology is shown using a neural network based biomarker association extraction approach for cancer classification [99,104]. These are complemented by linear biomarker association network, a fully connected neural network algorithm which assumes that the expression level of a biomarker is associated with some other biomarkers and can be estimated using a linear combination. This is briefly described next.

3.7 MACHINE-LEARNING APPROACHES IN SYSTEMS GENOMICS AND PHARMACOLOGY: BIOMARKER IDENTIFICATION BASED ON OMICS DATA Systems biomarker identification based on the machine learning has allowed understanding the importance of machine-learning and systems genomics approaches for multiomics data [105]. Machine learning assists humans to analyze large and complex dataset, and it can be used to interpret large genomic datasets [106]. The interpretation of these datasets helps to predict the structure and functions in biological sequence [107]. Moreover, machine learning has been used to address important problems in genomic medicine which determine how DNA variation can affect to development of different diseases [108]. For instance, we have previously reviewed machine-learning methods which can be applied for identification of single-nuclide polymorphism in our previous assay [109]. Also, machine learning plays a significant role in process DNA sequence and detects genetic disorders [63]. Consequently, this technique has been extended to a wide range of application in biology [110]. The gene expression data measures the activity level of the particular genes [111]. The gene expression analysis has been used in different areas in genetic studies. One key aspect is to identify potential disease biomarkers [106]. As a method of biomarker

COMPANION AND COMPLEMENTARY DIAGNOSTICS

3.7 MACHINE-LEARNING APPROACHES IN SYSTEMS GENOMICS

45

identification, classification of cancer types [111] and cancer detection [112] were previously proposed based on gene expression data [111]. Further, Tan et al. [113] examined the gene expression dataset to identify the key features in the breast cancer. The expression profile can be derived from two technologies such as DNA microarray and RNA-Seq [114]. To analyze these types of data, several methods have been applied such as biological [115], statistical [116,117], machine learning [106,113,118], and deep learning [111,112,119 121]. Wide varieties of machine learning methods have been employed in the current literature for identification of biomarkers. Generally, there are two main focus areas on machine learning: prediction or interpretation, and it consists of data preprocessing, supervised, semisupervised, and unsupervised methods [106]. In terms of principles, predicting phenotypes from genomic data is generally considered as a supervised machine learning problem [108]. Unsupervised methods are useful for finding the structure in a unlabelled dataset [106]. The common machine learning algorithms which are used to analyze sequence are SVM and RF, optimized evidence-theoretic K-nearest neighbor, and covariance discriminant algorithm [107]. In spite of conventional machine learning methods, deep learning methods have been employed to make an accurate prediction and finding hidden structure in the genomic dataset [64,120]. In typical RNA-Seq, experiments produce data sets of hundreds or thousands of genes with smaller samples [122]. The increment of the data complexity, dimensionally reduction method applies to simplify the analysis [122]. Further, extracting biological information of this high dimensionality feature space is often poorly understood or overlooked in the data modeling and analysis [123]. As feature selection methods, linear and nonlinear methods are suggested in the current studies. The gene-wise linear statistical model [117] and overdispersed Poisson statistical model [116] have been used to analyze different expression for RNA-sequencing studies. Moreover, the linear method called Least Absolute Shrinkage and Selection Operator (LASSO) and t-test with deep neural network have been proposed and the result has shown that LASSO outperforms the t-test with deep neural network. PCA has been introduced due to classification error, and joint likelihood of a gene subset cannot be reliably estimated when data dimensionality is high [123]. On the contrary, the nonlinear machine learning models have been introduced previously to find the hidden structure of the gene expression and complicated nonlinear interaction [120,122]. Stack autoencoders could be used to reduce the 20,000 genes into fewer genes dataset, and better result can be obtained using deep learning models with considering the biological knowledge [119]. Consequently, the stacked denoizing autoencoder has employed to transform high-dimensional, noisy gene expression data to latent representation, and results have shown that this method is useful when identifying the highly interactive genes [112]. Deep learning in biomedicine: Deep learning has its natural strength in (1) automatically discovering multiple levels of data representation and achieving a better prediction models, (2) use of the intermediate variables for different levels in model building, (3) the development of optimal prediction models, (4) handling big data to build multilayer and multilevel prediction models, (5) capable of connection all the input and out layers to achieve better models. Modeling in biomedicine has many challenges: (1) target mismatch due to the data having tumor bases results rather than survival timeline, (2) functional

COMPANION AND COMPLEMENTARY DIAGNOSTICS

46

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

mismatch due to loss of function important to patients, (3) data mismatch and selection bias due to sampling aspects, (4) environments changes and data variations, (5) reactive environment due to patients’ current status on a disease change, and (6) causality and variables that are considered in the model are not correlated and cause overlap and boost of the model [124]. Cancer modeling is always a challenge to system biomedicine approaches. Many algorithms such as SVMs and artificial neural network (ANN) are regularly used for binary and multiple subtypes classification. SVM is more suitable for two subtypes in dividing the plane as accurately to distinguish the predive analytics. ANN handles more subtypes through its capacity of multilayers inputs and models output as accurately as possible with high precision. The use of disease prognosis is based on the good quality of a medical diagnosis and (1) the prediction of cancer susceptibility (risk assessment), (2) the prediction of cancer recurrence/local control, and (3) the prediction of cancer survival, while building the models. ANN is used for more than 30 years for predictive biomedical models on various diseases, especially on different types and subtypes of cancer. ANN has the capacity to handle few data on a subtype against the other subtype of cancer which has more sufficient data, due to the collection of a particular data limitations and availabilities [67]. Deep learning can radically transform the human wellness in health care as it involves more technology, models, and data from various sources concurrently than the present learning models that handle only the static and older data. When patients’ data are interlinked to the model and the model is regenerated through deep learning, we can achieve a better accurate model and build the prediction model for the systems biomarker development and better treatment of different diseases.

3.8 CONCLUSION Rapid progress in experimental technologies of high-throughput data acquisition leads to deeper understanding of molecular and genetic mechanisms of tumorigenesis and discovery of more complex molecular taxonomy of cancer phenotypes characterizing by a high variety in patients’ responses to the targeted therapy. Further classification of cancer types according to tumor genotype and phenotype features of individual patients requires the development of the next generation of predictive biomarkers based on the integration of multiomics experimental and clinical data. The application of current computational and experimental methods and tools of systems biology allows the development of systems biomarkers of drug efficacy that significantly expands a concept of the well-established biomarkers identified before high-throughput era and takes into account the complex responsive network environment surrounding a drug target. We suggest, that systems biomarkers can be developed based on the computational framework, an in silico CDx assay, designed complementary to in vitro CDx assay based on large-scale omics data generated during all stages of drug-biomarker codevelopment. The identification of systems biomarkers based on the in silico CDx assay model is critically important at the current stage of the transition to phenotype-driven drug development, and their introduction can significantly contribute to the treatment of complex diseases such as cancer, diabetes, neurodegenerative, and cardiovascular diseases.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

REFERENCES

47

Acknowledgments AG acknowledges Scottish Informatics and Computer Science Alliance (SICSA) for partial support of his post.

References [1] Twomey JD, Brahme NN, Zhang B. Drug-biomarker co-development in oncology—20 years and counting. Drug Resist Updat 2017;30:48 62. [2] Kandoth C, et al. Mutational landscape and significance across 12 major cancer types. Nature 2013;502 (7471):333 9. [3] Wilken JA, Maihle NJ. Primary trastuzumab resistance: new tricks for an old drug. Ann NY Acad Sci 2010;1210:53 65. [4] Nahta R, Esteva FJ. Herceptin: mechanisms of action and resistance. Cancer Lett 2006;232(2):123 38. [5] Narayan M, Wilken JA, Harris LN, Baron AT, Kimbler KD, Maihle NJ. Trastuzumab-induced HER reprogramming in ‘resistant’ breast carcinoma cells. Cancer Res 2009;69(6):2191 4. [6] Nahta R, O’Regan RM. Evolving strategies for overcoming resistance to HER2-directed therapy: targeting the PI3K/Akt/mTOR pathway. Clin Breast Cancer 2010;10(Suppl. 3):S72 8. [7] Briffa R, et al. Multi-scale genomic, transcriptomic and proteomic analysis of colorectal cancer cell lines to identify novel biomarkers. PLoS One 2015;10(12):e0144708. [8] Jørgensen JT. The importance of predictive biomarkers in oncology drug development. Expert Rev Mol Diagn 2016;16(8):807 9. [9] Perez-Gracia JL, et al. Strategies to design clinical studies to identify predictive biomarkers in cancer research. Cancer Treat Rev 2017;53:79 97. [10] Henry NL, Hayes DF. Cancer biomarkers. Mol Oncol 2012;6(2):140 6. [11] Renfro LA, Mallick H, An M-W, Sargent DJ, Mandrekar SJ. Clinical trial designs incorporating predictive biomarkers. Cancer Treat Rev 2016;43:74 82. [12] Jørgensen JT, Hersom M. Companion diagnostics—a tool to improve pharmacotherapy. Ann Transl Med 2016;4(24):482. [13] Azmi AS, Wang Z, Philip PA, Mohammad RM, Sarkar FH. Proof of concept: network and systems biology approaches aid in the discovery of potent anticancer drug combinations. Mol Cancer Ther 2010;9 (12):3137 44. [14] Faratian D, et al. Systems biology reveals new strategies for personalizing cancer medicine and confirms the role of PTEN in resistance to trastuzumab. Cancer Res 2009;69(16):6713 20. [15] Carracedo A, Pandolfi PP. The PTEN-PI3K pathway: of feedbacks and cross-talks. Oncogene 2008;27 (41):5527 41. [16] Edelman LB, Eddy JA, Price ND. In silico models of cancer. Wiley Interdiscip Rev Syst Biol Med 2010;2 (4):438 59. [17] Klipp E, Wade RC, Kummer U. Biochemical network-based drug-target prediction. Curr Opin Biotechnol 2010;21(4):511 16. [18] Goltsov A, Langdon SP, Goltsov G, Harrison DJ, Bown J. Customizing the therapeutic response of signaling networks to promote antitumor responses by drug combinations. Front Oncol 2014;4:13. [19] Sharial M, Crown J, Hennessy BT. Overcoming resistance and restoring sensitivity to HER2-targeted therapies in breast cancer. Ann Oncol 2012;23(12):3007 16. [20] Stemke-Hale K, et al. An integrative genomic and proteomic analysis of PIK3CA, PTEN, and AKT mutations in breast cancer. Cancer Res 2008;68(15):6084 91. [21] Yarden Y, Sliwkowski MX. Untangling the ErbB signalling network. Nat Rev Mol Cell Biol 2001;2(2):127 37. [22] Schoeberl B, et al. Therapeutically targeting ErbB3: a key node in ligand-induced activation of the ErbB receptor-PI3K axis. Sci Signal 2009;2(77):ra31. [23] Samaga R, et al. The logic of EGFR/ErbB signaling: theoretical properties and analysis of high-throughput data. PLoS Comput Biol 2009;5(8):e1000438. [24] Normanno N, et al. The ErbB receptors and their ligands in cancer: an overview. Curr Drug Targets 2005;6 (3):243 57.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

48

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

[25] Brix D, Clemmensen K, Kallunki T. When good turns bad: regulation of invasion and metastasis by ErbB2 receptor tyrosine kinase. Cells 2014;3(1):53 78. [26] Landgraf R. HER2 therapy. HER2 (ERBB2): functional diversity from structurally conserved building blocks. Breast Cancer Res 2007;9(1):202. [27] Fla˚geng MH, Knappskog S, Haynes BP, Lønning PE, Mellgren G. Inverse regulation of EGFR/HER1 and HER2-4 in normal and malignant human breast tissue. PLoS One 2013;8(8):e74618. [28] Sassen A, et al. Cytogenetic analysis of HER1/EGFR, HER2, HER3 and HER4 in 278 breast cancer patients. Breast Cancer Res 2008;10(1):R2. [29] Alimandi M, et al. Cooperative signaling of ErbB3 and ErbB2 in neoplastic transformation and human mammary carcinomas. Oncogene 1995;10(9):1813 21. [30] Goltsov A, et al. Systems analysis of drug-induced receptor tyrosine kinase reprogramming following targeted mono- and combination anti-cancer therapy. Cells 2014;3(2):563 91. [31] Nagumo Y, Faratian D, Mullen P, Harrison DJ, Hasmann M, Langdon SP. Modulation of HER3 is a marker of dynamic cell signaling in ovarian cancer: implications for pertuzumab sensitivity. Mol Cancer Res 2009;7 (9):1563 71. [32] Nam S, et al. A pathway-based approach for identifying biomarkers of tumor progression to trastuzumabresistant breast cancer. Cancer Lett 2015;356(2 Pt B):880 90. [33] Duncan JS, et al. Dynamic reprogramming of the kinome in response to targeted MEK inhibition in triplenegative breast cancer. Cell 2012;149(2):307 21. [34] Garrett JT, et al. Transcriptional and posttranslational up-regulation of HER3 (ErbB3) compensates for inhibition of the HER2 tyrosine kinase. Proc Natl Acad Sci USA 2011;108(12):5021 6. [35] Grovdal LM, Kim J, Holst MR, Knudsen SL, Grandal MV, van Deurs B. EGF receptor inhibitors increase ErbB3 mRNA and protein levels in breast cancer cells. Cell Signal 2012;24(1):296 301. [36] Jørgensen JT. Companion diagnostics in oncology—current status and future aspects. Oncology 2013;85 (1):59 68. [37] Yap TA, Omlin A, de Bono JS. Development of therapeutic combinations targeting major cancer signaling pathways. J Clin Oncol 2013;31(12):1592 605. [38] Swain SM, et al. Pertuzumab, trastuzumab, and docetaxel for HER2-positive metastatic breast cancer (CLEOPATRA study): overall survival results from a randomised, double-blind, placebo-controlled, phase 3 study. Lancet Oncol 2013;14(6):461 71. [39] Franklin MC, Carey KD, Vajdos FF, Leahy DJ, de Vos AM, Sliwkowski MX. Insights into ErbB signaling from the structure of the ErbB2-pertuzumab complex. Cancer Cell 2004;5(4):317 28. [40] Lee-Hoeflich ST, et al. A central role for HER3 in HER2-amplified breast cancer: implications for targeted therapy. Cancer Res 2008;68(14):5878 87. [41] Chandarlapaty S. Negative feedback and adaptive resistance to the targeted therapy of cancer. Cancer Discov 2012;2(4):311 19. [42] Li J, Kim SG, Blenis J. Rapamycin: one drug, many effects. Cell Metab 2014;19(3):373 9. [43] Meric-Bernstam F, et al. PIK3CA/PTEN mutations and Akt activation as markers of sensitivity to allosteric mTOR inhibitors. Clin Cancer Res 2012;18(6):1777 89. [44] Noh W-C, et al. Determinants of rapamycin sensitivity in breast cancer cells. Clin Cancer Res 2004;10 (3):1013 23. [45] Seront E, Pinto A, Bouzin C, Bertrand L, Machiels J-P, Feron O. PTEN deficiency is associated with reduced sensitivity to mTOR inhibitor in human bladder cancer through the unhampered feedback loop driving PI3K/Akt activation. Br J Cancer 2013;109(6):1586 92. [46] Steelman LS, et al. Suppression of PTEN function increases breast cancer chemotherapeutic drug resistance while conferring sensitivity to mTOR inhibitors. Oncogene 2008;27(29):4086 95. [47] Komurov K, et al. The glucose-deprivation network counteracts lapatinib-induced toxicity in resistant ErbB2positive breast cancer cells. Mol Syst Biol 2012;8:596. [48] Mateo F, et al. Stem cell-like transcriptional reprogramming mediates metastatic resistance to mTOR inhibition. Oncogene 2017;36(19):2737 49. [49] Stewart-Ornstein J, Lahav G. Integrating genomic information and signaling dynamics for efficient cancer therapy. Curr Opin Syst Biol 2017;1:38 43. [50] Cavallo F, De Giovanni C, Nanni P, Forni G, Lollini P-L. 2011: the immune hallmarks of cancer. Cell 2011;60 (3):319 26.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

REFERENCES

49

[51] Shi Y, et al. Trastuzumab triggers phagocytic killing of high HER2 cancer cells in vitro and in vivo by interaction with Fcγ receptors on macrophages. J Immunol 2015;194(9):4379 86. [52] Salama AKS, Moschos SJ. Next steps in immuno-oncology: enhancing antitumor effects through appropriate patient selection and rationally designed combination strategies. Ann Oncol 2016;28(1):mdw534. [53] van Hasselt JGC, van der Graaf PH. Towards integrative systems pharmacology models in oncology drug development. Drug Discov Today Technol 2015;15:1 8. [54] Iyengar R, Zhao S, Chung S-W, Mager DE, Gallo JM. Merging systems biology with pharmacodynamics. Sci Transl Med 2012;4(126):126ps7. [55] Agoram BM, Demin O. Integration not isolation: arguing the case for quantitative and systems pharmacology in drug discovery and development. Drug Discov Today 2011;16(23 24):1031 6. [56] Zhang X-Y, Birtwistle MR, Gallo JM. A general network pharmacodynamic model based design pipeline for customized cancer therapy applied to the VEGFR pathway. CPT Pharmacometrics Syst Pharmacol 2014;3(1):e92. [57] Traynard P, Tobalina L, Eduati F, Calzone L, Saez-Rodriguez J. Logic modeling in quantitative systems pharmacology. CPT Pharmacometrics Syst Pharmacol 2017;6(8):499 511. [58] Eduati F, et al. Drug resistance mechanisms in colorectal cancer dissected with cell type specific dynamic logic models. Cancer Res 2017;77(12):3364 75. [59] Eduati F, Jaaks P, Merten CA, Garnett MJ, Saez-Rodriguez J. Patient-specific logic models of signaling pathways from screenings on cancer biopsies to prioritize personalized combination therapies. bioRxiv 2018;422998. Available from: https://doi.org/10.1101/422998. [60] Hickman JA, et al. Three-dimensional models of cancer for pharmacology and cancer cell biology: capturing tumor complexity in vitro/ex vivo. Biotechnol J 2014;9(9):1115 28. [61] Al-Lamki RS, Bradley JR, Pober JS. Human organ culture: updating the approach to bridge the gap from in vitro to in vivo in inflammation, cancer, and stem cell biology. Front Med 2017;4:148. [62] Jamieson LE, Harrison DJ, Campbell CJ. Chemical analysis of multicellular tumour spheroids. Analyst 2015;140(12):3910 20. [63] Manogaran G, Vijayakumar V, Varatharajan R, Malarvizhi Kumar P, Sundarasekar R, Hsu C-H. Machine learning based big data processing framework for cancer diagnosis using hidden Markov model and GM clustering. Wirel Pers Commun 2018;102(3):2099 116. [64] Min S, Lee B, Yoon S. Deep learning in bioinformatics. Brief Bioinform 2016;18(5):bbw068. [65] Camacho DM, Collins KM, Powers RK, Costello JC, Collins JJ. Next-generation machine learning for biological networks. Cell 2018;173(7):1581 92. [66] Lynch CM, et al. Prediction of lung cancer patient survival via supervised machine learning classification techniques. Int J Med Inform 2017;108:1 8. [67] Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, Fotiadis DI. Machine learning applications in cancer prognosis and prediction. Comput Struct Biotechnol J 2015;13:8 17. [68] Swinney DC. Phenotypic vs. target-based drug discovery for first-in-class medicines. Clin Pharmacol Ther 2013;93(4):299 301. [69] Blu¨thgen N, et al. A systems biological approach suggests that transcriptional feedback regulation by dualspecificity phosphatase 6 shapes extracellular signal-related kinase activity in RAS-transformed fibroblasts. FEBS J 2009;276(4):1024 35. [70] Hu H, et al. Feedforward and feedback regulation of the MAPK and PI3K oscillatory circuit in breast cancer. Cell Signal 2013;25(1):26 32. [71] Nagashima T, et al. Quantitative transcriptional control of ErbB receptor signaling undergoes graded to biphasic response for cell differentiation. J Biol Chem 2007;282(6):4045 56. [72] Nakakuki T, et al. Ligand-specific c-Fos expression emerges from the spatiotemporal control of ErbB network dynamics. Cell 2010;141(5):884 96. [73] Avraham R, Yarden Y. Feedback regulation of EGFR signalling: decision making by early and delayed loops. Nat Rev Mol Cell Biol 2011;12(2):104 17. [74] Khalil HS, Goltsov A, Langdon SP, Harrison DJ, Bown J, Deeni Y. Quantitative analysis of NRF2 pathway reveals key elements of the regulatory circuits underlying antioxidant response and proliferation of ovarian cancer cells. J Biotechnol 2015;202:12 30. [75] Zhang J, et al. TGF-β-induced epithelial-to-mesenchymal transition proceeds through stepwise activation of multiple feedback loops. Sci Signal 2014;7(345):ra91.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

50

3. SYSTEMS BIOLOGY IN BIOMARKER DEVELOPMENT FOR CANCER SIGNALING THERAPY

[76] Mo Q, et al. Pattern discovery and cancer gene identification in integrated cancer genomic data. Proc Natl Acad Sci USA 2013;110(11):4245 50. [77] Lee MJ, et al. Sequential application of anticancer drugs enhances cell death by rewiring apoptotic signaling networks. Cell 2012;149(4):780 94. [78] Ferretta A, et al. New insight into the role of metabolic reprogramming in melanoma cells harboring BRAF mutations. Biochim Biophys Acta 2016;1863(11):2710 18. [79] Xu AM, Huang PH. Receptor tyrosine kinase coactivation networks in cancer. Cancer Res 2010;70 (10):3857 60. [80] von Minckwitz G, et al. Trastuzumab beyond progression in human epidermal growth factor receptor 2positive advanced breast cancer: a German breast group 26/breast international group 03-05 study. J Clin Oncol 2009;27(12):1999 2006. [81] Olsen D, Jørgensen JT. Companion diagnostics for targeted cancer drugs—clinical and regulatory aspects. Front Oncol 2014;4:105. [82] Ferrer L, Dale JM, Karp PD. A systematic study of genome context methods: calibration, normalization and combination. BMC Bioinformatics 2010;11(1):493. [83] Krawczuk J, Łukaszuk T. The feature selection bias problem in relation to high-dimensional gene data. Artif Intell Med 2016;66:63 71. [84] Lehmann EL, Romano JP. Testing statistical hypotheses. Springer Science & Business Media; 2006. [85] Hall MA, Holmes G. Benchmarking attribute selection techniques for discrete class data mining. IEEE Trans Knowl Data Eng 2003;15(6):1437 47. [86] Ringne´r M. What is principal component analysis? Nat Biotechnol 2008;26(3):303. ˇ [87] Robnik-Sikonja M. Data generators for learning systems based on RBF networks. IEEE Trans Neural Netw Learn Syst 2016;27(5):926 38. [88] Osl M, et al. A new rule-based algorithm for identifying metabolic markers in prostate cancer using tandem mass spectrometry. Bioinformatics 2008;24(24):2908 14. [89] Baumgartner C, Lewis GD, Netzer M, Pfeifer B, Gerszten RE. A new data mining approach for profiling and categorizing kinetic patterns of metabolic biomarkers after myocardial injury. Bioinformatics 2010;26 (14):1745 51. [90] Guyon I, Weston J, Barnhill S, Vapnik V. Gene selection for cancer classification using support vector machines. Mach Learn 2002;46(1 3):389 422. [91] Enot DP, Beckmann M, Overy D, Draper J. Predicting interpretability of metabolome models based on behavior, putative identity, and biological relevance of explanatory signals. Proc Natl Acad Sci USA 2006;103(40):14865 70. [92] Saeys Y, Abeel T, de Peer Y. Robust feature selection using ensemble feature selection techniques. In: Joint European conference on machine learning and knowledge discovery in databases; 2008. p. 313 325. [93] Netzer M, et al. A new ensemble-based algorithm for identifying breath gas marker candidates in liver disease using ion molecule reaction mass spectrometry. Bioinformatics 2009;25(7):941 7. [94] Shin H, Sheu B, Joseph M, Markey MK. Guilt-by-association feature selection: identifying biomarkers from proteomic profiles. J Biomed Inform 2008;41(1):124 36. [95] Lee K, et al. Deep learning of mutation-gene-drug relations from the literature. BMC Bioinformatics 2018;19 (1):21. [96] Wang L, Wang Y, Chang Q. Feature selection methods for big data bioinformatics: a survey from the search perspective. Methods 2016;111:21 31. [97] Wu C, Schwartz J-M, Brabant G, Nenadic G. Molecular profiling of thyroid cancer subtypes using largescale text mining. BMC Med Genomics 2014;7(3):S3. [98] Li L, et al. Establishing reliable miRNA-cancer association network based on text-mining method. Comput Math Methods Med 2014;2014:746979. [99] Bravo A, Cases M, Queralt-Rosinach N, Sanz F, Furlong LI. A knowledge-driven approach to extract disease-related biomarkers from the literature. Biomed Res Int 2014;2014:253128. [100] Vlietstra WJ, et al. Automated extraction of potential migraine biomarkers using a semantic graph. J Biomed Inform 2017;71:178 89. [101] Mayer P, Mayer B, Mayer G. Systems biology building a useful model from multiple markers and profiles. Nephrol Dial Transplant 2012;27(11):3995 4002.

COMPANION AND COMPLEMENTARY DIAGNOSTICS

REFERENCES

51

[102] Bodenreider O. The unified medical language system (UMLS): integrating biomedical terminology. Nucleic Acids Res 2004;32(Suppl. 1):D267 70. [103] Unified medical language systems (UMLSs), www.nlm.nih.gov/research/umls/; 2018. [104] Kim S, Sen I, Bittner M. Mining molecular contexts of cancer via in-silico conditioning. Computational systems bioinformatics, vol. 6. World Scientific; 2007. p. 169 79. [105] Lin E, Lane H-Y. Machine learning and systems genomics approaches for multi-omics data. Biomark Res 2017;5(1):2. [106] Libbrecht MW, Noble WS. Machine learning applications in genetics and genomics. Nat Rev Genet 2015;16 (6):321 32. [107] Liu B. BioSeq-Analysis: a platform for DNA, RNA and protein sequence analysis based on machine learning approaches. Brief Bioinform 2017;bbx165:1 15. [108] Leung MKK, Delong A, Alipanahi B, Frey BJ. Machine learning in genomic medicine: a review of computational problems and data sets. Proc IEEE 2016;104(1):176 97. [109] Iddamalgoda L, Das PS, Aponso A, Sundararajan VS, Suravajhala P, Valadi JK. Data mining and pattern recognition models for identifying inherited diseases: challenges and implications. Front Genet 2016;7 (August):1 9. [110] Inza I, Calvo B, Arman˜anzas R, Bengoetxea E, Larran˜aga P, Lozano JA. In: Matthiesen R, editor. Machine learning: an indispensable tool in bioinformatics, 593. Totowa, NJ: Humana Press; 2010. p. 25 48. [111] Fakoor R, Ladhak F, Nazi A, Huber M. Using deep learning to enhance cancer diagnosis and classification. In: Proceedings of the international conference on machine learning, 28. 2013. [112] Danaee P, Ghaeini R, Hendrix DA. A deep learning approach for cancer detection and relevant gene identification. Pac Symp Biocomput 2017;22:219 29. [113] Tan J, Ung M, Cheng C, Greene CS. Unsupervised feature construction and knowledge extraction from genome-wide assays of breast cancer with denoising autoencoders. Pac Symp Biocomput 2015;20:132 43. [114] Stefano GB. Comparing bioinformatic gene expression profiling methods: microarray and RNA-Seq. Med Sci Monit Basic Res 2014;20:138 42. [115] Derveaux S, Vandesompele J, Hellemans J. How to do successful gene expression analysis using real-time PCR. Methods 2010;50(4):227 30. [116] Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 2010;26(1):139 40. [117] Ritchie ME, et al. Limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 2015;43(7):e47. [118] Pirooznia M, Yang JY, Yang MQ, Deng Y. A comparative study of different machine learning methods on microarray gene expression data. BMC Genomics 2008;9(Suppl. 1):S13. [119] Urda D, Montes-Torres J, Moreno F, Franco L, Jerez JM. Deep learning to analyze RNA-Seq gene expression data. Aliment Pharmacol Ther 2017;6(5):50 9. [120] Angermueller C, Pa¨rnamaa T, Parts L, Stegle O. Deep learning for computational biology. Mol Syst Biol 2016;12(7):878. [121] Chen Y, Li Y, Narayan R, Subramanian A, Xie X. Gene expression inference with deep learning. Bioinformatics 2016;32(12):1832 9. [122] Bzdok D, Altman N, Krzywinski M. Points of significance: statistics versus machine learning. Nat Methods 2018;15(4):233 4. [123] Clarke R, et al. The properties of high-dimensional data spaces: implications for exploring gene and protein expression data. Nat Rev Cancer 2008;8(1):37 49. [124] Wainberg M, Merico D, Delong A, Frey BJ. Deep learning in biomedicine. Nat Biotechnol 2018;36(9):829 38.

COMPANION AND COMPLEMENTARY DIAGNOSTICS