Molecular interaction networks and drug development: Novel approach to drug target identification and drug repositioning

Abstract

Conventional drug discovery requires identifying a protein target believed to be important for disease mechanism and screening compounds for those that beneficially alter the target's function. While this approach has been an effective one for decades, recent data suggest that its continued success is limited largely owing to the highly prevalent irreducibility of biologically complex systems that govern disease phenotype to a single primary disease driver. Network medicine, a new discipline that applies network science and systems biology to the analysis of complex biological systems and disease, offers a novel approach to overcoming these limitations of conventional drug discovery. Using the comprehensive protein–protein interaction network (interactome) as the template through which subnetworks that govern specific diseases are identified, potential disease drivers are unveiled and the effect of novel or repurposed drugs, used alone or in combination, is studied. This approach to drug discovery offers new and exciting unbiased possibilities for advancing our knowledge of disease mechanisms and precision therapeutics.

Abbreviations

ACE2

Angiotensin converting enzyme 2

ALDO

Aldosterone

BC

Betweenness centrality

BRAF^V600E

BRAF (gene) mutation

CAD

Coronary artery disease

CRBN

Cereblon

CM

Conditioned media

DNGR

Deep neural networks for graph representation

DTD

Drug‐target‐disease

Ev

Everolimus

GPSnet

Global positioning system network

HCM

Hypertrophic cardiomyopathy

JAK2‐STAT3‐COL4A2

Janus kinase‐2‐signal transducer and activator of transcription 3‐collagen 4A2

KCNH2

potassium voltage‐gated channel, subfamily H (eag‐related) member 2

MEK1^C121S

Mitogen‐activated protein kinase gene

MI

Myocardial infarction

MNM

Molecular network model

mTor

Mechanistic target of rapamycin

NEDD9

Neural precursor cell expressed developmentally down‐regulated protein 9

NKX2

Nk2 homeobox 5

Tem

Temsirolimus

TGF‐β

transforming growth factor‐beta

TNF‐α

Tumor necrosis factor‐alpha

PAH

Pulmonary arterial hypertension

PBMCs

Peripheral blood mononuclear cells

PDGFB

Platelet‐derived growth factor B

PGTS2

Prostaglandin endoperoxide synthase 2

PLX4032

Vemurafenib (PLX4032)

Pom

Pomalidomide

ROR‐ɤt

Receptor‐related orphan receptor‐gamma‐T

SARS

Severe acute respiratory syndrome

scRNA‐seq

Single‐cell RNA sequencing

SMAD3

Small mothers against decapentaplegic 3

1. CONTEMPORARY DRUG DEVELOPMENT AND ITS LIMITATIONS

It is widely recognized that the pharmaceutical industry has been central to the success of the biomedical enterprise in reducing morbidity and mortality of many diseases. Drug discovery has followed a well‐choreographed strategy derived from the early work of Ehrlich ¹ : A specific protein target is identified, and small‐molecule or biological therapeutics are developed that modify that target's function to prevent or attenuate manifestations of the underlying disease mechanism. While this approach has been successful for over a century, the drug development process has witnessed declining success over the last several decades with some analysts declaring that the industry is in the midst of an innovation crisis. ² That crisis is reflected in a decrease in the overall success rate of pharmaceutical agents, with an increase in the attrition rate for drugs in development, as well as an increase in the mean development time and the cost per new drug ultimately approved for use. Importantly, this crisis has occurred in the setting of increasing investment in research and development and, in particular, in new, cutting‐edge methods and technologies that would logically be expected to enhance the yield of approved drugs.

What accounts for this dichotomy between increasing investment in drug research and development and declining success? One possibility is an emphasis on seeking ‘blockbuster’ drugs that can be used for the broadest possible indication in the most inclusive population of patients with highly prevalent diseases. While this approach makes raw economic sense, it suffers from excessive risk and development costs, as most of these common diseases have many therapeutic options necessitating large clinical trials of new entities on a background of existing therapies. In addition, this development strategy represents the antithesis of more personalized, precision medicine. Schork highlighted this shortcoming by emphasizing that the top‐ten grossing drugs in the US only help between 4% and 25% of the patients treated with them, ³ illustrating what he termed ‘imprecision medicine.’

Another, and perhaps more critical, possibility is the continued adherence to the Ehrlichian search for a single drug target for each disease. Despite increasing recognition of the complexity of biological systems in health and disease, this driving principle remains central to the pharmaceutical industry and its regulatory agencies. Adverse off‐target effects, only some of which are routinely assessed during the drug screening process, are undoubtedly a consequence of the fallacy of the ‘magic bullet’ hypothesis of Ehrlich. While pharmacologists have long‐recognized that drugs are promiscuous in their interactions with protein targets, the magnitude of that promiscuity has only recently been assessed by Chartier and colleagues. ⁴ These investigators analyzed the extent to which the binding sites for 400 drugs are recapitulated in 14 082 binding pockets contained within 7895 different proteins with known tertiary structures. On average, they found 25 targets per drug, with 24 drugs having between 100–800 targets. This value represents a minimum, given that the set of tertiary structures of all human proteins is incomplete, and that other post‐translational (covalent) modifications of amino acid side chains are not considered. Clearly, this rich trove of drug targets offers an opportunity to explore both adverse off‐target effects, as well, and perhaps more important, as the therapeutic potential of (repositioning or repurposing) drugs for diseases for which they were not originally developed.

With this background, one can, for example, assess the consequences of repurposing drugs on the size of the effective drug space. Recent estimates suggest that there are ~3 × 10⁹ possible compounds comprising (organic) chemical space, among which ~3 × 10⁷ have been characterized from which ~1400 drugs have been developed. ⁵ These estimates indicate that only ~0.9% of all possible drugs (~1.5 × 10⁵) have been identified. Alternatively, if each drug has, on average, ~25 targets, the yield of the existing pool of approved drugs can be expanded to cover ~22% of potentially actionable drug targets in the human proteome, greatly reducing the de novo medicinal chemistry or chemical biology required for future drug discovery. To achieve this goal of improved, broadened drug use requires knowledge of all possible drug‐target interactions and their potential consequences for disease phenotypes, therapeutic mechanisms, and toxicities (which themselves are a consequence of undesirable drug‐target interactions), about which more will be discussed below.

2. NETWORK MEDICINE AND DISEASE MODULES

The field of network medicine is well‐poised as a discipline to address the often misleading excessive reductionism that underpins biomedical research and its associated drug discovery process. Network medicine was developed to provide a strategy for uncovering comprehensive mechanisms of disease and potential drug targets that modulate disease phenotype. This new field represents the integration of network science and systems biology as applied to biological systems in health and disease. In a concept paper in 2011, we proposed that within the universe of all physical protein–protein interactions (the interactome) in a cell, there exist subnetworks specific to each disease, so‐called disease modules ⁶ (Figure 1). This concept is predicated on four related hypotheses and organizing principles that link the interactome to human disease: (i) proteins involved in the same disease tend to interact, (ii) proteins involved in the same disease tend to cluster in connected subnetworks or disease modules, (iii) proteins in a disease module are often involved in the same biological process, and (iv) related diseases are located in the same neighborhood within the interactome from which unrelated diseases are excluded. ⁷ Subsequent studies by us ⁸ and others confirmed the correctness of this hypothesis, and have led to many fruitful efforts at identifying novel disease mechanisms by exploration of the unique, discrete disease modules in disorders ranging from preeclampsia ⁹ to asthma ¹⁰ to pulmonary arterial hypertension, ¹¹ among many others.

FIGURE 1.

Network medicine disease module hypothesis. The protein–protein interaction network (‘interactome’) is depicted on the left within which are subnetworks for specific diseases, denoted disease modules.

Using cerebrovascular disease as an example, ¹² Figure 2 depicts its disease module, which was derived from the interactome in the following way. We began by ascertaining the disease network ‘seed’ proteins, either from unbiased gene or protein expression screens or from the curated reductionist literature. We next mapped these proteins to the latest version of the comprehensive interactome, then applied a seed connector algorithm ¹² to determine whether or not the seed proteins form a discrete subnetwork in comparison to a randomly selected pool of proteins of similar number. In this case (as is the case in ~85% of all diseases studied to date), we found that the cerebrovascular disease module formed a distinct subnetwork in which the seed proteins were linked by not more than one additional connector protein in the interactome infrastructure. In effect, the interactome provides the missing protein links that facilitate the clustering of the seed proteins and define the disease module. Filling in these missing links, then, offers the opportunity to define pathways and interactions among potential disease proteins that were not previously recognized.

FIGURE 2.

Disease module construction. Disease module for cerebrovascular disease with blue nodes denoting seed (disease) proteins and gray nodes denoting connector proteins not more than one edge removed from another seed protein. [Reproduced with permission from Ref. 12.]

Defining specific disease modules in this way provides insight into disease mechanisms and, for the purpose of this perspective, novel drug targets and drug repurposing opportunities. ¹³ As an example of disease module‐based insight into disease mechanism, consider pulmonary arterial hypertension (PAH), a convergent phenotype that can be caused by several genetic and environmental factors. One of the key pathological features of PAH is perivascular fibrosis, which attenuates pulmonary vascular compliance, increases pulmonary vascular resistance, and thereby promotes increases in pulmonary arterial pressure, the phenotypic hallmark of the disease. As early as 2007, we proposed that disease could be classified based on the propensity of an individual to manifest pathological endophenotypes common to most diseases, including inflammation, thrombosis, fibrosis, and calcification. ¹⁴ Several years later, we demonstrated that in addition to disease modules in the interactome, there are also endophenotype modules each of which overlap with many specific disease modules. ¹⁵ In view of these findings, we explored the fibrosis module in PAH and defined interactome‐based determinants of adaptive (normal reparative) and pathological fibrosis. One protein common to both sub‐modules having high betweenness‐centrality, a network graph term used to define the importance of the node for the flow of information through the network, is the protein neutral precursor cell expressed developmentally downregulated 9 (NEDD9). NEDD9 interacts with small mothers against decapentaplegic 3 (SMAD3), which, in turn, is important for transforming growth factor‐beta (TGF‐β) signaling, mutations in the axis of which can cause PAH. We found that oxidation of ¹⁸Cys of NEDD9 (by aldosterone, levels of which are elevated in most individuals with PAH) impairs the SMAD3‐NEDD9 interaction, leading to increased NEDD9 binding to Nk2 homeobox 5 (NKX2‐5) with consequent increased NKX2‐5 binding to COL3A1, increased collagen III expression (Figure 3), and perivascular pulmonary fibrosis. Thus, the insight provided by the interactome‐based network analysis revealed a previously undisclosed critical determinant of the pathobiology of perivascular fibrosis in PAH.

FIGURE 3.

Novel drug target identification. (A) TGFβ‐dependent signaling pathway leading to perivascular fibrosis in pulmonary arterial hypertension, illustrating the role of aldosterone in promoting oxidation of NEDD9 (Ox‐NEDD9), thereby activating COL3A1 transcription and collagen III synthesis. (B) Proteins associated with adaptive fibrosis (blue), pathogenic fibrosis (red), or both (blue with red border) mapped to the consolidated interactome, defining the endophenotype module, the fibrosome. (C) Betweenness centrality (BC), a measure of importance of a particular protein (node) on information flow across a network and based on shortest path analysis, was calculated to identify NEDD9 as a critical node in the transition of endophenotype from adaptive fibrosis to pathogenic fibrosis. [Reproduced, in part, with permission from Ref. 11.]

3. NETWORK MEDICINE AND DRUG TARGET IDENTIFICATION

Network‐based identification of novel disease drivers within disease modules leads naturally to consideration of these drivers as potential drug targets. Validation of this approach can also be assessed by determining the proximity of known drug targets to the relevant disease module. In a interactome network study of acute myocardial infarction, ¹⁶ we incorporated drugs used to treat acute myocardial infarction (MI drugs); drugs that interact with MI drugs via common targets, transforming enzymes, and transporters (MI drug interactors and their MI‐related drug targets); drug targets (for the MI drugs and the MI drug interactors); and MI disease genes into the interactome, and demonstrated that MI drug targets are significantly proximate to the MI disease proteins comprising the MI disease module (Figure 4). Based on this outcome, we next constructed a bipartite network of MI‐related drug targets (targets of MI drugs and of MI drug interactors) and MI disease proteins, and identified 12 drug‐target‐disease (DTD) modules within which drugs used for other diseases are shown to affect targets that are near to the MI disease module, raising the possibility of their repurposing for acute MI (e.g., disulfiram, minocycline, and methimazole).

FIGURE 4.

Drug repurposing in myocardial infarction. The proximity relationships between myocardial infarction (MI)‐related drug targets and MI disease proteins in the interactome. (A) MI‐related drugs, drug targets, and MI disease proteins are mapped to the interactome. (B) Construction of a bipartite network of MI‐related drug targets and MI disease proteins, defining drug‐target‐disease modules. (C) Venn diagram illustrating overlap of MI‐related drug targets and MI disease proteins. (D) MI‐related drug targets and MI disease proteins have a significantly greater number of interactions (‘observed value’) than expected by chance. [Reproduced, in part, with permission from Ref. 16.]

We next turned to a population database‐centered approach to identifying potentially repurposable drugs for coronary artery disease (CAD), again using the interactome network as the basis for drug selection. ¹⁷ Using large healthcare databases that include over 220 million patients, we chose pairs of drugs used to treat diseases other than CAD; for each pair, one had a drug target within or near the CAD disease module, while the other did not. In the databases, we ascertained the incidence of CAD events over the course of one year. Comparing the anti‐epileptic agents carbamazepine and levetiracetam, after propensity matching and multiple adjustments to limit confounding, we found that carbamazepine, whose target is within the CAD module, is associated with an approximate 1.5‐fold increased risk of CAD compared with levetiracetam, whose target is distant from the CAD module. By contrast, comparing the immunomodulators hydroxychloroquine and leflunomide, we found that hydroxychloroquine, whose targets are within the CAD module, is associated with an approximate 0.75‐fold decreased risk of CAD compared with leflunomide. Experimental in vitro studies demonstrated that one mechanism by which hydroxychloroquine may provide this protective benefit is by suppressing inflammatory responses in the endothelial cell via its interaction with the toll‐like receptors, TLR7 and TLR9. ¹⁷

Drugs can also be repurposed to modulate pathological endophenotypes, as demonstrated for vascular calcification (Figure 5). In this example, we generated the vascular calcification module, or ‘calcificasome’, using the methods described above, then created a bipartite graph of drugs approved for other purposes and their known drug targets within the calcificasome. We identified everolimus and temsirolimus as drugs that target the mechanistic target of rapamycin kinase (mTOR), which resides within the calcificasome, as well as pomalidomide as a drug whose targets, tumor necrosis factor‐α (TNF), prostaglandin endoperoxide synthase 2 (PGTS2), and cereblon (CRBN), also reside or operate within the calcificasome. Experimental studies of vascular smooth muscle cell calcification then demonstrated that these three agents could suppress calcification significantly, again validating this interactome network‐based strategy for drug repositioning. ¹⁸

FIGURE 5.

Drug repurposing in vascular calcification. (A) The vascular calcification endophenotype module (‘calcificasome’), with node size correlating with degree or number of proteins to which a protein is bound, and node color indicating betweenness centrality (pink→purple, low→high). (B) Drug targets for everolimus/temsirolimus (upper) and pomalidomide (lower). (C) Human coronary artery smooth muscle cells were treated with everolimus (Ev), temsirolimus (Tem), or pomalidomide (Pom) in conditioned media (CM) for 10 days, and calcification assessed using Alizarin red staining (upper), semiquantitatively reported (lower). [Reproduced, in part, with permission from Ref. 18.]

Repurposing drugs for cancers represents yet another application of this drug development approach, and is based on two key elements of oncogenesis: the polygenic mutational basis for most malignancies, and the ready development of resistance to single drug therapies, including specific, targeted therapies. The former illustrates the complex genomic context within which cancers evolve, making them rife for a network medicine‐based analysis, while the latter demonstrates the futility of focusing on the conventional single drug target approach to cancer therapy, including personalized, targeted therapies. An early example of this limitation is shown in Figure 6, in a patient treated with the mutant v‐raf murine sarcoma viral oncogene homolog B1 (BRAF^V600E) inhibitor vemurafenib (PLX4032) to which the patient developed resistance as a result of a mutation in the downstream mitogen activated protein kinase kinase (MEK1^C121S) gene. ¹⁹ As a result of this mutation, this oncogenic pathway was reactivated and the tumor recurred. Recognizing this mechanism of resistance, Larkin and colleagues subsequently conducted a clinical trial using a combination of vemurafenib and the MEK inhibitor cobimetinib in previously untreated and unresectable melanoma patients, and demonstrated a clear survival benefit compared with the use of vemurafenib alone. ²⁰ This is but one of many examples of tumor relapse caused by selective pressure induced by a specific inhibitor in a key oncogenic pathway, limiting the longer‐term successes of many precise chemotherapies, and indicative of a need for rational combination therapies applied prior to the development of resistance. Network interactome‐based approaches offer an opportunity for identifying these rational combinations, rather than applying the usual semi‐empiric approaches, which are time‐consuming, may be associated with significant co‐morbidities, and are costly. To overcome some of these limitations of the conventional approach to cancer therapies, and to apply repurposing strategies effectively, we developed a genome‐wide positioning systems network (GPSnet) algorithm for drug repositioning by targeting disease modules based on the patient's individual sequencing and expression profiles, which are mapped to the interactome. ²¹ In this study, we analyzed whole‐exome sequencing and transcriptomes from 15 types of cancer in approximately 5000 patients. We were able to demonstrate that these disease modules reflecting differential gene expression can predict drug responses, and were also able to show the potential repurposing of 140 drugs approved for other purposes in some of these cancers, including, for example, niclosamide in the treatment of adenocarcinoma of the lung.

FIGURE 6.

Limitations of single drug target‐based therapy in cancer. A 38 year‐old man with mutant BRAF (V600E) melanoma is shown with extensive subcutaneous metastases (A) before treatment, (B) after 15 weeks of therapy with the mutant BRAF inhibitor vemurafenib, and (C) 23 weeks after treatment showing relapse, biopsy of which demonstrated a new mutation in the downstream kinase MEK1 (C121S) that sustained tumor recurrence. [Reproduced with permission from Ref. 19.]

Interactome network‐based drug repurposing for SARS‐CoV‐2 has also yielded promising results. In this case, the complexity of the interactome is further complicated by the interactions of the viral proteins with their human protein targets. As illustrated schematically in Figure 7, 26 of the SARS‐CoV‐2 proteins interact with 332 human proteins, of which 208 form a distinct disease module, the ‘covidome’, which can then be used to assess the proximity of approved drug targets to the covidome‐based disease drivers. To create a drug‐ranking algorithm, we used a combination of analytical approaches, including network proximity of drug targets to the covidome, network diffusion as a similarity metric for the effect of pairs of nodes on the network (including drug targets), and a graph convolutional network methodology. These three methods gave complementary information that greatly enhanced the predictive accuracy of the algorithm, which was assessed in a high‐throughput assay of viricidal activity and cytotoxicity using human (pulmonary or intestinal) epithelial cells and live SARS‐CoV‐2 virus. ²² Figure 8 demonstrates the results of this analysis, and illustrates the relationship between drug target proximity to the covidome and viricidal efficacy (strong, weak, or no effect). In addition, the results of a typical high‐throughput assay are shown, which indicates the ability of the algorithm to predict strongly antiviral drugs (compared with remdesivir as a positive control). A highly ranked drug in this screen, obatoclax (a BCL2 inhibitor), was then used in an in vivo assay in mice expressing human angiotensin‐converting enzyme 2 (ACE2) in epithelial cells (K18‐hACE2 trangenics), the receptor for the viral spike protein, showing significant suppression of PCR‐detectable viral titer in mouse lung. ²³

FIGURE 7.

Covidome. Schematic illustration of relationships among the SARS‐CoV‐2 viral proteins and their human (host) protein binding targets, which comprise a discrete subnetwork or Covid19 disease module, denoted the ‘covidome’, and the latter's interactome‐based relationship to drug targets whose drugs may be repositioned for treatment of Covid19 based on proximity calculations. [Adapted from Ref. 22.]

FIGURE 8.

The subnetwork formed by the drug targets of strong and weak drugs based on high‐throughput screening of those drugs in a dose‐dependent SARS‐CoV‐2 infectivity assay using human VeroE6 cells. Of the 77 total drugs that had strong or weak effects out of 918 (the remainder of which had no effect), purple denotes proteins targeted by strong drugs only; orange by weak drugs only, and pie charts illustrate proteins targeted by strong and weak drugs in which the chart‐based distribution reflects the proportional number of drugs in each category. (B) Drugs with no effect have a positive proximity z‐score to the covidome, which is interpreted as their being farther from the module than expected by chance, while strong and weak drugs have negative z‐scores, which is interpreted as their being closer to the module than expected by chance. (C) Viricidal effect of lead candidates predicted from in silico interactome‐based analysis on SARS‐CoV‐2 infection in primary human intestinal epithelial cells. After 3 days of incubation with virus and drug, cells were stained with viral N‐protein with a specific antibody (red) and host cell nuclei with Hoechst 33342 (blue). Bar graph illustrates viral staining relative to dimethylsulfoxide (DMSO) control. [Reproduced with permission from Refs. ^{22, 23}.]

The predictive yield of our network‐based algorithmic approach to drug repurposing is a marked improvement over conventional approaches. As a stochastic frame of reference, note that the standard ‘brute‐force’ screening method of 12 000 compounds proposed for drug repurposing for SARS‐CoV‐2 yielded positive results in 0.8%. ²⁴ By contrast, the network‐based algorithms applied here increased the yield in high‐throughput experimental screens ~ten‐fold, to 7.3%; with pharmacological curation (eliminating compounds with obvious toxicity, or with solubility or pharmacokinetic limitations), this predictive accuracy increased to 28%. ²² Notwithstanding this success, we must be mindful of the caveats of our analysis. First, we cannot determine a priori whether or not a computationally selected drug will improve or worsen the disease: in vitro testing remains essential to address this issue. Second, while repurposing drugs is assumed to be safe as these compounds have already passed toxicity assessment and population‐based clinical trial evaluation, new toxicities may emerge when treating a disease for which the drug was not originally developed (drug‐by‐disease interactions) (see below).

Deep learning approaches to network analysis represent yet another computational strategy that can be applied to drug target identification and drug repurposing. As an example, we have developed deepDTnet, a deep neural network model for graph representations (DNGR) as applied to heterogeneous drug‐gene‐disease networks embedding 15 types of chemical, genomic, phenotypic, and cellular network profiles. ²⁵ The workflow of deepDTnet and the network embedding and performance are shown in Figure 9. This algorithm was trained on 732 FDA‐approved small molecule drugs, and shows high predictive accuracy (AUC = 0.963) in identifying novel targets for approved drugs. As one example, deepDTnet predicted that topotecan, an approved topoisomerase inhibitor, can also inhibit the human retinoic acid receptor‐related orphan receptor‐gamma t (ROR‐γt), targeting of which showed potential therapeutic benefit in a mouse model of multiple sclerosis. ²⁵ As yet another computational approach that exploits multicellular network models with time‐dependent single‐cell RNAseq analysis following a perturbation, one can apply the notion of medical digital twins. Medical digital twins are computational disease models for drug discovery and treatment. Recently, we used this approach in the analysis of drivers of seasonal allergic rhinitis in individual subjects. ²⁶ Upstream drivers of the disease phenotype are assessed over time in peripheral blood mononuclear cells, giving insight into the patient‐specific disease determinants for more precise, targeted therapies. The strategy for this type of study is illustrated in Figure 10. ²⁶

FIGURE 9.

Deep neural network algorithm for drug target identification, deepDTnet. In this schematic illustration of the workflow of deepDTnet, 15 types of chemical, genomic, phenotypic, and cellular networks are embedded in order for the neural net to learn a low‐dimensional vector representation of the features for each node. The resulting feature matrices X and Y for drugs and drug targets, respectively, are then subjected to Pugh (PU)‐matrix completion by deepDTnet in order to define the best projection from the drug space onto the drug target space that optimizes the geometric proximity of the projected feature vectors of drugs to the feature vectors of their drug targets. DeepDTnet then infers new targest for a drug by virtue of its ranking by geometric proximity to the projected feature vector of the drug in the projected space (for details, see Ref. 25). [Reproduced with permission from Ref. 25.]

FIGURE 10.

Medical digital twins and drug target prioritization. (A) Peripheral blood mononuclear cells (PBMCs) from subjects with seasonal allergic rhinitis (SAR) (red) and non‐allergic controls (green) were stimulated with allergen (ragweed) or diluent control. (B) Major allergy‐related Th1/Th2 cytokines (INF‐γ, IL‐4, IL‐5, and IL‐13) were measured in the cell supernatants from SAR subjects (yellow) or non‐allergic controls (blue) over time of incubation. (C) ¹ Time‐dependent changes in single cell RNA‐seq (scRNA‐seq) of PBMCs following exposure to allergen. ² Multicellular network model (MNM) construction demonstrating predicted molecular interactions between cell types at each time point. ³ Ranking of upstream regulators by the number of cell types each is predicted to regulate at each time point. ⁴ Top‐ranking upstream regulator prioritization for those mediators that regulate the greatest number of cell types over the greatest number of time points, which is platelet‐derived growth factor B (PDGFB) in these experiments. ⁵ Experimental validation of two upstream regulators, IL4 and PDGFB, whose gene products were blocked by specific antibodies. For details, see Ref. 26. [Reproduced with permission from Ref. 26.]

Drug toxicities can also be predicted using interactome‐based strategies. Just as disease modules connote a subnetwork of proteins and pathways that govern disease phenotype, drug toxicity modules provide the same kind of mechanistic insight into pathways that govern adverse drug effects. As shown in a recent study, ²⁷ constructing drug toxicity modules and assessing their proximity to the disease modules for which a repurposed drug is being considered offer the opportunity to distinguish between those drugs whose therapeutic‐toxic ratio is acceptable versus those in which it may be adversely consequential and for which reason the drug should be avoided. For example, as shown in Figure 11, pitolisant, a drug approved for use in narcolepsy, has a drug target, the voltage‐activated potassium channel or hERG channel (KCNH2), which is contained within both the non‐ischemic cardiomyopathy disease module but also in the QT prolongation toxicity module. This shared target for (and implied proximity of) these modules would limit the likelihood that the drug could be repurposed for non‐ischemic cardiomyopathy. ²⁷

FIGURE 11.

Interactome‐based drug toxicity analysis. Schematic illustration of the localization of a target for the narcolepsy drug pitolisant, potassium voltage‐gated channel, subfamily H (eag‐related) member 2 (KCNH2), in both the (nonischemic) cardiomyopathy disease module and the long QT syndrome adverse effect module.

One final note relates to identifying individual variations in disease modules that can inform more precise drug target identification. We addressed this issue in detail in a study of hypertrophic cardiomyopathy (HCM) patients. ²⁸ Using a combination of differential gene expression comparing myocardial tissue excised at septal myectomy with normal myocardial tissue and interactome mapping of these differentially expressed genes within the HCM disease module, we were able to identify a wide variation of reticulotypes, or individual disease modules, that reflect different pathways that contribute to the this common, or convergent, phenotype (Figure 12). An analysis of these individual reticulotypes showed that among 18 patients analyzed in this way, extreme myocardial fibrosis in 2 appeared to be driven by the Janus kinase‐2‐signal transducer and activator of transcription 3‐collagen 4A2 (JAK2‐STAT3‐COL4A2) pathway, suggesting the potential for a specific therapy (JAK2 inhibition) in these patients.

FIGURE 12.

Patient‐specific interactomes (reticulotypes) in hypertrophic cardiomypathy (HCM). (A) Left ventricular (LV) myocardial biopsy specimens were obtained from rejected cardiac explants as healthy controls (N = 5) (C₁–C₅) and analyzed using RNA‐Seq, after which the Pearson correlation coefficient (r) was calculated for all gene (g) pairs. n, total combinations of pairwise correlations; m, total number of genes. (B) (Anterior) septal myectomy specimens were obtained from patients with HCM undergoing septal reduction surgery and analyzed using RNA‐Seq. The transcriptomic profile of an individual HCM patient was next added to the control gene expression matrix, and a new Pearson correlation coefficient (r′) was calculated for each gene pair. The HCM patient transcriptome was then removed from the matrix and the process repeated ad seriatim for the other HCM patients (n = 18). (C) (Step 1) Statistically significant differences in r and r′ coefficients were collated, and (Step 2) those significant gene pairs (g₁–g₂) were mapped to the interactome (Step 3). (Step 4) Gene pairs for which an interactome‐based interaction was identified were used to generate individual‐patient HCM interactomes (reticulotypes), examples of which from 5 patients are shown in (D). [Reproduced with permission from Ref. 28.]

4. CONCLUSIONS

Network medicine‐based approaches to drug target identification and drug development offer avenues for advancing therapeutics in the current era of big biological data. Drug target identification has only recently been viewed as a true systems‐based problem that warrants systems‐based solutions. ²⁹ The strategies I review here exploits the complexity of (macro)molecular interactions networks to provide those solutions. Doing so successfully requires the evolution of increasingly precise and sensitive tools to identify specific protein–protein interactions, such as cross‐linking mass spectrometry, cryo‐electron tomgraphy, ^{30 , 31 , 32} and proximity labeling. ³³ Coupled with the growing compendium of detailed protein data (e.g., the Human Protein Atlas) and deep learning‐based predictive algorithms of protein tertiary structure from primary sequence (e.g., AlphaFold), we are increasingly poised to explore the proteome for all possible protein–protein interactions in an unbiased and comprehensive way. Armed with these tools, the resulting molecular interaction network strategies reviewed here provide a means to address biological complexity, to understand disease mechanisms holistically, and, thereby, to identify potential drivers of disease phenotype that warrant consideration as drug targets. Furthermore, using the interactome template as a guide, rational repurposing of drugs, used alone or in combination, can be developed for optimal therapeutic efficacy. Coupled with increasingly precise delineations of individual disease networks or reticulotypes, precision therapeutics will become a growing reality.

DISCLOSURES

J.L. is scientific co‐founder and on BOD of Scipher‐Network Medicine Co., which uses network medicine analyses to identify disease biomarkers and potential therapies; consultant for Naring Health, its primary interest is biotech; and is on SAB for Applied BioMath, whose primary business is consulting.

Loscalzo J. Molecular interaction networks and drug development: Novel approach to drug target identification and drug repositioning. The FASEB Journal. 2023;37:e22660. doi: 10.1096/fj.202201683R

DATA AVAILABILITY STATEMENT

Not applicable, as this is a review, rather than original article.