We first outline the rationale of our novel control-hub-based method and present its primary steps. We then apply it to an integrated network constructed using human and SARS-CoV-2 PPI data and the data of drugs and drug targets. We compare our new method with nine existing gene selection methods, including the structural-controllability-based driver-node method, to show its performance in finding drug targets for Covid-19. We then examine the 65 drug targets and the corresponding drugs identified by our new methods, using the data and results in the literature for validation.
Total network controllability for drug repurposing
The primary concept of network structural controllability17, 18 is a control scheme for a network. It consists of control paths such that every node in the network can be reached or controlled by the head node of the control path to which the node belongs (Figure 1A). The head node is referred to as the driver or input node of the path. By exerting stimuli on the driver nodes, the network can be steered from any initial state to the designated state in a finite time. Structural controllability has been directly applied to repurposing drugs for treating Covid-19, where a small number of driver nodes targeted by drugs were used to find reusable drugs15, 16.
However, driver nodes are a double-edged sword and can also be exploited by viruses to infect the cell. Viral infections are exogenous stimuli to the cell via the interactions of viral proteins and host receptors. These can transform cells from normal to abnormal to accommodate viral replication and propagation. During SARS-CoV-2 infection, the viral spike protein S engages human receptor angiotensin-converting enzyme 2 (ACE2) to enter the host cell, triggering a series of adverse signaling cascades29.
Moreover, it is impractical to directly adopt structural controllability for controlling the cell or repurposing drugs. The control scheme is not unique (Figure 1A). An exponential number of control schemes may exist, and one control scheme may have as many as half of the nodes in the network as driver nodes. For example, one control scheme for the human PPI network25 (Table S1) contains 4,529 driver nodes, which are 49.8% of the 9,092 nodes in the network. Determining the best or an effective control scheme is a daunting task.
In light of these serious issues underlying the approaches to controlling the cell, we resorted to protecting the cell instead. We were motivated to identify critical genes, which, when perturbed, can render the cell uncontrollable by any control scheme or external stimulus on the driver nodes. Manipulating any of such critical genes can invalidate all the control schemes, so the cell is uncontrollable by undesired stimuli. To identify such critical genes, we extended structural controllability to total controllability by considering all control schemes and introducing a new concept of control hubs. A control hub is a middle node in one of the control paths of every control scheme (Figure 1B). Blocking a control hub will block at least one control path of every control scheme, making the overall network uncontrollable.
Therefore, control hubs are ideal drug targets for protecting the cell from being manipulated by viral infections. If the genes that viruses act on are known, the control hubs close to these nodes can be chosen as designated drug targets to increase drug efficacy.
Since the concept of control hubs is built atop all control schemes, a technical obstacle is the potentially exponential number of control schemes for a network. Finding all control schemes using the current best method, i.e., maximum matching30, is a computationally infeasible #P-complete problem24. To circumvent this difficulty, we developed an efficient, polynomial-time algorithm for finding all control hubs without computing all control schemes. The algorithm identified 1,256 control hubs in the human PPI network25, which are 13.8% of all the 9,092 genes and 27.7% of the 4,529 driver genes for the network (Table S1).
Control hubs can act as surrogates to reusable drugs, i.e., we focus on those existing drugs that can target control hubs. While in theory, any drug-targeted control hubs can be used, the ones closer to exogenous stimuli (i.e., viral proteins) are preferred over the distant ones since blocking the former may prevent the spread of external influences sooner and more effectively.
Finding drug targets for the treatment of viral infections
We capitalized on total controllability and control hubs and developed a drug-purposing method consisting of four major steps (Figure 1C, see Methods and Supplemental Method S3). The first is constructing a network to integrate information on human PPI, virus PPI, drugs, and their targets. We used the largest homogenous human PPI network25 (Table S1) and the data of PPIs between SARS-CoV-2 and human26, 27 (Table S2) and the data of drug targets28 (Table S3). The human PPI subnetwork and the virus PPI subnetwork are linked through the PPI between human and virus proteins, and the human PPI subnetwork and drug subnetwork are connected by the drug target information. The resulting network contains 9,092 nodes (proteins) from humans, 22 nodes from SARS-CoV-2, and 2,980 nodes of drugs. The overall network is relatively tight, with a total of 81,953 links.
The second step is to identify control hubs31. To focus on Covid-19, we left the technical details of our new method for finding control hubs to Methods and Supplemental Method S3. This control-hub finding method identified 1,256 control hubs in the network.
In the third step, to identify effective drug targets and drugs, we focused on the control hubs that were known targets of the existing drugs, categorically referred to as druggable control hubs hereafter. Among the 1,256 control hubs, 160 (12.7%) were drug targets (Figure 2A).
Druggable control hubs were not equally effective for treating SARS-CoV-2 infection. Some control hubs may directly interact with viral proteins and thus are ideal drug targets, whereas many others are far away from viral proteins in the human PPI network (Figure 1C). The closer a druggable control hub to virus proteins in the network, the more effective it should be for prohibiting viral infection.
Following this reasoning, in the fourth step, we examined the druggable control hubs in the community of proteins that were k steps away from the virus proteins in the PPI network, referred to as the k-step community, for convenience. A smaller k is preferred; the closer a control hub is to viral proteins, the more effective it is as a drug target to block viral infections. Two sets of enrichment tests, using the z-test, were performed to identify the best k-step community (see Methods). The first set of tests looked for the k-step community that was most enriched with control hubs among all k-step communities for different values of k, and the second set of tests assessed the enrichment of drug targets among the control hubs in the community chosen in the first test. The first z-test revealed that the 2-step community was most enriched with control hubs (z-score=5.28, p-value=1.3e-7, Figures 2B, S1A). It hosted 677 control hubs, among which 65 were drug targets (Table S4A). The second z-test confirmed that the 2-step community was also most enriched with druggable control hubs among all k-step communities (z-score=28.25, p-value=1.3e-175, for k=2, Figures 2C, S1B).
In the last step, we assessed if our novel control-hub approach was the method of choice for finding drug targets. In particular, we compared it with nine existing methods, including the driver-node-based method and eight popular node ranking methods. These included node-degree centrality, neighbor-degree centrality, betweenness centrality, load centrality, closeness centrality, and eigenvector centrality, as well as Page-Rank, and k-core32-41. To facilitate the comparison and better understand these methods, we compared them against a statistical model of drug targets in the 2-step community. Assuming that any protein in the 2-step community was equally likely to be a drug target, the drug-target enrichment for 677 (i.e., the number of control hubs in the 2-step community) randomly selected proteins in the community should follow an empirical normal distribution (Figure 2D). This empirical distribution was adopted as a statistical baseline model of drug-target enrichment. The enrichment of the 65 druggable control hubs in the 677 control hubs in the 2-step community substantially deviated from the baseline model (z-score=1.53, p-value=0.13; Figure 2D). Likewise, the drug-target enrichment for 677 driver nodes randomly chosen from 965 driver nodes in the 2-step community should also obey an empirical normal distribution (Figure 2D). The drug-target enrichment of our control-hub method was significantly better than that of the driver-node method (z-score=2.82, p-value=0.005). The driver-node method was slightly worse than the baseline model since the mean of the former was smaller than the mean of the latter (54.07 vs. 56.05; Figure 2D), and the two distributions were statistically indistinguishable (p-value = 0.98, c2-test; Figure 2D). We measured the drug-target enrichments of the top 677 nodes from the eight gene-ranking methods. Unfortunately, these methods all underperformed; their z-tests against the random baseline model all resulted in negative z-scores (Figure 2D). For instance, the Page-Rank method had a z-score=-1.89 with p-value=0.06.
This analysis showed that our novel control-hub method can identify the largest number of drug targets and candidate drugs for Covid-19 treatment.
Control hubs as drug targets for Covid-19 treatment
We examined the biological functions of the druggable control hubs to appreciate their role in SARS-Cov-2 infection and validate the new method using published results in the literature. Among all 160 druggable control hubs, three (RIPK1, CYB5R3, and COMT) directly interact with nonstructural proteins of SARS-CoV-226, 27 (Figures 3A, 3B, S2; Tables 1, S4A). RIPK1 can bind with viral nonstructural protein nsp1226, 27, the RNA-dependent RNA polymerase (RdRp) of SARS-CoV-242 (Figures 3A, 3B). nsp12 promotes viral replication and inhibits the host's innate immune response by suppressing the activity of interferon regulatory factor 3 (IRF3), which is key to interferon production43. Both CYB5R3 and COMT interact with the nsp7 protein of SARS-CoV-2 (Figures 3A, 3B), which forms a tetramer with viral nsp844 and functions as a cofactor of the viral RdRp, nsp1242. Since nsp12 and nsp7 are essential for viral transcription and replication, blocking the interactions of RIPK1 with nsp12, CYB5R3 with nsp7, and COMT with nsp7 can potentially inhibit or suppress viral replication.
RIPK1 encodes serine/threonine-protein kinase 1, plays a role in necroptosis, apoptosis, and inflammatory response, and mediates cell death and inflammation45. SARS-CoV-2 infection promotes the expression of RIPK1 in the lung of Covid-19 patients, and small-molecule inhibitors of RIPK1 can reduce the viral load of SARS-CoV-2 and proinflammatory cytokines in human lung organoids, indicating that the virus hijacks RIPK1-mediated immune response for its replication and propagation46. RIPK1 is targeted by Fostamatinib (Table 1, S4A; Figure 3A), a drug under intense scrutiny for treating SARS-CoV-2 infection47-52. Fostamatinib is an inhibitor of spleen tyrosine kinase originally approved for treating chronic immune thrombocytopenia. Fostamatinib is effective in a mouse model of acute lung injury and acute respiratory syndrome, symptoms observed in Covid-19 patients48. A clinical trial with a small sample of hospitalized Covid-19 patients (30 with fostamatinib versus 29 with placebo) showed that Fostamatinib could lower mortality, shorten the length of ICU stay, and reduce the disease severity of critically ill patients49.
CYB5R3 encodes NADH-cytochrome B5 reductase 3, a flavoprotein with oxidation functions. It is targeted by three drugs (Tables 1, S4A), two of which (NADH and Flavin adenine dinucleotide) are under clinical investigation for Covid-19 treatment. NADH is an energy booster for treating chronic fatigue syndrome and improving high blood pressure and jet lag, among many other symptoms. NADH, i.e., nicotinamide adenine dinucleotide (NAD)+ hydrogen (H), is the central catalyst of cellular metabolism, a chemical naturally produced in humans and plays a role in ATP production. The SARS-CoV-2 genome does not encode enzymes for ATP generation, and the virus needs to hijack host functions for viral synthesis and assembly. Therefore, NAD is a battlefield for viral infection and host immunity53. Indeed, coronavirus infection dysregulates the NAD metabolome, as indicated in a preclinical study54. Moreover, early phases 2 and 3 clinical trials showed that medication of NADH in a mixture of two metabolic activators could significantly shorten the time to complete recovery of SARS-CoV-2 infection55.
COMT encodes catechol-O-methyltransferase that can degrade estrogens, catecholamines, and neurotransmitters such as dopamine, epinephrine, and norepinephrine. It is targeted by 14 FDA-approved drugs, including Conjugated estrogens (Tables 1, S4A). Conjugated estrogens are a mixture of estrogen hormones for treating hypoestrogenism-related symptoms. Estrogen has been indicated as a susceptibility factor of SARS-CoV-2 infection56, as women are less susceptible to Covid-1957, 58 and mice with weaker estrogen receptor signaling due to respiratory coronavirus infection exhibit increased morbidity and mortality59.
Beyond the three druggable control hubs that directly interact with viral proteins, 19 druggable control hubs in the 2-step community engage more than one viral protein via another protein, and four of them (SLC10A1, SLC10A6, MUC1, and TTPA) are targeted by more than one drug (Tables 1, S4A; Figure S2). The potential of these four druggable control hubs for Covid-19 treatment is discussed in Supplemental Result S1.
In short, the 65 druggable control hubs within the 2-step community were enriched with biological functions related to cell (particularly leukocyte) proliferation, cellular response to (chemical) stress, regulation of apoptotic signaling, and response to nutrient levels (Figure 3C). All these results combined revealed the essential roles these control hubs might play in prohibiting the replication and proliferation of SARS-CoV-2. The results also revealed the essential immune-related signaling pathways induced by the virus and paved the way for understanding and explaining the therapeutic mechanisms of the drugs for Covid-19 treatment.
Drugs for the treatment of SARS-CoV-2 infection
The 65 druggable control hubs within the 2-step community were targeted by 185 existing drugs (Tables 2, S5; Figure 3D). As of June 2022, 38 were under clinical trials (https://clinicaltrials.gov/ct2/home). It is desirable to use drugs with multiple targets to gain treatment efficacy; the potency of a drug can be estimated by the number of control hubs it targets. Remarkably, 15 drugs target more than one control hub, and seven target more than two druggable control hubs (Tables 2, S5).
Among the seven drugs targeting more than two control hubs were Fostamatinib, NADH, and three dietary calcium supplements (Tables 2, S5). Fostamatinib is in phase 3 clinical trial after a promising phase 2 trial for Covid-19 treatment49. Experimental and clinical data showed that Fostamatinib inhibits neutrophil extracellular traps (NETs), which entrap and eliminate pathogens during viral and bacterial infections and may cause adverse injury to surrounding tissues by themselves or by increasing proinflammatory responses60 (Figure 3E). Activation and overreaction of innate and adaptive inflammatory responses during SARS-Cov-2 infection induce NETs, contributing to immunothrombosis in ARDS commonly seen in Covid-19 patients47, 61-63. Moreover, coherent antiviral therapeutic functions of Fostamatinib emerged after examining the functions of the control hubs that the drug targets (Figures 3D, 3E; Table S5). Among the ten control hubs that Fostamatinib targets, 7 (RIPK1, CLK2, CLK3, PAK5, STK3, PKN1, and CDK4) are serine/threonine type protein kinases, and two (BLK and YES1) encode Src family tyrosine kinases, all of which play essential roles in cell proliferation, cell differentiation, and programmed cell death64. CLK2 and CLK3 encode members of the serine/threonine type protein kinase family, and PAK5, STK3, PKN1, and CDK4 encode, respectively, one of the three members of the group II PAK family of serine/threonine kinases, serine/threonine-protein kinase 3, serine/threonine protein kinase N, and cyclin-dependent serine/threonine kinase. Plus, RIPK1 encodes receptor-interacting serine/threonine-protein kinase 1 and directly interacts with the viral RdRp nsp12, as discussed earlier. Interestingly, while not being a kinase, the remaining target COQ8A encodes a mitochondrial protein functioning in an electron-transferring membrane protein complex in the respiratory chain. Its expression is induced by the tumor suppressor p53 in response to DNA damage, and inhibition of its expression suppresses p53-induced apoptosis. Combined, the inhibitory function on NETs and kinase functions of 9 of the ten control hubs targeted by Fostamatinib suggested it to be potent for Covid-19 treatment by acting broadly on components of autoimmune, tumor repression, and inflammatory viral response pathways (Figure 3E).
NADH targets 5 control hubs, including CYB5R3 and NDUFB7. CYB5R3 encodes NADH-cytochrome B5 reductase 3, and NDUFB7 is a subunit of the multi-subunit NADH:ubiquinone oxidoreductase. NDUFB7 functions in the mitochondrial inner membrane and has NADH dehydrogenase and oxidoreductase activities. It has been reported that the NADH level was decreased in Covid-19 patients53, and coronavirus infection dysregulates the NAD metabolome54, so medication of NADH plays a role in attenuating the impact of virus infection.
The three dietary calcium supplements, Calcium Citrate, Calcium Phosphate, and Calcium phosphate dihydrate, target three control hubs, including S100A13 and PEF1, which are calcium-binding proteins. Several clinical studies have indicated a low serum calcium level as a prognostic factor of the mortality, severity, and comorbidity of SARS-CoV-2 infection65-67. As a side note, six vitamin E-related drugs targeting control hub TTPA (Table S5), which encodes a soluble protein that is a form of vitamin E (Tables 1, S5), have entered clinical trials for Covid-19 treatment. These results indicated that calcium, vitamin E, and many other micronutrients should be adopted as adjuvant therapy against viral infection.
In summary, the repurposed drugs fall into four major categories (Table S5), 1) antiviral and anti-inflammatory agents that are subscribed for virus infection and cancer treatment, 2) dietary supplements including NADH and Calcium that boost human immunity, 3) hormones, including conjugated estrogens, and 4) drugs acting on central nerve systems. Combined, the medicines in the first three categories help boost immunity to overcome viral infections' adverse stress and influence.