## Introduction

Though statistics vary between sites and are dependent on a host of comorbidities, between 20 and 60% of COVID-19 infected patients admitted to hospital develop acute respiratory distress syndrome (ARDS)1,2, with complications being a primary contributor to the mortality in up to 50% of these instances. Indeed, early evidence has suggested that COVID-19-induced ARDS may represent a specific variant thereof3,4. ARDS arises from complex interactions between components of the immune inflammatory response5,6 including complement activation7. This underlying inflammatory response appears especially aggressive in many cases of COVID-19 infection, with the self-aggravating release of increasing amounts of inflammatory cytokines fueling a run-away cytokine storm8,9. Though antiviral and immune-modulating agents are undergoing clinical trial, the sudden emergence and rapid global spread of COVID-19 has highlighted the quintessential value of rapidly identifying and redirecting tried and true pharmaceutical agents to immediately mitigate illness severity. This being said, the recent ad hoc use and subsequent failure of the antimalarial hydroxychloroquine in formal clinical trials10 suggests that important challenges remain if this is to be accomplished safely, effectively, and reliably. Current efforts have for the most part been focused on eliminating the root cause of infection with significant resources being applied to the characterization of SARS-CoV-2 molecular targets and the identification of novel antiviral compounds using molecular docking as the primary computational workhorse11,12. While conventional analyses of this type typically focus directly on individual virally encoded proteins, recent efforts have attempted to broaden their scope by mapping the pathogen-host interactome to identify downstream targets in the viral replication machinery that might be readily modulated using known compounds13,14. Though vital, these efforts are, by virtue of their single molecule resolution, both experimentally and computationally resource-intensive and time-consuming. An alternative and highly complementary means of reducing mortality in COVID-induced ARDS consist in appropriately modulating host immune response such that cytokine storm is averted. Unfortunately, while the molecular affinity approaches mentioned above are well-suited for identifying drugs binding with high specificity to a given target they require prior knowledge of what the appropriate target should be. Moreover, this time-independent analysis of binding affinity is ill-suited to describe the complex dynamics of immune cell signaling that lead to persistent cytokine storm where regulatory kinetics play a vital role in the selection of therapeutic targets.

Here, we propose a novel and expedient in silico approach for the evaluation and selection of drug repurposing candidates directed at model-predicted targets that formally account for the regulatory dynamics of the network as a whole. Utilizing both experimental and model-predicted affinities as well as literature-mined drug–target interaction data, we assemble and score multidrug therapies based on their expected regulatory actions on sets of concurrent molecular targets that act synergistically to disrupt cytokine storm in the context of coronavirus infection. Issues with data sparsity which are of special relevance in sudden and rapidly evolving health crises such as the COVID-19 pandemic, undermine the reliable use of data-driven approaches but can be circumvented by maximally leveraging the extraction of currently existing knowledge from the peer-reviewed literature. In previous work by our group15, we extracted prior knowledge from 2653 peer-reviewed publications to construct a network that included 18 host immune mediators documented to be of specific relevance to coronavirus infection, the latter being annotated as a disease entity in the Elsevier Knowledge Graph text-mining database16. These 18 immune mediators and the coronavirus pathogen were linked by 112 documented regulatory actions to create a network model that was capable of supporting biologically plausible immune response dynamics that included a persistent cytokine storm cascade17. Often lethal once triggered, this excessive and self-perpetuating host inflammatory response may persist in the absence of a pathogen as demonstrated recently in ref. 9. The objective of the current work therefore is to use the above-mentioned network model of host immune response to coronavirus infection as a basis for predicting intervention strategies capable of disrupting this run-away inflammatory cascade. Toward this, we conduct simulations of the host inflammatory response as part of an optimal search for companion drugs that would optimally compliment the actions of a concurrently administered antiviral, simulated here as an idealized reduction in viral titer. We focus in this first analysis on FDA-approved drugs as they might be more readily deployed but more importantly because there exists a broader more established base of clinical experience regarding their effects, compared to that of experimental compounds.

Our simulations highlighted several drugs and drug combinations that show promise in disrupting cytokine storm when administered as companions to an effective antiviral. In addition to recovering, known immunosuppressants like rapamycin and cyclosporine, as well as the corticosteroid dexamethasone, model predictions also correctly assigned a low score to hydroxychloroquine and chloroquine, two early contenders later proven to be much less effective than anticipated. Most importantly perhaps, our analysis pointed to several highly novel opportunities for significantly enhancing the actions of dexamethasone by pairing the latter with a select list of drugs that include antimicrobials such as ciprofloxacin or levofloxacin, and the statin simvastatin. Though these remain predictions, we propose that this model-based framework, because it is rooted in a broad body of prior knowledge, can be highly useful in focusing attention to drug candidates with significant potential and do so quickly with minimal data.

## Results

### Idealized interventions and suitability for repurposing

Of the 19 original network models15, we focus here specifically on model 18, which we believe supports the most biologically plausible predicted immune response dynamics based on criteria described in17. Stating the search for idealized intervention sets as a constraint satisfaction problem18, we identified 30 intervention solutions15 that prompt immune signaling predicted by the network model to migrate from a pattern of persistent immune hyperactivation (cytokine storm) to one that more closely resembles an idealized immune resting state. These solutions were computed under conditions of low viral load to highlight best case performance and because they would be used in concert with an effective antiviral treatment. The exact combinations of exogenous manipulations proposed in each of these theoretically optimal solutions si are listed in Supplementary Table 1 (Additional File 1) along with summary measures of optimality, namely the number of exogenous manipulations, the outcome residual distance from the target immune resting state, the number of transitions required to achieve this outcome as well as the overall pharmacologic actionability A(s) as it applies to the repurposing of the 144 candidate drugs. The latter is reported along with the significance of this value compared to that of a random intervention set, adjusted for multiple comparisons (Padj). These summary measures are also presented graphically in Fig. 1. Results show that 16 of the 30 optimal intervention sets score as significantly actionable (empirical null distribution P < 0.05; A > 7.25) in the 144 candidate drugs compared to a random intervention set. Interestingly these 16 significantly actionable solutions involve only slightly fewer immune targets on average (~6 targets) than their nonsignificant counterparts (~7 targets) despite this being a component of the actionability score. This similarity notwithstanding, the specific combinations of targets selected in each of these 16 solutions effectively exploit the regulatory network structure in a way that supports a much more expedient or efficient migration (~7 versus ~14 transitions) to outcome states where at least 85% of immune marker expression is restored to the baseline healthy resting state (~4 versus ~9 activation increments from an initial separation of 27). Of the 18 immune proteins in the network, TNF, IFNL1 (aka IL-29), and CSF3 are almost unanimously shared among the most actionable solutions. Indeed, IFNL1 and TNF appear in almost twice as many significantly actionable MIS solutions as they do in their nonsignificant counterparts. The two most actionable idealized interventions both consist of a joint inhibition of CSF3, CTSL, CXCL2, IFNG, and TNF, with one also inhibiting IFNL1. Interestingly recent work by9 supports a central role for type II interferon IFNG and TNF synergy in COVID-related cytokine shock. However, much less is known about a direct role of type III IFNL1 in a cytokine storm, though recent work proposes the latter as a major mediator of thromboinflammation19, an emerging clinical feature of COVID-19 infection20.

### Ranking individual drugs

The variability in performance across individual intervention solutions is further illustrated in Fig. 3. As expected, drugs at the top of the heatmap like rapamycin offer more consistent performance across those MIS solutions where treatment outcomes deviate by only two or three bits from the desired full- recovery profile. Indeed, even though drugs like raloxifene, sevoflurane (e.g., solution 25) and fenofibrate (e.g., solution 22) achieve higher adjusted enrichment scores in some isolated instances, they also present with no or even negative enrichment values in multiple other intervention solutions giving them a lower overall ranking. While most intervention solutions were supported reasonably well by at least one drug, this was not the case for solution 18 where all top-ranking drugs performed very poorly, presenting with uniformly negative enrichment scores. Intervention 18 (Additional File 1, Supplementary Table 1) is the single-most parsimonious of all significantly actionable intervention sets, consisting of only two-target manipulations, namely a downregulation of CD86 conducted in unison with an upregulation of TNF. As all other intervention sets involve four or more targets, broader-acting drugs are being favored overall, leading to a large overenrichment penalty for a two-target solution like that proposed under MIS 18. Of note, the importance of introducing drugs that more adequately align with solution 18 is undermined further by the latter’s poor outcome which is among the 3 worst recovery profiles predicted (i.e., 9 bits separation from full recovery).

### Ranking drug combinations

Ultimately, we would like a drug or drug pair to show the highest possible enrichment Eadj in at least one intervention solution (Max Eadj) and have it maintain this enrichment across as many of the best performing MIS solutions as possible, i.e., also show a high weighted median Eadj. Results presented in Fig. 5 show that most top-ranking drug pairs typically deliver a maximum enrichment Eadj equivalent or superior to that of single drugs. Moreover, all such pairs invariably maintain this performance more readily across those alternative MIS solutions producing more desirable outcomes. Indeed, while the combination of dexamethasone and medroxyprogesterone acetate offers only a slight improvement in maximum enrichment over medroxyprogesterone acetate alone (max Eadj 0.63 over 0.62), this improvement is much better sustained across those alternative intervention solutions that correspond to more favorable outcomes (WM Eadj of 0.52 over 0.36). Introducing simvastatin in combination with dexamethasone or levofloxacin offers the potential for even higher maximum enrichments of 0.67 and 0.70, respectively, albeit with moderate reductions in general applicability across intervention paths (WM Eadj 0.46 and 0.44, respectively, from 0.52). In contrast, combining simvastatin with the JAK1/2 kinase inhibitor ruxolitinib, a combination currently in Phase 2 trials (NCT04348695), scored poorly in both metrics and ranked 662nd among all combinations. Generally, one can reason that combining drugs that offer highly complementary drug actions will increase their target coverage and more readily support a broader range of alternative intervention solutions. Conversely, drug pairs that offer higher maximum performance but do so on a narrower range of solutions might be expected to perform more like single drugs as their actions overlap, limiting their ability to support alternative treatment paths.

Combination therapies offer the potential of broader overall drug action, however, they also carry the possibility for negative drug-drug interactions. While severity rating categories differ from one drug information resource to another22, recently proposed a standardized severity scale that reconciled definitions across multiple sources into three basic severity categories, namely Major, Moderate and Mild severity. In this work, we extracted drug interaction severities for our top-ranking drug pairs (Additional File 1, Supplementary Table 3) as documented in the DrugBank23 Drug Interaction Checker (OMx Personal Health Analytics Inc., Edmonton, AB, Canada) which is based essentially on these same three severity class definitions (https://dev.drugbank.com/guides/terms/severity). Of these ten top-ranking drug pairs, only five were assigned documented interactions in the DrugBank database. The severity of these interactions was Moderate for all but one pair consisting of dexamethasone and simvastatin. The latter was assigned a Major interaction severity, whereby concurrent administration of dexamethasone can increase elimination of simvastatin, decreasing serum concentrations and compromising its therapeutic effect24. Likewise, co-administering simvastatin with ruxolitinib (NCT04348695) can increase the latter’s serum concentration, affecting its tolerability and earning it a Major interaction assignment.

### Simulating proposed Interventions

It is important to recall that candidates from the initial set of 144 illness-relevant, pathway relevant, and model-relevant compounds were ranked on the basis of their respective ability to align with and support model-predicted and idealized sets of concurrent manipulations to specific immune mediator targets. As real-world drugs or combinations thereof may not necessarily align exactly with these idealized interventions, we conducted simulations to verify the suitability of the best-ranked candidates. Results presented in Figs. 6 and 7 show the gradual decrease in disparity (Manhattan distance) between the predicted immune marker co-expression profile at each iteration and the target signature at the desired immune resting state imparted by the simulated drug therapy and its eventual discontinuation. The initial state of the network is a persistent cytokine storm located at a Manhattan distance of 13 bits away from the immune resting state. Drugs actions are simulated in the context of a concurrently applied idealized antiviral resulting in the absence of a measurable viral load.

The simulated effects of administering dexamethasone in combination with the 4 companion drugs that provided the highest WM Eadj (Fig. 5), namely acetylcholine, medroxyprogesterone acetate, ciprofloxacin and rapamycin, are illustrated in Fig. 6. While all these drugs applied alone will eventually drive the immune network response to a stable recovery state, pairing dexamethasone with these companion drugs shows a noticeable positive synergy that in all cases delivers a much more expedient reduction in immune activation. Indeed, the predicted immune network response to all four drug pairs shows an immediate disruption of the cytokine storm state, rapidly achieving a residual distance from the immune resting state of only 1 bit within the first 4 iterations. This expediency is rivaled only by medroxyprogesterone acetate, with a repose time of 5 iterations, and is noticeably better than the responses of 8–12 iterations predicted for ciprofloxacin, acetylcholine and rapamycin. Importantly, all combinations produce a recovery that persists even after the intervention is discontinued. Different and more varied behaviors can be observed in drug pairs with decreasing values of WM Eadj that approach values obtained with single drugs (Fig. 5). For example, the pairing of cyclosporine with eltrombopag corresponds to an important reduction in WM Eadj or a loss in general applicability across multiple intervention paths that approaches that of a single drug. Accordingly, the simulated responses shown in Fig. 7a show a dominant effect of cyclosporine alone, with no additional benefit being contributed by the concurrent use of eltrombopag. At roughly the same reduction in WM Eadj, but with a much higher maximum enrichment in individual intervention, the pairing of simvastatin with levofloxacin provide an even greater emphasis on narrow performance at the expense coverage across paths to favorable outcomes. This even narrower focus not only eliminates any benefits of pairing but leads to a negative synergy where the actions of simvastatin alone dominate and where the latter is only hampered by the addition of levofloxacin (Fig. 7b). These examples suggest that as WM Eadj decreases, any benefit of pairing two drugs also decreases with the addition of a second companion drug having no effect or even counteracting the actions of the dominant drug.

## Discussion

While computational approaches, many of which are rooted in molecular docking methodology11, have been proposed to accelerate the data-driven screening of compounds targeting the COVID-19 pathogen or host proteins recruited during infection, little attention has been directed towards the in silico design of pharmacologic strategies for mitigating excessive host immune response, a key contributor to the severity of COVID-related ARDS and associated morbidity25. Moreover, while important resources like the VirHostNet database document interactions between virus and host proteins13, the resulting interaction networks remain static representations of molecular compatibility where drug targets are inferred on the basis of network structure alone. Similar principles of similarity have also been applied to transcriptomic profiles to construct drug-disease networks though once again candidate compounds are identified on the basis of network structure alone26,27. Here, we use a previously reported causal model of host immune regulation constructed by our group from published prior knowledge15 to implicitly account for and exploit the dynamic propagation of such network effects, and the resulting regulatory stability of the system, to identify subsets of dynamically coupled drug targets18. We then systematically apply a number of quantitative criteria to rank a broad range of candidate FDA-approved drugs based on their ability to jointly manipulate these subsets of immune mediators such that their corresponding downstream regulatory interactions are optimally leveraged to effectively disrupt a persistent ARDS-induced cytokine storm. The regulatory stability of this immune hyper-responsive state15 was such that only by concurrently modulating at least four host immune mediators would the host immune response be robustly directed to stand down in the context of a reduced viral load. Indeed, a closer examination of the 16 most pharmaceutically accessible intervention strategies revealed that drugs offering a broader coverage, in particular as it relates to the suppression of CSF3, IFNL1, TNF and to a lesser extent that of IFNG and STAT1, typically ranked higher. For these reasons many highly anticipated corticosteroids and antibody drugs did now fare as well as expected. For example, the antibody drug infliximab did rank among the top 20 drugs (ranked #14) with 7 documented targets. However, only two of these seven actions (TNF and IFNG) involved the five highly represented targets mentioned above, assigning it a lower rank. Similarly, prednisolone (ranked #31) counted five documented targets that included three of the five preferred targets. However, one of these three targets (CSF3) was regulated in the opposite direction to the action prescribed by the intervention sets leading to a penalty in rank.

Out of 144 candidate drugs, even the 10 top-ranking candidates when used alone offered no better than a Padj ~ 0.10 chance of enrichment above that of any single random immune modulator in directly harnessing at least one of the model-predicted optimal treatment mechanisms active in COVID-induced ARDS (p(Max Eadj = 0) ≥ 0.11). Not surprisingly, the actions of these individual drugs did not generalize well either. Indeed, the ability of any single drug to simultaneously recruit multiple alternate treatment pathways supporting the best outcomes was also highly variable, with none achieving significance over a random immune modulator (padj(WM Eadj) ≥ 0.14). Figuring prominently in this list are several drugs currently under active clinical investigation. Broadly studied in the general context of pulmonary distress, dexamethasone is currently the subject of at least 2 active Phase 3 clinical trials in COVID-induced ARDS specifically (NCT04836780; NCT04843761). Similarly, Phase 3 trials of the anesthetic sevoflurane, also studied in the broader context of ARDS, (NCT04355962; NCT04415060), the PPAR alpha agonist fenofibrate (NCT04661930), the calcineurin inhibitor cyclosporine (NCT04979884) and mTOR inhibitor rapamycin (Sirolimus)(NCT04948203) are currently active, with some recently completed. Also currently under active Phase 3 clinical study, we find acetylcholine maintenance or release by compounds such as nicotine (NCT04608201; NCT04598594) and pyridostigmine bromide (NCT04343963) ranking among the top-10 treatment strategies. In contrast, despite early interest in the antimalarials hydroxychloroquine and chloroquine, these aligned very poorly with our model-predicted intervention sets, ranking 66th and 96th out of the 144 candidate drugs. Consistent with this, subsequent data has not supported clinical effectiveness of the latter which together with broad off-target effects in the lung has led the US FDA to revoke Emergency Use Authorization (EUA) for CQ and HCQ as the known risks outweigh potential benefits of their use28,29.

As might be expected, we find both cyclosporine and rapamycin, known immune modulators commonly used to manage transplant rejection, ranking among the top five drugs in terms of their alignment with the best model-predicted interventions. The immunomodulatory, antiviral and tissue-protective properties of cyclosporine have recently been documented in the context of H1N1 and COVID-induced ARDS in several preliminary studies30,31; however, concerns remain regarding cyclosporine’s many side effects and the use of low-dose regimens are encouraged32. Similarly, though known to be an effective suppressor of Th17-mediated airway inflammation33 and T cell-driven lung injury34, investigation of rapamycin’s effectiveness and safety in treating COVID-induced ARDS remains in its early stages35. In sharp contrast with these well-known immune modulators, we also find drugs typically associated with the treatment of other conditions. For example, both simvastatin and fenofibrate are commonly used to treat dyslipidemia. Upregulation of the peroxisome proliferator-activated receptor-alpha (PPAR alpha) with compounds such as fenofibrate has been reported to also attenuate lung inflammation36 making it a candidate of legitimate interest in the treatment of COVID-induced ARDS and associated hypercoagulability37. However, results from a recent real-world observational study emphasize the need for rigorous randomized control trials before a clear benefit in a COVID setting can be confirmed38. Similarly, the anesthetic sevoflurane remains under active clinical study with limited evidence suggesting a contribution to host defenses engaging cytoprotective pathways around heme oxygenase-1 (HO-1) expression, a target of the SARS-cov-2 virus39,40. In a twist of irony, while smoking confers susceptibility to COVID-induced respiratory distress, there is mounting evidence that cholinergic agonists like nicotine41,42 and pyridostigmine bromide43 that promote acetylcholine expression not only offer significant anti-inflammatory actions44,45 but may also interfere with SARS‐CoV‐2 infection and viral replication.

Arguably, the best known of all these candidate drugs is the anti-inflammatory dexamethasone, with robust evidence supporting its efficacy and safety in the treatment of ARDS46. Indeed, recent molecular profiling has pointed to host immune response mechanisms dysregulated in COVID-induced ARDS that dexamethasone may be well-suited to modulate3. This is consistent with clinical observations of reduced all-cause mortality in cases of severe and critically severe ARDS reviewed in ref. 47, though these results appear to vary significantly based on subpopulation and timing of therapy31. However, evidence that its immunosuppressive effects lead to accelerated viral replication has fueled continued debate regarding its appropriate clinical use48. Some of these issues might be mitigated by combining dexamethasone with a companion drug in an effort to make it more transportable, more consistently robust, and potentially more efficient. Indeed, drug pairs were predicted in this work to perform substantially better in promoting the most robust and direct disruption of cytokine storm and to do so by concurrently harnessing multiple alternative pathways. Interestingly, dexamethasone was unanimously recruited into the top five pairs where it was combined with other high-ranking drugs. These imparted actions complementary to those of dexamethasone such as the downregulation of the immune activator CSF3. Acetylcholine agonists, medroxyprogesterone acetate, and ciprofloxacin ranked highest as companion drugs with identical intervention enrichment scores, followed by rapamycin and simvastatin. While acetylcholine agonists are under study, as noted above, this is less the case for medroxyprogesterone acetate and ciprofloxacin. Used in hormone replacement therapy and as a contraceptive, medroxyprogesterone acetate also exerts little-documented immune-modulatory actions, with high doses inhibiting the expression of inflammatory cytokines jointly affecting Th1, Th2, Th17, and Th22 immune signaling21, contributing to reductions in symptom severity and corticosteroid use in asthmatic subjects49. Fluoroquinolones such as the broad-spectrum antimicrobial ciprofloxacin, have been shown to also exhibit antiviral properties as well as significant immunomodulatory effects, including downregulation of IL-1 and TNF50 making these compounds of interest in the treatment of COVID infection and subsequent respiratory complications51. Similarly, clinical investigation of simvastatin in a COVID setting remains in its early stages. While inhibition by a statin of 3-hydroxy-3-methylglutarylcoenzyme A (HMG-CoA) reductase has been shown to interfere with the development of ARDS in an animal model52, clinical outcomes were initially found to be unchanged by simvastatin53. However, subsequent analysis of study data identified statin-responsive phenotypes54, prompting renewed interest in statins as modulators of thromboinflammatory pathways in COVID-induced ARDS55,56.

Unfortunately, while several of these top-ranking companion drugs are under independent study, none are yet being investigated in the context of a combination therapy. A possible exception to this would be the ongoing Phase 2 study pairing ruxolitinib with simvastatin (NCT04348695) which did not rank highly with the metrics used in this work. While this result in and of its own does not necessarily guarantee a poor outcome, it does suggest that other combinations might offer stronger synergies in leveraging multiple complementary drug-action pathways predicted to effectively disrupt ARDS cytokine storm. Indeed, the analysis presented in this work was never meant to replace more exhaustive simulation studies and experimental validation of drug responses. Nonetheless, this approach offers a computationally efficient quantitative framework for ranking individual drugs and drug combinations in a way that implicitly accounts for and exploits the network biology of interacting pathways using a mechanistic dynamic representation of broad community-wide peer-reviewed prior knowledge. While the metrics used here do not explicitly favor drug pairs expected a priori to deliver positive synergistic effects, it was interesting to observe that many such positive synergies seemed to emerge as a natural consequence of increased transportability across those multiple intervention paths predicted to produce better outcomes.

In this iteration, the proposed framework does not yet formally account for drug toxicity and known adverse events associated with drug combinations. Ongoing work is directed at including such penalties in the estimation of drug enrichment scores. In addition, this initial analysis does not explicitly include or leverage drug pharmacodynamic properties or drug dose-response data e.g., IC50 data, though such information is indirectly leveraged through literature-interpreted drug actions57. Where such data exist, the current framework could potentially benefit from concurrent use and validation against empirical drug data-driven approaches such as the Drug Atlas58 with the proviso that predictions made here include maximally synergistic drugs in addition to those that are maximally complementary in their actions. Similarly, recent digital repositories focused specifically on consolidating clinical experience and putative targets of COVID-19 therapeutics59,60 might also be leveraged to further inform the family of candidate drugs to be assessed in the context of host immune response dynamics captured here by our regulatory network model. Finally, in the current analysis, the selection of drugs useful in disrupting cytokine storm was limited to compounds having received FDA approval. The intention was to provide a sound basis for establishing credibility of the proposed drug selection method as more is typically known about the effects of such compounds. Moreover, repurposing of the most promising candidates might be facilitated by the fact that they are already in clinical use. Certainly, in future iterations of this analysis, experimental compounds would be included in the hope of further optimizing efficacy of treatment. Similarly, the optimal choice of targets and corresponding compounds might differ depending on the severity and phase of illness. While simulations of treatment response used here were conducted from an initial illness state corresponding to an established cytokine storm, this is not a requirement. Indeed, future work involving a broader search across a wider range of initial conditions might identify thresholds in severity where some drugs should be discontinued in favor of others.

Though opportunities exist for continued refinement of this approach, we contend that the current framework offers a first important step towards producing a reliable and computationally efficient means for the rapid repurposing of well-known compounds currently in clinical use and redirecting these towards the effective management of potentially severe comorbidities such as ARDS, thereby reducing fatalities from SARS-CoV-2 infection. This being said, it important to remember that the current model was adjusted to reproduce and capture host immune response to a SARS-CoV1 exposure, as this was the most readily available data at the time. Of course, new data sets continue to be shared that describe host responses to SARS-CoV2 specifically. Though we expect some degree of uniqueness in ARDS induced by this more recent pathogen, we believe that predictions presented here based on SARS-CoV1 would at the very least be relevant to the treatment of cytokine storm in SARS-CoV2 infection. This may be especially true given the coarser resolution afforded at the level of drug assignment. With these limitations in mind, we propose that the drugs and drug pairs identified nonetheless constitute a potentially promising set of initial candidates for further in silico, animal and clinical study.

## Methods

### A model of immune signaling in cytokine storm

As previously reported by our group15, a basic immune signaling network model was constructed from 19 molecular markers (nodes) with involvement in infectious pneumonia as documented in the Elsevier Knowledge Graph database16 (Elsevier, Amsterdam) using the Pathway Studio (Copyright © 2020 Elsevier Ltd. except content provided by third parties. Pathway Studio is a trademark of Elsevier Ltd.) suite of software tools61. This same database supported 112 regulatory interactions (edges) previously extracted from 2653 abstracts and/or full text of peer-reviewed journal publications using the MedScan natural language processing (NLP) engine62,63. The dynamic behavior of the network is governed by model parameter values that dictate the underlying transition from one network state to the next64 using an extension of discrete logic concepts put forward by Thomas and Thieffry65,66 and applied specifically to cell signaling systems by67 and others. More specifically, perception thresholds on incoming signals at each network node are used to mimic the actions of high and low-affinity receptors. Only those incoming signals that exceed their respective perception thresholds will be considered as active mediators of the target downstream node. Each possible combination of these active signals is assigned a respective logic weight that dictates the regulatory response at that node, i.e., increase, decrease or maintain the same expression level in the next iteration. In this way, each regulatory decision is context-specific, weighing the actions strong activators against those of weak inactivators and vice versa.

Acceptable sets of these perception thresholds and decisional logic weights (K-values) were identified here using a constraint-based optimization framework68 such that the network’s corresponding dynamic behaviors encompassed data describing the 72-h in vitro time course response of human Calu-3 cells to SARS-CoV infection69. Given the still limited availability of molecular data specific to COVID-19 infection and the similarity in clinical presentation to SARS70, we consider this an acceptable proxy for this exercise. These data were retrieved from the Gene Expression Omnibus (GEO accession number GSE33267) and normalized to the mean of the mock infection samples (control) at each timepoint and translated into discrete expression levels using Variational Bayesian Gaussian clustering71. As the complexity of the model exceeds that of the available experimental data, the parameter identification problem is said to be underdetermined. In this case, 19 unique models were identified with <5% departure from the experimental data, with 11 of these parameter sets explaining the time course data exactly (0% error). Of these, 3 parameter sets (models 15, 17, and 18) supported steady-state expression profiles that closely approximated an immune resting state and one considered representative of the persistent cytokine storm. In the case of one model in particular, the basin of attraction surrounding the inactive immune resting state conferred a sufficiently robust response to infection for cytokine storm to occur at a clinically realistic rate8 (~15%;) (26% predicted by model 18). As such, this model was retained as the basis for all simulated drug actions presented here. Additional details may be found in a concurrent report by our group17. These steps and the proposed drug selection process are illustrated as a workflow in Fig. 8.

### Minimal intervention set identification

As mentioned in the previous section, we use a discrete decisional logic to manage the flow of immune signals through the host regulatory network and direct its evolution from one immune state to the next. At any given iteration, each of the network’s immune mediator nodes is assessed for incoming signals expressed above their respective perception thresholds. As mentioned above, based on the specific combination of active upstream mediators and their mode of action (activator or inactivator), the state of a corresponding downstream node is predicted to either remained unchanged, increase or decrease in the next time step. Applying this to all nodes, the next state towards which the overall network should progress given the current state, namely the network image, is computed. According to the update scheme being used, this predicted change in state is applied to a single random node (asynchronous update) or to all eligible nodes simultaneously (synchronous update). As we were primarily interested in stable persistent behaviors like cytokine storm, simulations of immune response were conducted here using a synchronous update of all node states for reasons of computational efficiency. Using this framework, model predictions of host inflammatory response under different conditions were conducted as part of a multi-objective optimization problem directed at identifying the smallest or minimal sets of molecular targets in the immune network that when manipulated concurrently in a specific manner (i.e., upregulated or downregulated) would theoretically disrupt a fully established cytokine storm and return immune activation to a normal resting state either fully or partially. Here we describe a fully established cytokine storm as the co-expression pattern of immune mediators measured experimentally in the Calu-3 cell cultures 72-h after coronavirus infection69 (GEO accession number GSE33267). In addition, the baseline healthy resting state was specified as an idealized immune quiescent state where all immune mediator entities are expressed at a nominal activation level (e.g., inactive or 0 in the discrete space).

These Minimal Intervention Sets (MIS) were identified by solving a computationally efficient Constraint Satisfaction Problem (CSP)19 where the idealized manipulations of at most nine molecular targets at a time, or 50% of the network, were iteratively assessed. As reported in previous work by our group18, this optimization consisted of concurrently minimizing the number of immune nodes being targeted, the final distance to the target state achieved by the intervention (calculated using the L1-norm), and the number of state transitions required (efficiency) to reach this treatment stabilized state. This multi-objective optimization was stated as a constraint satisfaction problem (CSP)72 in MiniZinc73 and solved with the greedy solver Chuffed74.

### Ranking intervention sets as candidates for repurposing

By design, solving a constraint satisfaction problem allows for the recovery of all solutions which comply with the constraints stated. As mentioned above, these constraint-compliant solutions can be ranked further based on other performance metrics such as size or cardinality of the intervention target set, as well as the predicted expediency (efficiency) of response and reliability of outcome (robustness) as reported previously by our group18. However, despite their optimality in achieving a desired outcome, such idealized interventions (concurrent up or downregulation of specific targets) may not translate readily into clinical care using currently available drugs. Here, we propose a quantitative metric to assess the actionability A of an idealized MIS solution as a ready candidate for drug repurposing75. The A metric is composed of four component scores. The first of these, cardinality, is the number of prescribed targets in the MIS vector m.

$$C(m) = \left\| m \right\|_0$$
(1)

A second metric is the antagonist ratio. Because it is generally more common for drugs to exhibit inhibitory actions against molecular targets, MIS solutions with a higher ratio of antagonist-to-agonist actions are prioritized.

$$V\left( m \right) = \frac{p}{q}\left\{{\begin{array}{*{20}{l}} {p = \mathop {\sum}\nolimits_{i = 1}^n {[m_i = 1]} } \\ {q = \mathop {\sum}\nolimits_{i = 1}^n {[m_i = - 1]} } \end{array}} \right.$$
(2)

Where n is the number of targets in m, mi is the polarity of the target at index i, −1 represents an inhibition of the target, and +1 represents activation of the target.

A third consideration is a broad accessibility to equivalent pharmacotherapy, expressed here as the number of alternate drugs available to perform the same actions specified in a given MIS solution.

$$R\left( m \right) = \left( {\mathop {\prod}\nolimits_{i = 1}^n {\hat m_i} } \right)^{1/n}$$
(3)

Where n is the number of targets in m and $$\hat m$$ is the set of counts of all drugs that act with correct polarity on each element of m.

Finally, we attempt to account for off-target effects of all drugs applicable to an MIS, that is all drug actions inadvertently applied outside of those dictated in the MIS. Accordingly, the actionability of an MIS is penalized or discounted in proportion to the minimum number of documented off-target interactions, affected by all drugs at each target and cumulated across all component targets in the MIS (Eq. (4)).

$$T\left( m \right) = \left( {\mathop {\prod}\nolimits_{i = 1}^n {\min \left( {k_i} \right)} } \right)^{1/n}$$
(4)

Where n is the number of targets in m and k is a vector of n elements, each of which is itself a vector of length j. Each ki,j is the total number of interactions known for the drug at j. To be included in ki, j must interact with mi with the correct polarity.

These components are combined into a single aggregate score as shown in Eq. (5). Two tuning parameters, γ and τ, are used to adjust the relative weight of solution cardinality and the antagonist ratio, respectively. For this analysis, γ and τ were set to 1.0 to assign equal weight to all component scores. Of note, any MIS solution where the drug-action databases do not contain any drug known to interact with one or more targets is considered an unactionable solution or one that is not suitable for repurposing.

$$A\left( m \right) = \frac{{R\left( m \right) \ast V\left( m \right)^\tau }}{{C\left( m \right)^\gamma \ast T(m)}}$$
(5)

The statistical significance of an actionability score was estimated empirically from a null distribution of actionability for randomly generated mock MIS solutions. The null distribution was generated by repeatedly sampling a discrete uniform distribution along the integer interval [−1, 1], where each mock MIS solution vector is equal in length to the maximum possible cardinality allowed for in the current model (i.e., the number of allowable targets ≤ number of network nodes). The actionability scores corresponding to these artificial solution sets form the null distribution. Empirical P values for each real solution are then estimated as the proportion of artificial actionability scores greater than or equal to the actionability of each real solution. To account for multiple tests, the Benjamini–Hochberg76 correction is applied.

$$\hat p = \frac{{(r + 1)}}{{(n + 1)}}$$
(6)

Where n is the number of simulated solutions and r is the number of simulated solutions whose actionability is greater than or equal to the actionability of the actual solution. All actionability scores and null P values were computed using Python version 3.8.3 (2020–05–13) (https://www.python.org/).

### Creating a pool of problem-relevant candidate drugs

A broad preliminary set of candidate drugs with broad relevance to the problem of interest was assembled by applying three complementary surveys of the Elsevier (Amsterdam, NL) Knowledge Graph database16 (Fig. 2a). The first and perhaps broadest of these consisted of searching for immune modulators associated with at least one node in the network model developed by our group15, yielding 5120 drugs. A second more infection-focused survey included all FDA-regulated compounds known to act as agonists or antagonists of human proteins and protein families found characteristic of Coronavirus infection, Severe Acute Respiratory Syndrome Coronavirus 1 (SARS-Cov-1) and 2, (SARS-Cov-2; COVID-19) as well as the Middle East Respiratory Syndrome coronavirus (MERS-CoV), producing a set of 1211 drugs. A third more clinically motivated survey focused on drugs with targets implicated in the pathogenesis and progression of coronavirus-induced ARDS. This third set consisted of 1270 candidate drugs known as (a) affecting autophagy, or (b) being the object of past or ongoing clinical research for the treatment of SARS, MERS, or COVID-19. The intersection of individual results from each of these three complementary surveys, produced a consensus list of 144 FDA-regulated compounds, drawn from 94 Established Pharmacologic Classes (EPC), that were relevant to the pathogen, to the progression of infection and to host immune response in ARDS-induced cytokine storm. Of these, 67 are registered to interventional clinical trials in the United States as of July 27, 2020 (Fig. 2b). This set of drugs includes monoclonal antibodies, non-steroidal anti-inflammatories, antibiotics, statins, and other immunomodulatory compounds but the most widely represented are kinase inhibitors and corticosteroids (Fig. 2c).

Of note, in estimating the statistical significance for the enrichment of a drug Eadj with respect to a specific intervention solution (MIS) we compare against enrichment E0adj obtained in a background set of drugs serving as a null distribution. In order to avoid including an overwhelming majority of drugs for which no targets exist in the model, introducing a bias of the null distribution towards no enrichment, we define a background set consisting of only those 1900 drugs which like the candidate drugs (foreground) not only show affinity for at least one entity in the network but also modulate this target according to a well-defined mode of action, namely a specific upregulation or downregulation of the target by the drug. In the case of drug pairs, we establish a similar null distribution by randomly subsampling 50,000 instances from the over 1.8 million possible combinations of two drugs selected from these same 1900 candidates. In both cases, these sets of model-relevant drugs remain broad while also offering more conservative measures of the significance threshold for enrichment E0adj.

### Ranking drugs that optimally support an intervention

To facilitate the comparison of individual drugs or drug combinations that might support a specific model-predicted intervention to a greater or lesser extent, we propose as a quantitative ranking metric, the Adjusted Intervention Enrichment Score (Eadj). The Eadj score is a weighted measure of the degree to which set of drug actions enriches an MIS solution. First, a basic enrichment E is defined as the percentage of prescribed targets in the MIS solution that are acted upon by a given drug or combination of drugs. However, this basic measure of coverage does not account for an “overenrichment”, in which an entity unprescribed in the MIS solution is modulated by a candidate drug, nor does it account for conflicting modes of drug action, where a candidate drug acts on one or more targets in a manner contrary to that prescribed in the MIS solution. To account for such undesirable occurrences, two penalty scores, the overenrichment penalty, b, and the conflict penalty, c, are used to adjust the raw E score. It is this adjusted E score, Eadj, that is used to compare competing drugs and drug combinations.

Consider an MIS solution to be a vector, m, whose length is equal to the number of entities or nodes in a model and whose elements mi can hold one of three possible values18, namely −1, 0, or 1. A value of mi = 0 indicates no action on the corresponding model entity, a value of −1 indicates a required downregulation or inhibition, and 1 indicates a required upregulation or activation of that entity or node respectively. Similarly, the actions of a drug, d, can be represented by a vector of the same length as m and whose elements can hold the same value, but where those values are determined by known interactions between the drug and the elements of the corresponding model. The basic enrichment score E(d,m) is therefore the number of elements in d whose values are equal to their corresponding elements in m divided by the number of nonzero elements in m. The following equations, use Iverson Bracket Notation77 where statements in square brackets evaluate to a Boolean state of 1 if true and 0 otherwise.

Enrichment E(d,m) is defined as:

$$E(d,m) = \frac{{\mathop {\sum }\nolimits_{i = 1}^n \left[ {\left( {d = m_i} \right)\wedge{\left( {m_i \,\ne\, 0} \right)}} \right]}}{{\mathop {\sum }\nolimits_{i = 1}^n \left[ {\left( {m_i\, \ne \,0} \right)} \right]}}$$
(7)

where n is the number of elements in m. To account for overenrichment or direct off-target effects, a penalty, B is subtracted from the raw E score.

$$B(d,m) = \frac{{\mathop {\sum }\nolimits_{i = 1}^n \left[ {\left( {d_i \,\ne \,m_i} \right)\wedge{(m_i = 0)}} \right]}}{n} \cdot b_0$$
(8)

where b0 is a user-defined penalty weight. Larger values of b0 penalize more greatly for overenrichment. Another penalty accounting for contrary drug actions, C is also subtracted from the raw E score to account for conflicting actions on a target.

$$C(d,m) = \frac{{\mathop {\sum }\nolimits_{i = 1}^n \left[ {\left( {d_i\, \ne \, m_i} \right)\wedge{\left( {m_i \,\ne\, 0} \right)}} \right]}}{n} \cdot c_0$$
(9)

where c0 is a user-defined penalty weight. Larger values of c0 penalize more greatly for conflicting actions. Subtracting B and C from E yields the adjusted enrichment score, Eadj.

$$E_{adj}(d,m) = E(d,m) - B(d,m) - C(d,m)$$
(10)

Here, we estimate the statistical significance of an adjusted enrichment score Eadj for a given drug or combination of drugs with respect to a specific MIS solution by comparing it against the distribution of the enrichment scores E0adj obtained by randomly sampling drugs or combinations of drugs from a background set of 1900 drugs that exert a well-defined action on at least one network element, as described in the previous section. Empirical P values are estimated as the proportion of E0adj scores greater than or equal to each observed Eadj score. These raw P values are then corrected for multiple comparisons across the subset of 144 drugs of interest (both model and illness-relevant) using Benjamini–Hochberg76. This process is analogous in some ways to gene set enrichment analysis (GSEA)78, though the definition of enrichment and exact statistical methods differ.