Pharma R&D Today

Ideas and Insight supporting all stages of Drug Discovery & Development

Select category
Search this blog

Modeling Coronavirus – Cytokine Storm

Posted on September 4th, 2020 by in COVID-19

Figure 1. Classification of COVID-19 Disease States.
Adapted from Sidiqi, H.K. and M.R. Mehra, COVID-19 Illness in Native and Immunosuppressed States: A Clinical-Therapeutic Staging Proposal. J Heart Lung Transplant, 2020. [1]

(This piece was co-authored by Prof. Gordon Broderick, Clinical Systems Biology at Rochester General Hospital.)

The thing that kills you with coronavirus is not necessarily the infection itself but rather your own body’s response to the infection. Failure to clear the virus with an appropriate immune response can lead to pulmonary symptoms of increasing severity and which can even be fatal in a minority of cases (Figure 1). This late stage of a COVID-19 infection is characterized by a state of hyperinflammation or, as it is more commonly known, a “cytokine storm”.

Cytokine storms are not restricted to just coronavirus infections but can occur in a variety of deadly medical situations such as other viral respiratory infections (H5N1 influenza and SARS-CoV) as well as non-infectious conditions like graft-versus-host diseases (as can happen, for example, from stem cell transplants). Cytokine storms result from an uncontrolled and excessive release of pro-inflammatory signaling molecules called cytokines. Normally, cytokines are part of the body’s immune response to infection, but their sudden release in large quantities can cause multisystem organ failure and death. [2] The bottom line is that, once triggered, they are hard to stop and, as such, pose a unique danger to the patient.

Figure 2. Constructing a Network Model.
A. Proteins regulated by “coronavirus infection”.  B. Regulatory network of coronavirus-related proteins. Schematic representation of information flow through the coronavirus regulatory network, Red = an inhibitory signal, Green = an activation signal.

Elsevier’s In Silico Biology team, in collaboration with Dr. Gordon Broderick and his group at the Center for Clinical Systems Biology at the Rochester General Hospital in Rochester, NY., set out to see if we could model, at the molecular level, what is going on during a cytokine storm and use that model to suggest therapeutic interventions. Our first step was to use the prior knowledge maintained in Elsevier’s high-quality Biology Knowledge Graph to first identify the disease of “coronavirus infection” and then the proteins that are known to be regulated in response to it (Figure 2A). Next, we connect the dots! We then extract all the high-quality regulatory interactions between these 18 proteins as found in the Biology Knowledge Graph to assemble the corresponding immune signaling network.   Altogether that gave us a total of 128 relations (Fig. 2B), which together define a computable logic circuit which can be used to model information flow when coronavirus is activated in the network. Figure 2C is a schematic representation of this information flow as a biological circuit or, more formally, a protein regulatory network (PRN). By high-quality we mean that an individual relation linking model components is supported by many underlying independent observations in the peer-reviewed scientific literature. For this network, the median reference number = 14.5. And what that means practically is that, overall, this network represents a pretty strong scientific consensus regarding the “rules of engagement” between these specific proteins.

Figure 3. Aligning the Model Logic to the Data. 
A. What we see in patients: A gene expression dataset from a SARS-Cov2 related virus is used as a proxy of what might be happening in COVID-19 patients. B. Using machine learning to select sets of logical rules that allow the model regulatory circuit to best reproduce experimental observations. Out of 100s of potential logic models, we select a handful that best explain the experimental observations.

So that’s a good start, but – is it real?  And if so, is it relevant to this illness? In other words, does this network with its theoretical rules comply with what actually happens when we look at relevant experimental data describing coronavirus infection. The most straightforward way to test this is to assess whether or not actual observed experimental data can be predicted according to the network model. For this purpose, we located and downloaded a dataset of gene expression which measured the effects of a SARS virus infection of a human airway epithelial cell line across 72 hours (Figure 3A)[3]. We liked this data because of its complexity (10 different timepoints) and that fact that it is known that the SARS virus has the same propensity to trigger cytokine storm as does COVID 19 (also, at the time of this work, there were no relevant datasets for COVID-19).

Machine learning (ML) is used to eliminate logic rule sets that do not explain the available experimental measurements. As a byproduct of this process the values of the unmeasured model components are implicitly predicted using the preset rules identified to predict with high accuracy those experimental values which are measured and reported in the experimental dataset. The extent to which the ML gets the problem right is defined by how closely the predicted values approximate the experimentally observed values, i.e. the prediction error. Figure 3B illustrates this process over many iterations (in this case, close to 250). Gradually the prediction error decreases from 30% to less than 5% (the threshold below which the model is considered to be technically validated).  In this particular instance it can be seen that quite a few competing models (19) were generated with little or no error at all.

Once we have generated good candidate solutions (logical rule sets) for our network model, we can move on to the next phase of our project which is to analyze solutions supporting immune response behaviors which are most consistent with the known clinical pathology of coronavirus infection. We know that the vast majority of patients infected with COVID-19 do not develop cytokine storm (Fig. 1). And yet each of us walks around with the same basic network with the same fundamental rules albeit with a slightly different tuning. So why do only some patients “fall” into this condition?  Well, we know that some of the predetermining factors certainly involve the age of the patient and the presence of comorbidities but that is only half of the story.  The other half of the story tells us how the same network can support very different reactions in different people.

The secret lies in the understanding we can get from formal network theory (see In Silico Biology: Systems at the Edge of Chaos for a review of this topic) in which it has been convincingly demonstrated that in real world biology, network dynamics are constrained by design to support what are called attractors. These states are particular configurations that a given network dynamic tends to fall into. How often a network model returns to the same attractor is called its basin of attraction. Why that is important to us here is that in order to identify a credible model of cytokine storm we must be able to show that it can support both a normal and a pathological condition.  And their relative basins of attraction should be proportionately appropriate to epidemiological observations. In other words, we would expect the basin of attraction for cytokine storm to be relatively small compared to the basins of attraction for non-pathological states on the same network. This requirement is based upon the simple observation that only a relatively small percentage of the population ever falls prey to this condition even with the same type of infection. Conversely, a relatively inactive immune resting state should be broadly accessible in the absence of an active infection.

Figure 4. Attractor Mapping.
All attractors as identified by the different models under two separate conditions (A. Coronavirus Off, B. Coronavirus On) are shaded in blue, The red-shaded region are the location for the cytokine storm states supported by each model, and the green dot (“All”) is the healthy “resting” state (Immune Senescence). Bubble sizes are proportionate to the size of the basin of attraction for each model attractor.

Maybe a visualization might be helpful here to illustrate this point. Attractors existing in a higher dimensional space (expression of ~16 proteins) can be projected into a 2-dimensional representation using multidimensional scaling such that their relative position or mutual separation is maintained (Fig. 4). For example, different stable attractors supported by one of the more plausible candidate models are displayed in Figure 4, including  the “cytokine storm” state (red shaded region) and a hypothetical “rest” state (green dot) predicted by the model in question. The size of each dot represents the frequency with which that attractor was reached from a broad range of initial conditions using repeated simulations. In this case, 100,000 simulations were conducted for each model from starting positions where coronavirus (CV) was inactive (Fig 4A) or active (Fig4B). In the absence of active coronavirus infection, a subset of candidate models favor migration towards an inactive immune resting state in the majority of simulations. This is good! We don’t expect to experience a cytokine storm or other immune flare ups to occur just by random chance. However, in the presence of an active coronavirus infection, all models demonstrate significant and persistent disruption of immune signaling towards basins of attraction that closely resemble the cytokine storm state (Fig 1B, lower left-hand corner). Migration patterns such as these towards pathological and normal resting states are what we might legitimately expect from the dynamics of this disease in the general population.

Now that we have a credible disease model, we can move on to the next step of the In Silico Biology process – making predictions. Can we help rescue patients caught in a cytokine storm? The problem is that attractor states are stable regulatory cul-de-sacs. Without some outside intervention, once in it, you’ll be stuck there! 

Figure 5. Optimal drug assignments.
(A) Simulations of the immune response dynamics predicts an idealized target intervention set. (B) Drugs applied alone show enrichment on idealized target set. (C) Drugs in combination were assessed for the degree with which they supported this idealized intervention.

Based on the dynamics of the network model, we can begin to systematically identify the maximally concise set of therapeutic targets that effectively de-stabilize the illness condition and promote a rescue of proper regulatory equilibrium. To do this we employ another machine learning technique known as Answer Set Programming in order to identify optimized Minimal Intervention Sets (MIS) consisting of therapeutic targets that when agonized or antagonized in combination will de-stabilize the illness condition and favor of a return to a normal healthy resting state.  Figure 5 illustrate the results of this type of analysis using the coronavirus information network. This particular solution calls for ideally antagonizing (inhibiting) STATs 1&2, CD200R1, CTSB, and CTSL proteins while simultaneously up-regulating IFNG (Fig. 5A). Next, we screen drugs singly (Fig. 5B) or in combination (Fig. 5C) for their desired activity on the proposed target set. We don’t expect any one drug or even drug combination to do the entire job as specified. But what we are looking for is enrichment of drug activity which might begin to have the desired effect.  Enrichment scores help guide this process by indicating the most promising candidates in this regard. Of note here was the identification of Ruxolitinib alone or in combination (Fig.5 B&C) with an iron chelating drug (deferoxamine). Ruxolitinib is a well-known blocker of the JAK/STAT immune signaling pathways and is actively being evaluated for repurposing for COVID-19 treatment specifically for the symptoms of hyperinflammation [4] and is now rapidly moving forward into Phase III clinical trials. We also found that dexamethasone (once again, in combination with deferoxamine) was near the top of our list as well (Fig. 5C). Dexamethasone is a corticosteroid used in a wide range of conditions for its anti-inflammatory and immunosuppressant effects. After our analysis was completed, results from the UK-based RECOVERY trial were published and showed that the use of dexamethasone resulted in lower 28-day mortality among those who were receiving either invasive mechanical ventilation or oxygen alone [5]. We are pleased to see that many of our predicted results appear to be in good alignment with real world evidence.


Figure 6. Simulated administration of the lupus drug, ruxolitinib, shows broad direct (red shaded areas) and indirect down-regulation of immune mediators over-expressed in cytokine storm (A), with lasting stable resolution expected at low viral titers afforded by an idealized anti-viral (B).

In addition to the idealized target sets discussed above, another useful approach to drug prediction is to simply add the drug of interest directly into the network model based on its known pattern of interactions and then measure its effects on the network over a simulated time frame (Figure 6.). In this case we chose to model the effects of ruxolitinib (CXCL10 negative, IFNG negative, STAT1 negative, TNF negative) because it figures prominently in our initial results and because it is proceeding well in clinical trials for use as a treatment for hyperinflammation in COVID-19 patients. What our simulations suggest is quite interesting. Yes, in the presence of ongoing active viral infection, ruxolitinib appeared effective in combating the cytokine storm state (Fig. 6A) but only during the period of active drug administration. As soon as the drug was removed, all the molecular markers of hyperinflammation returned. However, the good news was that under conditions of low viral titer, as might be achieved through concomitant administration of an effective anti-viral medicine, a state of persistent remission was achieved (Fig. 6B). These results predict that ruxolitinib treatment of COVID-19 patients will be most effective with the co-administration of a strong antiviral treatment. A useful analogy might be to compare the difficulty of trying to stop a car using the brake with the accelerator pedal all the way down (ruxolitinib alone) as opposed to simultaneously lifting your foot off the accelerator pedal (the anti-viral treatment) while applying the brake (ruxolitinib). It’s pretty easy to see it would be best to do both!

We hope that through the general process outlined here, Elsevier’s In Silico Biology program will begin to open up a new way to rapidly develop and test hypotheses at the molecular level which in turn can inform both scientific research and as well as better clinical decision-making.

REFERENCES

1.            Sidiqi, H.K. and M.R. Mehra, COVID-19 Illness in Native and Immunosuppressed States: A Clinical-Therapeutic Staging Proposal. J Heart Lung Transplant, 2020.

2.            Farsalinos, K., A. Barbouni, and R. Niaura, Systematic review of the prevalence of current smoking among hospitalized COVID-19 patients in China: could nicotine be a therapeutic option? Intern Emerg Med, 2020: p. 1-8.

3.            Sims, A.C., et al., Release of Severe Acute Respiratory Syndrome Coronavirus Nuclear Import Block Enhances Host Transcription in Human Lung Cells. Journal of Virology, 2013. 87(7): p. 3885-3902.

4.            Caocci, G. and G. La Nasa, Could ruxolitinib be effective in patients with COVID-19 infection at risk of acute respiratory distress syndrome (ARDS)? Annals of Hematology, 2020. 99(7): p. 1675-1676.

5.            Group, R.C., et al., Dexamethasone in Hospitalized Patients with Covid-19 – Preliminary Report. N Engl J Med, 2020.

R&D Solutions for Pharma & Life Sciences

We're happy to discuss your needs and show you how Elsevier's Solution can help.

Contact Sales