DeepNEU: a machine learning platform for simulating Cytokine Storm and Coagulopathy that complicate severe COVID-19 to enable targeted drug repurposing

COVID-19 is a disease that results from infection with the SARS-CoV-2 virus. The disease is often complicated by cytokine storm and/or a coagulopathy These two complications appear to be responsible for much of the increased mortality seen in severe COVID-19. While available treatment options including systemic corticosteroids and heparin have been benecial in some ventilator dependent patients, mortality rates remain excessive. Several clinical trials using previously approved drugs and drug combinations are urgently underway to improve the current situation. The primary purpose of this project was to evaluate the DeepNEU stem-cell simulation platform by creating and validating computer simulations of articial lung cells infected with SARS-CoV-2 to enable rapid identication of therapeutic targets and drug repurposing for specically treating both the cytokine storm and coagulopathy associated with severe COVID-19. The data generated from this project conrm that (1) COVID-19 can be simulated in human alveolar type lung cells infected with SARS-CoV-2, (2) these simulated lung cells can develop features of the cytokine storm and coagulopathy that often complicate severe COVID-19, (3) the unsupervised machine learning system performed well in all COVID-19 simulations based on available published wet lab data and (4) the platform successfully identied potentially effective double drug combinations for both COVID-19 associated cytokine storm and coagulopathy for urgent clinical study. We conclude that while DeepNEU (v5.0) requires further validation, it is very likely that the most important application of this machine learning platform will be to improve our preparedness for serious future viral outbreaks and their life-threatening complications.


Introduction
As of this writing Coronavirus disease (COVID-19) resulting from SARS-C0V-2 virus infection has caused a worldwide pandemic, with >3,000,000 deaths and ~150,000,000 cases reported internationally (1). While the SARS-CoV-2 virus does have some similarities to the SARS virus of 2003, unfortunately SARS-CoV-2 represents a novel viral pathogen to the human hosts (2). Extensive efforts have resulted in effective vaccines and some antiviral therapies. However, vaccine development, testing and approval to combat emerging resistant strains across the globe will require ongoing efforts. Amongst the symptomatic patients with COVID-19 most will experience a mild respiratory tract infection, but some patients progress to more critical illness that may be characterized by cytokine storm, hyperactive immune response, and coagulopathy (3,4). Both cytokine storm and coagulopathy are associated with an increased risk of death (3,4) . Cytokine storm is characterized by an elevation of cytokines including several interleukins, interferons and tumor-necrosis factors (5). The excessive release of these immune mediators is injurious to host cells (6). In addition, the importance of COVID-19 associated coagulation abnormalities is becoming increasingly clear in that a signi cant proportion of patients with severe COVID-19 develop potentially lethal venous and arterial thromboembolic complications which sometimes go unrecognized (7).  such as IL-2, IL-6, IL-7, IL-10, GSCF, IFNγ, IP10, MCP-1, MIP1A and TNF-α, with the characteristic features of a cytokine storm (9). Normally, when a virus infects the body, the in ammatory response plays an important antiviral role, but when the anti-SARS-CoV2 in ammatory response is uncontrolled the resultant cytokine storm can be very damaging to tissues and organs in affected patients (10). Therefore, developing strategies for effectively suppressing cytokine storm is essential for limiting disease progression and reducing the mortality rate in patients with COVID-19.

The coagulopathy induced by SARS-Cov-2 infection
It has been reported that the pathology of the lungs in some patient with severe COVID-19 shows marked microvascular thrombosis and hemorrhage associated with extensive alveolar and interstitial in ammation and diffuse intravascular pulmonary coagulopathy (11). Patients at this phase of disease progression sometimes develop a hypercoagulable state and D-Dimer-based coagulation factors may be abnormal. At a more advanced stage of disease progression, D-dimer is signi cantly increased, along with a prolonged prothrombin time (PT) and gradual decreases of brinogen (FBG) and platelet count (12,13). It is also believed that COVID-19 can activate the coagulation cascade through various mechanisms, leading to severe hypercoagulability (14). Therefore, in addition to all necessary respiratory support and mitigating cytokine storm, the timely identi cation and treatment of coagulopathy is also crucial.
The DeepNEU machine learning platform for identifying therapeutic targets for severe  The DeepNEU platform is a validated hybrid deep-machine learning system with elements of fully connected recurrent neural networks (RNN), cognitive maps (CM), Support Vector Machines (SVM) and evolutionary systems (GA). Recently, DeepNEU (v5.0) has been used to simulate arti cially induced wild type and SARS-CoV2 infected Type 1 (AT1) and Type 2 (AT2) alveolar lung cells (aiLUNG and aiLUNG-COVID-19 respectively) derived from arti cially induced human pluripotent stem cells (aiPSC) (15).
The primary purpose of the current research project was to evaluate the DeepNEU stem-cell simulation platform by creating and validating computer simulations of arti cial AT1 and AT2 lung cells infected with SARS-CoV-2 to enable the rapid identi cation of therapeutic targets and drug repurposing for speci cally treating both the cytokine storm and coagulopathy associated with severe COVID-19.

Methods
The DeepNEU platform is a literature validated hybrid deep-machine learning system with elements of fully connected recurrent neural networks (RNN), cognitive maps (CM) support vector machines (SVM) and evolutionary systems (GA). The detailed methodology for simulation development and validation Page 4/24 plus the description of the current database (DeepNEU v5.0) used in these experiments has been previously described in detail (15)(16)(17).

The DeepNEU simulations
The main goal of this project was to extend our previous research into SARS-CoV-2 lung infection by evaluating a rationally de ned set of repurposed drugs and their combinations for the treatment of the cytokine storm and coagulopathy that often complicates this viral infection. As described previously (15),we rst created computer simulations (aiPSC) of human induced pluripotent stem cells (iPSC) and lung (aiLUNG) cells. Once validated with published peer reviewed data, the aiLUNG simulations were exposed to simulated SARS-CoV-2 infection by turning on extracellular Spike-RBP (RNA Binding Domain) in the presence of active Transmembrane Serine Protease 2 (TMPRSS2). Finally, several potential factor and combination of factor inhibitors were evaluated regarding their ability to ameliorate known features of the cytokine storm and coagulopathy often seen with severe COVID-19. A summary of the 17 single known drug simulations evaluated in the initial experiments are presented in Table 1 and the nal 15 known two drug combinations evaluated are summarized in Table 2. The same single and two drug combinations outlined below were applied to both the Cytokine Storm and Coagulopathy.   The nal predictions from the wild type aiPSC and aiLUNG simulations regarding the expression or repression of genes and proteins and presence or absence of phenotypic features were directly validated with published data as outlined previously (6). All experiments in this study were conducted in triplicate (N=3) using different initial conditions in the form of initial state vectors.
Detailed consensus genotypic and phenotypic pro les for both the cytokine storm and coagulopathy were generated from the peer reviewed literature. These features are summarized in Table 3 below and presented in graphic form in Figures 1A and 1B to contrast the highly signi cant differences between aiLUNG and aiLUNG-COVID-19 (two tailed Mann-Whitney u test P<0.0001). Table 3 COVID This test provides an exact probability, can compensate for prediction bias and is ideal for determining the statistical signi cance of experimental deviations from an actual distribution of observations that fall into two outcome categories (e.g., agree vs disagree). A p-value <0.05 is considered signi cant and is interpreted to show that the observed relationship between simulated and actual unseen wet lab data is unlikely to have occurred by chance alone. The pretest probability of a positive outcome prediction is 0.661 and the pretest probability of a negative prediction is therefore 0.339. This system bias was used when applying the binomial test to all simulation outcomes. For other between group comparisons the Mann-Whitney u test of signi cance was used (18). This nonparametric test was chosen because some of the data was not normally distributed.
To evaluate statistical signi cance and rank the treatment options, we used a multistep analysis of the repurposed single drug and two drug combination data as outlined below.
Step 1a: First we compared the aiLUNG-WT cell pro les to the aiLUNG-COVID-19 pro les using the Mann-Whitney u test to establish that these pro les were signi cantly different from each other.
Step 1b: The Mann-Whitney u test was then used to evaluate which treatment option pro les were signi cantly different from the aiLUNG-COVID-19 pro le. This would result in a form of clustering of the pro les relative to the uninfected aiLUNG pro le.
Step 2: The ranking of the predictions began with the calculation of Cosine Similarity (CS) for all interventions based on similarity to the aiLUNG or uninfected pro les. Cosine Similarity is a commonly used measure for comparing the similarity of two or more continuously valued vectors with the same number of elements. As similarity between the vectors increases, CS increases to +1 or maximum similarity. As CS similarity decreases away from the reference vector and becomes increasingly dissimilar, CS decrease towards -1 or maximum dissimilarity(19). We then used a simple mathematical transformation to derive Angular Cosine Distance (ACD) using the formula ACD = arcosine(CS)/Pi. ACD was selected because (1) it conforms with all four properties of a valid distance metric, (2) sample sizes are relatively small (N<20) minimizing any in uence of the curse of dimensionality and (3) it is a widely used and well validated metric for comparing bounded (-1 to +1) continuous valued vectors (20).
Step 3: Next, we ranked the cytokine storm and DIC like coagulopathy treatment options separately using the ACD metric.
Step 4: Next, the individual cytokine storm and coagulopathy data were sorted based on alphabetic order of the treatment options so that the two groups would be comparable.
Step 5: We then calculated the average of the ACD metric for cytokine storm and coagulopathy.
Step 6: Finally, the combined cytokine storm and coagulopathy data were sorted on the ACD metric to obtain the nal ranking of treatment options that are effective for both cytokine storm and coagulopathy outcomes. All nal repurposed drug combinations were selected such that the average ACD metric were greater than the ACD metric of the most effective single agent.

Results
The aiPSC and wild type (uninfected) aiLUNG simulations As reported previously both the unsupervised aiPSC simulations and the unsupervised aiLUNG simulations converged quickly (24 iterations) to a new system wide steady state without evidence of overtraining after 1000 iterations. The aiPSC simulations expressed the same human hESC speci c surface antigen and genomic pro le as both undifferentiated human embryonic stem cells (hESC) and induced pluripotent stem cells (iPSC) (15).The probability that all (N=15) of these aiPSC-WT outcomes were correctly predicted by chance alone using the binomial test is 0.0021.
The aiLUNG simulations produced similar genotypic and phenotypic expression pro les when compared with the human wild type (ATI and ATII) lung cell speci c factors taken from the literature (15). The probability that all (N=15) of these aiLUNG outcomes were correctly predicted by chance alone using the binomial test is 0.0021. Importantly, the data also indicate that the generation of aiLUNG cells from aiPSC produces a heterogenous population of alveolar cell precursors and more mature alveolar cells consistent with previous study (21).

Simulation of SARS-CoV-2-infected aiLUNG cells (aiLUNG-COVID-19)
The next step in the experiments was to expose the aiLUNG cells to simulated SARS-CoV-2 virus. For this simulated infection, the concept of SARS-CoV-2 viremia was activated (turned on). The viremia activates the viral life cycle beginning with the interaction of the viral Spike protein with its receptor protein Angiotensin-converting enzyme 2 (ACE2) and ending with exocytosis of new viral particles which completes the cycle by contributing new viral particles to the ongoing viremia (22). The SARS-CoV-2 genome consists of four structural genes, at least 6 non-structural genes and produces at least 10 proteins. The seventeen gene/protein expression pro le was compared with the uninfected aiLUNG simulations to assess the validity of simulated COVID-19. All gene/protein factors were expressed/upregulated in the aiLUNG-COVID-19 vs aiLUNG simulations. The probability that all (N=17) of these aiLUNG-COVID-19 simulation outcomes were correctly predicted by chance alone using the binomial test is 0.0009.
A phenotypic pro le of aiLUNG-COVID-19 was also developed from the published literature and has been described previously (15). These phenotypic features (N=8) include: New Extracellular Virus release, Spike-ACE2 Interface, Spike-RBD, TMPRSS2, Virus Clearance, Virus Intracellular RNA release, Virus Internalization and Virus Replication. The presence of all phenotypic features of COVID-19 was correctly predicted by the aiLUNG-COVID-19 simulations when compared with the aiLUNG simulations. The probability that all (N=8) of these aiLUNG-COVID-19 outcomes were predicted correctly by chance alone using the binomial test is 0.0364.
When we combined the genotypic and phenotypic pro les, the probability that all (N=25) features of simulated aiLUNG-COVID-19 were accurately predicted by chance alone using the binomial test is 0.00003.
Evaluation of the validated aiLUNG-COVID-19 simulations for repurposing combinations of known drugs that can mitigate COVID-19 associated cytokine storm and coagulopathy Comparison of the all predictions for >4100 factors from the aiLUNG-COVID-19 and aiLUNG simulations revealed a subset of factors that stood out as potential biomarkers. Based on the two tailed Mann-Whitney u test, the estimated p-values when comparing these aiLUNG-COVID-19 vs aiLUNG factors were highly signi cant at p=0.00001.
Evaluation of a subset of highly signi cant factors (N=17) were selected for further evaluation as previously outlined in (15). Inhibiting each of these factors in an iterative process produced variable effects in COVID-19 associated cytokine storm. Hydroxychloroquine (HCQ) was included in the initial evaluation because (i) it has multiple COVID-19 relevant cellular targets (23) (ii) it is already approved for other indications including malaria and in ammatory diseases, and (iii) early anti-COVID-19 results from at least one small trial appear promising (24). Interestingly, during our initial screening based on genotypic features, HCQ was the only effective single agent capable of signi cantly decreasing the cytokine storm.
In the cytokine storm simulations, the most effective single agent from the second screening run was an inhibitor of IL-1ab followed by inhibitors of IL-6, TNFa, COX2, IFN1ab and the addition of HCQ, when the output from aiLUNG cells and untreated aiLUNG-COVID-19 simulations were ranked using the ACD metric ( Figure 2A). In the experiments with the coagulopathy simulations the most effective single agent was an inhibitor of IL-1ab followed by inhibitors of IL-6, TNFa, COX2, IFN1ab and the addition of HCQ ( Figure 2B).
Based on these ndings a nal round evaluating N=15 two drug combinations, was carried out in triplicate. The individual two drug combinations are listed in detail in Table 2 above. Regarding the Cytokine Storm simulations, thirteen of the fteen two drug combinations evaluated were effective against COVID-19 using the two tailed Mann-Whitney u test based on the 14 viral target pro le outlined above. Ranking the effective combinations based on the ACD metric indicate that IL-1ab inhibition was a component of 6 of the effective combinations while TNFa inhibition and HCQ use were each a component of 5 of the effective combinations. Inhibitors of COX2 and IL-6 were components of 4 the effective the two drug combinations evaluated. Overall, the most effective 2 drug combination for ameliorating the cytokine storm appears to be HCQ plus an inhibitor of TNFa ( Figure 3A).
The analysis of the coagulopathy simulations revealed that thirteen of the fteen drug combinations evaluated were effective in COVID-19 using the two tailed Mann-Whitney u test based on the 17 viral target pro le outlined above. A nal ranking of these effective combinations based on the ACD metric indicated that TNFa inhibition and IL-1ab inhibition and HCQ application were components of 5 of the effective combinations, while inhibition of COX2 and IL-6 inhibition were each a component of 4 of the effective two drug combinations evaluated. Overall, the most effective double drug combination for ameliorating the coagulopathy appears to be a COX2 inhibitor plus an inhibitor of IL-1ab ( Figure 3B).
Importantly, we wished to determine which double drug combinations would potentially be most effective against both the cytokine storm and the coagulopathy associated with severe COVID-19. Based on the ACD metric ranking of comparable effective combinations, eight of the two drug combinations ranked above the best of the single drugs for potentially ameliorating both the cytokine storm and the coagulopathy complications. The nal ranking based on the average of two ACD metrics indicate that COX2 was a component of four effective combinations while IL-1ab inhibition and HCQ use were each an element of three of the effective combinations. Overall, the most effective double drug combination for ameliorating both the cytokine storm and the coagulopathy appears to be a COX2 inhibitor plus an inhibitor of IL-1ab followed closely by a combination of either a COX2 inhibitor plus an inhibitor of IL-6 or HCQ plus a TNFa inhibitor (Figure 4).

Comparing the E cacy of Single and Double Drug Combinations in Cytokine Storm and Coagulopathy
Finally, we wanted to determine if the data revealed any signi cant differences in e cacy between single drug and double drug combinations. A comparison of effective drugs and drug combinations indicated that (1) for the cytokine storm simulations the double drug combinations (N = 13, average ACD ± 95%CI = 0.046 ± 0.019) were generally more effective than single drugs (N = 7, 0.022 ± 0.005, two tailed Mann-Whitney u test p <0.01) and (2) for the coagulopathy simulations the double drug combinations (N=13, average ACD ± 95%CI = 0.103 ± 0.008) were also generally more effective than single drugs (N=7, 0.167 ± 0.041, two tailed Mann-Whitney u test p <0.01).

Discussion
Previously, we have evaluated the DeepNEU (v5.0) machine learning platform for simulating uninfected (aiLUNG) and SARS-CoV-2 infected differentiated Type 1 (AT1) and Type 2 (AT2) alveolar lung cells (aiLUNG-COVID-19).The primary purpose of this project is to extend our previous research into COVID-19 drug discovery, and to enable the rapid identi cation of therapeutic targets and drug repurposing speci cally for treating both the cytokine storm and coagulopathy frequently seen in patients with severe COVID-19.
In our recent publication (Esmail and Danter, 2020. Accepted)*, we presented data con rming that the DeepNEU (v5.0) platform can accurately simulate AT1 and AT2 lung cells based on the direct generation of human lung cells (iLUNG) from iPSC that is well documented in the peer reviewed literature (39,40).
Once the aiLUNG and aiLUNG-COVID-19 simulations were validated against the peer reviewed wet lab research, we applied them to therapeutic target identi cation and drug re-purposing. In the case of severe COVID-19, no re-purposed therapies are currently approved. The rational evaluation of currently licensed drugs to identify potentially effective therapies or simple double drug combinations may represent the most e cient path to improved patient outcomes when combined with respiratory support plus early and widespread testing.
The use of stem cells (iPSC) for targeted drug discovery has been well reported in the peer reviewed literature (52-54) and our approach to using the aiPSC derived aiLUNG-COVID-19 and aiLUNG simulations for COVID-19 speci c drug repurposing has been previously described in detail (15). In the current paper the same approach was used but in this project our drug repurposing efforts were focused speci cally on mitigating the cytokine storm and coagulopathy often seen with life threatening COVID-19. As of this writing the only widely available therapies showing promise are the early administration of systemic corticosteroids like dexamethasone and anticoagulation with heparin (55-57) Evaluating the aiLUNG-COVID-19 simulations for repurposing double drug combinations to mitigate COVID-19 associated cytokine storm and coagulopathy We had previously compared all genotypic features for >4100 factors from the aiLUNG-COVID-19 and aiLUNG simulations using the two tailed Mann-Whitney u test. A subset of seventeen factors emerged as potential therapeutic targets. This subset of highly signi cant (p=0.00001) genotypic factors (N=17) were selected for initial evaluation as previously outlined in (15). Inhibiting each of the seventeen factors in an iterative process produced variable effects on COVID-19 associated cytokine storm. As outlined above, the rational for including Hydroxychloroquine (HCQ) in the initial screening run was also proposed and interestingly HCQ was the only agent capable of signi cantly mitigating the cytokine storm during the initial screening. These signi cant results indicate that the COVID-19 associated cytokine storm and coagulopathy are more likely to be dependent on the host response to infection than on the SARS-CoV-2 genome itself. Therefore, anti-viral therapy primarily targeting the SARS-CoV-2 genome should not be expected to mitigate the cytokine storm and coagulopathy. In established and progressing COVID-19 other targeted approaches will be required (57-60).
To address this probable reality, we carried out a comprehensive review of the recent therapeutic literature to supplement our analysis and identi ed ten additional therapeutic candidates (11,19,(25)(26)(27)(28)(29)(30)(31)(32)(33)(34)(35)(36)(37)(38). These new candidates were con rmed by a more detailed statistical analysis of the aiLUNG vs aiLUNG-COVID-19 data and include Cathepsins B and L, COX2, IL-1ab, IL-6, Interferon ab, JAK1-3 and TNFa as summarized above in Table 4. This new data allowed us to evaluate a more extensive group of single drug candidates (N=16) in a second screening run. In the cytokine storm aiLUNG-COVID-19 simulations the most effective single agent from the second screening run was an inhibitor of IL-1ab followed by inhibitors of IL-6, TNFa, COX2, IFN1ab and the addition of HCQ, when the outputs from aiLUNG (uninfected) cells and untreated aiLUNG-COVID-19 simulations were ranked using the ACD metric (Figure 2A). Similarly, the experiments applying the drug candidates to the coagulopathy simulations also identi ed inhibitors of IL-1ab, IL-6, TNFa, COX2, IFN1ab and the addition of HCQ ( Figure 2B). These encouraging results allowed us to proceed with the evaluation of double drug combinations.
Based on these ndings a nal screening round was carried out to evaluate fteen double drug combinations. The individual double drug combinations evaluated are presented in detail in Table 2 above. From the cytokine storm simulations, thirteen of the double drug combinations evaluated were effective on the fourteen viral targets pro le outlined above. Ranking the most effective combinations according to the Angular Cosine Distance (ACD) from the uninfected pro le indicate that overall, the most effective double drug combination for mitigating the cytokine storm is HCQ plus an inhibitor of TNFa as presented in Figure 3A.
The analysis of the coagulopathy simulations also revealed that thirteen of the double drug combinations were effective based on the seventeen viral targets pro le outlined above. A nal ranking of these effective combinations based on the ACD metric indicated that overall, the most effective double drug combination for ameliorating the Coagulopathy appears to be a COX2 inhibitor plus an inhibitor of IL-1ab as presented in Figure 3B.
A central purpose of this project was to determine if any double drug combinations could mitigate both the cytokine storm and the coagulopathy associated with severe COVID-19. To that end we ranked the effective combinations based on their ACD metric and determined that seven of the double drug combinations ranked higher than the best of the single drugs based on their potential to alleviate both the cytokine storm and the coagulopathy complications. The nal ranking based on the average ACD metric indicated that overall, the most effective double drug combination was a COX2 inhibitor plus an inhibitor of IL-1ab followed closely by a combination of either a COX2 inhibitor plus an inhibitor of IL-6 or HCQ plus an inhibitor of TNFa ( Figure 4).
Importantly, all of these effective double drug combinations can be clinically evaluated because the approved drugs are currently available for urgent clinical study. For example, a single speci c COX2 Page 14/24 also widely available for use in North America. Although the results from clinical trials exploring the effectiveness of HCQ in preventing or treating COVID-19 have so far been disappointing, combining HCQ with a TNFa inhibitor has not yet been evaluated. Monoclonal antibody inhibitors of TNF including in iximab, adalimumab, certolizumab pegol and golimumab are available and are widely prescribed.
Etanercept, a circulating fusion protein TNF receptor inhibitor, is also available. A number of small molecules including thalidomide, lenalidomide and pomalidomide are also active as TNF inhibitors. Anakinra is a recombinant modi ed version of the human IL-1 receptor antagonist protein that has proven effective in life threatening sepsis and is available for clinical study in severe COVID-19 (31). Available monoclonal antibody inhibitors of IL-6 include tocilizumab and sarilumab, which are IL-6 receptor inhibitors and siltuximab which targets IL-6 itself.
Comparing the e cacy of single and double drug combinations in cytokine storm and coagulopathy were also generally more effective than single drugs (N=7, 0.167 ± 0.041, two tailed Mann-Whitney u test p <0.01).
These results support the conclusion that for serious diseases, drug combinations are generally more effective than single drugs (61,62). With a few exceptions like Gleevec for Chronic Myelogenous Leukemia (CML), this is particularly true in the majority of cancers, in life threatening sepsis and in situations where immune suppression is essential.
We remain aware that predictions from advanced computer simulations still require wet lab con rmation and this continues to be important for DeepNEU v5.0 as well. However, a key aim of this project was to make these ndings regarding the potential therapeutic bene t of novel double drug combinations for mitigating the complications of serious COVID-19 freely available to the global research community for wet lab validation at the very earliest opportunity. We also plan to validate these important ndings and we are currently looking for development partners with the goal of con rming them in animal models of severe COVID-19. We commit to making any additional information available at the earliest opportunity.

Conclusions/ Signi cance
The current results from our continued research and development of the DeepNEU platform have con rmed and extended our previous work (15)(16)(17). DeepNEU v5.0 accurately derived aiLUNG cells from aiPSC simulations that could be infected with simulated SARS-CoV-2 virus. These SARS-CoV-2 infected aiLUNG cells can reproduce the genotypic and phenotypic pro le typical of the cytokine storm and coagulopathy that too often complicate severe COVID-19. We also demonstrated that the aiLUNG-COVID- COVID-19 therapeutic potential for urgent animal and human trial validation. We also provide evidence that double drug combinations are expected to generally be more effective than single drugs for treating two common complications of severe COVID-19.
We believe it is inevitable that future viral outbreaks will occur and result in potentially lethal complications in large numbers of seriously ill people around the world. The rational process for viral therapeutic target identi cation and drug repurposing described in this manuscript requires the existence of a validated genome for the viral pathogen(s) in question. Although DeepNEU (v5.0) requires continued development and further validation, it is very likely that the most important application of this machine learning platform will be to improve our preparedness for serious future viral outbreaks and their life- maximum reduction in expression. The X-axis represents the expression level of each factor relative to the arbitrary baseline. Data represents mean of 3 experiments ± 99% Con dence Interval. the Angular Cosine Distance (ACD) from aiLUNG-WT. ACDs are calculated such that 0 = minimal distance and 1 = maximum distance relative to aiLUNG-WT. More effective treatment options will produce a response closer to aiLUNG-WT (close to 0). The X-axis represents the single or double drug combinations evaluated. Data represents mean of 3 experiments ± 99% Con dence Interval. * p<0.05 Figure 3 DeepNEU simulations of anti-coagulopathy e cacy (A) Anti-coagulopathy single drug e cacy against aiLUNG+COVID-19 cells. (B) Anti-coagulopathy double drug combinations e cacy against aiLUNG+COVID-19 cells. The vertical (Y) axis represents the single or double drug(s) e cacy based on the Angular Cosine Distance (ACD) from aiLUNG-WT. ACDs are calculated such that 0 = minimal distance and 1 = maximum distance relative to aiLUNG. More effective treatment options will produce a response closer to aiLUNG-WT (close to 0). The X-axis represents the single or double drug combinations evaluated. Data represents mean of 3 experiments ± 99% Con dence Interval. * p<0.05. with ACDs that were signi cantly below IL-1ab (p<0.05). Bars in blue represent treatment options that were not signi cantly different from IL-1ab. The vertical (Y) axis represents treatment e cacy based on