AI-guided discovery of the invariant host response to viral pandemics

We sought to define the host immune response, a.k.a, the “cytokine storm” that has been implicated in fatal COVID-19 using an AI-based approach. Over 45,000 transcriptomic datasets of viral pandemics were analyzed to extract a 166-gene signature using ACE2 as a ‘seed’ gene; ACE2 was rationalized because it encodes the receptor that facilitates the entry of SARS-CoV-2 (the virus that causes COVID-19) into host cells. Surprisingly, this 166-gene signature was conserved in all viral pandemics, including COVID-19, and a subset of 20-genes classified disease severity, inspiring the nomenclatures ViP and severe-ViP signatures, respectively. The ViP signatures pinpointed a paradoxical phenomenon wherein lung epithelial and myeloid cells mount an IL15 cytokine storm, and epithelial and NK cell senescence and apoptosis determines severity/fatality. Precise therapeutic goals were formulated and subsequently validated in high-dose SARS-CoV-2-challenged hamsters using neutralizing antibodies that abrogate SARS-CoV-2•ACE2 engagement. IL15/IL15RA were elevated in the lungs of patients with fatal disease, and plasma levels of the cytokine tracked with disease severity. Thus, the ViP signatures provide a quantitative and qualitative framework for titrating the immune response in viral pandemics and may serve as a powerful unbiased tool to rapidly assess disease severity and vet candidate drugs. One Sentence Summary The host immune response in COVID-19.


ABSTRACT (197 words)
We sought to define the host immune response, a.k.a, the "cytokine storm" that has been implicated in fatal COVID-19 using an AI-based approach. Over 45,000 transcriptomic datasets of viral pandemics were analyzed to extract a 166-gene signature using ACE2 as a 'seed' gene; ACE2 was rationalized because it encodes the receptor that facilitates the entry of SARS-CoV-2 (the virus that causes COVID-19) into host cells. Surprisingly, this 166-gene signature was conserved in all viral pandemics, including COVID-19, and a subset of 20-genes classified disease severity, inspiring the nomenclatures ViP and severe-ViP signatures, respectively. The ViP signatures pinpointed a paradoxical phenomenon wherein lung epithelial and myeloid cells mount an IL15 cytokine storm, and epithelial and NK cell senescence and apoptosis determines severity/fatality. Precise therapeutic goals were formulated and subsequently validated in high-dose SARS-CoV-2challenged hamsters using neutralizing antibodies that abrogate SARS-CoV-2•ACE2 engagement.
IL15/IL15RA were elevated in the lungs of patients with fatal disease, and plasma levels of the cytokine tracked with disease severity. Thus, the ViP signatures provide a quantitative and qualitative framework for titrating the immune response in viral pandemics and may serve as a powerful unbiased tool to rapidly assess disease severity and vet candidate drugs.

INTRODUCTION
As the rapidly unfolding COVID-19 pandemic claims its victims around the world, it has also inspired the scientific community to come up with solutions that have the potential to save lives.
In the works are numerous investigational drugs at various phases of clinical trials, from rationalizing 1 , to IRB approvals, recruitment and execution 2,3 , all directed to meet an urgent and unmet need -i.e., ameliorate the severity of COVID-19 and reduce mortality. Two obstacles make that task difficult-First, the pathophysiology of COVID-19 remains a mystery. The emerging reports generally agree that the disease has a very slow onset 4,5 and that those who succumb typically mount a 'cytokine storm' 4,6 , i.e., an overzealous immune response.
Despite being implicated as a cause of mortality and morbidity in COVID-19, we know virtually nothing about what constitutes (nature, extent) or contributes to (cell or origin) such an overzealous response. Consequently, treatment goals in COVID-19 have been formulated largely as a 'trial and error'-approach; this is reflected in the mixed results of the trials that have concluded 7 . Second, there is no established pre-clinical animal or human cell/organoid models for COVID-19; vetting the accuracy and/or the relevance of such models requires first an understanding of the host response in the disease.
We set out to define this aberrant host immune response in COVID-19 using machine learning tools that can look beyond interindividual variability to extract underlying gene expression patterns within multidimensional complex data. The approach was used across multiple cohorts of viral pandemics. The resultant pattern, i.e., signature, was subsequently exploited as a predictive model to navigate COVID-19. Findings not only pinpointed the precise nature of the cytokine storm, the culprit cell types and the organs, but also revealed disease pathophysiology, and helped formulate specific therapeutic goals for reducing disease severity. Key findings were validated in preclinical models of COVID-19 in Syrian hamsters and in the lungs and plasma of infected patients.

An ACE2-centric Study Design
To identify and validate an invariant (universal) gene signature of host response in COVID-19, we mined more than 45,000 publicly available datasets of viral pandemics across three species (human, mouse and rats) (Step 1; Fig 1). Three relatively widely accepted facts shaped our approach using Angiotensin-converting enzyme 2 (ACE2) as 'seed' gene in our computational studies: (i) ACE2 is the most well-known portal for SARS-CoV-2 entry into the host cell 8,9 ; its expression in cell lines correlates with the expression of innate immune genes 10 and susceptibility to SARS-CoV spike protein-driven entry 11,12 , and its depletion in mice abrogates SARS-CoV infection 13 ; (ii) ACE2 is a potent negative regulator of the renin-angiotensin aldosterone system (RAAS) 14 ; without such restraint, the RAAS contributes to exuberant inflammation in the setting of infections 15 ; and finally, (iii) although the mechanism through which ACE2 suppresses inflammatory response remains poorly understood, accumulating evidence indicates that infections perturb ACE2 activity, allowing for uncontrolled inflammation [16][17][18][19][20][21][22][23][24] .

As
Step 2 (Fig 1), we validated the signature in several human and mouse datasets of viral pandemics, and a subset of genes was identified and validated as indicators of disease severity.
The signatures were then validated in SARS-CoV-2-infected cells and tissues and to explore the nature, extent and cell of origin of host response in mild and fatal COVID-19.

As
Step 3 (Fig 1), the gene signatures were used to navigate the uncharted territory of COVID-19 and pinpoint immunopathologic mechanisms, which revealed the nature (IL15), source (airway epithelium), intensity (quantitative measure) and consequence (NK cell senescence) of the cytokine storm and helped objectively formulate precise therapeutic goals to reduce the severity of COVID-19.

As
Step 4 (Fig 1), the gene signature and the mechanism of action (IL15/IL15RA) was validated in lung tissues from SARS-CoV-2 challenged golden hamster using RNASeq and IHC.
In addition, a precise therapeutic goals on SARS-CoV-2 was validated in the golden hamster model. The mechanism of action (IL15/IL15RA) was also validated by ELISA in plasma and IHC in lung tissues from UCSD COVID-19 cohort participants (Supplementary Table S4-5).

A Shared Host Response Signature in Respiratory Viral Pandemics
Because publicly available transcriptomic datasets from SARS-CoV-2-infected samples are still relatively few, any conclusion drawn from so few samples using any computational methodology is likely to lack robustness. We chose to use an informatics approach, i.e., Boolean Equivalent Correlated Clusters (BECC) 25 , which can identify fundamental invariant (universal) gene expression relationships underlying any biological domain; in this case, we selected the biological domain of 'respiratory viral pandemics'. BECC enables comparison of the normalized expression of two genes across all datasets by searching for two sparsely populated, diagonally opposite quadrants out of four possible quadrants (high-low and low-high), employing the BooleanNet algorithm 26 . There are six potential gene relationships assessed by BooleanNet: two symmetric (Equivalent and Opposite; Fig 2A) and four asymmetric 26 . Two genes are considered "Boolean Equivalent" if they are positively correlated with only high-high and low-low gene expression values. Two genes are considered "Boolean Opposite" if they are negatively correlated with only high-low and low-high gene expression values. Asymmetric Boolean implications result when there is only one sparsely populated quadrant. The BECC algorithm focuses exclusively on "Boolean Equivalent" relationships to identify potentially functionally related gene sets. Once identified, these invariant relationships have been shown to spur new fundamental discoveries 27,28 , with translational potential 29 , and most importantly, offer insights that aid the navigation of uncharted territories where nothing may be known 30,31 .
We used GSE47963 [human airway epithelial (HAE) cultures with H1N1 and SARS-CoV infections; n = 438] as a 'test' dataset, which was comprised of human airway epithelial cell samples (HAE) infected in vitro with the causative agents of the 2009 'swine flu' (influenza A-H1N1; a triple recombination of human, avian, and swine influenza viruses [32][33][34] ) and the 2002 Severe acute respiratory syndrome (SARS-CoV-1) 35 outbreaks (Fig 2B). These datasets were chosen now, and other datasets were prioritized later in the study, e.g., H5N1 (the causative agent of the avian flu in 2006-06 36 ) and MERS-CoV (the causative agent of Middle East respiratory syndrome in 2012 37 ) based upon the fact that they all contributed to outbreaks that are characterized by acute respiratory syndromes with high case-fatality rates 8 .
ACE2 is used as a 'seed' to identify other genes that have 'Boolean Equivalent' and 'Boolean Opposite' relationships with ACE2. These genes were subsequently filtered using differential analysis on another dataset [GSE113211 (n = 118); Fig 2B] that profiled heterogeneous immunophenotypes of children with viral bronchiolitis (confirmed positive for the virus in ~100% patients; of which 25 % were infected with Influenza/Para-Influenza and 14.8% with human CoV).
Transcriptomes were analyzed in nasal mucosal scrapings (NMS) and PBMC samples taken during an acute visit (AV) and during a subsequent visit at convalescence (CV) 38 . 166 genes (Table S1; 1-1) retained the "Boolean Equivalent" relationship with ACE2 and their expression was downregulated during the convalescence visit. 26 genes (Table S1; 2-1) retained "Boolean Opposite" relationships with ACE2 and their expression were upregulated during the convalescence visit. All subsequent analyses were performed using the 166 -gene signature that had Boolean Equivalent relationship with ACE2 and that was down-regulated during a convalescent visit after acute viral bronchiolitis.
First the 166-gene signature was evaluated in the test dataset--it was used to rank order the samples and test for phenotype classification using a receiver operating characteristic curve [ROC curve; the area under this curve (AUC) represents degree or measure of separability] and displayed such classification using violin plots (Fig 2C-D). The signature classified the uninfected vs.
infected samples with reasonable accuracy in the setting of SARS-CoV-1 infection (ROC-AUC = 0.81, Fig 2C). It also classified perfectly in the setting of H1N1 infection (ROC-AUC = 1.00, Fig   2D). Good classification was observed between samples from the acute visit (AV) and

The ViP Signature Defines the 'Cytokine Storm' in Viral Pandemics
Reactome analyses on the 166 genes showed that the signature was largely enriched for genes within the immune system pathways, e.g., interferon and cytokine signaling, cellular processes that are critical for an innate immune response such as the ER-phagosome pathway and antigen processing and presentation, and finally the adaptive immune system (Fig 3A-C). In other words, the signature reflected a typical host immune response that is expected during any viral infection. This is not surprising because an overzealous host immune response, i.e., a 'cytokine storm' is shared among all respiratory viral pandemics (Influenza, avian and swine flu) 41 and severe COVID-19 patients who succumb to the disease 6 . However, there were 3 surprising factors: (i) This signature and reactome profile emerged using ACE2 as a 'seed' gene, which is not the receptor for influenza strains to enter into host cells. (ii) It is also noteworthy that despite filtration through two unrelated datasets (Fig 2B), one in vitro and another in vivo, and the reduction in the # of genes in the ACE2-equivalent cluster during such iterative refinement, the pathways/processes represented in the 166-gene cluster ( Table S1;  days, which can sometimes last up to ~10-14 d. Among datasets with curated samples representing other viral pandemics that are neither respiratory nor acute, we found that classification scores for RNA viruses were significantly better compared to DNA viruses in in vitro systems (Fig 3F top,   Fig S1A), especially for those that share clathrin-dependent endocytic methods to breach host cells (Table S2). However, the classification scores were indistinguishable between RNA and DNA viruses in in vivo studies (Fig 3F bottom, Fig S1A-B). These results indicate that the 166-gene signature is shared among all viral pandemics, and not specific to respiratory viral pathogens.
Notably, the 166-gene host response signature was specific for viral infections; it performed poorly in classifying bacterial lung infections compared to viral infections (p < 0.001; Fig 3G). The signature also implicated the epithelial and myeloid cells, but not ECs and fibroblasts contribute to host immune response because the classification scores were better for airway epithelial cells (AE) and dendritic cells (DC) compared to fibroblasts (FI) and microvascular endothelial cells (ME) (ROC-AUC: 0.66, 0.82 vs 0.43, 0.37; Fig 3H). These scores correlated well with ACE2 expression in these different cell types (p < 0.001; Fig 3H), raising the possibility that viral entry through engagement of ACE2 and the induction of ACE2-equivalent host genes may be intertwined. That myeloid cells are major contributors to this signature was confirmed in five independent datasets; the 166-gene signature distinguished 'reactive' (M1-polarized) macrophages in them all (Fig 3I).
Together, these findings indicate that the ACE2-equivalent 166-gene signature is specific enough to distinguish between bacterial and viral pulmonary infections but is of broader relevance.
The airway epithelial cells and cells of myeloid lineages (DCs and macrophages) appear to be major contributors to the ViP signature.

A 20-gene Subset within the ViP Signature Detects Disease Severity
To determine what constitutes 'severe/fatal' disease, we rank-ordered the 166 genes within the ViP signature for their ability to classify Influenza A/B-infected adult patients by clinical severity 45,46 (n = 154; Fig 4A). Severe disease was defined as intubation and mechanical ventilation due to poor oxygenation and/or death. A set of top 20 genes (Fig 4A; Table S1, 3-1; Table S3) was sufficient to classify healthy controls from infected patients (ROC-AUC = 1.00) as well as distinguish mild from severe disease with reasonable accuracy (ROC-AUC = 0.95) in the test cohort (Fig 4B). Reactome pathway analyses revealed that compared to the ViP signature, the 'severity'-related 20-gene cluster enriches a completely different set of cellular processes, i.e., DNA damage (especially induction of genes that are critical for base excision repair; BER), stressinduced senescence, neutrophil degranulation and changes in cell cycle (Fig 4C, Fig S2). We validated this signature side-by-side with the 166-gene ViP signature in three human datasets that included samples from mild vs. severe disease during the avian (H7N9), IAV (H3N1 and others) and the swine (H1N1) flu viral pandemics (Fig 4D, left). Both the 166-gene ViP signature and the 20-gene severity signature performed similarly when it came to classifying control vs. mild disease, but the latter performed significantly better in classifying mild vs. severe disease, and did so consistently in both validation datasets (ROC-AUC ranging from 0.8-0.9; Fig 4D, left).
The severity signature performed well also in a large murine lung dataset in which mice were intranasally infected with non-lethal (NL, control), sub-lethal (SL, mild) and lethal (L, We conclude that the 166-gene ViP signature that was initially built using in vitro infection datasets detects also the host immune response ('cytokine storm') in the complex in vivo systems where the response may be triggered by direct viral damage to the lung epithelium, but get propagated by feed-forward dysregulated immune response, both innate and adaptive.
Surprisingly, this 166-gene ViP signature was not associated with disease severity; instead, severity-associated 20 genes that regulate stress and senescence-associated repression of protein expression and DNA damage (Fig 4C). DAVID GO analyses on the 20-gene signature indicated that 3 biological processes, e.g., transcriptional repression, apoptosis, and intermediates within the type I IFN (IFNγ signaling) pathway (Fig 4E) indicative of cellular distress, senescence/aging and death are the determinants of severity/fatality.

The ViP Signatures are Induced in the Lung Epithelial and Immune cells in COVID-19
We next tested the ability of the ViP signatures to distinguish between SARS-CoV-2-infected samples and uninfected controls in 3 independent datasets, 2 of which were datasets generated from cells infected in vitro (Fig 5A-C) and one that was generated from lung samples from a fatal case of COVID-19 (Fig 5D). The signature perfectly classified infected from uninfected samples in them all (ROC-AUC 1.00; Fig 5A-B-D); of the 166 genes, both IL15 and IL15RA were notably elevated in infected samples (Fig 5A). We next tested the ability of these signatures to distinguish mild vs. fatal COVID-19 in single cell sequencing datasets from patient-derived lung samples (Fig 5I). The 166-gene signature was able to distinguish control vs. mild infection most effectively in macrophages, airway epithelium, CD4+ T cells and NK cells (Fig 5I, lower panel, lower row) and mild vs. severe disease in the epithelium and in NK cells (Fig 5I, lower panel, upper row). The 20-gene signature not only performed well in classifying control vs. mild infection in the same 4 cell types as above but also in B cells and CD8+ T cells (Fig 5I, upper panel, lower row). However, the 20-gene severity signature continued to perform most optimally in the epithelium (ROC-AUC = 1.00) and in NK cells (Fig 5I, upper panel, upper row).
Together, these findings show that the 166-gene ViP signature seen in other respiratory viral pandemics is conserved also in COVID-19. The cytokine storm (166-genes, which included IL15/IL15RA; Table S1) was induced in multiple cell types; however, the 20-gene ViP signature of disease severity and fatality was most prominently induced in two cell types: (i) the airway epithelial cells, known producers of IL15 after viral infections 48 , 49 and (ii) the NK cells which are known targets of physiologic as well as overzealous IL15 response 50,51 .

Viral infection and IL15 Induce, and Flu Vaccine Attenuates the ViP Signatures in NK cells
We next asked how the ViP signatures are impacted when NK cells are exposed to virus-infected epithelial cells. NK cells are known to lyse influenza virus-infected cells by direct cytotoxicity and antibody-dependent cellular cytotoxicity (ADCC); enhancing such NK cell function has been shown to control influenza virus infections 52 . Clearance of other viruses (HIV-1, other retroviruses, etc.) and cancer immunotherapies also leverage such NK cell-dependent ADCC 53,54 .
We analyzed a transcriptomic dataset (GSE115203) 55 generated from co-culture studies of human PBMCs (3 donors) with influenza (H1N1 Puerto Rico/08/1934)-infected airway epithelial cells (A549) (Fig 6A; top). PBMCs (from co-culture), or NK cells FACS-sorted from the PBMC were then analyzed by RNA Seq, and the study had confirmed NK cell ADCC responses were durably induced in this assay via type I IFN release from PBMCs. We found that both the 166-and 20gene ViP signatures were induced in PBMCs and in NK cells sorted from the PBMCs (Fig 6A; bottom left), indicating that NK cells in these co-culture models were sufficient to capture the observed host immune response in patients with COVID-19.
To test the role of IL15 in the induction of ViP signatures, we leveraged three datasetsone that used recombinant IL15 (PBMCs; GSE77601), another that used anti-IL1Rβ mAb (mouse skin biopsies; GSE45551) 56 , and a third study using the prototypic H3K27 demethylase inhibitor, GSK-J4; the latter was shown to inhibit NK cell effector cytokines in response to IL15 without impacting its cytotoxic killing activities (human, NK cells; GSE89484) 57 . Both ViP signatures were stimulated by IL15, but attenuated in the two other datasets where IL15's actions were blocked pharmacologically (Fig 6A; right). These findings indicate that IL15 could be necessary and sufficient to induce the ViP signatures.
Because two independent studies 58,59 recently showed that those vaccinated against influenza have lower odds of requiring intensive care, invasive ventilation and/or dying, we analyzed two transcriptomic datasets (GSE64655 60 and GSE133478 61 ) in which PBMCs from subjects vaccinated with seasonal trivalent or quadrivalent influenza vaccine (TIV/QIV) were collected and analyzed for NK cell activation. The first study showed that the NK cells continued to demonstrate progressive attenuation of both the 166-and 20-gene signatures rapidly within 7 days (Fig 6B, left). The second study, in which the NK cell-enriched and depleted fractions collected pre-and post(30 d)-vaccination were tested for their response to re-stimulation with IL15 (low dose, 0.75 ng/ml, 18 h); such stimulation is known to enhance NK cell activity [62][63][64] and promote viral clearance [65][66][67] . Both ViP signatures were attenuated post-vaccination in NK cellenriched fractions, but not in depleted fractions (Fig 6B, right). Because such post-vaccination attenuation happened in the setting of experimentally confirmed 61 enhancement of overall NK cell response, we conclude that attenuation of ViP signatures among recipients of TIV could continue to offer protection during re-challenge. Because such protection is seen in NK-cell enriched, but not depleted fractions, we conclude that the protection is mediated primarily via the preservation of functional NK cells.

An IL15-storm Originating in the Lung Alveoli Determines the Severity of COVID-19
We next analyzed the ViP signatures in transcriptomic datasets generated from multiple organs at autopsy. Both the 166-and 20-gene ViP signatures were predominantly enhanced in one organ, the lungs (Fig 6C-D); and IL15/IL15RA were also elevated in the lungs (Fig 6D). These findings indicate that the 20-and 166-gene signatures go together, and suggest a plausible cause and effect relationship. For instance, severity-related cellular events (such as epithelial and NK cell senescence) occur in the milieu of the organ that mounts the highest IL15-predominant cytokine response, i.e., lungs. We also found that IL15 and its receptor IL15RA were significantly increased in severe COVID-19 lungs (Fig 6E). These findings predict that an overzealous IL15-predominant cytokine response is the most consistent finding in the most severe cases of COVID-19 and that the lung epithelium is the likely source of such a storm.
These predictions were validated in a cohort of symptomatic COVID-19 patients who presented to the UC San Diego Medical Center with varying disease severity, ranging from mild to fatal (see Table S4). Plasma ELISA studies revealed that IL15 levels were significantly elevated during the acute compared to the convalescent visit (Fig 6F), and in whom the clinical presentation was moderate-to-severe compared to those with mild disease (Fig 6G). A sub-group analysis confirmed that while gender or age did not have a significant impact on plasma IL15 levels independently, the aged male (> 40 y) cohort had a significantly higher IL15 level than the young males (Fig 6H; left). No such pattern was noted among females. These findings are consistent with the fact that the gender gap in COVID-19-related deaths widens markedly with age 68 . Lungs collected during autopsies from patients who succumbed to COVID-19 (see Table S5) further confirmed that lung epithelial cells, especially the alveolar type II pneomocytes and alveolar immune cell infiltrates express high levels of IL15 and its receptor, IL15RA (Fig 7A-B).
Taken together, these findings support the following model of the immunopathogenesis of COVID-19 (Fig 7C): Airway epithelial cells and cells of the myeloid lineage and other immune cells are the primary source of the 166-gene cytokine storm, of which, IL15 is a component. It is possible, that the primary target of IL15, i.e., NK cells, when exposed to this storm for a prolonged period undergo damage, stress-induced senescence and apoptosis. Our model is consistent with prior studies showing that the airway epithelial cells (especially bronchial) constitutively express the IL15 and IL15RA/B genes and that viral infections 49 and IFNγ induced the synthesis and secretion of IL15 49 and that prolonged and excessive stimulation with IL15 is known to induce NK cell exhaustion 50,51 . These findings are consistent with the emerging reports that NK cells are significantly exhausted and reduced in cases of severe COVID-19 infection 69,70 and that such reduction was seen as early as 3-6 days after the onset of symptoms 71 . We conclude that fatal COVID-19 is characterized by a paradoxical immune response, i.e., suppression of epithelial and NK cell functions (immunosuppression) in the setting of a cytokine storm (overzealous immune response).

The ViP Signatures Formulate Therapeutic Goals, Track Treatment Efficacy
Previously we showed that the attenuation of the ViP signature was 'associated' with the acquisition of natural convalescence in several respiratory viral pandemics (Fig 2F-H); we now asked if they could serve as a readout of 'causality' in the setting of therapeutic intervention. We analyzed interventional studies in the setting of other viral infections that shared the ViP signature, i.e., HCV, HIV, Zika and Ebola (Fig 3H; Fig S1; Table S2). The 166-gene ViP signature classified HCV-infected liver biopsies treated or not with directly acting anti-viral agents (DAAs) (Fig 8A-C) and HIV-infected samples treated or not with anti-retroviral therapeutics (ART; ROC-AUC = 1.00; Fig 8D) with sufficient accuracy. In the case of Ebola, the ViP signature was somewhat effective in classifying crisis (i.e., acute) from convalescent PBMC samples (ROC AUC 0.64; Fig   8E), and previously described anti-Ebola therapeutic strategies (Topoisomerase depletion with siRNA 72 inhibited the signature in Ebola-infected alveolar epithelial cells (siTop; ROC AUC 1.00; Fig 8F) 72 . Finally, the ViP signature was accentuated in Zika infected human cortical neural progenitor cells (Fig 8G) and was effectively attenuated when these infected samples were treated with two investigational drugs that was found to be effective in inhibiting Zika infection. These findings suggest 'causality' between treatment (anti-viral therapies) and response (attenuation of the ViP signature), and imply that attenuation of the 166-gene ViP signature is a desirable therapeutic goal.
We next sought to determine if the SARS-CoV-2 virus is responsible for the host response and validate the ability of the ViP signatures to track therapeutic response. To this end, we used SARS-CoV-2-neutralizing antibodies whose design was inspired by monoclonal antibodies (mAbs) isolated from from convalescent donors 73 Fig 8I); (ii) their lungs were protected from overwhelming immune cell infiltration and obliteration of alveolar space (Fig 8J-K); (iii) expression of IL15 and IL15 receptor was significantly reduced compared to what was observed in the infected lungs (Fig 8J, L). These results validate the ACE-centric computational approach for identifying the ViP signatures, i.e., when ACE2•virus engagement was disrupted using antibodies, the signatures were suppressed. The findings also indicate that the reversal of the signature and the IL15 storm could be used as a readout of therapeutic efficacy.

CONCLUSION
The major and unexpected finding in this work is that all viral pandemics (regardless of their acuity, causative virus, case fatality rates and clinical presentation) share a common fundamental host immune response. Summarized below are our three major findings.
First, we defined an invariant 166-gene host response --the so-called "cytokine storm"that is surprisingly conserved among all viral pandemics. What was also unusual is that the signature emerged despite the rationalized use of ACE2 as a 'seed' gene to identify 'coronaviral infection-associated genes'. This suggests that while ACE2 may be the entry site for SARS-CoV-2, it is a prominently upregulated gene during host response to other viral infections. We conclude that the ViP signatures not just define the nature of the host immune response to viral pandemics, but also allows the tracking and quantification of the such response. Second, we define the precise nature of the cytokine storm and pinpoint the IL15 cytokine and its receptor, IL15RA as an invariant component. We demonstrate that systemic levels of IL15 tracks with disease severity among patients and that the levels are notably elevated in the aged male (the predisposed age group in COVID-19, as per reports worldwide). Using a combination of single cell RNA Seq and human lung histology, we also pinpoint the lung epithelial and myeloid cells as the key contributors to the ViP signature, and more specifically, IL15/IL15RA. Third, we found that a subset of 20-gene 'severe' ViP signature, indicative of stress-induced senescence, transcriptional repression, DNA damage and apoptosis is also shared among various viral pandemics. In patients with COVID-19, this signature was seen in lung epithelial and NK cells, which is intriguing because airway epithelial cells is a prominent source and the NK cells are a major target of IL15. Thus, the ViP signatures begin to paint a picture of 'paradoxical immunosuppression' at the heart of fatal COVID-19, in which, the observed NK cell exhaustion/depletion in severe COVID-19 69-71 could be a consequence of an overzealous IL15 storm, leading to their senescense and apoptosis.
In closing, given that the emerging pandemic is still largely a mystery to us in terms of how it picks its victims, the ViP signature we define here provides a computational framework for navigation in otherwise uncharted territory. While it is expected that the signature will be more effective and accurate when it is iteratively filtered using emerging COVID-19 datasets, we provide evidence for its usefulness now in formulating therapeutic strategies and rapidly screening for therapeutics. Because the ViP signature of host response is seen also in other viral pandemics tested, findings may also be relevant also in navigating management strategies in those pandemics. Hamster study with anti-CoV-2 therapy: Lung samples from 8-week old Syrian hamsters were generated exactly previously published study 11 . See Supplementary methods for details regarding IHC protocols.

Competing interests:
The authors declare no competing interests.

Data and materials availability:
All data is available in the main text or the supplementary materials.

Supplementary Materials: Includes
• Detailed Materials and Methods

• Supplementary Text-n/a
• Figures S1, S2 • Tables S1-5 • External Databases -None     Dataset is analyzed as a 'pseudo-bulk' of all cells or after selecting individual cell types using marker genes specifically expressed in these cell types.

Signature validation in COVID-19
Discovery and validation of an ACE2-centric, 166-gene transcriptomic signature for host immune response.