Abstract

   Epithelial immune responses govern tissue homeostasis and offer drug
   targets against maladaptation. Here, we report a framework to generate
   drug discovery–ready reporters of cellular responses to viral
   infection. We reverse-engineered epithelial cell responses to
   SARS-CoV-2, the viral agent fueling the ongoing COVID-19 pandemic, and
   designed synthetic transcriptional reporters whose molecular logic
   comprises interferon-α/β/γ and NF-κB pathways. Such regulatory
   potential reflected single-cell data from experimental models to severe
   COVID-19 patient epithelial cells infected by SARS-CoV-2. SARS-CoV-2,
   type I interferons, and RIG-I drive reporter activation. Live-cell
   image–based phenotypic drug screens identified JAK inhibitors and DNA
   damage inducers as antagonistic modulators of epithelial cell response
   to interferons, RIG-I stimulation, and SARS-CoV-2. Synergistic or
   antagonistic modulation of the reporter by drugs underscored their
   mechanism of action and convergence on endogenous transcriptional
   programs. Our study describes a tool for dissecting antiviral responses
   to infection and sterile cues and rapidly discovering rational drug
   combinations for emerging viruses of concern.
     __________________________________________________________________

   A phenotypic platform discovers drugs and their mechanisms of action by
   synthetic genetic tracing of antiviral innate responses.

INTRODUCTION

   Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is the
   etiological agent of the coronavirus disease 2019 (COVID-19) pandemic,
   which caused the death or long-term illness of millions of people while
   simultaneously imposing severe social and economic burden worldwide.
   COVID-19 is currently managed through massive vaccination, notably with
   the aid of novel mRNA vaccines ([40]1). However, SARS-CoV-2 displayed
   remarkable evolutionary potential and multiple strains emerged over the
   past 2 years, with the Omicron variant currently being the dominant
   strain globally. Immunity against SARS-CoV-2 strains is limited and
   continuously threatened by emerging variants ([41]1).

   Pharmacological treatments that reduce the risk of progression in
   COVID-19 patients are imperfect, and their efficacy is restricted to
   the early intervention ([42]2–[43]5). In particular, the direct
   intervention through antiviral drugs has been generally underwhelming,
   and the currently approved orally bioavailable prodrug of
   N4-hydroxycytidine (molnupiravir) showed modest efficacy in a
   placebo-controlled phase 2/3 trial ([44]2), indicating that combination
   therapies will be required to effectively control disease progression.

   One preferential entry site for SARS-CoV-2 is the nasal cavity, where
   the cells located in the upper respiratory tract, such as the nasal
   epithelial cells and alveolar epithelial type II cells, are targeted by
   the virus. Both share high coexpression of ACE2 and TMPRSS2, the
   putative main receptor and surface protease, respectively, mediating
   the entry of the several SARS-CoV-2 strains ([45]6). SARS-CoV-2
   infection leads to transcriptional changes associated with innate
   immune response in the epithelial cells, including chemokine secretion
   ([46]7), which is believed to trigger a cascade of proinflammatory
   innate immune cell recruitment (macrophages and neutrophils) and
   disease progression ([47]7). Therefore, the first line of defense
   against SARS-CoV-2 infection is the innate immune system of the upper
   respiratory tract epithelial cells. One key step of the epithelial
   cells innate immune response is the production of type I interferons
   (IFNs), which induce an antiviral state in neighboring cells and
   trigger the activation of the adaptive immune system to protect the
   host against viral infection. SARS-CoV-2 elicits activation of pattern
   recognition receptors (PRRs), such as cytoplasmic RIG-I/MDA5 ([48]8) as
   well as Toll-like receptor 2 (TLR2) ([49]9) and several other TLRs
   ([50]10). Upon binding to the viral nucleic acids, PRRs activate the
   nuclear translocation of two families of transcription factors (TFs),
   notably the IFN regulatory factors (IRFs), such as IRF3 and IRF7, and
   the nuclear factor κB (NF-κB) factors p65 and p50, which coordinate the
   antiviral and acute inflammatory cellular responses ([51]11). One of
   the subsequent interconnected events is the transcription of IFN genes,
   which in turn are secreted and engage the type I IFN receptors (e.g.,
   IFNAR1 and IFNAR2). IFN activation of the Janus kinase/signal
   transducer and activator of transcription (JAK/STAT) signaling pathway
   in both infected and neighbor cells leads to the transcription of
   IFN-stimulated gene (ISG) signaling that builds up an antiviral state
   in the cell ([52]6). To counteract host cell responses, viruses hijack
   the activation of the IFN pathway, limit the active antiviral response,
   and inhibit antiviral signaling ([53]12). This leads to the coexistence
   of viral replication and viral antagonism within the infected cell
   population. Hence, the transcriptional response of epithelial cells is
   one of the key events that inform on viral infection. Moreover, innate
   immunity can be activated or amplified by several other pathways,
   including purely sterile inflammatory environmental triggers, such as
   DNA damage inducers ([54]13).

   The pathways controlling the transcriptional response of epithelial
   cells to infection offer an outstanding opportunity for intervention.
   Whereas drugs with different mechanisms of action targeting the key
   nodes of the PRRs, NF-κB, and JAK/STAT pathways are available and offer
   an opportunity for creating combination therapies, there is lack in
   technologies to rationally design combination treatments with
   synergistic antiviral activity and less prone to resistance.
   Furthermore, phenotypic drug discovery platforms to identify novel
   drugs and combinations thereof are potential game changers ([55]14),
   but there are limited phenotypic readouts for SARS-CoV-2 infection, and
   RNA viruses in general. Recently, we developed a method to generate
   synthetic locus control regions (sLCRs) to enable genetic tracing of
   complex phenotypes, including inflammatory transcriptional responses
   ([56]15, [57]16). Using an evolved computational approach to design
   phenotype-specific sLCRs named logical design of synthetic
   cis-regulatory DNA [LSD ([58]17)], we generate sLCRs designed to
   intercept the transcriptional response of epithelial cells to
   SARS-CoV-2 infection using the earliest publicly available signatures
   in cell models ([59]18–[60]20). We characterized one reporter for
   genetic tracing of cellular response to SARS-CoV-2, which represents
   cell states including those described in upper airway tracts of severe
   COVID-19 patients ([61]7). This sLCR responds to cellular stimulation
   by the IFN pathway and of the PRRs. Hence, it is serving as a platform
   for unbiased drug discovery of synergistic modulators of epithelial
   immune responses from viral and sterile triggers and to study their
   mechanism of action.

RESULTS

Designing sLCRs for SARS-CoV-2 productive infection in epithelial cells

   To design SARS-CoV-2 reporters that would model infection response in
   epithelial cells, we first defined the input for LSD, namely,
   SARS-CoV-2 signature and TF genes ([62]Fig. 1A). We overlaid bulk and
   single-cell gene expression data from the earliest available reports in
   lung epithelial cells overexpressing the ACE2 SARS-CoV-2 entry receptor
   [A549 + ACE2 ([63]18)] and endogenously expressing ACE2 and the
   preferential SARS-CoV-2 protease TMPSSR2 [Calu-3 ([64]18)], as well as
   from primary gastrointestinal organoids ([65]19). In response to
   SARS-CoV-2 infection, each cell type expressed private and shared genes
   between 12 and 60 hours post-infection (hpi), and we selected signature
   and TF genes that were up-regulated in at least two of the five
   evaluated datasets. Pathway enrichment analysis revealed that both of
   the selected sets are enriched in IFN and acute inflammation response
   genes ([66]Fig. 1, B and C). As the selected TFs have the highest
   impact of CRE determination and ranking by LSD, we decided to
   systematically design SARS-CoV-2 genetic tracing reporters using a
   combinatorial selection of all TFs and signature gene lists. This
   resulted in 64 potential reporters (table S1).

Fig. 1. Generation of COVGT sLCRs from defined inputs using the LSD method.

   [67]Fig. 1.
   [68]Open in a new tab

   (A) Schematic representation of the SARS-CoV-2 sLCR generation. Bulk
   and single-cell expression profiles from the indicated studies were
   combined to infer an epithelial response signature to SARS-CoV-2 and
   define LSD input. (B and C) Signature selection to generate the
   SARS-CoV-2 sLCR. Left, heatmaps of ssGSEA enrichment in the indicated
   dataset for signature genes (B) and TFs (C). Clustering used Euclidean
   distance, complete linkage. Green denotes the genes meeting the
   threshold of activation in >2 different common genes among different
   datasets. Right, molecular hallmarks enrichment analyses.

   To select one of the above-designed reporters that would specifically
   mark productively infected single cells, we next made use of our
   recently developed phenotypic ranking approach [Materials and Methods;
   ([69]17)]. First, we aimed at defining genes that would be specific to
   SARS-CoV-2 infection response by comparative dimension reduction
   clustering of SARS-CoV and SARS-CoV-2 single cells from parallel
   single-cell RNA sequencing (scRNA-seq) experiments in one cell type. In
   Wyler et al. ([70]20), expression levels of antiviral genes in Calu-3
   were twice as high in SARS-CoV-2– compared to SARS-CoV–infected cells
   at the same time point and in the presence of similar amounts of viral
   RNAs ([71]21). Consistently, we found two cell clusters (C11 and C13)
   that appear specifically for the response to SARS-CoV-2 based on viral
   reads accumulation, with C11 having particularly high reads
   ([72]Fig. 2A). Next, we used our TF binding site (TFBS)
   enrichment–based phenotypic potential analysis to rank all our
   reporters ([73]17) and found that four sLCRs bear a homogeneous scoring
   higher than other specific or background reporters when both C11 and
   C13 were assessed (fig. S1). Our synthetic reporters COVGT1 to COVGT4
   markedly outperformed the phenotypic potential of endogenous IFN
   promoters used in broadly available reporters (e.g., Addgene, #102597,
   #17596, #30536, and #17598), suggesting that synthetic cis-regulatory
   elements have higher on-target potential than endogenous promoters of
   the human IFN-γ or IFN-β genes, which are expected to respond to
   SARS-CoV-2 infection (fig. S1).

Fig. 2. COVGT5 represents a productive SARS-CoV-2 infection and innate immune
response in silico.

   [74]Fig. 2.
   [75]Open in a new tab

   (A) Uniform Manifold Approximation and Projection (UMAP) representation
   of the Wyler et al. ([76]20) SARS-CoV/CoV-2 dataset. Colors highlight
   cell clusters (left), COVGT1 + COVGT4 signature scores (middle), and
   SARS-CoV-2 viral reads (right). (B) Violin plot ranking COVGT1 + COVGT4
   against SARS-CoV-2 patient samples. Signature and TF gene scores from
   COVGT1 + COVGT4 were calculated on annotated cell clusters from Chua
   et al. ([77]7). Enrichment scores (y axis) of distinct cell clusters (x
   axis) within epithelial (left) and immune (right) cell populations are
   shown. (C) In silico comparison of individual and combined COVGT sLCRs.
   Six combinations of sLCRs (black) are plotted against individual COVGT
   sLCRs (pink) and off-target reporters (yellow). Qscore (x axis) and
   TBFS coverage (y axis) predict specificity on the target phenotype. The
   dot size indicates the number of TFs captured from the input list. (D)
   Vector composition of COVGT5-mCherry. Representation of COVGT5 design
   from a combined COVGT4-COVGT1 sLCR driving mCherry expression.

   To advance reporters that may be representative of primary cell
   responses in patients, we ranked all the reporters against the
   nasopharyngeal and bronchial samples from 19 clinically
   well-characterized SARS-CoV-2 patients with moderate or critical
   COVID-19 and 5 healthy controls ([78]7). The combined signature
   underlying the design of two distinct sLCRs (COVGT1 and COVGT4)
   identified two primary cell populations: myeloid cells, notably
   inflammatory macrophages, and one subpopulation of ciliated epithelial
   cells with a distinctively strong IFN-γ response signature
   ([79]Fig. 2B; fig. S2, A and B; and table S1). These SARS-CoV-2
   response signature genes were retrieved in three independent cellular
   models for SARS-CoV-2 infection and are consistent with the pathway
   analysis of the signature and TF genes ([80]Fig. 1, B and C),
   suggesting that COVGT1 and COVGT4, cumulatively, could well represent
   primary host cells in which viral infection triggered an innate immune
   response. To aim for the highest activity potential, we combined COVGT1
   and COVGT4 that combinatorically marked a population of epithelial
   cells activated by SARS-CoV-2 in patients ([81]Fig. 2C) into one
   reporter, hereafter referred to as COVGT5 ([82]Fig. 2D).

COVGT5 responds to SARS-CoV-2 and triggers of epithelial innate immunity

   To validate the designed reporters, we genetically engineered Calu-3,
   A549, and 293T cells using transposon- or lentiviral-based vectors.
   These represent a limited but diverse set of lung cancer and
   immortalized kidney cell lines. We next tested the individual sLCR
   expression by fluorescence-activated cell sorting (FACS) and
   immunofluorescence and found that COVGT5 readily drove the fluorescent
   protein expression in response to a mix of interferons (IFNs) of all
   classes (hereon referred to IFN mix; [83]Fig. 3, A and B). This
   response was more pronounced in immortalized kidney cells compared to
   lung cancer cell lines. Calu-3 displayed the weakest response to IFNs
   compared to the other cell lines ([84]Fig. 3A and fig. S3, A to C). All
   tested sLCRs had low basal expression typically ([85]Fig. 3, A to C;
   fig. S3, A to C; and Materials and Methods) and were found to be
   inducible (fig. S3A), which is compatible with screening purposes.

Fig. 3. COVGT5 responds to SARS-CoV-2 and innate immunity activators.

   [86]Fig. 3.
   [87]Open in a new tab

   (A) Longitudinal measurement of 293T, A549, and Calu-3 COVGT5 reporter
   cells upon IFN mix treatment. Time after treatment is indicated on the
   x axis; each dot represents a single measurement. Mean mCherry
   intensity was calculated from technical triplicates (n = 3) measured by
   Operetta confocal imaging system, and the confidence interval was
   plotted. (B) Representative images of COVGT5-driven mCherry expression
   (magenta) in 293T. Images were taken 48 hours after IFN mix treatment.
   Scale bar, 200 μm. (C) Bar plot quantification of COVGT5-driven mCherry
   up-regulation. COVGT5 induction at 48 hours following IFN mix, 3p-hpRNA
   (0.1 and 0.5 μg/ml), and pI:C (1 and 5 μg/ml) treatment in 293T and
   A549 reporter cells was assessed by normalizing against untreated
   reporter cells. P values denote significance by one-way analysis of
   variance (ANOVA). (D) Fluorescence microscopy images of SCV2 infection.
   N-protein (red) was used as a proxy for productive SCV2 infection with
   Omicron, Delta, and mNG-engineered Wuhan strain in Calu-3 and 293T-ACE2
   reporter cells. Nuclei are counterstained with
   4′,6-diamidino-2-phenylindole (DAPI) (blue). (E) Longitudinal
   fluorescence microscopy images of COVGT5-driven mCherry expression
   (magenta). icSCV2-mNG [6600 plaque-forming units (pfu)/ml] (green) was
   used to infect 293T-COVGT to visualize productive infection. (F and G)
   Bar plot quantification of COVGT5-driven mCherry up-regulation in
   response to icSCV2-mNG (6600 pfu/ml) (F) and SCV2-Omicron (88 pfu/ml)
   (G). Measurement by Incucyte was taken at 120 hours post-infection
   (hpi). The relative mean fluorescent area was determined from technical
   replicates (n = 4). P values denote significance by one-way ANOVA.

   sLCRs may be used to discover external signaling leading to phenotypic
   state transitions ([88]15, [89]16). Thus, we next tested specific cues
   that may potentially drive the response of COVGT5 and be relevant for
   viral infection biology. COVGT5-modified 293T and A549 cell lines
   readily responded to individual stimulation by type I IFNs, IFN-α-2a,
   IFN-α-2b, and IFN-β, as well as synthetic double-stranded RNA
   transfection but not to IFN-γ, interleukin-28 (IL-28) (IFN-λ), tumor
   necrosis factor–α (TNF-α), or lipopolysaccharide (LPS) ([90]Fig. 3C and
   fig. S3D). The marked response to short, triphosphorylated stem-loop
   hairpin RNAs (3p-hpRNAs) as compared to the less specific inducer of
   PRRs, polyinosinic/polycytidylic acid (pI:C), indicated that
   RIG-I–mediated innate intracellular double-stranded RNA sensing
   ([91]22) is a key trigger of COVGT5 expression. In agreement with
   predictions made in silico ([92]Fig. 2C), head-to-head testing
   experimentally supported that COVGT5 was moderately but significantly
   more inducible than COVGT1 and COVGT4 in response to both IFN mix and
   3p-hpRNA in 293T (fig. S3E) and responds to a wider variety of triggers
   than IFN-β1 ([93]23) and IFN-γ ([94]24) promoters cloned within the
   same backbone, in both 293T and A549 cells (fig. S3F). Hence, we
   adopted COVGT5 for subsequent experiments.

   Next, we tested COVGT5 induction in response to SARS-CoV-2 infection.
   To this end, we engineered A549-COVGT5 and 293T-COVGT5 lines to
   overexpress human ACE2, the human receptor for SARS-CoV-2 entry.
   Extracellular ACE2 enzymatic activity confirmed the successful
   engineering (fig. S3G) and made both cell lines permissive for
   infection with all the strains tested (Wuhan, Delta, and Omicron;
   [95]Fig. 3D). Furthermore, introduction of ACE2 did not affect COVGT5
   reporter inducibility in 293T and A549 exposed to the IFN mix (fig.
   S3H). In a longitudinal live-cell imaging setting using the wild-type
   Wuhan SARS-CoV-2 strain engineered to express mNeonGreen [icSCV2-mNG
   ([96]25)], A549-ACE2-COVGT5 and 293T-ACE2-COVGT5 lines were permissive
   to infection and a measurable viral accumulation/replication starting
   from 72 to 120 hpi ([97]Fig. 3E). COVGT5 response appeared to be
   proportional to the extent of viral replication in our cell models. A
   high dose of icSCV2 caused a marked response in 293T-ACE2-COVGT5, which
   was comparable to IFN stimulation ([98]Fig. 3F). Consistently, in
   response to low-dose Omicron infection, we observed a moderate but
   measurable activation of the COVGT5 sLCR, which was delayed and reduced
   when compared to their response to IFNs ([99]Fig. 3G and fig. S3I).

   In our experiments, Calu-3 cells, despite endogenously expressing ACE2
   and the TMPSSR2 protease, albeit at low levels, and being initially
   adopted by the field as a model of choice for in vitro infection
   biology, proved to be less responsive to IFN signaling ([100]Fig. 3A
   and fig. S3, B and C). In addition, they appeared to be poorly
   permissive for infection with all the strains tested, which was
   confirmed by SARS-CoV-2 N-protein immunofluorescence ([101]Fig. 3D).
   Overall, our experiments show that COVGT5 responds to type I IFNs,
   double-stranded RNA, and SARS-CoV-2 viral replication, which are all
   triggers of epithelial innate immunity.

A platform for discovery of pharmacological modulators of cellular innate
immunity and response to SARS-CoV-2 infection

   To exploit the specific responses of COVGT5 to triggers of epithelial
   innate immunity, we next subjected our epithelial cell models to the
   validated signaling cues and systematically investigated
   pharmacological modulators of such response. To this end, we assembled
   a custom library of small molecules against some of the key targets and
   pathways, including targeted agents approved by the United States Food
   and Drug Administration (table S2). To screen for modulators of
   epithelial cell response to COVGT5 activation that may be used in
   clinics, we used a treatment scheme including three doses in the
   low-nanomolar to low-micromolar range (i.e., 0.03, 0.3, and 3 μM;
   [102]Fig. 4, A and B, and fig. S4A). In a 48-hour live imaging setting,
   we were able to rank selected drugs for their ability to enhance or
   block COVGT5 activation in response to both IFN mix and 3p-hpRNA and
   eliminate those drugs that would affect cell viability at low
   doses/short time points. The top ranked small molecules that blocked
   COVGT5 activation in both lines and in response to both triggers were
   well established and clinically relevant JAK inhibitors (JAKis)
   Baricitinib and Tofacitinib ([103]Fig. 4, B and C, and fig. S4B). The
   mechanism of action of both drugs is the inhibition of autocrine and
   paracrine cytokine signaling, with specificity for JAK1-JAK2
   (Baricitinib) and JAK3-JAK2 (Tofacitinib), in agreement with the
   JAK-mediated STAT1 activation caused by type I IFN stimulation of
   IFNAR. Since COVGT5 activation was efficiently blocked by both JAKis in
   both cell lines in response to the JAK1-/STAT1–independent trigger
   3p-hpRNA, the screen suggests that COVGT5 reports on signaling of both
   primary and secondary innate immune transcriptional responses in
   epithelial cells such as the activation of the antiviral and autonomous
   inflammatory genes.

Fig. 4. A platform for discovery of pharmacological modulators of cellular
innate immunity and response to SARS-CoV-2 infection.

   [104]Fig. 4.
   [105]Open in a new tab

   (A) Schematic representation of the platform. One hundred twelve
   pharmacological agents were tested on 293T, A549, and Calu-3 COVGT5
   reporter cells using IFN mix and 3p-hpRNA (0.1 μg/ml) as inducers and
   high-content image acquisition and quantification as readout. (B)
   Circular heatmaps representing screening results of 293T (left) and
   A549 (right) COVGT5 reporter cells. Drugs were tested in concentrations
   of 30, 300, and 3000 nM at 24 and 48 hours after induction in technical
   replicates (n = 4). Z scores calculated for cell viability (inner
   circle) and mCherry intensity (outer circle) are plotted, and drugs are
   ordered based on hierarchical clustering for both parameters. (C)
   Spider plots highlighting COVGT5 modulators. Amplifiers (Topotecan and
   Etoposide) and inhibitors (Tofacitinib) were identified from screening
   data in 293T (teal) and A549 (purple). Each data point is represented
   as a corner of the plot with the indicated scale ranges (note for color
   coding: red, activation; blue, inhibition). (D) Representative images
   showing modulatory drug effect. Topotecan (300 nM) and Tofacitinib (300
   nM) were used on top of the IFN mix in 293T reporter cells. Scale bar,
   200 μm. (E) Longitudinal measurement (left) of COVGT5 modulation in
   virus infection setting. 293T-COVGT5 were infected with SCV2-Omicron
   (88 pfu/ml) and treated with 10 μM Baricitinib, 10 μM Tofacitinib, 10
   μM Etoposide, 200 nM BX-795, or 2 μM ruxolitinib. Polynomial regression
   model (right) of longitudinal measurement pronouncing drug effects on
   COVGT5 induction. Measurement and quantification were performed using
   Incucyte factoring mean fluorescence intensity and fluorescent area.
   (F) Independent bar plot quantifications of 293T-COVGT5 induction by
   SCV2 infection. SCV2-Omicron (6600 pfu/ml) or icSCV2-mNG were used with
   10 μM Baricitinib, 10 μM Tofacitinib, and 10 μM Etoposide. Measurement
   and quantification were performed using Incucyte at 120 hpi. The
   relative mean fluorescent area was determined from technical replicates
   (n = 4). P values denote significance by one-way ANOVA.

   Among the top amplifiers of COVGT5 activation in response to the
   aforementioned triggers and across multiple cell lines and doses, we
   scored several modulators of genome integrity, with the notable case of
   DNA damage inducers (e.g., TOP1 inhibitor, Topotecan, and TOP2
   inhibitor, Etoposide; [106]Fig. 4, B and C, and fig. S4B). The main
   hits of the screens were separately validated in the reporter cell
   lines, and specific COVGT5 expression was confirmed ([107]Fig. 4D).

Etoposide and Baricitinib/Tofacitinib antagonistically modulate COVGT5
response to productive SARS-CoV-2 infection

   To validate the response observed in the screen using SARS-CoV-2, we
   next infected 293T-ACE2-COVGT5 with the Omicron strain and
   simultaneously treated these cells with inhibitors of either TOP2
   (Etoposide), JAK (Baricitinib, Tofacitinib and ruxolitinib), or
   TBK-1/IKK-ε/PDK-1 (BX-795). The latter blocks noncanonical IκB kinase
   and IRF-3 activation during innate immune sensing. During the
   longitudinal monitoring of cell responses, 293T-ACE2-COVGT5 displayed a
   measurable activation of the COVGT5 sLCR in response to Omicron
   infection, with a moderate upward trajectory at the predefined end
   point of 156 hpi. This activation mode appeared to be specific for the
   viral infection setting since the cellular response to the IFN
   signaling reached the plateau at 48 hours after cytokine stimulation
   and lasted until 156 hpi (fig. S4C). In line with the results of our
   screen, COVGT5 expression was abrogated by all the independent JAKis
   tested (Ruxolitinib, Baricitinib, and Tofacitinib), as well as by the
   TBK-1/IKK-ε/PDK-1 inhibitor BX-795 ([108]Fig. 4E). Instead, Etoposide
   imparted a moderate amplification of the COVIGT#5 signal when
   administered in parallel to Omicron ([109]Fig. 4E), and a nonlinear
   regression model supported the hypothesis that Etoposide and JAKi
   antagonistically modulate COVGT5 response to productive SARS-CoV-2
   infection ([110]Fig. 4E). Similar antagonism was observed for the
   SCV2-mNG strain ([111]Fig. 4F and fig. S4C). Of note, in 293T cells
   exposed to icSCV2-mNG, similar viral replication levels in individual
   treatment arms excluded that JAK or topoisomerase inhibitors regulated
   COVGT5 through viral replication, suggesting that they mainly affect
   cellular responses through their mechanism of action (fig. S4D). Thus,
   our data support the use of an sLCR to detect responses to viral
   infection as a drug discovery platform to identify pharmacological
   modulators of innate immune responses.

Synergistic and antagonistic pharmacological modulation of COVGT5 underscored
drug mechanism of action and convergence on endogenous transcription

   The targets of individual JAKis are biochemically well established, and
   the consequences of signal transduction observed in our experiments
   aligned with expectations. Cell exposure to chemotherapeutics such as
   Topotecan and Etoposide leads to far less predictable targets and
   consequences. We therefore tested whether our reporter could provide
   evidence supporting the mechanism of action of the drugs controlling
   its regulation.

   DNA damage induces ISGs and the IFN-α and IFN-λ genes via NF-κB
   activation ([112]13), which is consistent with both the predicted
   outcome of TOP1 and TOP2 inhibition by Topotecan and Etoposide,
   respectively. However, small molecules may also have unrelated,
   off-target effects. To address whether or not DNA damage is the actual
   trigger of the SARS-CoV-2 infection–like response observed, we set out
   to conduct both selective pairwise synergy screens and a comprehensive
   transcriptomic analysis ([113]Fig. 5A). In 293T-COVGT5 and A549-COVGT5
   exposed to IFN mix or 3p-hpRNA, Bliss synergy scoring of the dose
   escalation for Topotecan and Etoposide determined that the combined
   treatment synergistically amplified COVGT5 expression, in line with
   their predicted induction of DNA damage ([114]Fig. 5B and fig. S5A).
   The extent of synergism was cell line specific and reflected the extent
   of response to IFN mix ([115]Fig. 5B and fig. S5A). Conversely, JAKi
   combination synergistically restricted COVGT5 activation ([116]Fig. 5B
   and fig. S5A), indicating that synergism reveals mechanistical
   convergence. Thus, we next substituted either Topotecan or Etoposide
   with ionizing radiation (IR), which is a direct method to damage DNA.
   IR amplified activation of COVGT5 in dose-dependent manner on top of
   3p-hpRNA transfection, and in both lung and kidney cell lines, starting
   from a dose of 1 Gy (fig. S5B). Bliss synergy scoring of the dose
   escalation of each combination indicated that IR activation is
   epistatic to Topotecan or Etoposide ([117]Fig. 5C), supporting that
   cooperative DNA damage induction occurred. In contrast, Bardoxolone, an
   NRF2 activator, antagonistically regulated COVGT5 when combined with
   Topotecan (fig. S3C), indicating that drugs with distinct mechanisms of
   action do not synergistically amplify cellular responses to 3p-hpRNA.

Fig. 5. Synergistic and antagonistic pharmacological modulation of COVGT5
underscored the drug mechanism of action.

   [118]Fig. 5.
   [119]Open in a new tab

   (A) Schematic depiction of approach to define mechanistic basis of
   COVGT5 regulation. (B) Dose escalation of Topotecan against Etoposide
   (top) and Baricitinib against Tofacitinib (bottom). 293T-COVGT5
   received IFN mix induction followed by combinatorial drug treatment.
   Mean mCherry intensity was calculated from technical triplicates (n =
   3) and normalized to non–drug-treated cells. Quantification was
   performed by high-throughput FACS measuring mean mCherry fluorescence
   intensity. Bliss synergy scores inform on synergism between Topotecan +
   Etoposide and Baricitinib + Tofacitinib in modulating COVGT5 response.
   (C) Mean Bliss synergy score calculated from all single drug
   combination data points. Topotecan + IR and Etoposide + IR in addition
   to drug combinations from (A) in 293T and A549 reporter cells were
   performed in technical triplicates (n = 3). Quantification was
   performed by high-throughput FACS. P values denote significance by
   one-way ANOVA. (D) GSEA plot of IFN mix- and 3p-hpRNA-treated 293T and
   A549 RNA-seq data (n = 3 biological replica) using the
   COVGT5-coregulated gene signature. COVGT5 inducers are the denominator
   of the GSEA against the effect of Topotecan + Etoposide amplification
   (top) or Baricitinib + Tofacitinib (bottom). (E) Bar plot showing the
   expression of the indicated genes in the indicated cell lines by
   RT-qPCR. Note the conserved response of Stat1 expression in response to
   IFN mix and IR in mouse and hamster cells. P values denote significance
   by one-way ANOVA. (F) Schematic representation of proposed mechanisms
   of action for COVGT5 modulation. Pathways for induction and
   amplification by Topotecan, Etoposide, and IR (left) and COVGT5
   inhibition by Baricitinib and Tofacitinib (right) (strength of
   inhibition arrows indicate specificity) are proposed. ns, not
   significant.

   Type I IFN responses involve activation of TFs such as IRF3 and IRF7
   ([120]26) and negative feedback by proteins including SOCS1 and USP18
   ([121]27). However, in response to viral infection, only a few
   endogenous genes are strictly dependent on individual factors, as
   exemplified by experiments in IRF3 knockout cells ([122]28).
   Consistently, knockdown of IRF3, IRF7, and SOCS1 in 293T and A549 cells
   did not markedly affect COVGT5 activation in response to IFN mix (fig.
   S6, A and B), despite IRF3 and IRF7 binding sites being included among
   the input.

   To determine whether COVGT5 expression reflects the activation of a
   broad or cell type– and trigger-specific gene expression program, we
   next systematically dissected COVGT5-coregulated genes by RNA-seq in
   FACS-purified cells ([123]Fig. 5A and fig. S6C). Principal components
   analysis (PCA) indicated that COVGT5 activation and its modulation by
   antagonistic JAK and TOP1/2 inhibitors dominate gene expression
   variation together with the expected cell type–specific programs (fig.
   S6D). We then defined 454 genes coregulated with COVGT5 in a cell type–
   and trigger-independent manner, and 104 genes were antiregulated (fig.
   S6, E to G, and table S3). COVGT5-coregulated genes were generally
   connected to response to IFNs and viruses, as well as antigen
   presentation (fig. S6, E and F). The substantial similarity between
   COVGT5-coregulated genes in cell lines and IFN signatures obtained from
   primary models (fig. S6F) ([124]18) supports the broad validity of our
   findings. Consistent with functional inference experiments, IRF3, IRF7,
   SOCS1, or USP18 was either lowly expressed or not regulated in our
   models (table S3). COVGT5 signature genes were significantly amplified
   by low-dose TOP1/2 and depressed by JAKis, respectively ([125]Fig. 5D).
   Selected genes from the COVGT5 signature such as IFIT3, IFI44, OAS1,
   and STAT1 were confirmed by reverse transcription quantitative
   polymerase chain reaction (RT-qPCR) in A549 and 293T cells (fig. S7).
   Therefore, unbiased gene expression profiling corroborated our drug
   screen findings and connected COVGT5 to an endogenous gene expression
   program.

   Stat1 induction by human IFN mix was observed in both mouse NIH3T3 and
   Chinese hamster ovary (CHO)–K1 cells and amplified by an IR dose of 3
   Gy ([126]Fig. 5E), which is on the higher end of the dosage compatible
   with clinically applicable low-dose radiotherapy (LDRT) ([127]29).
   Hence, Stat1 conserved induction could be used as an endogenous
   molecular beacon in preclinical animal models for modulation of type I
   IFN responses and viral infection.

   Together, our data suggest that DNA-damaging agents and JAKi
   antagonistically modulate epithelial innate immunity in response to
   SARS-CoV-2 infection ([128]Fig. 5F). These leads were enabled by
   synthetic genetic tracing for SARS-CoV-2 viral infection in epithelial
   cells applied in a drug discovery setting and feature its application
   to infection biology more broadly.

DISCUSSION

   COVID-19 pandemic has seen no single drug or combination proven to
   effectively halt disease progression, and the effect of the massive
   worldwide vaccination campaign is waning. Additional threats may come
   from near-future surges of this pandemic by SARS-CoV-2 subvariants, and
   other pandemic-scale infectious diseases, which are recurrent in
   history. To grapple with these pressing issues, we report the COVGT
   sLCRs and the validation of the approach to design such reporters.
   Together, these form a useful resource for investigating the cellular
   response to SARS-CoV-2 entry and replication as well as an invaluable
   tool for high-throughput screening of therapeutics. Combining reporters
   that measure viral entry and replication with those that measure
   cellular transcriptional responses is timely ([129]30). The flexibility
   to use such systems in any chosen model, potentially including
   preclinical models in vivo, is a unique feature of our system. In
   addition, the adaptability of our system to establish in vitro
   infection models allows the rapid testing of wild-type and evolving
   pathogens stains.

   Whereas heterogeneous responses to viral infection are likely rooted in
   biology ([130]30), future experiments with this and other reporters
   will likely benefit from using cellular models that are physiologically
   more accurate than cell lines. Through our experiments, 293T cells were
   found to be a more responsive model for triggers of innate immunity and
   viral infection, while A549 displayed intermediate IFN and RIG-I
   responses, as well as permissiveness of SARS-CoV-2 infection. On the
   other hand, we observed limited response in Calu-3, a lung cancer cell
   line adopted at the start of the pandemic, and later dismissed as an
   effective model. Hence, synthetic genetic tracing may be used to
   rapidly screen for the most optimal cellular models for infection
   biology studies.

   The COVGT5 reporter characterized here was independently and strongly
   activated by type I IFNs and RIG-I agonists in cells and outperformed
   publicly available reporters both in silico and in vitro, and its
   expression was connected to endogenous gene expression. RIG-I, a major
   receptor for viral mRNA sensing, was validated head-to-head in primary
   cells and cell lines ([131]31), supporting the validity of our
   findings. COVGT5 activation in our epithelial cell models may be
   modulated by bioactive agents found through in vitro image-based
   screens. An example is our discovery of DNA damage inducers and JAKis,
   which act as rheostats of epithelial immunity. Of note, the
   identification of a cell type– and trigger-independent signature of
   endogenous innate immune responses suggests that COVGT5 activity may be
   conserved across various tissues and cell types. On the basis of our
   past experience with other synthetic reporters ([132]15, [133]16),
   identifying and elucidating critical mechanistic nodes of
   context-specific COVGT5 activation will require phenotypic functional
   screens at a genome scale or focused on pathways, such as the signature
   genes we report.

   IFN response plays a critical role in the first lines of defense
   against viral infections. It involves the transcription of IFN genes
   first and later of ISGs, which in turn lead to an antiviral state in
   both infected and neighboring cells. Viral infections are detected
   through PRRs, including RIG-I–like receptors. Notably, SARS-CoV-2
   evolution through variants of concern involved increased IFN response
   evasion ([134]32), and a synthetic agonist stimulation of RIG-I
   protects mice from acute and chronic SARS-CoV-2 infection, even in the
   absence of the adaptive immune system ([135]33). This underscores the
   importance of epithelial innate immune responses to restrain SARS-CoV-2
   infection and supports the potential of pharmacological modulators of
   such response as a therapeutic strategy to manage COVID-19 and viral
   infections more broadly. Our COVGT5 sLCR designed on the
   transcriptional response of lung epithelial and enterocyte organoids to
   SARS-CoV-2 infection could be explained by activation of IFN- and
   RIG-I–like signaling. This provides an additional asset for our study
   as it offers a cellular reporter for viral infection biology and a
   molecular tool for cellular and molecular dissection of these pathways.

   Given the antiviral immunity’s pivotal signaling through JAK/STAT
   pathway, JAKis were already clinically tested as therapeutics for
   COVID-19. Retrospectively, JAKi appeared potentially helpful to manage
   disease severity ([136]3). Likewise, Topotecan, one of the other drugs
   uncovered by our COVGT5 reporter as an amplifier of innate immunity,
   was recently found as treatment to blunt the hyperinflammatory response
   to anti–SARS-CoV-2 treatment with successful validation in preclinical
   models ([137]34). More generally, topoisomerase 1 and 2 inhibitors are
   potent modulators of innate immune activation ([138]35, [139]36). In
   mice, Topotecan and Etoposide ameliorate responses to colitis
   ([140]37), and Etoposide partly protects from sepsis ([141]38), thereby
   featuring an anti-inflammatory function at systemic level. An ongoing
   clinical trial is testing Topotecan in COVID-19 patients to assess the
   safety of a single dose and its potential to blunt inflammation
   systemically ([142]NCT05083000). Our experiments are focused on
   epithelial cell–intrinsic responses to triggers of innate immunity and
   may be helpful to interpret the results of this trial as they suggest
   that DNA damage modulation may also be effective in priming an
   antiviral state in vivo or amplifying it. This may occur in both
   infected and neighboring cells as a function of cellular state or
   productiveness of the infection. However, preclinical testing in vivo
   will be important to investigate the impact of cell type–specific and
   systemic responses.

   Combination therapies were frequently tested in COVID-19 trials,
   particularly using antiviral drugs as anchor, since those were deemed
   insufficient if administered at disease progression ([143]5).
   Remdesivir, the first antiviral to show mild effectiveness against
   COVID-19, appeared to cooperate with Baricitinib ([144]5). For JAKis
   and Dexamethasone, a glucocorticoid receptor agonist that did not
   consistently modulate our reporter across the screens and validations,
   trials suggested that drug administration should occur in a particular
   window of time to promote beneficial effects, or else these drugs could
   be detrimental ([145]3, [146]4). In our experiments, all JAKis
   abrogated cellular response to viral replication, providing support for
   their use for such purposes. We arrived at our conclusion through a
   rapid screening of a curated set of drugs. This approach is
   significantly more scalable than the use of clinical trials and
   artificial intelligence, which reached a similar conclusion ([147]39).
   Moreover, our platform is uniquely placed to discover drug combinations
   synergistically acting toward cellular response to viral infection in
   preclinical models, which may later be tested during disease
   progression. Alternatively, rather than abrogating epithelial immunity,
   combinations of drugs that promote an antiviral state and cooperate
   with antiviral drugs could be tested to potentiate disease control. As
   supplementing IFN-β-1a in combination with Remdesivir did not improve
   outcomes and was associated with adverse effects in a
   placebo-controlled phase 3 trial ([148]40), innate immunity
   amplification will require the identification of safe combinations. Our
   screen also initially revealed Rapamycin as a potential epithelial
   immunity amplifier drug. However, this drug was not validated in later
   experiments. Whether targeted IR may reduce the impact of viral
   shedding in selected organs is an intriguing hypothesis that stems from
   this study. Low-dose radiation therapy (LDRT) has been studied as a
   possible treatment option during the COVID-19 pandemic, yet its
   effectiveness is still debated ([149]41, [150]42). Future combinations
   should focus on approved drugs with higher safety profiles or involve
   substantial preclinical testing. DNA damage-inducing agents have the
   potential to worsen cellular responses to an infection ([151]43), and
   heightened IFN responses have been linked to severe COVID-related
   disorders such as vasculitis, “COVID toes,” and Kawasaki-like disease.
   Nevertheless, our finding that STAT1 induction by IR within epithelial
   cells appears to be conserved in mouse and hamster cells, if used at
   doses slightly stronger than conventional LDRT, offers a possibility of
   testing LDRT for its potential to be an effective treatment in widely
   used preclinical models and compare outcomes across species.

   The onset of the COVID-19 pandemic resulted in substantial global
   morbidity and mortality, and led to worldwide economic collapse due to
   the lack of early preparedness. Our approach took 4 weeks from the
   discovery of the first transcriptional signatures to validate the
   reporter and offers significant potential for optimization, thus making
   synthetic genetic tracing more available for infection biology in the
   future.

MATERIALS AND METHODS

Datasets

   The data to generate the signature genes and TFs listed in [152]Fig. 1
   were downloaded from Gene Expression Omnibus, accession [153]GSE148729
   and [154]GSE147507, or the Supplementary Materials (see below).

LSD reporter design

   COVGT reporters were designed using the LSD method, as described
   ([155]17). Briefly, the LSD algorithm takes a list of PWMs, a list of
   marker genes of a target phenotype, and the reference genome of the
   organism of interest, and it generates a list of naturally occurring,
   putative cis-regulatory elements used to assemble the synthetic
   reporter. The algorithm can be divided into three steps. In step I, LSD
   generates a pool of potential CRE with a fixed length within
   user-defined regulatory landscapes [default is a 150–base pair (bp)
   window sliding with a 50-bp step]. In step II, LSD assigns TFBSs to the
   CRE pool using FIMO (default --output-pthresh 1e-4 --no-qvalue) and
   creates a matrix of putative CREs × TFBS. In step III, LSD ranks and
   selects the minimal number of CREs representing the complete set of
   TFBS. For that purpose, it uses an algorithm to sort and select the
   best CRE based on the overall TFBS affinity and diversity among input
   TFs showing high affinity for the CRE. Starting from the ranked CREs,
   LSD selects the highest-ranking CRE defined by the sum of the affinity
   score [−log[10](P value)] and TFBS diversity (number of different
   TFBS). Subsequently, it removes the selected CRE and the corresponding
   TFBS from the CRE × TFBS matrix and repeats the selection. This
   continues until either none of the CRE or of the TFBS is left. In the
   ranking, priority is given to CREs proximal to known transcription
   start site (TSS) based on 5′ CAGE data (ENCODE) to increase the chances
   of successful transcriptional firing using the same strategy as above.
   Last, LSD returns an ordered list of the selected CREs, together with a
   representation of the TFBS scores ([156]Fig. 1A). LSD is available at:
   [157]https://gitlab.com/gargiulo_lab/sLCR_selection_framework. All
   COVGT sLCRs are available from Addgene either with destabilized GFP or
   additionally with mCherry as fluorophores
   ([158]https://www.addgene.org/Gaetano_Gargiulo/; Addgene IDs #201447,
   #201448, #201446, #201449, #201451, #201452, #201450, #201453, #201454,
   #201455). The COVGT5 plasmid #201454 pPB-Neo-COVGT5-d2EGFP-mCherry was
   primarily used in this study.

Cis-regulatory potential assessment of the SARS-CoV-2 sLCRs

   To generate a common SARS-CoV-2 signature and define the specific TF
   ([159]Fig. 1A), we integrated different transcriptional profiles
   available at the moment of this publication. We first downloaded
   Blanco-Melo et al. ([160]18) ([161]GSE147507), Wyler et al. ([162]20)
   ([163]GSE148729; [164]www.mdc-berlin.de/singlecell-SARSCoV2), and
   Lamers et al. ([165]19) (differentially expressed tablets). Then, we
   reanalyzed Blanco-Melo et al. ([166]18) and Wyler et al. ([167]20)
   using the same parameters as in the publications. Blanco-Melo et al.
   ([168]18) and Lamers et al. ([169]19) gene signatures and TF were
   selected using log[2]FC > 1 and adjusted P value (P[adj]) < 0.05. Wyler
   et al. ([170]20) gene signatures were identified by using the
   FindMarkers function (Seurat v.3.1, R v.3.6) and selected using P[adj]
   < 0.05 as a threshold. scRNA-seq TF selection was generated by
   selecting exclusive highly expressed (quantile >75%) TF genes.
   Integration of the datasets to construct the common gene signatures was
   done using at least three common genes between experiments, and two
   common TF genes to define the TF input. Further analysis of TF using
   Ingenuity Pathway Analysis (IPA) generated the final set of TFs and
   TFBS used in the different comparisons (fig. S1 and table S1).
   Previously published reporters carrying promoters of the human IFN-γ or
   IFN-β genes were included as benchmark (Addgene, #102597, #17596,
   #30536, and #17598). SARS-CoV-2 sLCR pool was designed using a
   combination of individual and common datasets and the LSD framework.
   The inference of the specific phenotypic potential () of each reporter
   was defined as the linear correlation between the Qscore
   (affinity-score) and the SignScore (phenotypic-score). The
   affinity-score (QScore) was calculated for each reporter as the sum of
   the TFBS-affinity using FIMO (default --output-pthresh 1e-4
   --no-qvalue) given a specific set of TFBS (e.g., using as reference IPA
   selected COVGT#1 TFBS) and normalized by the sLCR sequence length and
   the ratio of observed/expected TFBS [Σ (FIMO-score) × sequences
   length)/(observed/expected TFBS)]. In contrast, the phenotypic score
   (SignScore) was defined as the ssGSEA-enrichment value calculated on
   target expression profiles (ssgsea.norm=FALSE) using the signature
   genes, and normalized by the ratio of observed/expected TFBS
   [ssGSEA-score × (ratio observed/expected TFBS)]. Heatmaps and gene set
   enrichment analysis (GSEA) ([171]Fig. 1, B and C) were generated using
   common genes and selection with pheatmap v.1.0.12 and piano v.2.0.2,
   respectively. Hg19 assembly was used to extract references and generate