High-Throughput Screening Normalized To Biological Response: Application To Antiviral Drug Discovery

Abstract

The process of conducting cell-based phenotypic screens can result in datasets from small libraries or portions of large libraries, making accurate hit picking from multiple datasets important for efficient drug discovery. Here, we describe a screen design and data analysis approach that allows for normalization not only between quadrants and plates but also between screens or batches in a robust, quantitative fashion enabling hit-selection from multiple datasets. We independently screened the Microsource Spectrum and NCI Diversity Set II libraries using a cell-based phenotypic HTS assay that uses interferon stimulated response element (ISRE)-driven luciferase-reporter assay to identify interferon (IFN) signal enhancers. Inclusion of a per-plate, per-quadrant IFN dose-response standard curve enabled conversion of ISRE activity to effective IFN concentrations. We identified 45 hits based on a combined z-score ≥ 2.5 from the two libraries, and 25 of 35 available hits were validated in a compound concentration-response assay when tested using fresh compound. The results provide a basis for further analysis of chemical structure in relation to biological function. Together, the results establish an HTS method that can be extended to screening for any class of compounds that influence a quantifiable biological response for which a standard is available.

Introduction

High-throughput screening (HTS) systems using cell-based phenotypic assays are often more complex and costly than comparable HTS using target-based biochemical assays. As a consequence, cell-based approaches might involve screening several small libraries of compounds or subsets of a large library in batches rather than a single large screen. Alternatively, a small library might be subject to an initial screen to direct the subsequent selection and screening of additional libraries. Further, cell-based assays are generally more susceptible to batch-to-batch or day-to-day variability when compared to target-based assays. Given these challenges for cell-based assay systems, there is a need for methods to accurately hit-pick across multiple screens and thereby enable the inherent advantages of phenotypic screening for first-in-class drug discovery.¹

We encountered these difficulties with cell-based screening in our work to discover new antiviral drugs. We developed an HTS system to identify small molecules that enhance the endogenous interferon (IFN) signaling pathway as a means to inhibit viral replication without the toxicities of exogenous IFN administration (Fig. 1).^2-6 The system uses a cell line engineered to express a luciferase reporter under IFN-stimulated response element (ISRE) regulation, and was validated in a pilot screen of the Johns Hopkins Clinical Compound Library (Fig. 1B).⁵ To identify additional hits for drug development, we proceeded to screen two additional compound libraries, but realized the need to normalize the data not only between plates in an individual screen, but also between screens in order to pick the most active hits from both screens together in a robust, quantitative fashion, which cannot be achieved using conventional analytical approaches for HTS data that are largely based on outlier identification. Plate median and B-score normalization methods do not account for agonist concentration-response behavior or its variance from run-to-run. Similarly, scaling data points to maximum and minimum values from positive and negative controls does not account for the differences in variance, since it assumes that there is no heterogeneity in the non-linear variance of the dose-response curve from run-to-run.

Figure 1.

Scheme for discovery of small molecule IFN signal enhancers (SMISEs) as broad-spectrum antiviral drugs. (A) IFN signal biology. STAT: Signal transducer and activator of signaling; ISG: IFN stimulated gene. (B) Drug discovery strategy based on biology depicted in A. Primary screening relies on promoter activity, hits from which are subjected to validation assays based on ISG expression, and STAT activity. It is inclusive of structurally similar compounds for initiation of SAR. Select compounds are subjected to antiviral experiments in cell lines and primary human tracheal epithelial cell infection model that is predictive of in vivo efficacy. 112×77mm (300 × 300 DPI)

Here, we report a solution to these problems, the development of a quantifiable biological response HTS design and data analysis approach that includes a replicated standard curve for each quadrant on each plate. Screening data is fitted to the concentration-response curve, converting the assay readout (luciferase reporter luminescence) to a standardized biological response (effective IFN-β concentration). This provides a simple, precise normalization and data representation, thereby enabling hit picking across multiple screens and evaluating compound activity in units that are directly relevant to the biological response.

Materials and Methods

Screening Libraries and Chemical Compounds

Aliquots of the Spectrum library (MicroSource Discovery Systems, Gaylordsville, CT) and NCI Diversity Set II library (National Cancer Institute – Developmental Therapeutics Program, NCI-DTP, Frederick, MD) were obtained from the Washington University Chemical Genetics Screening Core. Library compounds were diluted to 5 mM in DMSO in 96 well plates, sealed using a Flexiseal plate heat sealer (K Biosciences, Beverly, MA), and stored at −20 °C. Plates were thawed, equilibrated to room temperature, and centrifuged before use. IFN-β was obtained from PBL Interferon Source, (Piscataway, NJ). Compounds for hit confirmation and further analysis were obtained from Sigma Aldrich (St. Louis, MO), MicroSource Diversity Systems, VWR (Radnor, PA), Specs (Cumberland, MD), Princeton Bimolecular Research (Monmouth Junction, NJ), Toronto Research Chemicals (Toronto, CA), MP Biomedicals (Santa Ana, CA), Santa Cruz Biotechnology (Dallas, TX), or TimTec Drug Discovery (Newark, DE).

ISRE Activity-Luciferase Reporter Assay and Automation

The ISRE activity assay using 2fTGH-ISRE-CBG99 cells that stably express CGB99 green luciferase under control of ISRE promoter activity was described previously.⁵ For the present experiments, cells were plated in 384 well plates and allowed to grow for 13-15h, then treated with compound and/or IFN-β for 7-11 h, the window for maximal assay signal. After simultaneous cell lysis and addition of luciferin using the SteadlyLite plus reagent (Perkin Elmer, Waltham, MA), cells were incubated for 40 min at room temperature. Luminescence was then read using a Synergy 4 plate reader (BioTek, Winooski, VT). Screening was performed in a fully automated screening facility that contains a Caliper Sciclone ALH 3000 workstation (Perkin Elmer) and EL406 washer (BioTek) for liquid handling, an automated Liconic incubator (Thermo Scientific) for storage of plates at 4 °C, an automated Cytomat incubator (Thermo Scientific) for cell culture, a separate hotel for storage of plates at room temperature, a Synergy 4 plate reader, a Caliper Twister II, and a Beckman Sagian Orca robotic arm on a linear rail (Beckman Coulter, Fullerton, CA), all enclosed in a customized BSL2 laminar flow hood.

HTS Protocol

The Spectrum and NCI Diversity Set II libraries each contained 2000 compounds. Each library was plated in 96-well plates (N=25/library), and was screened separately with a 1-month interval between screening runs. Each compound was represented on duplicate plates at four different concentrations (0.24, 1.2, 6, and 30 μM), all in the presence of IFN-β (15 U/ml) (Supplementary Fig. S1A). Control wells with a range of IFN-β concentrations (0-200 U/ml) were included on the two outer columns of each plate as positive controls. The IFN-β concentration in the compound test wells was adjusted to maintain biological response despite batch-to-batch variations in IFN-β activity, and the IFN-β concentrations in the control wells were adjusted so that more data points were collected at the initial inflexion of the IFN-β concentration-response curve rather than higher concentrations that were unlikely to be achieved by any compounds.

Screening was performed as described previously⁵. First, compound dilution plates containing 0.24, 1.2, 6, and 30 μM compound in media with 15 U/ml IFN-β were made and stored at 4 °C. Corresponding DMSO concentrations in final treatment solutions were 0.6, 0.12, 0.024 and 0.0048% in these compound solutions respectively, well below the assay tolerance limit of 1% DMSO⁵. The IFN-β control solutions were pipetted into the first and last columns of these dilution plates. Next, cells were plated onto duplicate assay plates and maintained at 5% CO₂ and 37 °C. Then, assay plates and compound dilution plates were transferred to the Sciclone liquid handler to treat cells with compound or control solutions. The assay plates were incubated at 5% CO₂ and 37 °C for 11 h, after which the luciferase assay was performed by aspiration of media followed by addition of the Steadylite reagent, incubation at room temperature for 40 min, and measurement of luminescence. Taking into account the 11 min rate-limiting step of reading each plate, the screening schedule was designed to provide equivalent time periods for incubation, treatment, and luciferase assay for each assay plate. Total screening time for each compound library was approximately 42 h.

HTS Data Analysis

The screening data from each compound library was analyzed using two methods. For the first method, data from each library were treated individually. The CellHTS2 software package was used for statistical analysis.^7-9 Raw data were normalized using plate median methods, and variance adjusted using B-score normalization.^{7, 10} Next, z-score transformations were applied to center each data set across the entire individual screen. Replicates for a given compound at a given dose (N = 2 for each dose/compound combination) were then mean summarized. A z-score threshold of ≥ 2.5 for any compound concentration was used to identify hits from each library.

For the second method of analysis, the two screening data sets were normalized to combine the data sets before computing z-scores. Raw luminescence values were fit to the IFN-β-concentration-luciferase activity response curve on a per-plate, per-quadrant basis (Supplementary Fig. S1B), converting the data to effective IFN-β concentrations. The data was converted using the four-parameter concentration-response curve as described previously with the log(agonist concentration) versus response, variable slope algorithm in GraphPad Prism 5 software (La Jolla, CA) where Y=Bottom+(Top-Bottom)/(1+10^((LogEC50-X)*HillSlope)).⁵ The automatic outlier analysis algorithm which uses robust regression and outlier (ROUT) removal method¹¹ was used to exclude control outlier wells during reference curve fitting. Thereafter, the two replicates of the effective IFN-β concentrations for each compound concentration were mean summarized for all test compounds. Z-score transformation was then applied to these average effective IFN-β concentration values to center and scale the data across both screens. Compounds were then ranked based on a maximum z-score (for any of the four concentrations tested per compound) in this combined analysis, and hits picked on the criteria of a maximum combined z-score ≥ 2.5.

Hit Confirmation and Potency Estimation

Hits from the primary HTS were confirmed by measurement of ISRE activity over a range of compound concentrations in the presence of IFN-β (15 U/ml). Compounds were tested in quadruplicate in a 12 point, 2-fold dilution series starting at a concentration of either 25, 50, or 125 μM, depending on the DMSO stock concentrations. DMSO concentrations were maintained at 0.5% for all compound dilutions. Data was fit to a log(agonist concentration) versus response curve to calculate compound potency, as defined by the half-maximal effective concentrations (EC₅₀). Structurally similar compounds were included when available.

High-content STAT1/STAT2 Nuclear Translocation Assay

Compound effects on IFN signaling were further validated with an assay of IFN-stimulated nuclear translocation of STAT1 and STAT2. For this assay, 2fTGH cells (7500 cells per well in 96-well black tissue culture treated plates) were treated with compound for 18 h and then with compound plus IFN-β (15 U/ml) for an additional 20 min or 2 h at 37 °C. For assay validation experiments in which treatment time with IFN was varied, saturation concentrations for IFN-β and IFN-γ were used. To assess STAT1 and STAT2 translocation, cells were fixed in 4% paraformaldehyde and then were immunostained for STAT1 using mouse-anti-human STAT1 91/84 (BD Biosciences, San Diego, CA) and STAT2 using rabbit-anti-human STAT2 (Thermo Scientific, Rockford, IL) for 1 h followed by detection with goat-anti-mouse Alexa Fluor 660-conjugated-IgG (Life Technologies, Carlsbad, CA) for STAT1 and donkey-anti-rabbit Alexa Fluor 555-conjugated-IgG (Life Technologies) for STAT2. Cell nuclei were stained with Hoechst33342 (Thermo Scientific) and then were imaged using an ArrayScan VTI High Content Imaging system (Cellomics/Thermo Scientific, Pittsburgh, PA) with a 10× objective. Images were analyzed using vendor software as follows. To measure STAT fluorescence intensity in the nucleus, a circle was defined 1 pixel inside the nuclear-cytoplasmic boundary. Similarly, STAT fluorescence in the cytoplasm was measured in a ring of 1 pixel thickness, the inner edge of which was defined by the nuclear-cytoplasmic boundary. STAT nuclear translocation was then calculated for each cell as the difference in pixel-averaged fluorescence intensity between the circle defined in the nucleus and the ring defined in the cytoplasm in fluorescence units. Values are reported as the average STAT nuclear retention values for all cells imaged per well for 3-4 wells (15 images and 700 cells per well). Values are negative for untreated cells, since average STAT2 fluorescence intensity is higher in the cytoplasm than the nucleus at baseline.

IFN-Stimulated Gene (ISG) Expression Assays

Compound effects on IFN signaling were also assessed with assays of ISG expression. For these assays, 2fTGH cells were treated with compound, with or without IFN-β for 12 h, and then lysed using the Cells-to-cDNA lysis buffer (Life Technologies). After DNAse treatment, the lysates were used to generate cDNA using the High Capacity Reverse Transcription Kit (Life Technologies). The qPCR assays for 2′-5′-oligoadenylate synthetase 1 (OAS1), myxovirus resistance 1 (MX1), and IFN-induced protein with tetratricopeptide repeats 1 (IFIT1) mRNA expression were performed in accordance with MIQE guidelines.¹² Expression of ornithine decarboxylase antizyme (OAZ1) mRNA was used as the normalizer, as described previously.⁵ For IFIT1 mRNA quantification, the RTPrimerDB assay ID 4052 was used, with the following forward primer, reverse primer, and probe respectively: 5′-cgctatagaatggagtgtcca-3′, 5′-tttcctccacacttcagca-3′, and 5′-/56-FAM/aatagactgtgaggaaggatgggcc-3′.¹³ Primers and probes were obtained from Integrated DNA Technologies (Coralville, IA). A cDNA vector was used as IFIT1 standard (Clone ID: 4250452, Thermo Open Biosystems, Huntsville, AL). Assay reagents and standards used for the OAS1, MX1, and OAZ1 expression assays were described previously.⁵ All qPCR assays were run on a 384-well LightCycler 480 (Roche Applied Science, Indianapolis, IN) real-time PCR system in 3 μl duplicate reactions. Quantification cycle values were computed using a second derivative maxima algorithm with the LightCycler 480 software.

Virus Infection Assays

Compounds with IFN signal enhancing activity were assessed for antiviral activity in 2fTGH cells and primary-culture human tracheobronchial epithelial cells (hTECs). 2fTGH cells cultured in 96 well plates (2,000 cells per well) were treated with compound or IFN-β either for 6 h before virus infection, and then after for 18-20 h, or only after virus infection. Virus was inoculated with cells for 1 h at 37 °C. Thereafter, cells were washed, and medium containing 2% FBS added, before collection for viral titer. hTECs were isolated and cultured as described previously.¹⁴ For the present experiments, hTECs were grown to confluence under submerged conditions (4-8 d) in BEGM medium and then under air-liquid interface (ALI) conditions in ALI medium as described previously.¹⁵ Differentiation of cells was verified after 21 days by visualization of beating cilia using a Nikon Ti-S inverted phase-contrast microscope enclosed in a customized environmental chamber maintained at 37 °C. The average cilia beat frequency was measured to be 7-13 Hz. For infection, the basal compartment media was changed to medium containing only 0.5 mg/ml BSA and 5×10⁻⁸ M retinoic acid, and 16 h later, virus was added to the apical compartment for 2 h at 34 °C. Cells were then washed, and basal medium was replaced with medium containing compound, IFN and DMSO, or DMSO alone.

Viruses included encephalomyocarditis virus (EMCV), strain VR-129B (ATCC, Manassas, VA) at multiplicity of infection (MOI) of 0.01, human rhinovirus (HRV-A1) was used at MOI of 0.1 (obtained from J. Gern, University of Wisconsin), and respiratory syncytial virus (RSV Long, GenBank accession number AY911262, obtained from R. Brazas, Duke University) at MOI 2. Standard plaque assay protocols were followed to calculate plaque-forming units (PFU).¹⁶ For EMCV, virus particles in cell supernatant were lysed by heating for 10 min at 95 °C and then quantified using a 1-step qPCR assay for the EMCV 3D gene as described previously.⁵ For HRV-A1, 1-step qPCR was performed using primers 5′-cctccggcccctgaat-3′, and 5′-aaacacggacacccaaagtagt-3′, and probe 5′-/56-FAM-ctaaccttaaacctgcagcca-3′, complementary to the 5′ UTR for HRV-A1, as described previously.¹⁷ For RSV Long qPCR, the F1 and F2 primers and RS-F3 probe were used as described by Mentel et al.¹⁸ Purified RNA from virus stocks were used as standards. Virus particles in the apical washes were normalized to expression of OAZ1.

Results

Primary HTS for ISRE Activity

We anticipated run-to-run variance of the primary HTS assay for ISRE activity based on cell stock, IFN batch, IFN storage conditions and time after dilution. The IFN-β concentration-ISRE activity responses from the first plates of the NCI Diversity Set II screen and the Spectrum library screen demonstrate that while the coefficient of variance per data point does not change significantly from experiment-to-experiment, the absolute values and slope of the log(agonist)-response curve do vary between experiments (Fig. 2A). Normalization of the raw data by conversion to a percentage maximum activity using the minimum activity (0 U/ml IFN-β) and maximum (200 U/ml IFN-β) controls cannot account for the change in the Hill slope or the right-shift of the entire activity response curve near the first inflexion point (Fig. 2B). In other words, normalization to a minimum-maximum activity scale is insufficient to characterize the inherent heterogeneity in experiment-to-experiment variance in different IFN-β concentration regimes.

Figure 2.

IFN-β concentration-response for ISRE activity. (A) Effect of IFN-β on ISRE activity for each compound library. (B) Corresponding normalization of the raw data to a minimum defined by the 0 U/ml IFN-β treatment control and maximum defined by the 200 U/ml IFN-β control. 119×168mm (300 × 300 DPI)

When we screened the Spectrum library and the NCI Diversity Set II libraries, we first analyzed the two library data sets individually as we described previously.⁵ For this type of analysis, data from each library screen was normalized separately using plate median and B-score normalization to adjust for plate-to-plate variance, and obtaining z-scores for each individual set. In this analysis, four data points corresponding to four test concentrations represent each compound, and compounds are ranked by maximum z-score at any concentration for ease of visualization (Fig. 3A). Based on the criteria of a maximum z-score ≥ 2.5 for any compound concentration, we obtained 9 hits from the NCI library and 23 hits from the Spectrum library, for a total of 32 hits out of 4000 compounds screened, yielding a hit rate of 0.8%.

Figure 3.

Data from screening the NCI Diversity set II and the Microsource Spectrum libraries. (A) ISRE activity represented as z-scores from independent z-score normalization analyses from the two different screens. (B) Raw ISRE activity expressed as effective INF-β concentrations by fitting raw data to the ISRE-IFN-β concentration response curve from control wells on a per plate, per quadrant basis. Each point represents the average of duplicate readings. Dotted arrows reflect effective IFN-β concentrations for the highest z-score from each library. (C) Combined z-score normalization of both sets of data after conversion to effective IFN-β concentrations. Each compound is represented by four points, each corresponding to one compound concentration. Compounds are ranked by the maximum z-score at any concentration tested. Dotted line represents z-score ≥ 2.5 cutoff. The number in parentheses represents number of hits for a given analysis. 198×231mm (300 × 300 DPI)

While the 9 compounds selected from the NCI library, and the 23 compounds selected from the Spectrum library are the best-performing out of the 2000 compounds queried per individual set, we questioned if these 32 compounds were indeed the best performing of the 4000 small molecules tested in total. To answer this question, we would need to analyze the two sets of screening data as a single set. Therefore, we needed to implement a type of analysis that would require normalization that did not rely solely on statistics but also could take into account the heterogeneity of the run-to-run variance of the assay. However, we also recognized that the two library screens produced different responses to IFN-β (Fig. 2). For example, a compound that increased ISRE activity by 5% compared to IFN-β treatment alone in one screen was not equivalent to a compound that increased activity by 5% in the second screen, because the Hill slope of the IFN-β concentration-ISRE activity response is different (Fig. 2). In fact, since we know the activity of IFN-β depends greatly on incubation time after dilution, there might not only be screen-to screen but also plate-to-plate variances in response. To address these issues for normalization, we fit the test data to the IFN-β concentration-ISRE activity luminescence curve on a per-plate, per-quadrant basis. Since we included 8 replicates of each IFN-β concentration control, 4 on each of the left-most and right-most columns, we had two replicates of each IFN-β concentration on a particular quadrant (Supplementary Fig. S1B). Thus, we could fit data from test compounds to four individual curves representing each quadrant on the plate, to eliminate any pipetting or compound dilution effects. This method of normalization resulted in conversion of the raw data represented in luminescence units to effective IFN-β concentrations. The parameter estimates and goodness of fits for the log(IFN-β concentration) versus response curves are provided in Supplementary Table S1. While the goodness of fits were excellent (r² = 0.986 on average for the NCI library, and 0.994 for the Spectrum library), there was significant difference in the magnitude and variance of the fit parameters (top, bottom, logEC₅₀, and Hill slope) both within and between the two libraries (Supplementary Fig. S2). These differences cannot be adjusted for by conventional minimum-maximum or plate median normalization. However, these non-linear, non-uniform variances are accounted for by the use of our curve-fitting approach, which leads to more accurate assessment of compound activity, and predictably improves hit picking. The effective IFN-β concentration values for both libraries after mean summarization of duplicate values per compound concentration are shown in Fig. 3B. In terms of effective IFN-β concentration, the compound with the highest z-score from the Spectrum library shows greater activity in terms of effective IFN-β concentration compared to the compound from the NCI library with the highest z-score. On analyzing the data to identify the source of this discrepancy, we discovered that compounds from the NCI library had lower mean activity and a narrower range of activities than compounds in the Spectrum library (Supplementary Fig. S3). This resulted in smaller standard deviations, and higher z-scores for compounds in the NCI library compared to the Spectrum library for a given magnitude of biological activity. Computing z-scores on the set of combined effective IFN-β concentrations allows comparison across the entire distribution of activities, and eliminates the inflation of the z-scores seen when analyzing the NCI library separately. The difference in activity distribution between the NCI and Spectrum libraries suggests that the 32 hits picked through analyses of each library separately were not the 32 most active compounds of the total 4000 compounds tested. When the combined dataset of effective IFN-β concentrations was transformed to z-score values, we again used a z-score threshold of 2.5 to select active compounds, identifying 45 hits (hit rate of 1.125%), only two of which hits were from the NCI library, again reflecting the higher activity of compounds in the Spectrum library in our assay (Fig. 3C). Thus, we avoided selection of less active compounds from the NCI library as hits, selecting additional, more active hit compounds from the Spectrum library. Further, normalization of test data to quantifiable biological response allowed elimination of outliers among the control wells, without having to exclude data from an entire plate.

Hit Confirmation

Of the 45 hits called from our combined analysis, we selected 35 for further evaluation. Of the 10 remaining hits called, 8 were not selected for further exploration because they were not readily available from commercial sources, and the 2 others were ribavirin and doxorubicin. Ribavirin is an FDA-approved antiviral therapeutic, and was previously reported to augment IFN signaling.^{19, 20} Our previous work shows that doxorubicin acts as an IFN signal enhancer.⁵ As a result, all 35 compounds selected for further analysis (indicated with green dots in Fig. 3C) were derived from the Spectrum library. These compounds represented several structural classes including CNS monoamine neurotransmitter-like compounds, flavone-like compounds, statins, salicylates, large and lipophilic, lipophilic acids, and small size phenols. When these compounds were obtained from outside the library, and subjected to hit confirmation using the same ISRE activity assay across 12 concentrations, we found that 25 of the 35 compounds were confirmed as hits based on concentration-dependent increases in ISRE activity.

Hit Validation and Expansion using ISRE Activity and Orthogonal Assays

We started evaluating structural classes represented by the confirmed hits using orthogonal assays for IFN signaling, by assessing the activities of structurally similar small molecules in each structural class. In particular, we found that 6 of 10 statins tested significantly enhanced IFN-β-driven ISRE activity with varying efficacy and potency (Fig. 4A, B). The IFN signal-enhancing property of these statins was confirmed when these compounds also increased ISG expression in 2fTGH cells treated with statin and IFN-β (Fig. 5A, and data not shown). Thus, each of three ISGs with known antiviral activity (OAS1, MX1 and IFIT1) showed significant increases in expression in response to statin plus IFN-β treatment (Fig. 5A).

Figure 4.

Certain statins enhance IFN-driven ISRE activity in a concentration dependent manner. (A) Statin concentration-ISRE activity response curves for the 6 active statins (10 tested). The lines represent fits of the data to log(agonist)-response curves. Cells were co-stimulated with 15U/ml IFN-β. (B) Corresponding potencies for the ISRE activity from (A). 178×437mm (300 × 300 DPI)

Figure 5.

The IFN signal enhancing property of statin c is validated in orthogonal assays for (A) IFN-driven antiviral gene expression, and (B) STAT2 nuclear translocation. (A) Cells were treated with statin and IFN for 12 h before assaying for OAS1, MX1, or IFIT1 mRNA expression. p-values for individual comparisons are from Bonferroni post-tests from repeated measures two-way ANOVA, comparing the 0 μM statin c condition to the other drug concentrations. For OAS1, overall significance: IFN, p<0.0001, statin dose, p<0.0001; interaction, p=0.0108. For MX1, overall significance: IFN, p<0.0001, statin dose, p<0.0001; interaction, p<0.0001. For IFIT1, overall significance: IFN, p<0.0001, statin dose, p<0.0001; interaction, p=0.0072. (B) For nuclear translocation of STAT2, cells were treated with statin for 20h before stimulation with IFN-β for 20min or 2h as indicated, and then prepared for staining and imaging. Data were analyzed using one-way ANOVA, overall p-value < 0.0001 for both IFN-β treatment times. p-values indicated are from Dunnett’s post-test with each dose compared to the no treatment control. For comparisons with DMSO alone treatment, *p<0.05, **p<0.01, ***p<0.001. 125×91mm (300 × 300 DPI)

To further confirm the effect of select statins on IFN signaling, we also assessed STAT2 nuclear retention as a measure of IFN-β-dependent STAT2 activation. After type 1 IFNs bind the receptor, STAT2 binds to the receptor, enabling recruitment of STAT1 and release of phosphorylated STAT1-STAT2 heterodimer in the cytoplasm (Fig. 1A). The STAT1-STAT2 heterodimer recruits IRF9, and this complex translocates to the nucleus and binds with the ISREs, yielding ISG transcription. Prolonged STAT1 phosphorylation may even be a feature of increased antiviral activity.⁶ To determine whether statins might enhance STAT1-STAT2 translocation as a measure of IFN signal enhancement, we developed a high-content imaging assay for monitoring STAT1/STAT2 nuclear translocation (Supplementary Fig. 4A). Consistent with the known biology, IFN-γ selectively drives STAT1 translocation, while IFN-β drives STAT1-STAT2 translocation with a maximal effect at 20 min (Supplementary Fig. 4B). Statin treatment caused a significant increase in STAT2 nuclear retention at 20 min and persistent retention at 2 h in 2fTGH cells when given in combination with IFN-β (Fig. 5B). Taken together, these findings validate the IFN signal enhancing activity of select statins.

Validation of IFN Signal Enhancers for Antiviral Activity

To validate our hits in a more biologically relevant context, we assessed the antiviral activity of select statins. In the first set of experiments, we found that statins with potent ISRE enhancing activity caused a concentration-dependent decrease in the level of two picornaviruses, EMCV and HRV-A1, in 2fTGH cells (Fig. 6A). Statins caused maximal decreases of 93-96% in EMCV level and 52-70% decrease in HRV-A1 level. The IC₅₀ values for antiviral activity correlated with potencies for ISRE enhancing activities (Fig. 6B and 4B).

Figure 6.

IFN signal enhancing statins show improved antiviral activity. (A) Dose-dependent antiviral activity of statins a, b and c in 2fTGH cells when introduced after infection with EMCV (MOI 0.01) and HRV-A1 (MOI 0.1). (B) Potencies of statins for inhibition of picornavirus replication, obtained by fitting data in A. (C) Antiviral activity of statin b in primary hTECs cultured on air-liquid interface, infected with HRV-A1 (MOI 0.01), or RSV Long (MOI 2). hTECs were treated with either 5 μM statin b, 100 U/ml IFN-β and DMSO, or DMSO alone after infection. For comparisons with DMSO alone treatment, *p<0.05, **p<0.01, ***p<0.001. 122×86mm (300 × 300 DPI)

We then moved to test the antiviral activity of statins against common respiratory pathogens picornavirus HRV-A1 and paramyxovirus RSV in primary-culture human airway epithelial cells (hTECs), which are highly predictive of in vivo findings.^{14, 21} We found that statin b treatment of infected hTECs caused a decrease in viral titers for HRV-A1 (by 83%) and RSV (by 89%), effects similar to that of IFN-β treatment (100 U/ml) (Fig. 6C), suggesting that statin b exerts a biologically relevant antiviral effect. Together, these results indicate that our screening strategy was able to identify compounds as IFN signaling enhancers at a level that translates to corresponding functional activity as antiviral compounds.

Discussion

In this study, we developed and applied a new HTS design and analysis method that enables data integration and hit picking across multiple phenotypic screens. Central to this is the addition of a standard curve that allows transformation of assay data to biologically relevant units, allowing both cross-screen normalization and interpretation of results in their functional context. We validated the approach with confirmatory testing of representative hit compounds, and an extended analysis of statin analogs as a structural class with significant IFN signal enhancing and antiviral activity. We found that the most potent statins for augmenting IFN-β-driven ISRE activity showed significant efficacy for ISG expression and STAT2 activation. Moreover, statin enhancement of IFN signaling translated to antiviral activity against EMCV and HRV-A1 picornaviruses in the 2fTGH cell line (the parent screening line) and against HRV-A1 and RSV in primary human airway epithelial cells cultured under physiologic conditions.

Co-treatment with IFN-β increased but was not required for statin antiviral effect. However, IFN-β co-stimulus was required for statin-induced enhancement of ISRE activity and ISG expression. Two factors might contribute to this discrepancy. First, viral replication results in activation of the pattern recognition receptors, and subsequent production of endogenous IFN, which may drive sufficient IFN signaling for augmentation by statin. Second, IFN-independent antiviral effects of statins may be responsible, consistent with the observation that statins, more formally 3-hydroxy-3-methylglutaryl-CoA (HMG-CoA) reductase inhibitors, can inhibit viral entry and replication by HMG-CoA reductase-inhibition dependent decreases in mevalonate and subsequent decreased prenylation of host proteins.^22-24 While others have reported the inhibition of sterol synthesis by type 1 IFN signaling in macrophages^{25, 26}, ours is the first evidence that statin treatment results in enhancement of IFN signaling. In evaluating some statins, Ikeda et al. did not find type 1 IFN signal enhancement.²³ Discrepancies between our results and theirs may be due to different drug incubation times, which statins were tested, and cell type. Further, statins may have multiple modes of antiviral action, which could explain the discordance between observed antiviral capacity of statins and their HMGCR inhibitory potencies.²³ Exactly how statins augment IFN signaling has yet to be determined, and both this and a full understanding of statin antiviral activity are the subject of future studies. Based on these data, however, we conclude that statins show promise both as treatment for active viral infections, and as prophylaxis in times of epidemic threat, in the elderly, or even in young children to prevent virus-induced development of chronic respiratory disease²⁷.

The term ‘qHTS’ has been used to describe screening of compounds at multiple concentrations, after which compound concentration-response curves are fit to calculate IC₅₀/EC₅₀ values, depending on whether a gain- or loss-of-function screen is performed.^{28, 29} In this type of analysis, raw data is scaled to a positive control (a high concentration for a positive control compound) and a negative control. In the present study, rather than scaling linearly, we fit each test data point to a complete biological response curve for the phenotype of interest (ISRE regulated gene transcription) produced by the native agonist (IFN-β). Accordingly, we have designated this methodology quantifiable biological response HTS (qbrHTS), as it requires normalization of screening data to the response generated by a biologically relevant molecule or ligand. In addition to enabling hit picking across multiple screens, this methodology also reduces the impact of systemic drift, potentially eliminating the need for inter-plate calibrators in large screens. Our analysis allowed us to eliminate weaker hits called by analysis of data for individual screens. Thus, when the data for two library screens were considered together, some of these hits could be eliminated from the hit list since they were now revealed to be significantly less active. Moreover, after performing the analysis, we recognized that there is an additional benefit to a combined analysis. Not only are weaker or less active hits eliminated from validation and subsequent work, but also stronger or more active compounds are picked as hits that would have been excluded from the list of hits based on individual analyses at any given statistical threshold. In the present case, 9 hits were called from the NCI library and 23 hits from the Spectrum library based on independent analyses. Using the combined qbrHTS analytical approach, 2 hits were called from the NCI library and 43 from the Spectrum library, increasing the total number of hits from 32 to 45. The hits in the two types of analyses are quite different, thereby leading to improved prioritization and resource allocation of post-screening development efforts.

There is another benefit to normalization by curve fitting to a biological response. In our case, the expression of screening data as effective IFN concentrations provides an index that is more applicable to a biological system. This type of information can provide more relevant insight into functional implications for hit compounds. For example, interferon activity is more readily extrapolated to potential in vitro and in vivo antiviral activity than a percentage increase in luminescence units for ISRE activity. The biological readout instead enables the investigator to make more informed decisions in designing validation experiments. This type of normalization to the response curve for biologically relevant ligand is especially useful when a good tool compound is not available for use as a positive control, as would be the case when screening for first-in-class drugs, such as IFN signal enhancers. The method also offers advantages when screening for relatively small increases in gain-of-function due to its robustness. We found that it was unlikely that a small molecule would produce the magnitude of response as great as the native ligand at high concentrations, particularly at the transcriptional level. However, a maximal response might be undesirable due to the known autoimmune and toxic side effects of IFN treatment.^{2, 4} Further, the balance between type 1or 3 versus type 2 IFN signaling has to be preserved to ensure adequate antiviral, as well as bacterial host defense states.³⁰ Thus, screening for compounds that result in smaller increments of ISRE activity would be advantageous, and would be readily assessed by the present approach.

In summary, conventional normalization and z-score based hit-selection on a per library basis does not allow accurate analyses across multiple runs or screens. Herein, we describe an alternative approach that we have designated qbrHTS, wherein screening data is converted to a biologically relevant readout by fitting to a concentration-phenotype response curve produced by a relevant ligand. This approach allows for the exclusion of less active compounds, and enrichment of more active compounds in the hit list based on a more accurate analysis of data from multiple screens achieved by assessing activity across the combined dataset rather than status as an outlier within an individual screen. As such, it helps avoid costly validation of lower activity compounds and adds value to post-screening drug discovery efforts. Given the number of possible batch effects in a screening campaign, we feel that application of this approach for other targets or phenotypes, while requiring some additional effort in the assay development and screening stages, will enhance the reliability and accuracy of hit selection, ultimately leading to better rates of hit-to-lead development.