key: cord-0271590-dw6areec
authors: White, Brandon; Komalo, Ben; Nicolaisen, Lauren; Donne, Matt; Marsh, Charlie; DeVay, Rachel M.; Nguyen, An M.; Cousin, Wendy; Heinrich, Jarred; Van Trump, William J.; Plott, Tempest; Fuller, Colin J.; Nguyen, Dat; Chen, Daniel; Bucher, Delia; Tyrra, Sabine; Haynes, Laura; Kuchel, George; Goronzy, Jorg; Larbi, Anis; Fulop, Tamas; Heiser, Diane; Schwandner, Ralf; Elabd, Christian; Kamens, Ben
title: A Multi-Phenotype System to Discover Therapies for Age-Related Dysregulation of the Immune Response to Viral Infections
date: 2020-07-30
journal: bioRxiv
DOI: 10.1101/2020.07.30.223875
sha: 0a1a56c75f0d5130bf82204b3d75f06f0826bef2
doc_id: 271590
cord_uid: dw6areec

Age-related immune dysregulation contributes to increased susceptibility to infection and disease in older adults. We combined high-throughput laboratory automation with machine learning to build a multi-phenotype aging profile that models the dysfunctional immune response to viral infection in older adults. From a single well, our multi-phenotype aging profile can capture changes in cell composition, physical cell-to-cell interaction, organelle structure, cytokines, and other hidden complexities contributing to age-related dysfunction. This system allows for rapid identification of new potential compounds to rejuvenate older adults’ immune response. We used our technology to screen thousands of compounds for their ability to make old immune cells respond to viral infection like young immune cells. We observed beneficial effects of multiple compounds, of which two of the most promising were disulfiram and triptonide. Our findings indicate that disulfiram could be considered as a treatment for severe coronavirus disease 2019 and other inflammatory infections.

Aging increases risk of infection progressing to severe disease and death. 1 The cause of this increased risk is not well understood, but it is known that the immune system undergoes many changes as it ages. Some changes impact the innate immune system, such as the dysregulation of inflammatory responses-otherwise known as "inflamm-aging"-which may disrupt innate antiviral responses by increasing the resting level of circulating proinflammatory cytokines. [2] [3] [4] Other changes associated with age may impair the adaptive immune system, ranging from relative reductions in naïve T cells and increases in memory/effector T-cell populations to alterations in signaling pathways, epigenetics, and effector functions. [5] [6] [7] [8] [9] [10] [11] Collectively, these age-related alterations in the immune system are referred to as "immunosenescence." 1, 6 Yet, while immunosenescence and inflamm-aging have been linked with pathology, both may also have beneficial effects. 2, 12 Maintaining a balance between the beneficial effects of these aging processes while targeting those that are detrimental requires an ability to measure and model a complex interplay of immune states and phenotypes. Such a model could be a powerful system for finding potential therapies for the dysregulated viral response in older adults.

Coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) exemplifies this increased risk of infection leading to hospitalization or death in older adults. [13] [14] [15] The disease has disproportionally affected older adults who suffer from the highest morbidity and mortality. [13] [14] [15] A study of patients hospitalized with SARS-CoV-2 infection demonstrated that age at admission was associated with a greater mortality risk than other factors such as obesity and chronic cardiac disease (age 70-79 years [HR, 9.6] ; age ≥80 years [HR, 13.6] ; obesity [HR, 1.4] ; chronic cardiac disease [HR, 1.3] ). 15 In China, the case fatality rate was 0.4% for patients age 40-49 years and 15% for patients ≥80 years of age 14 ; in Italy, 96% of COVID-19 deaths occurred in people ≥60 years of age, while only 1% of patients under the age of 50 years died. 13 Some believe that the increased severity of COVID-19 in older adults may be exacerbated by the aging immune system. 16 The importance of treating the host immune response was recently demonstrated by a randomly controlled trial of the anti-inflammatory glucocorticoid dexamethasone. Among hospitalized patients with COVID-19, the use of dexamethasone lowered 28-day mortality in patients receiving invasive mechanical ventilation (average age 59) and those receiving oxygen without ventilation (average age 66). 17 This tremendous burden of infection in older adults is not limited to COVID-19. Between 12,000 to 61,000 deaths occur as a result of seasonal influenza in the United States each year, and approximately 90% of these deaths occur in patients ≥65 years of age. 18, 19 Between February and May 2020, the combination of deaths related to pneumonia, COVID-19, or seasonal influenza among all age groups accounted for 14% of deaths in the United States. 20 Of these deaths, 9% were related to pneumonia, 7% to COVID-19, and 1% to influenza. 20 Between 2009 and 2014, adults ≥60 years of age accounted for 70% of the sepsis cases reported in the United States. 21 The mortality rate for sepsis was 16% for patients 60−79 years and 18% for patients ≥80 years of age. 21 In another study, patients >80 years of age with sepsis had an inhospital mortality rate of 47%. 22 We have developed a system to discover and target the immune mechanisms that change with age and lead to dysregulated responses to infection. This system combines data generated from a high-throughput in vitro laboratory with machine learning to model the dysfunctional viral response in older adults. We trained the system to accurately detect the differences between young and old viral immune responses by exposing peripheral blood mononuclear cells (PBMCs) from young and old donors (N=89) to vesicular stomatitis virus encoding a red fluorescent protein (rVSV-ΔG-mCherry). This virus was chosen because of its ability to model the innate immune activation pathways of prevalent respiratory RNA viruses and because it is safe to use in a high-throughput, biosafety level one laboratory. 23, 24 The virus infected the monocytes and macrophages, which subsequently created a highly inflamed environment for the lymphocytes and other PBMCs. The differences between the young and old immune response were captured in our system's multi-phenotype aging profile of viral response, which contained dozens of features measuring changes in cell composition, physical cell-to-cell interaction, organelle structure, cytokine levels, and other hidden complexities that contribute to age-related dysfunction. Many of these features showed significant differences between young and old viral responses. Since these aging profiles can be generated from a single well, we used them to assess the effect of 3428 bioactive compounds on the viral immune response of older adults ( Figure 1 ). We ranked these compounds by combining the features in the multiphenotype aging profiles into scores that measured their beneficial effect on immune aging and disease-specific phenotypes. Among others, the ranking revealed disulfiram and triptonide as two promising compounds that restored multiple aspects of the viral immune response of old PBMCs to a younger state. The scalability and generalizability of our system enables a novel characterization of the aging immune system that could be used to help sustain the health of older adults over their lifetime. 

The first feature we focused on in our multi-phenotype aging profile was cell composition. We built a scalable method for classifying immune cell types within bulk PBMCs based on cell morphology, because traditional methods for measuring cell composition were unable to scale to the needs of our high-throughput system and recent studies have shown that functional changes in cells can affect their shape while changes in shape can in turn alter their cellular function. [25] [26] [27] To highlight a diverse range of morphological changes, the cells were stained with cell painting palettes that broadly characterized a variety of subcellular compartments such as the nucleus, cytoskeleton, mitochondria, and cell membrane. 28 To understand and verify the morphological differences between immune cell types, we first isolated PBMCs from whole blood and then further isolated T cells (CD3 + ), B cells (CD19 + ), natural killer (NK) cells (CD56 + ), and monocytes (CD14 + ) using immunomagnetic negative selection based on customarily defined cell surface markers. Isolated cells were plated, stained with Cell Painting Palette 1 (Supplemental Table S1 ), and imaged. Within these images, we observed and quantified notable differences between the shapes of the isolated immune cell populations. As shown in Figure 2 , the T cells and B cells shared the most similarities to each other; the T cells had a rounder and more uniform morphology, while the B cells had a more oval-shaped form with small areas of highly concentrated phalloidin staining on either end. The NK cells diverged from the tightly spherical form of other lymphocytes and showed more elongated and irregular forms.

The monocyte populations were the largest in size and encompassed podosomes, dendrites, and kidney-shaped nuclei ( Figure 2 ).

human bulk PBMCs using immunomagnetic negative selection and stained with Hoechst (blue), ConA (red), Phalloidin (green), WGA (far red). The boxed areas represent the single cell shown in the zoomed image. Scale bar = 10 µm (overview) and 5 µm (zoom). (b) Normalized averages for a subset of morphologic features that differentiate cell populations. ConA, concanavalin A; MAD, mean average deviation; NK, natural killer; PBMC, peripheral blood mononuclear cell; WGA, wheat germ agglutinin.

Using the observed phenotypic variation in isolated cells, we designed a computational model to classify single cells within a mixture of PBMCs. We built the model on PBMCs exposed to rVSV-ΔG-mCherry so that the multi-phenotype aging profile could accurately measure the compositional changes between rVSV-exposed PBMCs from young and old donors. These cells were stained with Cell Painting Palette 2, which revealed similar subcellular compartments as Cell Painting Palette 1, but included the membrane potential sensitive dye, MitoTracker Orange.

The infected cells also expressed a mCherry-rVSV protein that had an overlapping emission spectrum with the MitoTracker causing them to be visualized in the same channel (Supplemental Table S2 ). Image embeddings, a statistically optimized representation of an image, were generated on an expert-curated training set of >61,000 single cell images from over 8 experimental runs using 31 plates seeded with cells from 105 distinct donors. The embeddings were then used to train a classification model. The model classified a single cell in an image of bulk PBMCs as a T cell, B cell, NK cell, dead cell and infected monocyte, or uninfected monocyte (Figure 3a ). To validate the accuracy of the classification, we built an expert-labeled test set of >10,000 single cell images from PBMCs. When we evaluated the classifier on this data set, the model achieved 91% accuracy (data not shown). We also took single cell images and ran a principal component analysis on several hundred features derived from CellProfiler that were not used to train the model. Our model-predicted classification demonstrated clear clustering on the first two principal components (PC) of our analysis ( Figure   3b ). Because immune cell populations vary significantly between individuals, the model was run on PBMCs derived from several different individual donors to see if this variance could be captured by the classifier. 29 Predicted cell compositions not only recapitulated the large interdonor variance but also fell within the expected range of PBMC compositions (Figure 3c ). 30 Figure 3 . (a) Images of PBMCs and cells predicted to be T cells, B cells, NK cells, uninfected monocytes, or dead cells or infected monocytes. The cells were stained with the Cell Painting Palette 2 (Hoechst, blue; Phalloidin, green; MitoTracker/mCherry, red); WGA; ConA), and the predictions were derived from a model trained on image embeddings. Scale bar = 10 µm (PBMCs) or 5 µm (predicted cells). (b) PC analysis plot of 1200 morphologic features computed on single cells within bulk PBMCs exposed to rVSV-ΔG-mCherry. PCs were generated using the average feature value, then colored by cell type labels from a model trained on image embeddings of cell types. (c) The predicted PBMC composition of 24 donors exposed to rVSV-ΔG-mCherry at 10× MOI. The predictions were derived from a model trained on image embeddings computed from single-cell crops, labeled by cell type. ConA, concanavalin A; MOI, multiplicity of infection; NK, natural killer; PBMC, peripheral blood mononuclear cell; PC, principal component; rVSV-ΔG-mCherry, recombinant vesicular stomatitis virus expressing a red fluorescent construct; WGA, wheat germ agglutinin.

To provide an independent validation of our method, we correlated the cell composition predictions from our morphological model with those identified using flow cytometry (Supplemental Figure S1 ). The cell composition predictions from our model significantly correlated with those determined by flow cytometry for T cells (r=0.73; P=.007) and NK cells (r=0.92; P<.001). The dead cell and infected monocyte composition correlated significantly with all dead and infected myeloid cells derived from flow cytometry (r=0.62; P=.02; Figure 4) . A Bcell marker was not included in the flow cytometry panel because of the mCherry overlap, so a correlation for the B-cell population was not calculated. While flow cytometry was able to accurately differentiate live and dead cells at different granularities, the morphological properties of dead cells were very similar to infected monocytes in the images, which made it difficult to differentiate live and dead infected cells using our system. Overall, the cell type classification model we developed provides a scalable and reproducible method to incorporate cell composition and other phenotypic features into our multi-phenotype aging profile. . Correlation between the predicted composition of PBMCs exposed to rVSV-ΔG-mCherry and the composition determined by flow cytometry analysis for (a) T cells, (b) NK cells, and (c) infected and dead myeloid cells. Statistical analysis was performed via the 2-tailed t test. NK, natural killer; PBMC, peripheral blood mononuclear cell; rVSV-ΔG-mCherry, recombinant vesicular stomatitis virus expressing the fluorophore mCherry.

Dysregulated immune responses are a central driver of disease progression in severe infection of older adults. [31] [32] [33] [34] To develop aging profile features that specifically measure phenotypes related to viral infection, we profiled the responses to viral load in PBMCs from donors ranging from 30 to 60 years old. We observed notable shifts in cellular composition in virally exposed PBMCs; specifically, there was a significant increase in dead cells and infected monocytes and a significant decrease in uninfected monocytes at 0.1×, 1×, and 10× multiplicity of infection (MOI; P<.001 for all MOIs; Figure 5a ). The NK cell population decreased moderately but significantly for all viral conditions as compared to unexposed conditions (P<.001 for all MOIs;

Figure 5a) but no difference was observed when comparing different viral loads. The T-cell and B-cell populations did not undergo significant changes due to viral load. We ran an experiment to test cytokine production of PBMCs when exposed to rVSV at 1×, 5×, 7× and 10× MOI. In contrast, monocyte chemoattractant protein 1 (MCP1; P<.001) significantly decreased upon 10× viral exposure compared to 1× exposure ( Figure 5b ). The significant differences in these cytokines were also reinforced by similar dose dependent trends at 1×, 5×, 7×, and 10× MOIs ( Figure 5b ). Given the robust changes in PBMC composition and cytokine production upon viral exposure, we incorporated these features into our multi-phenotype aging profile to understand how they differ with age. To more broadly measure how the immune response changes to different viral loads, we trained machine learning algorithms to learn the differences in the cellular immune response by titrating rVSV into PBMCs at different MOIs. We first generated embeddings on single cell images from PBMCs that were exposed to rVSV at 0.1×, 1×, or 10× MOI, and then stained with Cell Painting Palette 2. The embeddings were then used to train a machine learning classifier to discriminate the cellular responses to rVSV at a range of MOIs. PBMCs from 24 donors were evenly distributed across a 384-well plate assay to minimize bias due to plate location then exposed to rVSV at three different MOIs. We used stratified four-fold cross-validation to evaluate the classifier's performance on each well. When we evaluated the model trained on monocytes only, the classifier was able to differentiate between unexposed, 0.1× MOI, 1× MOI, and 10× MOI with 91% accuracy across all four folds ( Figure 6a ). We trained a model on T cells but we were unable to distinguish between the unexposed cells and the cells exposed to 0.1× and 1× MOI conditions. However, our model learned significant differences between the T-cell response to 10× MOI cells and unexposed cells. The model was able to distinguish these conditions at an area under the receiver operating characteristic curve (AUC) of 0.92 across the four folds ( Figure 6b ). T cells, unlike monocytes, are not directly infected by the rVSV and therefore their response could be dependent on modifications of the environment after viral exposure which might explain why no difference is observed at lower MOIs. These validated models of cell response to viral loads can be used to observe whether the phenotype of old PBMCs looks more similar to the higher viral load phenotype compared with young PBMCs. 

We used our multi-phenotype aging profile to conduct a novel characterization of the differences between young and old viral immune responses across dozens of features (Table 1) . We generated the aging profiles for PBMCs from 49 young (<35 years of age) donors and 40 old (>60 years of age) donors. Demographics of the donors reflected the standard population of older adults, with old donors having a higher number of comorbidities, particularly hypertension, and taking more medications than young donors (Supplemental Table S3 ). All 89 donors were plated across two 384-well plate assays, with the two age populations evenly distributed across the plate to account for plate position effects. The donors were exposed to rVSV at 10× MOI and stained with Cell Painting Palette 2. Compared with young donors, PBMCs isolated from old donors had an increased percentage of dead cells and infected monocytes but showed no significant change in the percentage of uninfected monocytes (P<.001 and P=.53, respectively; Figure 7a ). The percentage of T cells decreased in both age groups with a significant decrease in T cells from old donors compared with young donors (P<.001; Figure 7b ). In contrast, a different experiment comparing unexposed PBMCs from young and old donors did not show a significant difference in T cells (P=.11). More work is necessary to understand these mechanisms, but the decreased T-cell population could recapitulate, at least in part, the lymphopenia observed in severely infected patients. 38, 39 Percentages of T cells identified from unexposed and rVSV-ΔG-mCherry-exposed (at 10× MOI) PBMCs, presented according to donor age. Statistical significance (via two-tailed t test) between young and old donors after viral exposure is indicated (NS, nonsignificant; ***, P<.001). The shading around the trend lines indicates the 95% CIs of the fitted lines. MOI, multiplicity of infection; rVSV-ΔG-mCherry, recombinant vesicular stomatitis virus expressing a red fluorescent protein.

Decreased T-cell levels in older individuals may be driven by cytokine-induced apoptosis or pyroptosis. This has been demonstrated to occur via induction of proinflammatory cytokines (eg, TNFα) and defects in the nuclear factor-κB (NF-κB) signaling cascade. 40, 41 To understand this relationship in aging, we compared the cytokine levels in our multi-phenotype aging profiles of young and old donors. No significant changes in cytokine production were observed between young and old donors in unexposed PBMCs ( Figure 8a ); there was also little difference in cytokine production of these cells when exposed to rVSV at 1× MOI (Supplemental Figure S2 ).

However, when PBMCs were exposed to 10× MOI, significant differences in young and old cytokine production were observed. The production of proinflammatory cytokines significantly Cell composition, cytokines, and unbiased machine-learning-driven aging models are only a fraction of the features in our multi-phenotype aging profile. One big advantage of cell imaging is the ability to generate point patterns to model how interactions change with various perturbations. 37 

to T-cell activation and aggregation of macrophages can contribute to cytokine production. 35, 36 To this end, we computed an interaction score using spatial statistical analysis to model non-Poisson multitype point distributions of specific cell types. The point patterns were assessed for complete spatial randomness using quadrate tests and distance methods. Spatial point pattern distributions of PBMCs were found to be homogenously distributed with strong evidence of interpoint interactions (Supplemental Figure S3 ). When we tested changes in cell-to-cell interaction between young and old donors, we found that the lymphocyte-to-monocyte interaction score was significantly increased in old donors compared with young donors ( Figure   10a ). The virus-related features in our aging profile also showed significant differences between young and old responses. Our T-cell viral load model predicts the T-cell response to the amount of viral load (0.1×, 1×, and 10× MOIs), and the same model predicted that the T-cell response of old donors had a higher apparent viral load than young donors ( Figure 10a ). Together, the many features within our multi-phenotype aging profile provide a comprehensive understanding of several differences between young and old donors.

In ). These general changes in mitochondrial structure between young and old cells are consistent with reported mitochondrial mechanisms that have been causally linked to the agerelated T-cell senescence and decline in T-cell function. 45 Compounds that alter the mitochondrial behavior in cells from old donors, or any of the other signatures we modeled in our system, may improve the response to infection in older adults. 46 Altogether, there are many significant differences between young and old viral immune responses as measured by dozens of features in our multi-phenotype aging profile. Normalized averages for features measuring the intensity distribution of mitochondria membrane potential in T cells from old and young donors exposed to rVSV-ΔG-mCherry at a 10× MOI. AUC, area under the receiver operating characteristic curve; MAD, mitochondrial-associated protein degradation; MOI, multiplicity of infection. rVSV-ΔG-mCherry, recombinant vesicular stomatitis virus expressing a red fluorescent construct; STD, standard deviation.

To find promising therapeutics for the purpose of rejuvenating immune responses in older adults, we generated our multi-phenotype aging profile for each of 3428 bioactive compounds.

Of the 3428 compounds, approximately 1900 have been approved and marketed in at least one country. PBMCs from 35 young and old donors were seeded in 384-well plates and exposed to rVSV, then the PBMCs from old donors were treated with a single compound or vehicle to be used as controls. Each individual treatment and control well generated its own multi-phenotype aging profile containing dozens of age-related features. Every feature in the aging profile was classified as a "hit" or "miss" based on the magnitude of its distance from old vehicle-only controls. All labeled features were combined into scores for the aging profile of each compound by weighting each feature based on its relevance to immune aging or COVID-19. For our COVID-19 cytokine score, we increased the weights of cytokines reported to be increased in patients with severe COVID-19 so that compounds moving these cytokine levels in a beneficial direction ranked more highly. For our immune aging score, we weighted the features based on the consistency and magnitude of the differences between young and old donors. Generally, this method allows us to tailor scoring in order to focus on any immune dysfunction in any specific disease by altering weights of the features. For example, a disease like cancer could have a score that more heavily weighs the cell-to-cell interaction features to find potential therapeutics that improve T-cell trafficking.

These immune aging and COVID-19 cytokine scores allowed us to evaluate thousands of compounds for their therapeutic potential by measuring how each one affects dozens of agerelated phenotypes in rVSV-exposed PBMCs. All compounds were ranked by our COVID-19

cytokine and immune aging scores, then the top 196 highest-scoring compounds were validated in a follow-up dose-response study. Through this process, we identified two compounds, disulfiram and triptonide, with the potential for rejuvenating the aging immune system ( Figure  Figure 11 . (a) Immune aging score and COVID-19 cytokine score plotted for the top 40 compounds from our high throughput screen of 3428 bioactive compounds. Ranking defined by the sum of the immune aging score and the COVID-19 cytokine score. (b) A ranking of our immune aging score for our top 40 compounds. COVID-19, coronavirus disease 2019.

Triptonide, a diterpenoid from the medicinal herb Tripterygium wilfordii Hook F, 47 had notable effects on virally exposed PBMCs from old donors. Tripterygium compounds generally have demonstrated anti-inflammatory, immunosuppressive, and antiviral effects in animal models and in vitro studies. [47] [48] [49] Studies have shown that the anti-inflammatory effects of Tripterygium compounds target both T cells and macrophages. For T cells, activation is suppressed through inhibition of NF-κB transcription. 50 For macrophages, IL-1β, IL-6, and TNFα production are decreased and expression of the anti-inflammatory cytokine IL-37 is increased. 51, 52 Consistently, we observed significant decreases in proinflammatory cytokines produced by triptonide-treated rVSV-exposed PBMCs relative to old vehicle-only control PBMCs. After normalizing the cytokine levels by the total number of cells, MCP1, IL-1β, and TNFα all showed significant differences at one or more concentrations tested compared with vehicle-only controls ( Figure 12 ). The most notable of these was MCP1, which showed a significant, dose-dependent reduction when compared with vehicle-only controls ( Figure 12 ). These results confirm what was previously seen in the literature for the Tripterygium compounds and validate the ability of our aging profile to identify potential immunomodulatory compounds. 47, 48 Triptonide also shifted the T-cell metabolic features from an older profile to a younger profile. There was a concentration-dependent decrease in the percentage of reticular T-cell mitochondria ( Figure   13a ), consistent with reports that indicate that activated T cells have more fragmented mitochondrial networks, and activated T cells may be more prevalent in responsive, relatively young immune systems. 53 The immunosuppressive and anti-inflammatory effects of Tripterygium compounds have also been linked to their ability to induce apoptosis in T cells. 54 We saw that triponide-treated

PBMCs had a significant increase in the percentage of dead cells and infected monocytes and a decrease in the percentage of T cells when compared to vehicle-only controls. For the highest three concentrations, there was an average of 21% more dead and infected monocytes and 19% less T cells relative to vehicle-only controls (data not shown). This was a common trend for many of our top scoring compounds.

The predictions from our T-cell aging models also significantly changed after treatment with triptonide. Compared with old vehicle-treated controls, PBMCs from old donors treated with various doses of triptonide shifted the model's readout from an older phenotype to a younger phenotype (Figure 13b ). At a low concentration, the T-cell age score of triptonide-treated Figure 13c ). These "off-age" effects are not definitively detrimental but show that the compound induces effects that are not directly related to age. In our compound screen, we observed all possible effect combinations; some compounds only showed "on-age" effects, some showed both "on-age" and "off-age" effects, some showed neither, and some induced only "off-age" effects. As an example, another compound, dimethyl fumarate, did not improve the "on-age" effects, as indicated by "on-age" scores that were very close to those of old vehicle-only controls (Figure 13c ).

Taken together, the multi-phenotype aging profile of old rVSV-exposed PBMCs treated with triptonide showed restored phenotypes in cytokines, reticular mitochondria, and the T-cell predicted age at multiple concentrations. A deeper understanding of the aging mechanisms behind triptonide could shed light into the dysfunction of immune viral responses in older adults. 

In addition to triptonide, the small molecule disulfiram also demonstrated significant beneficial effects on our multi-phenotype aging profile of PBMCs from old donors exposed to rVSV.

Disulfiram is an approved alcoholism aversion therapy that blocks the reaction converting acetaldehyde, an alcohol metabolism intermediate, into acetate. 55, 56 These effects are complemented by several anti-inflammatory mechanisms, including an ability to inhibit inflammasome-mediated pyroptotic cell death. 57, 58 Within our aging profiles, disulfiram significantly reduced several proinflammatory cytokines when compared with old vehicle-only controls from old donors. After normalizing the cytokine levels by the total number of cells, MCP1, IL-1β, IL-6, and TNFα were all significantly decreased for at least two concentrations ( Figure 14 ). The reduction in MCP1 was concentration-dependent ( Figure 14) . Similarly, IL-1β was significantly reduced at 0.1 µM, and this effect was maintained for all subsequently higher concentrations of disulfiram ( Figure 14) .

These anti-inflammatory effects rank disulfiram highly in our COVID-19 cytokine score, showing its potential to improve older viral immune responses to infection.

Disulfiram treatment induced similar compositional changes as the other top compounds. On average, the three highest concentrations of disulfiram increased dead cells and infected monocytes by 12% and decreased T cells by 13%. This trend is in the opposite direction of a young phenotype but consistent with disulfiram's ability to induce apoptosis through redox-related processes. 59 The viral features within our multi-phenotype aging profile were also significantly improved after treatment with disulfiram. Vehicle-and disulfiram-treated PBMCs were exposed to 10× MOI rVSV, and at several concentrations, disulfiram caused the cells to appear like they were responding to a lower viral load. In particular, compared with old, vehicle-only controls, Figure 15a ). Importantly, these phenotypes restored by disulfiram are both found at higher viral loads and are significantly different between young and old viral immune responses.

T cells from rVSV-exposed PBMCs treated with disulfiram also exhibited a significantly Figure 15b ). Similar trends were seen with the "on-age" and "offage" scores. Compared with vehicle-only controls, the "on-age" T-cell score significantly shifted Figure 15c ). The "off-age" score also had a significant effect in the samples exposed to concentrations of disulfiram above Figure 15c ).

Altogether, these beneficial shifts in age-related phenotypes demonstrate the strong potential of disulfiram for treating dysfunction in the old viral immune response. 

Aging is the most critical risk factor for morbidity, mortality, and progression of severe infectious disease. 21,60-62 While the COVID-19 pandemic has drawn attention to the vulnerability of older adults, the heightened risk faced by these patients is not limited to this pandemic. 63 Older adults face an increased risk of death from seasonal influenza and sepsis, and the ever-increasing aging population across the globe presents a challenge to protect this at-risk population. To help address this challenge, we developed a system that produces a multi-phenotype aging profile by combining high-throughput lab automation and machine learning algorithms. This is a novel way to study immune cell responses to viral infection that allows for the rapid characterization of these responses in multiple dimensions (eg, cell composition, cell-to-cell interactions, cellular features, cytokine production, and hidden complexities) and enables us to screen for compounds that can have an impact on how immune systems of older patients respond to viral infections.

There were some limitations to our approach. First, we could not differentiate between dead cells and those infected with rVSV given their morphological similarities. To resolve this, we grouped these cells into one class called "dead cells and infected monocytes." Although we were still able to identify multiple significant effects, future work will focus on building more granular cell types that differentiate between infected cells and cells that died from apoptosis, Another limitation is the potential for confounders to unknowingly bias our machine learning models. A significant amount of our top compounds saw a compositional shift which was in the opposite direction of a young phenotype. For many of these compounds, dead cells and infected monocytes significantly increased while T cells significantly decreased. We hypothesized that this increase was mostly due to cells undergoing compound induced apoptosis then being labeled as dead cells and infected monocytes. While our T-cell machine learning aging models were not trained or evaluated on cells from this class, the ability of these models to capture subtle trends from high dimensional data leaves them vulnerable to bias by technical effects. This requires special tooling and attention to ensure known confounders, such as plate, donor, and batch effects, are not biasing the models and that proper awareness is brought to new potential confounders. The many features in our multi-phenotype aging profile allow us to broadly profile compound effects leading to the identification of some features, such as apoptosis, which could be potential confounders. We do not know if an increase in apoptosis impacted the model or whether the impact would be a negative artifact rather than a real beneficial aging phenotype. One hypothesis is that the increase in apoptosis caused differential cytokine release which induced an early stimulation of the innate immune system so that it was more prepared to mount a better viral response and resolution. 64 More work is needed to test these results for specific mechanistic hypotheses and in vivo efficacy.

Despite these limitations, we conducted a successful screen on 3428 bioactive

compounds. Through this screen, we identified leading candidates for treatment of severe infection by assigning scores to many features within our aging profile based on their relevance to immune aging and COVID-19 cytokine pathology. Within the top 10 ranked compounds from our screen, disulfiram, a long-used therapy for patients with alcohol dependency, and the diterpenoid triptonide, obtained from a woody vine commonly found in Eastern and Southern

China, were two interesting compounds that robustly rejuvenated aspects of the viral immune response in old rVSV-exposed PBMCs. The reported anti-inflammatory mechanisms of triptonide in combination with findings from our study indicate that this compound shows promise for dysfunctional aging immune responses. However, additional research is needed to ensure its safety before use in humans for aging-related indications. Tripterygium compounds have demonstrated liver and genital toxicity in animal models. 65 While the toxicity of triptonide may make it unsuitable for the clinic, further understanding of its mechanisms could inform a search for similar compounds that would be more suitable for human therapeutic use. Further studies are in progress to better understand the mechanisms by which triptonide modulates immune cells in the hope to identify specific compounds that target similar age-related mechanisms.

In contrast to triptonide, disulfiram has over 60 years of clinical use in humans. It has been used as an alcohol aversion therapy through its inhibition of aldehyde dehydrogenase. 56, 66 Over these 6 decades, disulfiram has an established and well-understood safety profile with no significant evidence of clinical toxicity in the absence of ethanol. 66 Our multi-phenotype aging profile of disulfiram demonstrated significant anti-inflammatory and rejuvenating effects. It reduced the proinflammatory cytokines MCP1, IL-1β, IL-6, and TNFα, and rejuvenated several other features in our aging profile such as T-cell age score, viral load response, and mitochondrial morphology. These mechanisms could be related to the recently discovered ability of disulfiram to block the final step in inflammasome-mediated pyroptosis and cytokine release. 67 Disulfiram blocks gasdermin D (GSDMD), a pore-forming protein that plays a critical role in the terminal step of inflammasome-mediated pathways by facilitating the release of proinflammatory cytokines. [68] [69] [70] Cleaved GSDMD similarly creates pores on the cell membrane of neutrophils to allow for the release of neutrophil extracellular traps (NETs), a process known as NETosis. 71, 72 Both of these mechanisms make disulfiram an attractive treatment for hyperinflammatory infections such as COVID-19, Acute Respiratory Disease Syndrome (ARDS), and sepsis. 58, 73, 74 Disulfiram has already been shown to protect mice from lethal lipopolysaccharide-induced septic shock. 67 There is mounting evidence that targeting NETosis and inflammasome activation could be a potential treatment for severe COVID-19. Markers in the sera of patients with severe COVID-19 indicate the presence of NETs, and an autopsy of a lung specimen from a patient with COVID-19 showed extensive neutrophil infiltration. 75, 76 In addition, proinflammatory cytokines in patients with severe COVID-19 were significantly higher than in moderate cases. 77 This includes elevated levels of IL-1β which is typically a result of inflammasome activation. 78 CoV-1, which may disrupt the replication and IFN suppression mechanisms of these viruses. [79] [80] [81] Disulfiram was shown to inhibit the main protease of SARS-CoV-2 by demonstrating high inhibitory concentration values but the direct inhibition of SARS-CoV-2 viral replication by

disulfiram has yet to be demonstrated. 79 Recently, a clinical trial was initiated at the University of San Francisco to test how three days of disulfiram treatment changes adverse events, disease severity, SARS-COV-2 viral load, and inflammatory cytokine production in patients with early COVID-19 (ClinicalTrials.gov Identifier: NCT04485130).

We plan to test the most promising and safe compounds discovered by our multiphenotype aging profile of viral infection for their ability to impact clinical care in sepsis, influenza, ARDS, and COVID-19. At the same time, we hope to further develop our system by building the capability to generate aging profiles of other age-related diseases such as autoimmune conditions, cancer, and vaccine responses. If successful, our current and future work could be used to improve the health of many older adults.

All study protocols were evaluated and approved by the WCG Institutional Review Board, and all donors provided informed consent. Eighty-nine healthy donors underwent screening at the Discovery Life Sciences Donor Clinic (Huntsville, AL). No donors were excluded based on comorbidities, alcohol use, smoking, or current medication.

Whole blood was collected into EDTA tubes and then diluted with an equal volume of phosphate-buffered saline (PBS) + 2% fetal bovine serum (FBS) and layered over Ficoll using 

Each 384-well plate assay layout was designed to account for potential bias in experimental conditions due to plate location; conditions were plated such that every condition was included in every row or column. To generate data on isolated cell types, T cells, monocytes, macrophages, dendritic cells, B cells, NK cells, and PBMCs were evaluated as separate conditions. To evaluate effects due to viral load, we plated unexposed cells and cells exposed at 0.1×, 1×, and 10× MOI as different conditions (described below). To analyze differences between age groups, we plated 48 individual donors per assay plate, with 8 replicates per donor. Two plates were required to evaluate all donors. Donors were stratified by age group (≤35 vs ≥60 years). The age groups were considered as separate conditions such that the interspersed plating resulted in a balance between young and old donors across all plate locations.

Cells were thawed from the gas phase of liquid nitrogen by immersing the cryotube in a 37 °C water bath until a thin layer of ice was present. The cells were transferred to 10 mL of prewarmed culture medium (refer to Supplemental Methods for all media recipes) and

immediately centrifuged for 10 min at 138×g. The cell pellet was gently washed with 10 mL of culture medium, resuspended in cold plating medium at a density of 4×10 5 cells/mL, and 25 µL were added to the assay plate (filled with 15-µL prewarmed plating medium) using a 96-well dispensing head to achieve a final cell density of 10,000 cells per well. Cells were kept cool during the entire plating process. Assay plates were centrifuged for 1 min at 138×g and incubated for 30 min at 37 °C with 5% CO 2 to allow for cell adhesion to the plastic surface of the assay plate.

After 30-min cell adhesion, 10 µL of 5× trigger medium (including rVSV-ΔG-mCherry, DMSO, test compound, and FBS) was added to the assay plate using a 384-well pipetting head to achieve a final concentration of rVSV-ΔG-mCherry at different MOIs, 0.1% DMSO, 10% FBS, and 0.33 µM or 5.3 µM compound concentration. The assay plate was centrifuged for 1 min at 138×g and incubated for 24 h at 37 °C with 5% CO 2 .

After 24 h, cell supernatant was removed and evaluated to determine cytokine levels using the Block buffer (50 µL/well; Supplemental Methods) was added to each well, and the assay plate was incubated for 1 h at room temperature.

For the cell painting assay, Cell Painting Palette 1 or 2 (Supplemental Methods) was set up in block buffer. The palette was similar to the cell painting assay described by Bray et al. 28 Block buffer was removed from the cells and 25 µL/well of the CPMM palette was added. After an overnight (~18 h) incubation at 8 °C, the assay plate was washed 4× with 50-µL/well PBS/ Antibiotic Antimycotic and subjected to image acquisition. The plate was stored under refrigerated conditions for further use.

Image acquisition was conducted on an Operetta CLS TM high-content analysis system (PerkinElmer, Waltham, MA) with binning 1, using a 40× water immersion objective in wide-field mode with Z-stack of 4 planes and distance steps of 1.0 µm. Four to 16 fields of view and six channels were captured.

Cytokine levels in the cellular supernatant were evaluated using the FirePlex- Figure 1 ).

Initial data preprocessing was structured as a directed acyclic graph, with image transformation at each step in the graph execution; transformed images were passed along as inputs to the next step. For the first step, the input images, acquired as multiple confocal Z-layers per field of view, were projected onto a single field of view by taking the maximum value of each pixel across all Z-layers for a given field. The projected images were saved both at the original source resolution and at a down-sampled half-resolution. For each channel, an illumination correction function was computed using the half-resolution images and then applied to the full-resolution and half-resolution images to generate illumination-corrected versions of the images. 82 For each channel, the mean and standard deviation of the pixel value distributions were computed over the projected images. The full-resolution and half-resolution images were then z-score normalized using these precomputed means and standard deviations of the plate for each channel to generate pre-normalized versions of the images.

Using the normalized images, the data processing method continued to feature generation. A 

Every image, and each cell within the image, was run through a semi-automated quality control process. The image metrics computed by CellProfiler were visualized and analyzed per plate to identify technical artifacts that arose from issues such as location, operator error, and poor sample quality (Supplemental Figure S4 ). Each plate was reviewed by an operator to check for quality across cell count, intensity, blur, focus, and outlier metrics. Plates that had strong negative or positive correlations between quality control metrics and plate location were marked for further review. These metrics were used to flag problematic issues, refine assay protocols, and inform interpretation of results, but at this point were not used to exclude data in an automatic fashion.

The 

We used the embeddings and features created to train classifiers that predicted whether an immune response was indicative of a young or old immune system. Data-generation methods were optimized to reduce bias by using balanced plate layouts, robust quality control processes, automated liquid handling, and a large, diverse number of samples. Embeddings were generated for each single cell and then aggregated by taking the median of all cells within a field of view. We trained the classification models on these field-level embeddings using XGBoost, an implementation of gradient-boosted decision trees. 72 As described in the Results, we performed stratified four-fold cross-validation, ensuring that each fold had equal proportions of young and old donors, and that each donor appeared in only one-fold.

Supplemental Figure S1 . Representative flow plots for cell composition discrimination. Events were first gated on FSC/SSC profile to distinguish dead cells and debris from live myeloid and live lymphoid cells. Myeloid morphology cells were then stratified by CD14 and CD11c expression to distinguish monocytes (CD14 + ). Lymphoid cells were first stratified by CD3 and CD19 expression to distinguish T cells (CD3 + ) and B cells (CD19 + ). Double-negative cells from the lymphoid morphology fraction were used to distinguish NK cells (CD3 -CD19 -CD56 + ).

Supplemental Figure S2 . Cytokine levels from the supernatant of PBMCs obtained from 89 old and young donors and exposed to rVSV-ΔG-mCherry at 1× MOI. IFN, interferon; IL, interleukin; MCP, monocyte chemoattractant protein 1; MOI, multiplicity of infection; PBMC, peripheral blood mononuclear cell; rVSV-ΔG-mCherry, recombinant vesicular stomatitis virus expressing a red fluorescent construct; TNF, tumor necrosis factor.

Supplemental Figure S3 . (a) Spatial Kolmogorov-Smirnov plot to identify inconsistencies with a homogeneous Poisson distribution. (b) Significant departures from the expected distribution (homogeneous) would indicate a heterogeneous distribution that is not independent of the spatial landscape. CSR, complete spatial randomness. • All centrifugation steps were performed at room temperature (22 °C) 

Medical writing and/or editorial assistance was provided by Rebecca Brady, PhD, of The Lockwood Group

Data interpretation

Software creation: Ben Komalo, Lauren Nicolaisen, Charlie Marsh, Jarred Heinrich

Drafting and editing the manuscript: Brandon

Christian Elabd, and Ben Kamens are coinventors on patents filed for this work and pending to Spring Discovery

Ralf Schwandner report personal fees from Spring Discovery. Anis Larbi is the co-founder of

Ralf Schwandner is the founder and CEO of Assay.Works GmBH. Delia Bucher, Sabine Tyrra, Tamas Fulop, and Diane Heiser have nothing to disclose

Aging, frailty and age-related diseases

Immunosenescence and inflamm-aging as two sides of the same coin: friends or foes? Front Immunol

Aging is associated with chronic innate immune activation and dysregulation of monocyte phenotype and function

Inflammation and trajectory of renal function in community-dwelling older adults

Age-related decrease in TCR repertoire diversity measured with deep and normalized sequence profiling

Age-related deregulation of naive T cell homeostasis in elderly humans

Ageing and infection

Epigenetic drift in the aging genome: a ten-year follow-up in an elderly twin cohort

Age-related changes in memory and effector T cells responding to influenza A/H3N2 and pandemic A/H1N1 strains in humans

Successful and maladaptive T cell aging

Longterm maintenance of human naive T cells through in situ homeostasis in lymphoid tissue sites

Immunobiography and the heterogeneity of immune responses in the elderly: a focus on inflammaging and trained immunity

Age specific mortality rates COVID-19

Novel Coronavirus Pneumonia Emergency Response Epidemiology Team. The epidemiological characteristics of an outbreak of 2019 novel coronavirus diseases (COVID-19) in China

Features of 16,749 hospitalised UK patients with COVID-19 using the ISARIC WHO Clinical Characterisation Protocol

Aging immunity may exacerbate COVID-19

Centers for Disease Control and Prevention. Estimates of deaths associated with seasonal influenza-United States

Centers for Disease Control and Prevention. Provisional death counts for coronavirus disease (COVID-19): daily updates of totals by week and state

Incidence and trends of sepsis in US hospitals using clinical vs claims data

Outcome of very old patients admitted to the ICU for sepsis: a systematic review

TLR7 and TLR8 activate distinct pathways in monocytes during RNA virus infection

Recognition of single-stranded RNA viruses by Toll-like receptor 7

Single-cell morphology encodes metastatic potential

Functionally-relevant morphological profiling: a tool to assess cellular heterogeneity

Cell form and function: interpreting and controlling the shape of adherent cells

Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes

Variation in the human immune system is largely driven by non-heritable influences

Human immune system variation

Immunology of COVID-19: current state of the science

The many faces of the anti-COVID immune response

The trinity of COVID-19: immunity, inflammation and intervention

Pathological inflammation in patients with COVID-19: a key role for monocytes and macrophages

Duration of antigen receptor signaling determines T-cell tolerance or activation

An activating NLRC4 inflammasome mutation causes autoinflammation with recurrent macrophage activation syndrome

Global survey of the immunomodulatory potential of common drugs

Persistent lymphopenia after diagnosis of sepsis predicts mortality

Lymphopenia and risk of infection and infection-related death in 98,344 individuals from a prospective Danish population-based study

Increased TNF-α-induced apoptosis in lymphocytes from aged humans: changes in TNF-α receptor expression and activation of caspases

Role of NF-kappaB signaling pathway in increased tumor necrosis factor-alpha-induced apoptosis of lymphocytes in aged humans

Reduction and functional exhaustion of T cells in patients with coronavirus disease 2019 (COVID-19)

Viral and host factors related to the clinical outcome of COVID-19

Mitochondrial heterogeneity in stem cells

Mitochondrial dysfunction and the aging immune system

T cells with dysfunctional mitochondria induce multimorbidity and premature senescence

Immunosuppressive and anti-inflammatory mechanisms of triptolide, the principal active diterpenoid from the chinese medicinal herb Tripterygium wilfordii Hook. f

Advances in pharmacological activities of terpenoids

Triptolide inhibits human immunodeficiency virus type 1 replication by promoting proteasomal degradation of Tat protein

Immunosuppressant PG490 (triptolide) inhibits T-cell interleukin-2 expression at the level of purine-box/nuclear factor of activated T-cells and NF-kB transcriptional activation

Modulation of IL-37 expression by triptolide and triptonide in THP-1 cells

Triptolide induces anti-inflammatory cellular responses

Mitochondrial dynamics controls t cell fate through metabolic programming

Triptolide induces apoptotic death of T lymphocyte

From disulfiram to antabuse: the invention of a drug

Disulfiram suppresses NLRP3 inflammasome activation to treat peritoneal and gouty inflammation

Gasdermin D (GSDMD) as a new target for the treatment of infection

Disulfiram induces apoptosis in human melanoma cells: a redox-related process

Infectious disease hospitalizations in the United States

Severe sepsis and septic shock in the elderly: an overview

Mortality attributable to 9 common infections: significant effect of influenza A, respiratory syncytial virus, influenza B, norovirus, and parainfluenza in elderly persons

Covid-19 and immunity in aging populations -a new research agenda

Apoptotic cells and innate immune stimuli combine to regulate macrophage cytokine secretion

The targeting of 14-succinate triptolide-lysozyme conjugate to proximal renal tubular epithelial cells

The safety of disulfiram for the treatment of alcohol and cocaine dependence in randomized clinical trials: guidance for clinical practice

FDA-approved disulfiram inhibits pyroptosis by blocking gasdermin D pore formation

Gasdermin D is an executor of pyroptosis and required for interleukin-1β secretion

Inflammasome-activated gasdermin D causes pyroptosis by forming membrane pores

Cleavage of GSDMD by inflammatory caspases determines pyroptotic cell death

Gasdermin D plays a vital role in the generation of neutrophil extracellular traps

Noncanonical inflammasome signaling elicits gasdermin D-dependent neutrophil extracellular traps

Inflammasomes and pyroptosis as therapeutic targets for COVID-19

Circulating gasdermin-d in critically ill patients

Targeting potential drivers of COVID-19: neutrophil extracellular traps

Neutrophil extracellular traps in COVID-19

Clinical and immunological features of severe and moderate coronavirus disease 2019

Clinical features of patients infected with 2019 novel coronavirus in Wuhan

Structure of M pro from SARS-CoV-2 and discovery of its inhibitors

Disulfiram can inhibit MERS and SARS coronavirus papain-like proteases via different modes

The SARS coronavirus papain like protease can inhibit IRF3 at a post activation step that requires deubiquitination activity

Pipeline for illumination correction of images for high-throughput microscopy

CellProfiler 3.0: next-generation image processing for biology

Supplemental Figure S4. Power log-log metrics computed on the Hoechst channel for (a) a plate that passed quality control and (b)

Replicate experiment measuring cytokine levels from the supernatant of rVSV-ΔG-mCherry-exposed (at 10× MOI for 24 h) PBMCs obtained from 89 donors from young and old adult populations

Acquisition of data was provided by Assay.Works GmbH (Regensburg, Germany), funded by Spring Discovery. Sample collection and processing was provided by Discovery Life Sciences