Drug-induced phospholipidosis confounds drug repurposing for SARS-CoV-2

Screening for drugs that don’t work In the battle against COVID-19, drugs discovered in repurposing screens are of particular interest because these could be rapidly implemented as treatments. However, Tummino et al. deliver a cautionary tale, finding that many leads from such screens have an antiviral effect in cells through phospholipidosis, a phospholipid storage disorder that can be induced by cationic amphiphilic drugs (see the Perspective by Edwards and Hartung). There is a strong correlation between drug-induced phospholipidosis and inhibition of severe acute respiratory syndrome coronavirus 2 replication in cells. Unfortunately, drugs that have an antiviral effect in cells through phospholipidosis are unlikely to be effective in vivo. Screening out such drugs may allow a focus on drugs with better clinical potential. —VV

The outbreak of COVID-19 has inspired multiple drug repurposing screens to find antiviral therapeutics that can be rapidly brought to the clinic (1). To date, over 1,974 drugs and investigational drugs have reported to have in vitro activity against SARS-CoV-2 (1) (Fig. 1). Since almost all of these act against human targets, and might be unlikely to be viable against a novel virus (2), the question of mechanism of action arises.
Our interest in this question was motivated by the discovery that human sigma receptors were candidates for modulating SARS-CoV-2 infection (3), and that drugs and reagents like chloroquine, haloperidol, clemastine, and PB28-all with nanomolar affinity against one or both sigma receptors-had cellular antiviral IC50 values in the 300 nM to 5 μM range. Subsequently, we investigated over 50 different molecules with a wide range of affinities at these receptors. While this found molecules with relatively potent antiviral activity, there was little correlation between receptor potency and antiviral efficacy in cells (figs. S1 to S3 and table S1). Whereas drugs like amiodarone, sertraline, and tamoxifen had mid-to high-nM antiviral IC50s, other potent sigma ligands, such as melperone and DTG, were without measurable antiviral activity. Intriguingly, the antiviral sigma drugs were all cationic at physiological pH and relatively hydrophobic, while those that were inactive against the virus were often smaller and more polar. This cationic-amphiphilic character was shared by many of the hits emerging from other phenotypic screens ( Fig. 1 and fig. S4), suggesting it was this physico-chemical property that might explain cellular antiviral activity, instead of a specific on-target activity (4).
If the cationic-amphiphilic nature of these molecules led to antiviral activity in vitro, rather than their target-based Drug-induced phospholipidosis confounds drug repurposing for SARS-CoV-2 activities, one would expect this physical property to reflect a shared cellular mechanism. Indeed, cationic amphiphilic drugs (CADs) can provoke phospholipidosis in cells and organs (5). This side effect is characterized by the formation of vesicle-like structures and "foamy" or "whorled" membranes (5,6), and is thought to arise by CAD disruption of lipid homeostasis. CADs accumulate in intracellular compartments such as endosomes and lysosomes where they can directly or indirectly inhibit lipid processing (5). Modulation of these same lipid processing pathways is critical for viral replication (7), and inhibiting phospholipid production has previously been associated with inhibition of coronavirus replication (8). CADs have in vitro activity against multiple viruses including Severe Acute Respiratory Syndrome, Middle East Respiratory Syndrome, Ebola, Zika, Dengue, and filoviruses (9), though CAD-induction of phospholipidosis has only been proposed as an antiviral mechanism for Marburg virus (10). Finally, among the most potent known phospholipidosis inducers are amiodarone (11) and chloroquine (12, 13), which are also potent inhibitors of SARS-CoV-2 replication in vitro (14)(15)(16), while drugs from SARS-CoV-2 phenotypic screens, such as chlorpromazine (17) and tamoxifen (16), are also known to induce phospholipidosis (18). As an effect that rarely occurs at concentrations lower than 100 nM, that does not appear to translate from in vitro to in vivo antiviral activity, and that can result in dose-limiting toxicity (19), phospholipidosis may be a confound to true antiviral drug discovery.
Here, we investigate the association between phospholipidosis and antiviral activity against SARS-CoV-2 in cell culture. This apparently general mechanism may be responsible for many of the drug repurposing hits for SARS-CoV-2, and an extraordinary amount of effort and resources lavished on drug discovery against this disease. We explore the prevalence of this confound in SARS-CoV-2 repurposing studies, how phospholipidosis correlates with inhibition of viral infection, and how to eliminate such hits rapidly so as to focus on drugs with genuine potential against COVID-19, and against new pandemics yet to arise.

Correlation of phospholipidosis and antiviral activity
To investigate the role of phospholipidosis in antiviral activity in vitro, we tested 19 drugs for their induction of this effect in A549 cells using the well-established NBD-PE staining assay (20). Here, the vesicular lipidic bodies characteristic of the effect may be quantified by high content imaging ( Fig.  2A). Three classes of drugs and reagents were initially investigated: A. Sigma-binding antiviral CADs we had discovered, like amiodarone, sertraline, chlorpromazine, and clemastine (nine total); these molecules are predicted or known to induce phospholipidosis; B. Analogs of these CADs that no longer bound sigma receptors, but were still antiviral (four total); these molecules are predicted to induce phospholipidosis despite their lack of sigma binding; and C. Sigma-binding, non-antiviral drugs, like melperone and DTG, that are more polar than classic CADs (two total); these molecules are predicted not to induce phospholipidosis. Of the nine sigmabinding CADs that were antiviral (class A), six of which were also reported in literature phenotypic screens as inhibitors of COVID-19, eight induced phospholipidosis, consistent with the hypothesis (Fig. 2, A and B, and figs. S5 and S6). The only non-phospholipidosis inducing antiviral from this set was elacridar, a promiscuous P-glycoprotein inhibitor; this drug may therefore be active via another mechanism. Intriguingly, analogs of the potent sigma ligand PB28 that had lost their sigma-binding activity but remained CADs (ZZY-10-051 and ZZY-10-061; Fig. 2, B to F, and figs. S5 to S8), did induce phospholipidosis, as did the antipsychotic olanzapine and the antihistamine diphenhydramine, which are weak sigma receptor ligands but are structurally related to potent sigma receptor ligands like chlorpromazine and clemastine. Finally, melperone and DTG, which are potent cationic sigma receptor ligands but are not antiviral, did not induce phospholipidosis (Fig. 2, A and B, and figs. S5 and S6; class C). These results do not prove phospholipidosis as the antiviral mechanism but are consistent with the phospholipidosis hypothesis.
If phospholipidosis is responsible for antiviral activity, then other molecules known to induce phospholipidosis should be antiviral. We tested three CADs for antiviral activity, including ebastine, ellipticine, and Bix01294, all of which are reported to induce phospholipidosis (21) (Bix01294 and ebastine have also been reported as drug repurposing hits against SARS-CoV-2 (22)). We further tested azithromycin, also reported to induce phospholipidosis (23), but having different physical properties from typical CADs. We first confirmed phospholipidosis-inducing activity for these molecules ( Fig. 2B and figs. S5 and S6). All four molecules were next shown to be antiviral, as ever in this work using live virus assays (e.g., SARS-CoV-2 strain Be-taCoV/France/IDF0372/2020; Methods), with IC50 values in the 400 nM to 3 μM range, overlapping with the activities of other CADs we and others have identified for SARS-CoV-2 (22) (fig. S6). This too was consistent with the antiviral phospholipidosis hypothesis.
For phospholipidosis to explain antiviral activity, we might expect a correlation between concentration-response curves for phospholipidosis and for antiviral activity. We compared concentrations that induce phospholipidosis to those that inhibit SARS-CoV-2 for each drug individually. Most correlations were high-not only did antiviral activity occur in the same concentration ranges as phospholipidosis, but R 2 values, ranging from 0.51 to 0.94, supported a quantitative relationship between the two effects ( Fig. 3A). We then

Concurrent measurement of viral infection and drug induced phospholipidosis
In the previous experiments, drug-induced phospholipidosis and drug antiviral activity were measured separately. To measure the two effects in the same cells at the same time, we dosed cells with either 1 or 10 μM of five characteristic CADs (amiodarone, sertraline, PB28, hydroxychloroquine (HCQ), and Bix01294), followed by a mock or SARS-CoV-2 infection, and quantified phospholipidosis and the accumulation of viral spike protein ( Fig. 4A and fig. S9). Compared to DMSO, drug treatments led to substantial increases in NBD-PE aggregates, indicating increased phospholipidosis ( fig.  S9). At 1 μM drug concentrations, SARS-CoV-2 spike protein was readily stained, and one could visualize both spike protein and phospholipidosis in the same cells (yellow puncta), suggesting at this low concentration of drug-often close to the antiviral IC50 value-both phospholipidosis and viral infection co-occur, though even here viral staining was reduced relative to the DMSO treated controls. As drug concentration rose to 10 μM, viral spike protein staining dropped while staining for phospholipidosis increased (fig. S9); there was nearly complete loss of spike protein signal with a concomitant increase in phospholipidosis (Fig. 4A) for all treatments.
In seven-point concentration-response curves for amiodarone, sertraline, and PB28, viral staining monotonically decreased as phospholipidosis increased (Fig. 4, B and C).
We then further filtered for drugs that topologically resembled known phospholipidosis inducers (18,21) using an ECFP4-based Tanimoto coefficient ((Tc) ≥ 0.4) (table S2). Sixty percent of the 310 drugs passed the cLogP and pKa threshold; 34% also resembled a known phospholipidosis inducer ( Fig. 1 and figs. S4 and S10). Although the two physical property filters do not capture atypical phospholipidosis inducers such as azithromycin, they do capture 16 of the other 18 CADs we had already tested (missing only the medium phospholipidosis inducers olanzapine and ellipticine); intriguingly, nine of these, including amiodarone, sertraline, chlorpromazine, Bix01294, clemastine, and benztropine also appeared in at least one of the 12 other repurposing studies. To probe the reliability of this association, we tested another five drugs that passed our filters, and had been reported as antiviral against SARS-CoV-2, for their induction of phospholipidosis. Not only were all five active in the NBD-PE assay, but we confirmed SARS-CoV-2 antiviral activity for these drugs ( fig. S10). Additionally, these molecules fit into the sigmoidal model relating phospholipidosis to viral load (salmon points overlaid with sigmoidal model; Fig. 3B). Finally, we note a preliminary identification of 30 CADs, 19 of which overlap with the literature-derived SARS-CoV-2 list, active against other viruses including Middle East Respiratory Syndrome and Severe Acute Respiratory Syndrome (33), Ebola (34)(35)(36), Marburg (36,37), Hepatitis C (38), and Dengue (39) (table S3). It may be that most drugs repurposed against many viruses are CADs whose antiviral activities can be attributed to phospholipidosis.

Animal efficacy for repurposed drugs
Though phospholipidosis is considered a drug-induced side effect, it remains possible that it can be leveraged for antiviral efficacy. Accordingly, we tested four of the repurposed, phospholipidosis-inducing drugs most potent against SARS-CoV-2 in vitro, amiodarone, sertraline, PB28 and tamoxifen (5,18), for efficacy in a murine model of COVID-19 (40). In the same model, we also tested elacridar, which does not induce phospholipidosis (Fig. 2B) and remdesivir, which is unlikely to induce phospholipidosis at concentrations relevant to its antiviral activity. All molecules had relatively long half-lives, especially in the lung where tissue Cmax values often exceeded 10 μM after a 10 mg/kg dose, or 10 to 1000 times higher than their in vitro antiviral IC50, suggesting that exposure would be high enough for efficacy (tables S4 to S8). Guided by the pharmacokinetics of each drug, mice were dosed either once (amiodarone and elacridar) or twice per day (remdesivir, PB28, tamoxifen, and sertraline), for three days. Two hours following the first dose, mice were intranasally infected with 1 × 10 4 PFU of SARS-CoV-2 and lung viral titers were measured after a three-day infection period. Notwithstanding their high lung exposure, the four phospholipidosis-inducing drugs had little effect on viral propagation in the mice. Conversely, remdesivir reduced viral load by two to three orders of magnitude. While the cationic non-phospholipidosis drug elacridar had a modest antiviral effect, it did not rise to statistical significance (Fig. 5) and mice given elacridar doses higher than 3 mg/kg exhibited toxicities that limited further study.
Because phospholipidosis is typically an in vivo side effect that appears after chronic dosing, we then pre-treated mice with five-fold higher concentrations (50 mg/kg) of amiodarone over twelve days prior to a 3-day infection period. Even here, no diminution of viral titer was observed in mouse lungs after infection, and amiodarone offered no protection from infection-induced weight loss or from pulmonary inflammation and epithelial necrosis, as measured by histopathology ( Fig. 5 and fig. S11). We note that foamy vacuolation and whorled vacuoles that are the hallmarks of phospholipidosis were not seen in lung and spleen by light or transmission electron microscopy. It is thus possible that this treatment was not long enough to induce protective phospholipidosis. Still, the in vitro activities of the phospholipidosisinducing drugs did not readily translate in vivo, and drugs whose antiviral activity arises due to phospholipidosis seem non-viable for clinical progression.

Discussion
The emergence of COVID-19 has motivated intense effort to repurpose drugs as SARS-CoV-2 antivirals. An extraordinary number of diverse, apparently unrelated hits have emerged (1). A key observation from this work is that many, perhaps most of these are active in antiviral assays via induction of phospholipidosis ( Fig. 1 and figs. S4 and S10). This disrupts lysosomal lipid catabolism and trafficking, which may in turn disrupt the double membrane vesicles that the virus creates and on which it depends for propagation. Quantitatively, there is a close in vitro correlation between drug-induced phospholipidosis and antiviral activity, both drug-by-drug and over the set of drugs tested here (Fig. 3). The effect is predictive: molecules that induce phospholipidosis are antiviral over the same concentration range, irrespective of whether they are cationic amphiphilic drugs (CADs) or not (e.g., azithromycin), while molecules that are related by target activity to the CADs, but are more polar and do not induce phospholipidosis (e.g., melperone and DTG), are not antiviral. Unfortunately, CAD induction of phospholipidosis, at least at the potencies observed here, does not translate in vivo (Fig. 5). More encouragingly, this study illuminates a method to rapidly identify confounds in cellular antiviral screens, allowing one to eliminate them from further study and to focus on those molecules with true potential.
Although the molecular mechanisms for the antiviral effects of phospholipidosis remain unclear, certain associations may be tentatively advanced. SARS-CoV-2, like many viruses, subverts the cell to produce double membrane vesicles in which it replicates (41)(42)(43). Disruption of lipid homeostasis by the induction of phospholipidosis may disrupt these vesicles, reducing viral replication. The disruption of lysosomal (44) and endosomal (45) compartments and CAD-induced shifts in compartmental pH (46) may further affect viral entry and propagation (47). Accordingly, targeting the endosomal-lysosomal pathway has been suggested as a viable strategy against SARS-CoV-2 infection (48), but developing potent and targeted inhibitors remains challenging.
The cost to the community of investments in what appears to be a confound merits consideration for future pandemics. According to the DrugBank COVID-19 dashboard (49), which draws from U.S. and international clinical trials, putatively antiviral CADs have been promoted into an astonishing 316 Phase I to Phase III clinical trials against COVID-19. While 57% of these study the phospholipidosis-inducing CADs hydroxychloroquine (Fig. 3A, top row) or chloroquine, that still leaves 136 trials across 33 other predicted or known phospholipidosis-inducers. Using conservative estimates (50,51), the expense of the clinical trials component alone, over the last year, for phospholipidosis-inducing CADs may be over $6 billion US dollars (table S9).
Certain caveats merit airing. First, the correlation between antiviral activity and phospholipidosis, as strong as it is, does not illuminate the mechanism by which phospholipidosis is antiviral. Phospholipidosis is itself only partly understood, and there are no good genetic or chemical ways to either inhibit its induction by drugs nor to promote it by target-selective reagents. Second, predicting whether a molecule will induce phospholipidosis remains challenging, and even non-CAD molecules can induce it. Thus, we have chosen conservative criteria to predict phospholipidosis-inducers, which may miss many drugs. Third, phospholipidosis is a confound that only affects drugs repurposed for direct antiviral activity-it is irrelevant for drugs like dexamethasone (52) and fluvoxamine (53) that have been repurposed for immunomodulation in COVID-19, nor is it relevant for CADs whose antiviral activity is well-below the concentration range where phospholipidosis occurs. Fourth, our estimates of the clinical trial costs of phospholipidosis-inducing CADs are obviously rough. Finally, we do not exclude exploiting phospholipidosis therapeutically, though we suspect that would have to proceed through a more target-directed mechanism than that of the CADs studied here.
These caveats should not obscure the central observation of this study. Many drugs repurposed for antiviral activity against SARS-CoV-2 are cationic amphiphiles, and despite their diverse structures and multiple targets, many likely have their antiviral effects via a single shared mechanism: phospholipidosis. Both because of the side effects with which it is associated, and the limited efficacy to which it leadsrarely better than 100 nM in vitro-drugs active due to phospholipidosis are unlikely to translate in vivo (Fig. 5). Many resources will be saved by counter-screening for phospholipidosis in simple cellular assays (20), allowing investigators to focus on drugs with genuine promise as antivirals.  . Images of DMSO and a nonphospholipidosis inducing molecule (melperone) are included for reference. Thresholds for determining phospholipidosis power are shaded in dark grey (low phospholipidosis), light gray (medium phospholipidosis) and no shading (high phospholipidosis). B. Pooled DIPL amounts (mean ± SD) at the highest non-toxic concentration tested for each drug. Results were pooled from three biological and three technical replicates and were normalized to amiodarone (100%) from the control wells in the same experimental batches. C. Structures of PB28 and its analog ZZY-10-051, the latter of which is inactive on the sigma receptors. D. Viral infectivity (red) and viability (black) data for PB28 (square) and ZZY-10-051 (circle) in A549-ACE2 cells. Data shown are mean ± SD from three technical replicates. E. Fractional binding of PB28 and ZZY-10-051 against Sigma-1 (purple; S1R) and Sigma-2 (maroon; S2R) normalized to a buffer control at 1.0 in a radioligand binding experiment. Data shown are mean ± SEM from three technical replicates. PB28 is a strong ligand of both Sigma-1 and Sigma-2 and has high displacement of the radioligands, whereas ZZY-10-051 is unable to displace the radioligands to a high degree at 1 μM. F. Dose response curves for PB28 (blue) and ZZY-10-051 (gold) show that these closely related analogs both induce phospholipidosis.