Landscape of the 403 Cayman antiviral library

The Cayman antiviral library contained 403 compounds categorized into 27 distinct groups based on their antiviral activity (Cayman Chemical Company 2023). The largest category was “Undefined antiviral association,” with 151 compounds or 37.5% of the library. The next largest group targeted human immunodeficiency virus (HIV) with 72 compounds or 17.9% of the library (Fig. 1). Other significant categories targeted Hepatitis B (HBV), Hepatitis C (HCV), influenza, HSV, or were antimalarials, broad-spectrum antivirals, protease inhibitors. Each ranged from 9 to 29 compounds, totaling 139 compounds or 34.5% of the library. The remaining 17 categories contained fewer than 4 compounds each.

Fig. 1

Landscape of Cayman antiviral library (403). Cayman documented the compounds mechanism of action into 29 activity classes, with the largest number of compounds belonging to the top 3: Undefined antiviral association, HIV, and HCV (151, 72, and 29, respectively)

The mechanisms of action for these compounds were divided into 17 classes, with inhibitors, undefined mechanisms, and prodrug inhibitors being the most prevalent, numbering 189, 137, and 28, respectively. Of the 403 compounds, 114 have undergone FDA evaluation, receiving designations such as prescription, discontinued, over the counter, or a combination of prescription and discontinued. Notably, 15% (17 compounds) have been discontinued for various conditions including hypercholesterolemia, hepatitis, Paget’s disease, and viral infections like HSV-1, HCV, HBV, and HIV. Three compounds were available over the counter, including Docosanol, Loratadine, and N-acetylcysteine. The 94 compounds currently prescribed targeted viral infections such as HIV, HCV, HSV, and HBV.

These compounds in the library are believed to operate through 17 distinct mechanisms, including activation, agonism, antagonism, antioxidation, chelation, DNA synthesis induction, and various types of inhibition. Among these were 189 inhibitors, 30 prodrugs, and 137 compounds with undefined mechanisms. Of the 137 with undefined mechanisms, 11 have been FDA-approved for malaria, osteoporosis, HBV infections, seizures, actinic keratosis, sickle cell anemia, intermittent claudication, and lupus, reflecting the library’s potential against a wide spectrum of diseases. HIV and HCV were the most targeted, by 27 and 12 compounds, respectively. The library especially emphasized HIV antiviral therapy, with 21 compounds currently prescribed and 6 discontinued.

A review of the Cayman antiviral library annotations indicated that 114 of 403 compounds (28.3%) were Food & Drug Administration (FDA) approved. Of these 17 compounds (4.2%) were previously approved by the FDA but are now discontinued (6 discontinued were treating HIV infections, 3 HCV), while only 2 compounds (0.5%) are currently available over the counter (OTC). Most (95 compounds, 23.6%) were prescription only. Most prescription compounds treated HIV and HCV (29), malaria (6), HBV (6) and influenza (3).

25% of Cayman library identified as toxic by CHS

Toxicity screening of a 403-compound antiviral library was performed using the Aseda Sciences Systematic Cell Health Screen (CHS), an in vitro assay designed to evaluate cellular stress across multiple parameters (Bieberich et al. 2021). The CHS assay quantitatively measures eight key cellular stress indicators: cell morphology, membrane integrity, reactive oxygen species (ROS) production, glutathione levels, nuclear membrane integrity (assessed via two distinct metrics), cell cycle progression, and mitochondrial membrane potential. These measurements are combined to generate a CHI, which estimates the toxicity potential of each compound, with higher CHI values (approaching one) suggesting increased toxicity. The CHI serves as a pseudo-probability metric that assesses how closely a compound’s phenotypic pattern matches those found in a training library of known toxic compounds with various mechanisms of action.

The initial screening, conducted at ten different concentrations, indicated that 74.2% (299) of the compounds were low-risk, based on a CHI threshold of 0.5 (Fig. 2). Further analysis identified 104 compounds (approximately 25% of the library) as high-risk under this criterion. Applying a more stringent CHI threshold of > 0.75 narrowed the selection to 55 compounds. Saikosaponin A exhibited the highest toxicity potential (CHI = 0.958), followed closely by NH125 and gallotannin (CHI values of 0.95 and 0.935, respectively), all of which are plant-derived compounds used in the treatment of inflammatory conditions and cancer.

Fig. 2

CHI breakdown for Cell Health Screen (CHS). The CHI is an indicator of toxicity for compounds. As CHI approaches 1, this indicates high toxicity. A CHI of > 0.5 has a higher probability of toxicity (104), and the most active are those with > 0.75 (55) to the right of the dotted line

Further categorization of the activity of these 104 high-risk compounds was distributed across 27 distinct activity bins, with their mechanisms of action falling into six primary targets: agonists, antagonists, antioxidants, inhibitors, ligands, and prodrug inhibitors, with some remaining undefined. Most of these compounds were characterized as inhibitors (65) or had undefined targets (31), underscoring the prevalence of inhibition as a key mechanism of antiviral activity. Of the 189 total inhibitors in the library, the CHS assay identified 34% as toxic. Inhibition as an antiviral mechanism was not measured in this study. Still, an indirect association between inhibitor efficacy and developmental toxicity to zebrafish was plausible and may have been the driver of toxicity in this group.

Developmental zebrafish identifies hazard associated with an additional 20% of Cayman library

Of the 104 compounds that exhibited a CHI exceeding 0.5, ZBEscreen identified 60 (57.7%) as toxic, manifesting either adverse morphologic effects (DEVTOX) or aberrant embryonic and larval photomotor responses (EPR and LPR, respectively) (Fig. 3a; Table S2). The other 44 compounds were only a hit in CHS, but not in zebrafish. Of these 44, 10 were FDA and European Medical Association (EMA) approved — nine currently authorized for use by both, and one withdrawn by EMA (Doxycycline hyclate). These drugs were approved primarily for infections (HIV, HBV, HCV and bacterial) except for Dabigatran etexilate used for Venous thromboembolism. For five of these prescribed compounds, severe liver injury has been reported when taken in combination with other drugs, demonstrating that CHS correlates with in vivo toxicity. The lack of hit concordance between CHS screen and ZBEscreen can be attributed to differences in metabolism and compound solubility.

Fig. 3

Bioactivity profiles. Comparison of the active calls in (A) between CHI > 0.5 and in zebrafish, and (b) breakdown of the zebrafish hits across the three assays

If the CHI cutoff was increased to 0.75, the zebrafish assay identified 34 compounds as toxic. It was observed that 49 compounds had CHI values ranging between 0.5 and 0.75, while 55 compounds exceeded a CHI of 0.75 (Fig. 2). The zebrafish developmental assay detected toxicity from an additional 10% of compounds (79 in total) that did not register as toxic in the CHS assay. Specifically, 64 out of the 79 compounds induced developmental toxicity, primarily evidenced by increased mortality rates. This demonstrates the advantages of the additional behavioral and developmental measurements that can be observed in a live, whole organism model, adding depth to the toxicity coverage when used in combination.

In aggregate, the embryonic zebrafish model flagged 139 compounds as hazardous. The study’s aim was to use the ZBEscreen to rapidly identify hazardous compounds for further mechanistic investigation. Initially, all 403 compounds were screened only at 100 µM and the 162 hits in that screen were subjected to concentration–response testing for abnormal morphology and behavior readouts. This approach facilitated the identification of a benchmark concentration (BMC) for further testing. The concentration–response testing of 162 hits yielded 139 toxic hits in at least one of the three assays (Fig. 3b) and 25 compounds that were a hit in all three assays. There were 21 compounds that only elicited morphologic defects or morbidity, nine that only elicited abnormal EPR behavior, and eight compounds that only elicited abnormal LPR behavior.

This refined approach to early hazard detection underscored the utility of combining cellular and whole-organism models to provide actionable safety data well upstream of mammalian and clinical testing.

Advantages and limitations of using BMC10 in aebrafish toxicity screening

The Benchmark Concentration 10 (BMC10) is a widely used metric for regulating compounds, as it incorporates the concentration–response relationship and accounts for variability within the data, providing a robust estimate of the concentration at which a specific adverse effect is observed in 10% of the population. However, the BMC10 method requires extensive, high-quality concentration–response data and involves parametric statistical modeling, which can limit its practicality for routine screening compared to simpler approaches such as the Lowest Effect Level (LEL). Due to these constraints, BMC10 was applied to the morphologic data but not to the behavioral data. If the LEL had been used for morphology, an additional five compounds would have been identified as toxic. Of these, CHS would have detected 3 (Saikosaponin A, Tyrphostin A9, and 20(S)-Ginsenoside Rh2), none of which have documented antiviral activity.

The ZBEscreen is multidimensional with several BMC/LELs for each compound, which made it difficult to compare to the CHS. To account for this, the ZBEscreen LEL was computed across all four of the assay endpoints. For the 139 zebrafish toxic compounds, 29 different antiviral compound classes were used (Fig. 4a). The class with the highest median ZBEscreen LEL was Herpes Virus with a median LEL of 75 µM. Thus, half of the compounds in this class had a ZBEscreen LEL higher than 75 µM, and the other half lower than 75 µM. The antiviral compound class with the lowest median ZBEscreen LEL was Broad-Spectrum Antivirals with a median LEL of 20 µM. These findings suggested that Herpes Virus compounds were less toxic than Broad-Spectrum Antiviral compounds.

Fig. 4

Distribution of compound activity across antiviral classes. Distribution of (A) Zebrafish Lethal Effective Concentration (LEL) and (B) Cytotoxicity Index (CHI) for different antiviral compound classes. The boxplot depicts the median (horizontal line within the box), interquartile range (IQR, represented by the box), and whiskers extending to 1.5 IQR from the median. Outliers are plotted as individual points

For the 104 compounds with a CHI > 0.5, the antiviral compound class with the highest median CHI was Broad-Spectrum Antivirals (Fig. 4b). The median for this class was 0.85. The classes with the lowest median CHI were Dengue Virus and Corona Virus with median CHI approximately 0.7.

Power of both model systems

In the Cayman antiviral library, 183 compounds (45.5% of the library) were hits in either or both assay systems. Most of these compounds had undefined antiviral associations (39.9%) and anti- HIV and -HCV made up the next two largest groups (17.5% and 7.7%, respectively) (Figure S1).

CHS detected toxicity from 44 compounds that ZBEscreen did not at 100 µM. The target activity of the 44 was primarily undefined (17) and secondary anti-HIV (13). The ZBEscreen detected toxicity associated with 79 compounds that CHS did not detect.

Sixty compounds were active in both CHS and ZBEscreen (Fig. 5), with 23 compounds (38%) having an unknown mechanism. HIV and HCV (each with 5) were next, then influenza and protease inhibitors (4 each). Both broad-spectrum and general antivirals made up 6 members. The remainder, with 2 members or fewer, were anti-malarial, anti-zika, anti-HSV, anti-parasitic, anti-dengue, anti-coronavirus, anti-HBV, and anti-herpes.

Fig. 5

Inter-relationship between 60 common compounds hits between CHS and ZBEscreen and their activity class. A total of 60 compounds were toxic in both in vitro and in vivo and classified into 14 activity classes. Each color represents one of these parameters, and the line indicates the relationship between two parameters. The width of each parameter is a count of the number of relationships

The limited concordance between the CHS and ZBEscreen datasets was expected. The CHS assay is specifically designed to monitor rapid, acute responses for the parent compounds that serve as proxies for general acute cellular toxicity. In contrast, the ZBEscreen captures more complex and subtle interactions between toxic compounds and the physiology of whole organisms. While ZBEscreen is expected to detect many toxic effects identified by CHS, the reverse is less likely, as the cellular model used in CHS lacks the physiologic complexity and metabolic competency of the ZBEscreen to detect certain toxic actions. The physiologic dynamics of the processes occurring during vertebrate development cannot be replicated by even the most comprehensive set of in vitro assays. The CHS assay utilized HL60 cells, assessing 12 endpoints related to cell health, including cell morphology, membrane integrity, reactive oxygen species (ROS) production, glutathione depletion, nuclear membrane integrity, cell cycle progression, and mitochondrial membrane potential (Bieberich et al. 2021; St Mary et al. 2023)}. In contrast, ZBEscreen examined the outcomes of complex-signaling pathways over time, from early body patterning to metabolic competence and the interactions among tissues and organs during development. Although the developmental endpoints in ZBEscreen are less process-specific than the in vitro endpoints, they query a far broader range of biologic targets and events than CHS, while also incorporating metabolism and potentially capturing the toxic effect of metabolites. Metabolism is also a potential contributor to the differences between compounds toxic in the CHS screen versus negative in the ZBEscreen.

Moreover, the inclusion of photomotor behaviors in ZBEscreen introduces an additional dimension that cannot be replicated in vitro. The limited overlap between the two screens reflects their partially orthogonal designs and intended functions. Consequently, each system detected unique toxicities. When used together, these two screening approaches provide a more comprehensive method for detecting toxic effects than when either screen is used in isolation.

Identifying key compounds with strong cellular impact through CHI and zebrafish assay analysis

There were 12 compounds with a CHI > 0.9, indicating an exceptionally detrimental impact on cell health. Notably, 11 of these 12 compounds were also hazardous in the zebrafish assay. The sole exception was Timosaponin AIII (CAS: 41059-79-4), a steroidal saponin that has potential therapeutic effects in cancer treatment, neuroprotection and anti-inflammatory roles. It has been used in traditional Chinese medicine but with a known cytotoxic potential (Han et al. 2018).

Four of the CHI > 0.9 compounds are established medications currently on the market: cepharanthine (CAS: 481-49-2), mefloquine hydrochloride (CAS: 51773-92-3), pyronaridine phosphate (CAS: 76748-86-2), and gallotannin (CAS: 5424-20-4) and all except mefloquine hydrochloride have demonstrated low toxicity profiles. Mefloquine hydrochloride, primarily used for malaria treatment, has shown potential for neurologic side effects in animal studies (Nevin 2014) including motor and respiratory impairment in small fish (Maaswinkel et al. 2015). Neurotoxicity was also observed in organotypic cultures (Yu et al. 2011).

Of the remaining seven highly toxic CHS compounds, six have yet to progress to market due to their toxicity in cell culture models. Cyclosporin C is structurally similar to Cyclosporin A, a known immunosuppressant on the market. The compound has promise for its potential to prevent organ transplant rejection and to treat autoimmune diseases (Ray et al. 2003). The last compound, simeprevir sodium salt (CAS: 1241946-89-3), was on the market to treat hepatitis, but was discontinued by FDA.

Self-organizing map of the antiviral library

With the multi-dimensional data, a SOM with six parameters: CHI, zebrafish morphology, EPR, LPR-light, LPR-dark, and the SMILES for each compound, showed patterns and relationships from 200 neurons per axis.

The SOM is an unsupervised data clustering methodology (Kohonen 1982). It has been widely used in drug screening and discovery applications (Tanrikulu et al. 2013). SOM_Fit values were computed and used as the color bar of Fig. 6A. They indicated a well-separated SOM (< 0.5) of 6 neighborhoods (Fig. 6B), which suggested that structure influenced both ZBEscreen and CHS outcomes. The largest neighborhood (blue and green background) was more dispersed, which suggested variability in the response and/or a range of compound structures that were less responsive in CHS and ZBEscreen. The neighborhood with mostly orange and red clusters represented a high density of similar compounds, which suggested conserved structure–activity relationships.

Fig. 6

Self-organizing map. A SOM map displaying the fit quality for the ZBEscreen, CHI and the SMILES for each compound as nodes in the grid. Each hexagonal unit represents a SOM node, and the color intensity indicates the fit quality, where red indicates a poorer fit (higher error) and blue indicates a better fit (lower error). The hexagons are overlaid with circular markers representing individual samples, and the colors of the markers represent data point distribution over the map. Dense clusters of markers with a strong gradient between adjacent nodes. B The SOM with identified clusters outlined. The contours define specific regions of interest, indicating groupings of similar patterns across compounds. Neighborhoods 1 through 6 are marked, and the distinct boundaries between clusters demonstrate meaningful divisions. The color scale below the plot indicates the fit quality, with lower values (in blue) corresponding to better model fit

SOM neighborhood 1 illustrated the clustering of the most toxic compounds in the CHS assay. This neighborhood was highly connected with 2, showing those toxic in CHS and ZBEscreen. Neighborhood 3 compounds were toxic in zebrafish only, and those that were more densely clustered had more structural similarity. Neighborhoods 4, 5, and 6 were not toxic in the ZBEscreen but moderately toxic in CHS (CHI ~ 0.2), with neighborhood 6 having the most structural similarity.

Neighborhood 2 contained the most variable toxicity and the compound with the highest SOM_fit (0.5), arbidol hydrochloride (CAS: 1317070-25-0) with a CHI of 0.79 (morbidity, mortality and behavioral effects) in zebrafish at ~ 3 µM. Surrounding compounds had similar toxicity profiles such as shikonin (CHI 0.88) which elicited mortality in zebrafish at 3.46 µM by 24 hpf, and saikosaponin A (CHI 0.95) which elicited zebrafish mortality at ~ 3 µM by 24 hpf.

Arbidol hydrochloride is an antiviral drug known as umifenovir in Russia and China, for the treatment and prevention of influenza and oral viral infections (Boriskin et al. 2008). It inhibits viral fusion with the host cell membranes and targets hemagglutinin protein of influenza viruses, preventing the virus from entering host cells and inhibiting its replication. With the coronavirus pandemic in 2020, the COVID-19 antiviral efficacy of arbidol was evaluated in China (Zhao et al. 2022) where it was reported to accelerate the recovery of COVID-19 patients without significant adverse effects. In vitro test showed arbidol reduced hepatitis B virus DNA in extracellular supernatant (Li et al. 2021). It has not been approved by the US FDA or the European Medicines Agency (EMA), and the toxicity data reported herein suggest that the safety profile of arbidol may be considerably different from that of Chinese claims.

A neighbor to arbidol on the SOM was shikonin, a phytochemical with anti-inflammatory and antioxidant properties like arbidols. It is a major chemical component of Lithospermum erythrorhizon roots and may attenuate neuroinflammation and neurodegeneration (Liu et al. 2023). The compound has demonstrated neuroprotective properties in mouse brains attributed to its antioxidant properties (Wang et al. 2010). Shikonin can accumulate in mitochondria and generate ROS, resulting in a variety of adverse effects (Boulos et al. 2019), which may explain the toxicity observed in CHS and ZBEscreen. CHS measured mitochondrial membrane potential and impact to this endpoint would result in a higher CHI. Likewise, a potent inducer of mitochondrial ROS would be anticipated to adversely impact vertebrate development.

In the same SOM activity neighborhood, saikosaponin A was identified. It has traditionally been used in Chinese and Japanese medicine for modulating the immune system and as an anti-inflammatory (Lee et al. 2024) and is major component in the roots of Bupleurum sp. plants. Saikosaponin has antiviral, anti-fibrotic, and antioxidant properties and is used to treat liver disease, viral infections, and cancer therapy (Huang et al. 2024).

These three SOM neighbors were structurally dissimilar yet induced similar toxicity profiles and were anti-inflammatory and neuroprotective. This particular outcome did not support any sort of a predictive aspect of the dataset. However, assuming the growth of such a dataset to several thousand structurally diverse compounds, integrated structure–activity information would begin to enable hazard prediction.

Zebrafish non-toxic library members were highly correlated with over-the-counter or prescribed antiviral drugs

We anticipated that OTCs and prescribed antivirals in the library would largely test benign in the ZBEscreen. Of the 263 compounds categorized as non-toxic, 241 were without morphological effects at 100 µM while 22 elicited abnormal behavior. The library had 114 compounds with FDA approval as “RX”, “OTC”, and 17 that were discontinued for various reasons after FDA approval (Figure S2A). Only three compounds were toxic hits in ZBEscreen: amodiaquine, simeprevir and probucol, each a discontinued drug. Amodiaquine was approved as a prodrug to treat malaria but was found to cause reproductive toxicity in rats (Niu et al. 2016). Simeprevir was approved to treat hepatitis C but was later found to cause severe liver toxicity in humans (Diseases 2012). Probucol was an approved antioxidant to treat hypercholesterolemia but was withdrawn in 1995 due to cardiotoxicity (Hirata 2021). Of the other 14 discontinued compounds, no adverse morphologic affects were observed at 100 µM and, because of this, they were not tested for abnormal behavior effects. Of the 17 discontinued drugs, only simeprevir was identified as a hit in CHS. The others had CHI values under 0.34.

From the library, 97 compounds, representing 24.1% of the total, are currently prescribed pharmaceuticals. Of these, 78 exhibited no morphological effects in the ZBEscreen while 19 were identified as toxic hits. A total of 17 were toxic in CHS with a CHI > 0.5. Seven of the currently prescribed pharmaceuticals were a toxic hit in both CHS and ZBEscreen. The compound with the highest CHI (0.93) was mefloquine, which has reported liver and CNS toxicity, although still used to treat malaria. In embryonic zebrafish, mefloquine had a BMC10 of 8.88 µM for ‘any effect’ and elicited abnormal swimming behavior at 3.7 µM. The other 6 currently prescribed but herein toxic hits in CHS also have known nervous system impacts (trifluoperazine, efavirenz, ivermectin), cardiotoxicity (doxorubicin, imatinib), gastrointestinal toxicity (niazoxanide). The only over-the-counter drug was loratadine, an antihistamine used to treat allergies. It was not identified as toxic in the CHS but was in zebrafish with a BMC10 of 1.23 µM. Other loratadine studies in developmental zebrafish demonstrated reduced heart rate and body length (Teixido et al. 2019) and altered cardiac β-adrenergic receptor activity (Steele et al. 2009). Decreased heart rate by competitive inhibition of muscarinic receptors in mammals has been noted when exposed to antihistaminic compounds.

Regulatory scrutiny by the EMA has been applied to 67 of these commercially available compounds, aligning with the FDA on 87% of the approval or withdrawal decisions (59 compounds). Within this subset, seven compounds were withdrawn from the European market. Only one compound discontinued by the FDA is still authorized for use by EMA (Figure S2B). Of eight chemicals discontinued by both agencies, only simeprevir was toxic in both CHS and ZBEscreen (CHI = 0.90, ZF any effect LEL = 3.60 uM). The other 7 had CHI < 0.34, and no LEL in zebrafish. During large randomized controlled trials of simeprevir, no link to increased instances of liver injury were detected, but after approval and wide scale use, a pattern of injury was observed. ZBEscreen and CHS both identified simeprevir as toxic. Perhaps it would never have made it to the clinical stage, let alone become a prescription drug, with better safety testing early in its development.

Of the 51 drugs approved by both EMA and FDA, 43 were non-toxic in ZBEscreen and 35 were non-toxic in CHS assays. The intersection of non-toxic compounds in both zebrafish and CHS assays underscored the thorough if inefficient, pre-market safety evaluation process. Those evaluations eventually got to the right place, but likely more slowly and at a much greater cost than the developers would have incurred had they availed themselves of the early testing we described here.