Assessing Bioactivity-Exposure Profiles of Fruits and Vegetables in the BioMAP Profiling System

Abstract

The ToxCast program has generated in vitro screening data on over a thousand chemicals to assess potential disruption of important biological processes and assist in hazard identification and chemical testing prioritization. Few results have been reported for complex mixtures. To extend these ToxCast efforts to mixtures, we tested extracts from 30 organically grown fruits and vegetables in concentration-response in the BioMAP® assays. BioMAP systems use human primary cells primed with endogenous pathway activators to identify phenotypic perturbations related to proliferation, inflammation, immunomodulation, and tissue remodeling. Clustering of bioactivity profiles revealed separation of these produce extracts and ToxCast chemicals. Produce extracts elicited 87 assay endpoint responses per item compared to 20 per item for ToxCast chemicals. On a molar basis, the produce extracts were 10 to 50-fold less potent and when constrained to the maximum testing concentration of the ToxCast chemicals, the produce extracts did not show activity in as many assay endpoints. Using intake adjusted measures of dose, the bioactivity potential was higher for produce extracts than for agrichemicals, as expected based on the comparatively small amounts of agrichemical residues present on conventionally grown produce. The evaluation of BioMAP readouts and the dose responses for produce extracts showed qualitative and quantitative differences from results with single chemicals, highlighting challenges in the interpretation of bioactivity data and dose-response from complex mixtures.

Introduction

Political, scientific, financial, and ethical pressures have all contributed to global efforts to transition away from animal-based toxicity testing toward incorporation of in vitro and alternative testing approaches to inform chemical safety. To build confidence in applying in vitro tools to this task, a range of issues need to be considered to relate activity of compounds in in vitro assays to potential human health ef-ects and real-world exposures. In the 2007 National Research Council Report, “Toxicity Testing in the 21st Century”, a clear distinction was made between a biological activity and a pathway perturbation leading to an adverse response, with the latter being used to define a ‘toxicity pathway’ (National Academies of Sciences, 2017; NRC, 2007). Linking in vitro biological activities with in vivo outcomes is a strategy that has been used extensively in drug discovery and provides a framework for validating strategies to assess chemical modes of action (Kallenberg and Heeringa, 2015; Taha et al., 2014). However, validation of in vitro tools designed to predict therapeutic effects of drugs for which intake doses, systemic concentrations and biomarker studies can be evaluated using clinical study data is more straightforward than validation of in vitro tools designed to predict human toxicological effects of environmental, commercial and industrial chemicals, where controlled study in vivo data are lacking. As new in vitro assays come into practice for measuring molecular and cellular responses, un-certainty in the relationship between measurable bioactivity and potential adversity needs to be carefully considered, particularly when considering risk from complex environmental exposures.

Assessing the toxicity and human health concerns with complex mixtures of chemicals remains a major public health concern and new methods and concepts are needed to more adequately assess possible risks of complex mixtures. The EU report on “Toxicity and Assessment of Chemical Mixtures” emphasized the importance of developing techniques to triage the universe of chemical mixtures to which humans and environmental species are exposed (Scientific Committee on Health and Environmental Risks et al., 2012). Although this report was not focused on foods or food products, the authors recommended the testing of real world mixtures in various biological systems. The 2007 report “Toxicity Testing in the 21st Century: A Vision and A Strategy” noted that the use of in vitro testing with mixtures could support grouping of chemical mixtures according to their biological response, and recommended testing over a wide range of treatment concentrations to develop an intelligent, focused approach to the problem of assessing risks in human populations (NRC, 2007).

Here, we applied this testing strategy to complex mixtures derived from foods that are commonly consumed and are regarded as safe (organic fruits and vegetables) and compared these responses to those obtained with environmental chemicals (Berg et al., 2006; Berg et al., 2010; Houck et al., 2009; Kleinstreuer et al., 2014). The comparison of biological activity across a wide range of chemicals—exogenous, endogenous, and naturally occurring compounds - should provide some context to the relevance of responses that occur in high-throughput in vitro testing systems for safety assessment decisions.

In this study, the biological activities of fruit and vegetable juices were assessed in the BioMAP profiling suite (DiscoverX, South San Francisco, CA), a testing battery that represents one of the platforms employed by the US EPA ToxCast screening program (Kavlock et al., 2012). Juices extracted from organically grown fruits and vegetables, confirmed to have no quantifiable levels of agrichemical residues, heavy metals, or naturally occurring toxins, were tested in this plat-form. The bioactivities, concentration-response relationships, and out-puts from bioinformatic and predictive tools were used to evaluate patterns of activity. The findings were also compared to platform results with several other chemicals screened as a part of the ToxCast program to provide context and guide dataset interpretation. Activity of both organic juice extracts and several food use agrichemicals were evaluated in a concentration-response format in order to derive an estimate of lowest effective concentration (yielding a response in 10% of tested assays; LEC-10) and to determine the trends in responses elicited by the mixtures compared to single compound agrichemical exposures.

Materials and Methods

Test agent preparation

Organically grown fruits and vegetables (Table 1) were purchased from grocery stores in and around Durham, North Carolina (USA). Fruits and vegetables selected for analysis were those previously studied in in vitro bioactivity screening systems (Amakura et al., 2003; Charles et al., 2002; Jeuken et al., 2003; Kassie et al., 1996; Renwick et al., 1984). When the items were not available due to the growing season, substitutes from the same plant family were selected. Extracts were prepared within 3 days of purchase. After weighing, each item was rinsed prior to processing. Most items were processed in a Breville BJE820XL Dual Disc Juicer. The resulting juices were filtered through Whatman #4 paper prior to centrifugation at 10,000×g for 15 min. The supernatants were then filter-sterilized through a 0.22 μM vacuum filter in a laminar flow hood, and the final volume was recorded to calculate the volume (μL) yield per mass (g) processed. Aliquots were then pre-pared and stored at temperatures lower than −70 °C until experimental use. Due to the varied nature of the items tested and the resulting juices, some required additional filtration and centrifugation steps or use of a blender or coffee grinder for processing. Additional details, including measures of osmolality and pH levels of the filtered juices are provided in Supplementary Table S1A. Assuming a density of 1 g/mL, the weight of the initial produce and the juice were used to calculate yield based on the initial weight of the fruit or vegetable (see Supplementary Table S1B). Organic solids (pulp) were removed during the extraction process.

Table 1.

Fruit and Vegetable Extracts Evaluated.

Produce Family	Test Agent ID	Processing Procedure
Rosaceae	(Gala) Apples	Juicer
	(Red) Grapefruit	Peeled; juicer
Fabaceae	Bean Sprouts	Juicer
	(Steamed) Edamame	Blender
	Green Beans	Juicer
	Snow Peas	Juicer
	Soybeans	Grinder
Amaranthaceae	(Red) Beets	Juicer
	Spinach	Juicer
Ericaceae	Blueberries	Juicer
Brassicaceae	Broccoli	Juicer
	Cabbage	Juicer
	Cauliflower	Juicer
	Kale	Juicer
Apiaceae	Carrots	Juicer
	Celery	Juicer
	Parsley	Juicer
Allioideae	Garlic	Blender
	(Vidalia) Onions	Juicer
Zingiberaceae	Ginger	Blender
Vitaceae	(Green) Grapes	Juicer
Rutaceae	Oranges	Peeled; juicer
	Strawberries	Juicer
Bromeliaceae	Pineapple	Peeled; juicer
Solanaceae	Yukon Gold Potatoes, u	Juicer
	Yukon Gold Potatoes, p	Peeled; juicer
	(Red Bell) Peppers	Juicer
	(Red) Tomatoes	Juicer
Cucurbitaceae	(Butternut) Squash	Peeled; juicer
Convolvulaceae	Sweet Potatoes	Peeled; juicer

Agrichemical residue, heavy metal and mycotoxin analyses

To determine whether the fruits and vegetables had detectable levels of agrichemicals or heavy metals, all items underwent an expanded screening for 480 agrichemicals and five heavy metals (antimony, arsenic, cadmium, lead, and mercury) at Medallion Labs (Minneapolis, MN). Where relevant, a subset of items was screened for mycotoxins: ginger was assessed for the presence of aflatoxins (B1, B2, G1, G2), deoxynivalenol (by high pressure liquid chromatography), fumonisins (FB1, FB2), and ochratoxins. Tomatoes, carrots, grapes and apples were assessed for the presence of alternariol, alternariol-monomethylether, altenuene, tenuazoic acid, and tentoxin. In addition, apples and grapes were assessed for patulin. Testing for antimony, arsenic, cadmium, and lead were performed using inductively coupled plasma mass spectrometry (ICP-MS). Except where noted differently, all other analyses were conducted using liquid chromatography-mass spectrometry (LCMS). A complete listing of screened analytes and associated LOD/LOQ in-formation is available at www.MedallionLabs.com, with data relevant to this study provided in Supplementary Table S1. Medallion Labs maintains A2LA accreditation to ISO/IEC 17025 for the specific tests listed, in A2LA Certificate #2769.01.

General experimental design using BioMAP

The BioMAP assay panel used in this study is comprised of 8 different primary cell models intended to replicate human disease states of vascular inflammation and immune activation. More details on cell culture and experimental design can be found in the literature (Berg et al., 2006; Houck et al., 2009). Guided by knowledge of relevant disease biology and mechanisms, either single cell types or defined mixtures of cells are stimulated with combinations of cytokines, growth factors and other stimuli to activate specific signaling pathways (Table 2), including wound healing and immune and inflammatory signals associated with internal or external stimuli. Based on these outputs, it is important to note that the cells in these systems and the resulting bioactivities following xenobiotic exposure may not accurately capture the behaviors and responses representative of cells in an environment that is not similarly stimulated.

Table 2.

Summary of BioMAP assays used in this study.

BioMAP System	Primary Human Cell Type	Stimuli	Disease/Tissue Relevance	Readouts
3C	Venular endothelial cells	IL-1b + TNF-a + IFN-g	Vascular Biology, Cardiovascular Disease, Chronic Inflammation	E-selectin, IL-8, VCAM-1, ICAM-1, MCP-1, MIG, HLA-DR, TM, TF, uPAR, EC Proliferation, SRB, Vis
4H	Venular endothelial cells	IL-4 + Histamine	Asthma, Allergy, Oncology, Vascular Biology	VCAM-1, Eotaxin-3, MCP-1, VEGFRII, uPAR, P-selectin, SRB
LPS	Peripheral blood mononuclear cells + Endothelial cells	TLR4	Cardiovascular Disease, Chronic Inflammation, Infectious Disease	E-selectin, IL-1a, IL-8, TNF-a, PGE2, VCAM-1, MCP-1, CD40, M-CSF, Tissue Factor, SRB
SAg	Peripheral blood mononuclear cells + Endothelial cells	TCR	Autoimmune Disease, Chronic Inflammation, Immune Biology	IL-8, MCP-1, E-selectin, MIG, CD38, CD40, CD69, PBMC Cytotox., T cell Proliferation, SRB
BE3C	Bronchial epithelial cells	IL-1b + TNF-a + IFN-g	COPD, Respiratory, Epithelial Biology	IL-1a, IP-10, MIG, HLA-DR, uPAR, MMP-1, PAI-1, TGFb1, SRB, tPA, uPA
CASM3C	Coronary artery smooth muscle cells	IL-1b + TNF-a + IFN-g	Vascular Biology, Cardiovascular Inflammation, Restenosis	IL-8, IL-6, SAA, MCP-1, VCAM-1,MIG, HLA-DR, M-CSF, uPAR, TM, TF, LDLR, SMC Proliferation, SRB
HDF3CGF	Fibroblasts	IL-1b + TNF-a + IFN-g + EGF + bFGF + PDGF-BB	Tissue Remodeling, Fibrosis, Wound Healing	IL-8, VCAM-1, IP-10, MIG, M-CSF, Collagen III, EGFR, MMP-1, PAI-1, Fibroblast Proliferation, SRB, TIMP-1
KF3CT	Keratinocytes + Fibroblasts	IL-1b + TNF-a + IFN-g + TGF-b	Skin Biology, Psoriasis, Dermatitis	IL-1α, MCP-1, ICAM-1, IP-10, MMP-9, SRB, TIMP-2, uPA, TGFb1

To enable comparison of this work with those obtained for single compounds in the ToxCast program at the U.S. EPA NCCT, the testing and data analysis procedures were consistent with those used in previous assessments using the BioMAP platform (Houck et al., 2009; Kleinstreuer et al., 2014). Briefly, point of departure assessments utilized lowest effective concentration (LEC) instead of curve-fitting and AC50 derivation.

Juices were screened at multiple concentrations, using a half-log format. The original extraction was defined as 100% juice. Dilution of the juices were made directly with the cell culture media. Concentrations of juice to which cells were exposed are expressed throughout the manuscript as “% juice”, which is the % volume of the media comprised of 100% juice. Testing proceeded in two phases. In the first phase, tested concentrations ranged from 0.46% to 12.5% juice in media (0.46, 1.39, 4.17, and 12.5%). These concentrations were selected based on review of previously published data (Charles et al., 2002) and in consultation with DiscoverX scientists with expertise in the technology across a wide range of test agents. In the second phase, data from the first studies were used to refine dose selection; concentrations ranged from 0.0057% to 0.46% juice in media (0.0057, 0.017, 0.051, 0.154, and 0.46%). Aggregate volumes of media and juice added to the sample wells were equivalent across all concentrations.

Estimating molar concentrations of juice active ingredients in cell culture

The conversion of the percent juice concentrations to molar equivalents was completed to more easily compare results from this study to the archived ToxCast chemical data. Because these juices are complex mixtures with limited information on their chemical makeup, it was necessary to make a few key assumptions that, while admittedly simplistic, made it possible to derive a rough estimate of the molar concentrations of potential “active ingredients” in the tested media. These intentionally conservative assumptions were: 1) fiber and large carbohydrates (cellulose, etc.) are maintained in the pulp that remained in the blender or filter after processing and therefore are not present in the extract; 2) all other compounds present in the fruits and vegetables are water soluble and therefore equally distributed in the extract; 3) mass balance of the ingredients can be tracked by the measured weights of fruit/vegetable in the blender and juice yield, and the difference of these two measures, which represents pulp content remaining in the blender and filters (based on the assumption that all chemicals of interest are water soluble and any residual active compounds would only be recoverable with organic solvents and additional processing)(Altemimi et al., 2017).

Volume fractions in test media were converted to molar concentrations in the following manner using the information in the USDA Food Composition database (https://ndb.nal.usda.gov/ndb/). Assuming a density of 1 g/mL, the juice volumes were converted to mass quantities (e.g., 0.46% juice = 0.46 g juice/100 mL media). Total juice was converted to mass fraction of “active ingredient” based on the composition of the produce item. Active ingredient mass fraction is defined as the difference between unity and the sum of water, fiber, and nutritionally unavailable carbohydrate mass fractions. This definition, which includes nutritionally available sugars as well as protein, lipids, minerals, phytochemicals, and other components, is inclusive and may overestimate the net concentrations of active ingredients.

Assuming a representative molecular weight of 300 g/mol—which serves as an approximation for both available sugars (primarily mono-and disaccharides) and phytochemicals—the active ingredient mass fractions were converted to molar concentrations (see Supplementary Table S1b). This value of 300 g/mol is consistent with a median molecular weight determined during a review of 48 phytochemicals screened in ToxCast (median MW of 302.22 with a standard error of 5.19 - see Supplementary Table S7).

Assay readouts and endpoints

Assay readouts were measured mostly using enzyme-linked immunosorbent assay (ELISA) as described (Melrose et al., 1998). A total of 174 different assay endpoints (i.e., 87 assay readouts assessed for increase and decrease) were measured across the eight different primary cell models (Table 2). Evaluation of cell viability and toxicity are built into each cell model. Cytotoxic responses were determined by the sulforhodamine blue (SRB) assay, a measure of total cell protein, across all cell models, where log10 transformed values < −0.3, representing a 50% decrease in total protein, were considered to be overtly cytotoxic (Houck et al., 2009). In addition, peripheral blood mononuclear cell viability was measured, and microscopic visual inspection was used to assess cellular phenotype and morphologic changes indicative of cellular toxicity. The assay data were generated under contract by DiscoverX in South San Francisco, CA.

Data outputs, similarity mining, multivariate analyses and mechanism model classification

Assay readouts for each sample were divided by the relevant mean vehicle control and then log10 transformed as previously described (Houck et al., 2009). Significance prediction envelopes (95%) calculated from historical controls were used to determine activity in each assay endpoint. In this study, an endpoint was considered to be “active” at any concentration where the response fell outside of the confidence envelope and was not associated with overt cytotoxicity. Lowest effect concentrations (LECs), defined as the lowest concentration at which a significant up- or down-regulated effect was observed, were calculated for each juice-assay readout. The LEC-10 was defined as the lowest concentration at which a significant effect was observed for a minimum of 10% (i.e., at least 9 of the 87 assay readouts) of the BioMAP assays screened.

Prior to profile similarity analyses, overtly cytotoxic test agent profiles were removed. Remaining profiles were compared against the reference compound library generated by DiscoverX, with a correlation metric comprised of a combination of similarity metrics in addition to Pearson’s correlation as previously described (Berg et al., 2010; Houck et al., 2009). Two-way hierarchical clustering was performed on the log10-transformed LECs for the juices using Ward linkage and Euclidean distance metric.

Support vector machines (Cortes and Vapnik, 1995) or support vector networks are supervised machine learning models which utilize a set of training examples already marked with distinct categories and a model that assigns new examples to one of the categories. SVM can perform both linear and non-linear classification. A set of 28 mechanism classifiers developed using SVM and a reference dataset of BioMAP profiles has been previously described (Berg et al., 2013).

Agrichemical residue evaluation

Using data from the U.S. Department of Agriculture’s (USDA) Pesticide Data Program (PDP; https://www.ams.usda.gov/datasets/pdp/pdpdata), residue data were gathered for a subset of ToxCast Phase I chemicals commonly used in crop protection. As not all agri-chemicals are analyzed annually, the USDA revisits a subset every 3 to 5 years. To compensate for this and to ensure data across a wide range of agrichemicals and time periods, residue levels were collated over a 12-year period (2003–2015). Monitoring data were collected on all fresh produce, including: number of samples, number of samples with detected residues, range of residue concentrations detected (ppm), the range of limits of detection for measurements (ppm), EPA tolerance level (ppm), maximum detected value, and the year the commodities were measured. A detailed listing of the collated data is provided in Supplementary Table S2.

BioMAP bioactivity data derivation for ToxCast food-use agrichemicals

We compared the bioactivity profiles and the estimated exposures required to elicit an equivalent bioactivity for six crop-protection agents commonly found on commercial produce. Chemicals selected had 1) previously published BioMAP activity screening data generated as a part of the ToxCast program (Houck et al., 2009; Kleinstreuer et al., 2014); 2) registered uses on food crops (https://www.ams.usda.gov/datasets/pdp); and 3) residue data available from the USDA PDP (Dis-coverX/BioSeek BioMAP profiles for all fruits and vegetables at all concentrations are provided in the Supplementary Table S8).

Comparison of juice and agrichemical bioactivities

Approaches such as in vitro to in vivo extrapolation (IVIVE) utilize a process called reverse dosimetry to estimate the hypothetical external (oral) dose that would yield in vivo steady-state blood concentrations equivalent to nominal concentrations in the in vitro test media (Wetmore et al., 2015; Yoon et al., 2012). This approach provides a method for putting in vitro measures of activity into context for human exposures where the chemicals and kinetic parameters are known (or can be measured in vivo). This approach is not readily expanded to complex mixtures such as the natural product extracts studied here, however, where we lack information on (a) the composition of each extract, and (b) key biokinetic parameters for individual components. Nevertheless, human exposure to complex mixtures is a routine phenomenon, and methods need to be developed for characterizing their risk to health.

We developed a metric to compare the bioactivity potentials of produce extracts to those of specific agrichemicals commonly in their production. This alternative dose metric integrates exposure information with bioactivity of produce extracts and residue compounds to account for substantial differences in the intake rates between the produce extracts and the agrichemicals. The metric also includes adjustment of mass fraction of produces and the residue compound. Overall, the purpose of this metric was to compare bioactivity between produce extracts and agrichemicals adjusted by mass fraction. It should be noted that this does not explicitly account for internal dosimetry and is therefore a fundamentally different metric from the oral equivalent dose.

Residue compound exposures are quantified for individual produce items via the USDA PDP. Using information from this database, we define an intake-adjusted bioactivity ratio, Br,p for each residue compound (r) found on each produce type (p):

$$B_{r,p}=\frac{L_p}{L_r}\frac{I_{r,p}}{I_p},$$

where Lp and Lr are the LEC-10 values (expressed as molar concentrations) for the produce extract and the residue; Ip is the mass fraction of active ingredients in produce type p; and Ir,p is the mass fraction of residue compound r found on produce type p. Thus, the intake-adjusted bioactivity ratio becomes a composite measure of bioactivity that normalizes for exposure. A value for Br,p of 1.0 indicates equal contributions of residue and produce to the calculated bioactivity; a value of 0.1 would indicate that the residue has only 10% of the activity of the produce. Overall, increasing values of Br,p indicate increasing contributions of the agrichemical residue to the total expected bioactivity of the produce/residue mixture. This approach does not explicitly ac-count for in vivo pharmacokinetics and therefore comparing potency based on this metric implicitly assumes bioavailability of the produce extract and the agrichemical are similar as data is not available to determine otherwise.

Results

Evaluation of fruit and vegetable bioactivities

The produce extracts showed extensive bioactivities across all primary cell models and many assay endpoints screened in the BioMAP profiling platform. Analysis of the number of endpoints significantly up-or down regulated showed that the lowest number was 36 significant responses for two items (of 174 possible) – garlic and pineapple (Fig. 1; Table 3). Eight items had between 91 and 100 total active endpoints, while 25 of the 30 tested had between 71 and 110 total active end-points. Strawberry and spinach had the highest number of assay end-point responses, at 127 and 114 per juice, respectively. For the results presented here, assay activity observed at concentrations where overt cytotoxicity was observed (i.e., SRB measurements <−0.3) were omitted. See Supplementary Table S3 for log ratio data, LECs, and cytotoxicity information associated with the assay data.

Figure 1. Produce extracts induce more bioactivity than individual compounds.

Distributions of the number of hits in BioMAP assays for (a) fruit and vegetable extracts, (b) the complete ToxCast library, and ToxCast (c) Phase I and (d) Phase II compounds. Produce extracts resulted in hits in more assays than did pure compounds in the ToxCast library (average number of hits 87 and 20, respectively). This result is consistent for both the Phase I and Phase II release of ToxCast.

Table 3.

Fruit and Vegetable Juice Bioactivity, Potency, and Cytotoxicity Profiles

Test Agent	LEC- 10a (%, v/v)	LEC-10 (μM)	Hits at LEC-10 (↑↓) (Total Hits)	Lowest Cytotoxic Concentration (%v/v)	Lowest Cytotoxic Concentration (μM)	Cell Model(s) with Cytotoxicity at Lowest Cytotoxic Concentration	Bioactivity Ratiob
Apples	0.02	75	13 (3↑ 10↓) (105)	4.17	15713	LPS, SAg, 3C	245.12
Bean Sprouts	0.05	132	19 (9 ↑ 10↓) (97)	12.5	33056	All but 3C	245.10
Beets	0.15	495	13 (4↑ 9↓) (88)	1.39	4586	BE3C, CASM3C, HDF3CGF	9.02
Blueberries	0.46	1808	19 (14↑ 5↓) (96)	12.5	49143	LPS, Sag	27
Broccoli	0.46	929	38 (17↑ 21↓) (88)	1.39	2808	LPS, Sag	3
Cabbage	0.02	36	10 (7↑ 3↓) (101)	4.17	7444	Sag	245.12
Carrots	0.15	361	10 (5↑ 5↓) (96)	None	None	None	NA
Cauliflower	0.05	84	15 (7↑ 8↓) (95)	4.17	6978	KF3CT, SAg, 3C	81.71
Celery	0.15	149	9 (4↑ 5↓) (87)	None	None	None	NA
Edamame	0.05	216	20 (15↑ 5↓) (77)	None	None	None	NA
Garlic	0.05	230	24 (4↑ 20↓) (36)	0.15	689	All	2.94
Ginger	0.15	300	9 (5↑ 4↓) (42)	0.46	927	HDF3CGF	3.01
Grapes	0.46	2652	19 (7↑ 12↓) (98)	1.39	8007	HDF3CGF	3
Grapefruit	0.46	1242	14 (9↑ 5↓) (76)	12.5	33746	All but 3C	27
Green Beans	0.15	310	13 (6↑ 7↓) (105)	None	None	None	NA
Kale	0.02	68	12 (7↑ 5↓) (74)	4.17	14067	HDF3CGF, SAg	245.12
Onions	0.46	980	22 (6↑ 16↓) (81)	4.17	8878	Sag	9
Oranges	0.05	179	9 (4↑ 5↓) (93)	12.5	44655	All but BE3C and HDF3CGF	245.10
Parsley	0.05	120	21(7↑ 14↓) (100)	12.5	30020	HDF3CGF	245.10
Peppers	0.05	97	11 (4↑ 7↓) (87)	4.17	8073	HDF3CGF	81.71
Pineapple	0.05	185	11 (2↑ 9↓) (38)	0.46	1712	CASM3C, HDF3CGF, KF3CT	9.08
Potatoes (peeled)	0.46	751	28(19↑ 9↓) (103)	12.5	20401	CASM3C, 4H, HDF3CGF, SAg	27
Potatoes	0.02	33	9 (6↑ 3↓) (98)	12.5	20401	HDF3CGF, KF3CT	735.29
Snow Peas	0.05	226	11 (6↑ 5↓) (104)	None	None	None	NA
Soybeans	0.02	136	14 (6↑ 8↓) (85)	None	None	None	NA
Spinach	0.02	37	26(10↑16↓) (114)	1.39	2580	CASM3C, HDF3CGF	81.71
Squash	0.15	226	9 (3↑ 6↓) (88)	None	None	None	NA
Strawberries	0.02	43	9 (7↑ 2↓) (127)	None	None	None	NA
Sweet Potatoes	0.05	134	23 (6↑ 17↓) (76)	12.5	33608	All but KF3CT	245.10
Tomatoes	0.05	71	12 (2↑ 10↓) (69)	12.5	17808	BE3C, CASM3C, LPS, SAg, 4H	245.10

^athe lowest concentration at which at least 10% of the assays showed significant bioactivity.

^bthe ratio of the lowest cytotoxic concentration divided by the LEC-10.

To facilitate comparisons of the fruit and vegetable extracts used in this study to the ToxCast compounds, we converted the % juice values to molar concentrations using an approximate molar unit mass of 300 g (see Materials and Methods for description of calculations and key assumptions). Considerable uncertainty exists regarding the chemical makeup of the extracts, the extract constituents that would contribute to the in vitro bioactivities, and the assumed molecular weight of the active ingredients. These assumptions may overestimate the con-centration of the active ingredients in the extracts. Nonetheless, there is value in this approach in that it allows a rough estimation of the molar concentrations of the “active ingredients” in the extracts and enables comparison of these complex mixtures to the ToxCast chemicals that are tested individually based on molar concentrations.

The breadth of the bioactivities of the extracts was greater than that observed across a collection of over 1000 chemicals screened within ToxCast in Phases I and II, where 50% of the chemicals had a maximum of 6–10 endpoint responses per test agent with a maximal testing concentration of 40 μM for the Phase I and 80 μM for the Phase II chemicals (Houck et al., 2009; Kleinstreuer et al., 2014). However, the estimated molarity of the tested extract concentrations was substantially higher than the individual ToxCast chemicals. Produce samples with estimated molar concentrations < 40 μM had an average of 8 endpoint responses.

Table 3 provides a listing of estimated LECs at which a minimum of 10% of the assays were significantly changed (LEC-10). These LEC-10s ranged from 33 to 2650 μM with four extracts demonstrating an LEC-10 below 50 μM: cabbage, potatoes, spinach and strawberries. In addition to the LEC-10s, the total number of significant responses for each ex-tract is also provided. Bioactivity ratios were also used to frame the activities with regard to specificity, as determined by distance between the concentration at which activity was observed and concentrations at which toxicity was observed. These ratios, derived for each item by dividing the lowest concentration at which cytotoxicity was observed by the LEC-10, showed that values ranged from approximately 3 (broccoli, garlic, ginger, and grapes) to 735 (potatoes). Bioactivity ratios could not be derived for eight items since cytotoxicity was not observed at any of the tested concentrations (Table 3). The use of LEC instead of AC50 as a potency metric (as typically employed with other ToxCast assay platforms), is consistent with previously published assessments of the ToxCast BioMAP profiling data (Houck et al., 2009; Kleinstreuer et al., 2014).

Multivariate analyses and hierarchical clustering of BioMAP endpoints

We used hierarchical clustering via Ward’s method to identify similarities in bioactivities between 1) the fruit and vegetable extracts and the ToxCast Phase I and II chemicals previously analyzed (Houck et al., 2009; Kleinstreuer et al., 2014); and 2) the fruit and vegetable extracts themselves. To better compare the potencies of the ToxCast compounds with the produce extracts, we estimated the molar concentrations of active ingredients in the extracts from the dilution series used in the assays (see Materials and Methods). The hierarchical clustering of the produce extracts and ToxCast Phase I and II chemicals showed a tendency to group extracts and chemicals separately (Supplementary Fig. S1). When constrained to the maximum tested con-centration in the ToxCast program (40 μM), produce extracts and chemicals appeared more similar; however, it is difficult to draw firm conclusions given the number of extracts below the 40 μM limit and the simplistic juice molar concentration estimation (Supplementary Fig. S2). Clustering of the fruit and vegetable extract data alone is provided in Fig. 2 where all cytotoxic responses have been removed (leaving 154 endpoints). LEC-10s observed across the 154 BioMAP endpoints (or 77 assay readouts without cytotoxic hits, increased or decreased; depicted in the columns) in the extracts (rows) revealed distinct profiles. On average, endpoints representing decreased responses were more likely to be active compared to endpoints with increased responses (55.5%versus 35.0%). However, endpoints representing increased responses that had significant activity were active at lower concentrations (geo-metric means 0.38% vs. 0.64%) (green clade in hierarchical tree). Activity across all eight cell models was, for the most part, equally distributed across the samples.

Figure 2. Multivariate analyses of fruit and vegetable extract bioactivity profiles.

Hierarchical clustering of produce extracts (rows) by activity in the BioMAP primary cell models (columns). Only those endpoints without cytotoxic responses are included, leaving 77 of the original 87 readouts, with increased or decreased response (154 columns). Lowest effective concentration for a produce extract in each of 154 endpoints is shown as color saturation. Produce extracts that did not produce a significant change in an endpoint are colored white. Data on the vertical axis are sorted via hierarchical clustering of log-transformed LEC values, performed using Ward’s linkage and Euclidean distance. Interestingly, more produce items have activity across endpoints representing decreased responses (55.5%) versus those representing increased responses (35.0%). However, endpoints representing increased responses that had significant activity were active at lower concentrations (geometric means 0.38% vs. 0.64%). While there are groups of produce items that cluster together, they did not cluster by taxonomic family.

Cytotoxicity evaluation

Each of the eight cell models employed had measurements for direct or indirect markers of cytotoxicity. These endpoints include visual inspection of morphologic changes, changes in cellular proliferation and changes to cellular viability, as indicated by the SRB assay for total protein. Across the thirty fruit and vegetable extracts, eight had no evidence of overt cytotoxicity at any concentration or in any of the cell models (carrots, celery, edamame, green beans, snow peas, squash, and strawberries). Alternately, ginger, garlic and beets were cytotoxic across multiple cell models at the top three concentrations tested. Ginger elicited toxicity at concentrations as low as 0.15% (300 μM). The HDF3CGF cell model appeared the most sensitive, with cytotoxicity observed across nineteen extracts (Table 3).

Composite concentration response evaluation

A notable distinction between the fruit and vegetable extract bioactivities and those of the ToxCast chemicals was the number of assays that exhibited composite concentration responses, i.e., bioactivity significantly up and downregulated at various concentrations for the same assay endpoint. All but one of the extracts had at least one assay with composite concentration-response curves compared to only two of the ToxCast Phase I chemicals (azoxystrobin, a fungicide commonly used on wheat, banana’s and grapes; and pyraclostrobin, also a fungicide used on almonds, grapes, strawberries and tomatoes); 19 extracts had 10 or more assays that exhibited this pattern (Table 4). Strawberry, spinach and apple extract had the largest number of composite responses across 33, 25, and 25 assay endpoints, respectively. However, some of the differences in the composite concentration response frequency may be due to the differences in the concentration response range tested. Although both types of composite responses were observed, the majority (83% overall) had increased bioactivity at the lower concentration followed by decreases at higher concentrations. Any concentrations at which overt cytotoxicity was observed were re-moved to reduce the likelihood that decreased activity was secondary to cytotoxicity (see Supplementary Table S4 for more information on the number of composite concentration responses for each extract in each of the primary cell models and the ToxCast chemicals).

Table 4.

Evaluation of Composite Concentration Responses in Juice Bioactivity Profiles

Test Item	Number of composite concentration responses	Number with lower concentration increased (fraction, %)
Strawberries	33	31 (94)
Spinach	25	18 (72)
Apples	25	21 (84)
Grapes	19	16 (84)
Snow Peas	17	9 (53)
Potatoes (peeled)	16	14 (88)
Bean Sprouts	16	14 (88)
Green Beans	16	15 (94)
Oranges	16	13 (81)
Cabbage	15	12 (80)
Beets	14	13 (93)
Parsley	14	14 (100)
Cauliflower	14	14 (100)
Potatoes (unpeeled)	12	11 (92)
Celery	12	9 (75)
Carrots	11	9 (82)
Broccoli	11	11 (100)
Blueberries	10	10 (100)
Soybeans	10	4 (40)
Edamame	9	6 (67)
Squash	8	4 (50)
Peppers	8	7 (88)
Tomatoes	7	6 (86)
Sweet Potatoes	6	5 (83)
Onions	4	4 (100)
Grapefruit	4	3 (75)
Kale	3	2 (67)
Ginger	2	2 (100)
Garlic	1	0 (0)
Pineapple	0	0 (NA)
Total	358	297 (83)

BioMAP Profiling Analyses

Fig. 3 displays BioMAP profiles depicting bioactivities noted for eight of the fruit and vegetable extracts. This particular set – bean sprouts, broccoli, celery, edamame, green beans, squash, strawberries, and sweet potatoes – was selected as representative of the more complete set of materials. Summaries of the major observed bioactivities (proliferation, immunomodulation, inflammation, and tissue re-modeling) and the endpoints associated with each of these bioactivities are also displayed. For most of the agents, a mix of treatment-dependent up and downregulation of activities was observed, particularly for bean sprouts, green beans, celery and squash, and predominantly within the 3C, 4H, and LPS cell models. Alternately, broccoli, strawberries, and sweet potatoes primarily decreased activities. Anti-proliferative activities, indicated by thick gray arrows, were evident across a few of the cell models at various extract concentrations. Overt cytotoxicity was noted at the two highest tested concentrations for broccoli; leading to the removal of results from these concentrations in the analysis.

Figure 3. BioMAP Activity Profiles Across Fruit and Vegetable extracts.

At left, representative profiles generated based on log ratio outputs from the BioMAP system for bean sprouts, broccoli, celery, edamame, green beans, squash, strawberries, and sweet potatoes. Endpoint activities for each concentration across the eight cell models are displayed. Endpoints outside 95% confidence interval (gray band) are annotated. Extract concentrations with more than a single hit are shown. Thick gray arrows indicate anti-proliferative activities within a particular cell model; thick black arrows indicate cytotoxicity, as defined by SRB evaluation. At right, assay hits are categorized by the phenotypic activity they represent. See Supplementary Figure S1 for additional profiles.

Supervised similarity search findings

Table 5 provides a list of the significant supervised similarity search hits (Pearson r > 0.7) that were returned following comparisons to the BioMAP reference profile database. Fig. 4 shows the bioactivity profiles across a representative subset of these extracts, displaying the endpoint-level similarities that contributed to the search findings. At the time, we interrogated the BioMAP reference profile database, it contained pro-files for over 2000 compounds including drugs, drug candidates, and chemicals screened as part of the ToxCast Phase I and II chemical libraries (https://actor.epa.gov/dashboard/). Nine different extracts re-turned at least 1 hit, with similarities to agents including plant-derived polyphenols (beets: tannic acid); kinase inhibitors with cancer chemotherapeutic potentials (kale and sweet potatoes: Janus kinase (JAK) inhibitors); organic peroxides (grapes: tert-butyl perbenzoate) and fungicides (broccoli: mancozeb). The extract concentrations that caused these activities ranged from 0.05% (230 μM) for garlic (similar to oxi-bendazole at 10 μM) to 4.17% (11,175.6 μM) for sweet potatoes (similar to INCB-018424/ ruxolitinib at 10 μM). Overall, 6 of the 11 hits identified occurred at the 0.46% concentration.

Table 5.

Supervised similarity search hits from BioMAP reference compound library

Test Agent (concentration, %)	Compound Hit (concentration)	Compound Activities And/or Uses	r
Broccoli (0.46)	Mitomycin C (370 ng/mL)	alkylating agent, DNA cross-linker	0.806
Broccoli (0.46)	Mancozeb (40 uM)	Broad-spectrum fungicide; anti-proliferative effects in multiple cell types	0.774
Beets (0.46)	Tannic acid (4.1 uM)	polyphenol, component in anti-allergen sprays	0.765
Kale (1.39)	ZM449829 (3.3 uM)	JAK3 inhibitor tool compound binds to ATP site on enzyme	0.831
Sweet Potatoes (4.17)	INCB-018424 (10 uM)	Ruloxitinib, JAK 1/2 kinase inhibitor that improved survival in myelofibrosis patients	0.813
Green Beans (1.4)	15-deoxy-delta12,14-Prostaglandin J2 (3.3 uM)	PPARγ ligand that promotes adipocyte differentiation, induces heme oxidase (HO-1) and is known to inhibit the NFκB pathway	0.839
Green Beans (0.46)	Tannic acid (4.1 uM)	polyphenol, component in anti-allergen sprays	0.804
Strawberries (0.46)	tert-Butyl perbenzoate (40 uM)	organic peroxide used as a radical initiator to induce polymerization	0.796
Garlic (0.05)	Oxibendazole (10 uM)	Antihelminthic	0.731
Grapes (0.46)	tert-Butyl perbenzoate (40 uM)	organic peroxide used as a radical initiator to induce polymerization	0.718
Apples (0.46)	Tin(II) Chloride (30 uM)	HIF-1α inducer	0.705
Sweet Potatoes (1.4)	15-deoxy-delta12,14-Prostaglandin J2 (3.3 uM)	PPARγ ligand that promotes adipocyte differentiation, induces heme oxidase (HO-1) and is known to inhibit the NFκB pathway	0.705
Blueberries (1.4)	TCMTB (1.5 uM)	TCMTB (2-(Thiocyanomethylthio) benzothiazole) is an antimicrobial pesticide	0.691
Bean Sprouts (0.46)	FICZ (12 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.68
Cauliflower (0.46)	XL147 (1.1 uM)	XL147 is an orally available inhibitor of PI3 kinase	0.664
Spinach (0.15)	FICZ (12 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.661
Ginger (0.15)	DuP 128 (1.1 uM)	DuP 128 is an inhibitor of acyl-CoA:cholesterol acyltransferase (ACAT)	0.655
Potatoes, p (0.46)	Histamine (10 uM)	Histamine is an agonist of Histamine Receptors	0.636
Squash (1.4)	Benz[a]anthracene (40 uM)	Benz[a]anthracene is a polycyclic aromatic hydrocarbon (PAH) and AhR (aryl hydrocarbon receptor) agonist	0.628
Soybeans (1.4)	Tris(2-ethylhexyl) phosphate (13 uM)	Tris(2-ethylhexyl) phosphate (TEHP) is a plasticizer	0.627
Carrots (1.4)	FICZ (37 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.622
Pineapple (0.15)	OEA (10 uM)	OEA (Oleoylethanolamide) is a dietary fat-derived lipid and activator of PPARα	0.62
Potatoes, u (0.46)	Tin(II) Chloride (30 uM)	Tin(II) Chloride is a metal containing inducer of HIF-1α	0.613
Cabbage (0.46)	FICZ (12 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.61
Parsley (1.4)	Folpet (40 uM)	Folpet is a broad spectrum, non-systemic fungicide	0.608
Onions (1.4)	Colchicine (1.1 uM)	Colchicine is a microtubule inhibitor	0.595
Celery (1.4)	Bilirubin (10 uM)	Bilirubin is a breakdown product of heme metabolism	0.584
Broccoli (0.46)	Benz[a]anthracene (1.5 uM)	Benz[a]anthracene is a polycyclic aromatic hydrocarbon (PAH) and AhR (aryl hydrocarbon receptor) agonist	0.581
Snow Peas (0.46)	FICZ (37 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.573
Kale (0.46)	FICZ (12 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.565
Tomatoes (1.4)	FICZ (12 nM)	FICZ is an agonist of the aryl hydrocarbon receptor (AhR)	0.543
Oranges (1.4)	IL-6 R alpha (1 nM)	Soluble IL-6Ralpha is component of the tripartate IL-6R signaling complex	0.511
Peppers (1.4)	2,4-Dinitrophenol (40 uM)	2,4-Dinitrophenol is an uncoupler of oxidative phosphorylation	0.504
Grapefruit (1.4)	UO126 (370 nM)	UO126 is an inhibitor of MEK kinase	0.459

Figure 4. BioMAP Unsupervised Similarity Search Results.

Results shown are from unsupervised searches for biologically similar compounds from the BioMAP® reference database of > 3000 compounds, biologics, approved drugs and experimental agents. Profiles generated based on log ratio outputs, comparing produce extracts (red) to reference compound library hit (blue) identified following similarity analyses. Apple, beet, broccoli, grape, green bean, kale, sweet potato, and squash extracts had significant similarity to compounds in the BioSeek profile library. Endpoints common to both the produce extract and the reference compound are annotated and detailed in the table at right.

Additional similarity hits for extracts (0.6 < r < 0.7) are noted in Supplementary Table S5. These hits occurred at concentrations ranging from 0.05% to 12.5%. Six of the extracts – bean sprouts, cabbage, carrots, kale, spinach, snow peas – returned similarities to 6-formylindolo[3,2-b]carbazole (FICZ), a tryptophan photoproduct identified as an endogenous AhR ligand (Oberg et al., 2005). Additional hits of potential interest for follow-up include noted similarities of blue-berries with 2-thiocyanomethylthiobenzothiazole (TCMTB), an anti-microbial agrichemical; onions with colchicine, a microtubule inhibitor; parsley with folpet, a broad-spectrum fungicide; soybeans with the plasticizer tris (2-ethylhexyl) phosphate (TEHP); and squash with polycyclic aromatic hydrocarbon benz[a]anthracene.

SVM mechanistic model hits

The pairwise similarity search above identified a list of reference profiles that are the most correlated to the test profile. This similarity can result from a match with the primary targets, or may be due to secondary or non-specific mechanisms. The use of classifiers, developed through machine learning, can help identify whether an observed similarity is due to the primary mechanisms. Thus, profiles were further analyzed by testing through SVM Mechanism Model Classifiers (Berg et al., 2013, see Materials and Methods). Three extracts – edamame, squash and sweet potatoes – returned hits that indicated classification of profiles into defined mechanisms: histone deacetylase activity and JAK inhibitors (Table 6). Additional hits that did not achieve significance but may warrant further investigation are provided in Supplementary Table S5. The profile of sweet potatoes (at 4.17% con-centration), which the similarity search had identified as highly similar to the profile for INCB-018424 (ruloxitinib, a JAK inhibitor), was also classified by the SVM mechanism models as a JAK inhibitor. The pro-files for squash (at 12.5% concentration) and edamame (at 12.5%concentration), which were not similar to any profiles in the BioMAP reference database (at correlation coefficients above 0.7), were classified by SVM models as HDAC inhibitors. Since extracts are mixtures of bioactive substances, it may be that other materials contributing to the overall profile shape resulted in a low similarity to HDAC inhibitors by the pairwise similarity search analysis.

Table 6.

Mechanistic Hits using SVM Mechanism Model Prediction Tool.

Test Agent (concentration, %)	Molar Concentration (μM)	Mechanism	Decision Value
Squash (12.5)	18875	Histone deacetylase (HDAC) (1.11 μM)	0.547
Edamame (12.5)	53994	Histone deacetylase (HDAC) (1.11 μM)	0.458
Sweet Potatoes (4.17)	11211	Janus Kinase Inhibitors (JAK Inhibitor) (1.11 μM)	0.588

Intake adjusted bioactive ratios using food use agrichemicals from ToxCast

ToxCast

Concentrations of the “active ingredients” in extracts at the LEC-10 shown in Table 3 can be compared against concentrations of the food use agrichemicals that caused activity in the BioMAP assays performed during Phase 1 of ToxCast (Houck et al., 2009) (Supplementary Table S6). Estimated concentrations of extract constituents associated with bioactivity (LEC-10) ranged from 33 μM to > 2000 μM. Potatoes, cabbage, spinach, and strawberries were the most bioactive (33, 36, 37, and 43 μM), followed by kale, tomatoes, apples and cauliflower (68, 71, 75, and 84 μM). Grapefruit, blueberry and grape extracts were the least bioactive with estimated LEC-10 values of 1242, 1808, and 2652 μM. The LEC-10 values for food use agrichemicals ranged from 1.5 to 40 μM. Zoxamide, a fungicide used on grapes and potatoes, and pyraclostrobin were the most bioactive with LEC-10 values of 1.5 μM(Wetmore et al., 2015; Wetmore et al., 2012). The remainder of the agrichemicals had LEC-10s of 13.3 or 40 μM (40 μM is the highest concentration tested) (see Table S2 in the Supplementary material). Overall, the estimated bioactive concentrations for phytochemicals were higher than agrichemicals. Only three extracts (potatoes, cabbage, and spinach) had bioactivity at concentrations lower than the maximum tested con-centration for the agrichemicals in the ToxCast screening workflow (40 μM). Thus, while the extracts induced changes in many more assay endpoints than the agrichemicals (Fig. 1), they were tested at concentrations that are estimated to be far higher than the individual agrichemicals. It is unknown whether the agrichemicals, if tested at higher concentrations, would induce broader suites of responses similar to the extracts. However, it was also observed that at higher concentrations, several of these compounds did produce cytotoxic hits in the assay.

An important distinction between the produce and the food use chemicals being evaluated here is the expected intake volume. To better relate how the bioactivity of the fruit and vegetable extracts in these in vitro assays compares to regulated agrichemicals used in crop protection, we used a metric that adjusts bioactivities by relative intake. The intake-adjusted bioactivity ratio Br,p is the ratio of expected agrichemical and produce bioactivity modified by the ratio of agrichemical and produce intake (see Materials and Methods). The intake-adjusted bioactivity ratio is a combined metric for exposure (intake) and bioactivity, not an oral equivalent dose. While this latter metric would account for pharmacokinetic properties, it is not possible to make these calculations with the available data. We computed this ratio for each combination of the produce items considered here and sixteen commonly used agrichemicals (Supplementary Table S6). For every produce item and agrichemical pair, the intake-adjusted bioactivity is less than one. Only six calculated intake-adjusted bioactivity ratios were > 0.001 (Table 7). The produce items with the highest intake-adjusted bioactivity ratios are grapes, cabbage, broccoli, blueberries, kale, and spinach. Agrichemicals with the highest intake-adjusted bioactivity ratios are pyraclostrobin, endosulfan, fludioxonil, and azoxystrobin.

Table 7.

Produce/agrichemical pairs with intake-adjusted bioactivity ratios above 10⁻³.

Produce	Agrichemical	Agrichemical LEC-10 (Lr, μM)	Produce LEC-10 (Lp, μM)	Mass fraction agrichemical in produce (Ir,p)	Mass fraction AI in produce (Ip)	Intake-adjusted bioactivity ratio (Br,p)
Grapes	Pyraclostrobin	1.5	2650	1.50E-06	1.68E-01	1.57E-02
Cabbage	Endosulfan	13	36000	5.22E-02	8.20E-08	4.35E-03
Broccoli	Pyraclostrobin	1.5	930	3.40E-07	5.76E-02	3.66E-03
Blueberries	Pyraclostrobin	1.5	1810	1.13E-01	2.60E-07	2.79E-03
Kale	Pyraclostrobin	1.5	68	9.47E-02	4.10E-06	1.96E-03
Broccoli	Fludioxonil	13	930	1.20E-06	5.76E-02	1.49E-03
Spinach	Pyraclostrobin	1.5	40	2.90E-06	5.39E-02	1.43E-03
Cabbage	Pyraclostrobin	1.5	36000	5.22E-02	3.00E-09	1.38E-03
Broccoli	Azoxystrobin	13	930	1.00E-06	5.76E-02	1.24E-03
Blueberries	Azoxystrobin	13	1810	1.00E-06	1.13E-01	1.24E-03

Discussion

In the ten years since the release of the NRC Report Toxicity Testing in the 21st Century (NRC, 2007), scientists in the U.S. and abroad have leveraged advances in in vitro testing, bioinformatics, and computational approaches to generate data that can inform assessment of potential human health hazards. The U.S. EPA’s ToxCast effort has generated an enormous amount of data in cell-based and cell-free systems for more than a thousand chemicals. A related effort underway in Europe, Horizon 2020, the Framework Programme for Research and Innovation, is dedicating over €80 million from 2014 to 2020 to stimulate innovative research, technological developments and translational efforts in the biomedical sciences.

As in vitro assays are reductionist models of in vivo systems, it is critical to understand the context of effects seen in vitro in order to appropriately use the resulting data. Obvious examples of important differences include simplified exposure conditions, i.e. direct addition of chemical to target cells, and lack of xenobiotic biotransformation and transport processes in certain cell models. While long term goals are to design assay systems that minimize such liabilities, in the short term it is useful to fully characterize existing systems to better understand and use experimental findings from them. A key approach for evaluating these assays is testing of reference compounds with known human effects as was done in previous work with the BioMAP platform (Berg et al., 2006; Berg et al., 2010; Houck et al., 2009; Kleinstreuer et al., 2014). Here, we expanded that effort by testing complex mixtures de-rived from commonly consumed foods regarded as safe (organic fruits and vegetables) and compared these to previously tested environmental chemicals.

Evaluating the biological activity and concentration-response trends for the fruit and vegetable extract mixtures

A comparison of the fruit and vegetable extracts with ToxCast chemicals indicated significant qualitative and quantitative differences in responses. First, the produce extracts showed activity across many more assay endpoints than the single ToxCast chemicals. On the sur-face, the produce extracts may appear to be more biologically active. However, when converted to a molar basis, the produce extracts were tested at 10 to 50-fold higher concentrations than the single ToxCast chemicals. When constrained to the maximum tested concentration for the ToxCast chemicals, the produce extracts did not show activity in as many assay endpoints. Nonetheless, when the full data set were analyzed using hierarchical clustering, the produce extracts and ToxCast chemicals showed distinct bioactivity profiles. On a quantitative basis, the produce extracts were on average 193-fold less potent (based on LEC-10s) than agrichemicals in the ToxCast library that are typically used on the corresponding fruit and vegetable crops.

Concentrations where cytotoxicity is elicited in many of the produce extracts are found to be significantly higher than the major metabolite concentration observed in human after oral consumption of that pro-duce as reported in literature. For example, ginger elicited toxicity at a concentration as low as 0.15% (300 μM) where it is reported (Yu et al., 2011) that the major metabolite concentration observed in humans after oral consumption of 2 g ginger is 0.03 μM which is 10,000 times lower than the observed concentration for cytotoxicity in our analysis.

BioMAP assays, like several other assays employed in ToxCast, are evaluated for effects in both the negative and positive directions. Whereas the vast majority of compounds evaluated to date through the ToxCast program showed activity in one (or neither) of these two directions, produce extracts showed activity in both directions in 14%(358 of 2610) of the assays. This phenomenon, described here as a composite concentration response, manifested typically as an increase in activity at low concentrations followed by decreases at higher concentrations. Although some of the differences in the composite con-centration response frequency may be due to the differences in the concentration response range tested, this pattern is consistent with the fact that these extracts are complex mixtures of multiple bioactive substances. Each phytochemical would be expected to have its own individual concentration-response characteristics and the union of these individual curves could result in complex composite concentration-response behaviors resulting from additive, inhibitory or synergistic interactions among the phytochemicals (Efferth and Koch, 2011). At this time, the mechanism of action behind these composite concentration responses is not clear and deserves further study, perhaps using both the extracts and several defined mixtures of bioactive phytochemicals. The difficulty in interpreting composite concentration response curves, and delineating the underlying mechanisms of action, highlight the potential challenges that will be encountered in the analysis of chemical mixtures.

Using the BioMAP platform to infer mechanistic activities of the fruit and vegetable extracts

The BioMAP platform panel employed here is comprised of 8 human primary cell-based systems each primed with cytokines, growth factors, and pathway stimuli to allow concurrent activation of multiple signaling networks (Berg et al., 2006; Berg et al., 2010; Kleinstreuer et al., 2014). Analysis of phenotypic changes following xenobiotic treatment and comparison of the resulting bioactivity profiles of reference com-pounds with known activities can provide a broad characterization of the biological activity across multiple cell types and pathways as well as identify similarities to reference compounds and infer mechanisms (Berg et al., 2006; Berg et al., 2013). For example, a comprehensive reference database has permitted the assignment of several relevant toxicity-related mechanisms including mitochondrial dysfunction, microtubule inhibition, lysosomal inhibition, AhR agonism, immunosuppression, and others (Berg et al., 2013). Data mining of the BioMAP data has also identified associations of assay endpoints with clinical effects and enabled mechanistic insights. Increased levels of tissue factor (CD142) in the BioMAP vascular inflammation model was associated with drug-induced thrombosis-related side effects, with analysis leading to a mechanism involving the process of autophagy in vascular cells (Berg et al., 2015).

In the fruit and vegetable extracts, some extracts showed significant similarity to selective pharmaceutical agents (Table 5). For example, kale extracts showed a similar bioactivity profile to the JAK3 inhibitor ZM449829 while sweet potatoes showed similar bioactivity to the JAK1/2 inhibitor INCB-018424. In contrast, other produce extracts showed similarity to compounds with more mixed mechanisms of action. For example, broccoli showed a similar bioactivity profile to mitomycin C. The originally proposed mechanism of action for mitomycin C is related to DNA alkylation; however, it has also been suggested that redox cycling, inhibition of rRNA, and thioredoxin reductase inhibition may contribute to the biological effects, suggesting a complex mix of mechanisms (Paz et al., 2012). Broccoli extracts showed a similar bioactivity profile to mancozeb. Mancozeb is known to increase re-active oxygen levels in immune cells leading to cell death (Srivastava et al., 2012), but, it also inhibits NFkB activation leading to down-regulation of immune-related responses (Corsini et al., 2006). However, while mancozeb inhibits mitochondrial function and metabolism; similarities between broccoli and mancozeb do not necessarily indicate that components of broccoli will have the same effects following in vivo exposure. Indeed, the FDA approved drug, metformin, used therapeutically for diabetic patients, is also classified as a mitochondrial inhibitor. Likewise, broccoli showed similar response patterns to the DNA damaging compound mitomycin C, which is an anti-cancer chemotherapeutic with applications in the treatment of stomach, pancreatic, and several other cancers.

Several similarities were noted between the responses to produce extracts and the responses to various endogenous and exogenous ligands of nuclear receptors (Table 5). The endogenous PPAR-γ ligand 15-deoxy-Δ12,14-Prostaglandin J2 showed a high degree of bioactivity profile similarity, as measured by Pearson’s correlation coefficient (r), to green beans (r = 0.84) and sweet potatoes (r = 0.70). The endogenous AhR ligand 6-formylindolo (3,2-b)carbazole (FICZ) showed similarity to a variety of produce extracts, including bean sprouts (r = 0.68), spinach (r = 0.66), carrots (r = 0.66), cabbage (r = 0.62), snow peas (r = 0.57), kale (r = 0.57), and tomatoes (r = 0.54). The endocannabinoid oleoylethanolamide is a natural ligand of the PPAR-α receptor and showed similarity to pineapple bioactivity (r = 0.62). Celery showed similarity to bilirubin (r = 0.58), which has been shown to activate both the pregnane X receptor (PXR) and the constitutive androstane receptor (CAR). The similarities between bioactivity of produce extracts and nuclear receptor ligands highlight that activation of these receptors cannot be directly interpreted as an indicator of toxicity. Indeed, PPAR-γ modulates metabolism and the AhR pathway is important in development, immune responses and tissue repair. Due to the extensive mechanistic characterization of lipophilic xenobiotics such as TCDD, AhR agonists have been mostly studied in terms of associated toxicity. However, recent efforts have identified agonists with anti-inflammatory and other therapeutic activities.

Analysis of endpoint-specific bioactivities indicate a consistent downregulation of targets associated with immunomodulatory, inflammatory and tissue remodeling activities across both the extracts and chemicals. Focusing on inflammation-related activities, strawberries, possessing one of the highest phytochemical contents across the extracts assessed, elicited downregulation across 14 endpoints in all 8 cell models (with many endpoint activities present and conserved across the models; Fig. 3). Alternately sweet potatoes, possessing the lowest phytochemical content, showed similar downregulation of 9 responses across all of the cell models. The extensive hits for anti-inflammatory activities using this particular platform for phytochemical mixtures is not necessarily surprising, particularly when one considers the high content of vitamin A precursors ɑ-carotene and β-carotene in sweet potatoes, and the high anthocyanidin content in strawberries. Inhibition of VCAM-1, IP-10, and MIG in the HDF3CGF system, is a signature that is common among antioxidants and Nrf2 activators, such as tannic acid, myricetin, and dimethyl fumarate (Gillard et al., 2015; Krajka-Kuzniak et al., 2015). Increased levels of Tissue Factor (TF/CD142) in the BioMAP 3C system, observed with bean sprouts, is associated with vascular autophagy and thrombosis, important processes in wound healing and angiogenesis (Berg et al., 2015). These potential modes of action that mimic endogenous processes, together with similarity search results showing similarity of extracts to chemicals associated with both toxic and therapeutic response, demonstrate that it is currently challenging to distinguish potentially therapeutic (nutraceutical) activities from toxicities in these in vitro classification types of assays.

Intake-adjusted bioactivities

The lack of biokinetic information for these mixtures presents a major challenge for interpretion of the in vitro responses to these pro-duce extracts. These extracts are uncharacterized mixtures, consisting of multiple (mostly unknown) compounds, each with different ab-sorption, distribution, and clearance properties. Thus, standard IVIVE approaches for approximating circulating concentrations of test sub-stances for specific dosage rates (mg/kg/day), becomes difficult if not impossible with these mixtures. The approach presented here assessed some level of integrated activity based on percent juice required to elicit a response in the test system and then provides a measure of relative activity of the natural extract in relation to that of pesticides normally used on particular produce. This comparison puts the potency of the extract and pesticides onto a common measure of intake but is clearly different from more quantitative IVIVE approaches that calculate an oral equivalent dose (OED) for a steady-state dosing (Wetmore et al., 2012). Due to the relatively small amounts of agrichemicals ap-plied to the crops on a mass basis, the intake-adjusted bioactivities of the agrichemicals were in every case < 1% of those observed for the produce extract. As the intake-adjusted bioactivity ratios do not account for differences in internal kinetics and bioavailability, these results do not provide a direct comparison of potency (i.e., oral equivalent dose). However, they do indicate that the differences in exposure would likely ensure that the agrichemicals would have considerably lower biological activity in vivo than the fruit and vegetables themselves when consumed as part of a healthy diet.

Higher-throughput testing with mixtures versus individual compounds

A variety of new approach methodologies (NAMs) are now available to replace animal tests and optimally place safety testing on a course to collect more relevant risk-related information using human cells and human cellular components. The NRC (2007) discussed these assay readouts in relation to determining if test compounds affected toxicity pathways, defined as normal biological pathways where a perturbation could lead to adverse consequences if the perturbation were sufficiently large and maintained for a sufficient duration of exposure. The report also discussed that a critical component in applying these new assays would be in understanding the risk context. The differences in context and approaches for using these NAMs with mixtures versus single compounds needs to be more fully discussed and refined. Some efforts describe development of tiered approaches for use in specific decision contexts (e.g., compound prioritization) that rely on collection and analysis of readouts from simpler assays (Thomas et al., 2013). Fig. 5 shows how NAM-based screening approaches and dosimetry evaluations differ for mixtures compared to individual compounds. With single test compounds, initial evaluations could include computational models for determining thresholds of toxicological concern (TTCs), read-across for hazard identification, and high-throughput exposure modeling for prioritizing compounds for further testing. With mixtures, initial evaluations are likely to be more restricted to considerations of any work with similar mixtures, such as those derived from other foodstuffs, herbal medicines, active botanicals, etc. Initial dosimetry comparisons for mixtures would have to be placed on an intake adjusted measure of dose, as we have done here. After prioritization level assessments, more directed second-tier assays could use various NAM technologies to look at the responses in various assays - moving from broadly based assessments of responses such as transcriptomics in ex-posed cells, to assays with broad biological coverage, such as BioMAP, and then on to testing with NAMs for specific toxicity pathways. Higher-throughput characterization with mixtures could test defined mixtures or components if they were known and available. Dosimetry comparisons would remain based on intake adjusted doses or be refined as key mixture components are identified and measured. With identification and quantification, IVIVE could be performed with major bioactive components from the mixtures.

Figure 5. Challenges in conducting higher throughput assessments for mixtures using new approach methodologies.

The three dashed boxes show the steps from de novo assessments based on limited knowledge of biological targets or similarity to other mixtures through high-throughput characterization of the mixture using various in vitro assays and then on to interpretation of results in a risk-related context. In the absence of information on individual components, dosimetry with mixtures becomes limited to intake adjusted measures of activity and observed dose-response behaviors of the assay readouts will reflect the contribution of multiple bioactive components. The nature of the mixture, the potency of the response with respect to intake adjusted measures of dose and public health concern would determine the need to move from the mixture to examining individual subfractions or individual components.

BioMAP for fruit and vegetable extracts

The relationship of observed bioactivity in in vitro assay platforms and possible adverse responses during human exposures is far from obvious, depending, in part, on the nature of the interactions evaluated by the test platform and the relationship of the interactions measured in the test to possible adverse outcomes. The BioMAP technology provides a great breadth of information on altered cell signaling to compare unknowns with patterns of activity of a broad library of previously tested compounds across 8 cell systems. In this way, results with the BioMAP platform are intended to support comparisons to other com-pounds as indicators of possible avenues for therapeutic intervention and for broad information on possible pathway targets. They do not, however, directly identify key events associated with specific adverse outcome pathways, nor are they direct proxies for in vivo toxicity.

Agrichemicals used in the cultivation and processing of produce are designed to be biologically active for their intended purpose. These bioactivities, however, can lead to adverse effects in humans and animals when the compounds are ingested at sufficiently high intake levels and the exposures persist for a sufficiently long time. Nonetheless, the determination of specific bioactivities, by itself, should not be regarded as a marker for adversity under all use conditions. This caveat becomes especially clear when comparing the agrichemical residue and produce extracts.

Our results with produce extracts, which have broad activities in this assay platform, provide a clear example of the care necessary in interpreting these perturbations as an indicator of likely risks to hu-mans. Fruits and vegetables are well-established to have positive effects on health, although the full suite of bioactivities of these extracts across test platforms have not been systematically examined. In the BioSeek platform, the contributions of bioactivities of the produce material, calculated with our Br,p ratios, greatly exceed the contributions of agrichemical residues for expected exposures.

Conclusions

In this study, the biological activity of complex mixtures, i.e., fruit and vegetable extracts, were characterized using the BioMAP assays and the results were compared with agrichemicals tested in ToxCast. The intent of the comparison was to provide better context for the bioactivities ob-served in the ToxCast chemicals since the produce extracts represent substances to which humans are routinely exposed and are generally considered safe. We were also able to consider the manner in which high throughput assays could be used to explore and quantify bioactivity of mixtures. The fruit and vegetable extracts clearly affected multiple BioMAP endpoints. This reinforces earlier conclusions that bioactivity does not necessarily equate with an adverse response (Judson et al., 2011; Karmaus et al., 2016; Wetmore et al., 2013). An understanding of this distinction is important for toxicity testing in the transition from apical, histological endpoints to molecular and pathway-based responses. How-ever, the results also demonstrated that the produce extracts had qualitatively and quantitatively different bioactivity profiles than the single chemicals. Many of the observed differences are likely due to the fact that the extracts were tested at higher concentrations on a molar basis and are complex mixtures of bioactive phytochemicals that may result in distinct bioactivity profiles due to interacting components. Dose-response behaviors in these extracts are likely to arise from the superposition of responses to diverse compounds. The results highlight the potential challenges that will be faced when the new high-throughput toxicity testing approaches are applied to complex mixtures. Equally important is con-sideration of pharmacokinetics in relation to target tissue concentrations and how this information can be used to both guide interpretation of in vitro results and to relate real-world exposure scenarios to those exposures anticipated to initiate a response.

Supplementary Material

Supplement1

Supplementary Table S1 - Agrichemical/mycotoxin/metal analyses; juicing methods, yields, molar conversions, pH, osmolality, extract phytochemical content

Supplementary Table S2 - USDA Agrichemical Residue Collated Results

Supplementary Table S3 - LogRatio values, LEC values from F/V work; Sig Envelopes; Cytotoxic hits

Supplementary Table S4 - Details of composite concentration response analysis.

Supplementary Table S5 - Complete Listing of SVM Hits

Supplementary Table S6 - LEC-10 concentrations for food-use pesticides

Supplementary Table S7 - Molecular weight determined during a review of 48 phytochemicals screened in ToxCast

Supplementary Table S8 - DiscoverX/BioSeek BioMAP profiles for all fruits and vegetables at all concentrations

Highlights.

• Bioactivity of juices extracted from 30 organic produce samples was assessed using the BioMAP panel of in vitro assays.

• Produce samples were 10 to 50-fold less potent than pure compounds on an estimated molar basis.

• Produce samples exhibited extensive assay bioactivities when compared to pure compounds screened similarly in EPA’s ToxCast Program, but when constrained to the maximum testing concentration of the ToxCast chemicals, the produce extracts did not show activity in as many assay endpoints.

• Produce extracts displayed more complex concentration-response behavior than pure compounds.

• A dose metric is proposed to rescale bioactivity values for food and associated food-use agrichemicals for relative intake.