Mixture toxicity and gene inductions: Can we predict the outcome?

Abstract

As a consequence of the nature of most real‐life exposure scenarios, the last decade of ecotoxicological research has seen increasing interest in the assessment of mixture ecotoxicology. Often, mixtures are considered to follow one of two models, concentration addition (CA) or response addition (RA), both of which have been described in the literature. Nevertheless, mixtures that deviate from either or both models exist; they typically exhibit phenomena like synergism, ratio or concentration dependency, or inhibition. Moreover, both CA and RA have been challenged and evaluated mainly for acute responses at relatively high levels of biological organization (e.g., whole‐organism mortality), and applicability to genetic responses has not received much attention. Genetic responses are considered to be the primary reaction in case of toxicant exposure and carry valuable mechanistic information. Effects at the gene‐expression level are at the heart of the mode of action by toxicants and mixtures. The ability to predict mixture responses at this primary response level is an important asset in predicting and understanding mixture effects at different levels of biological organization. The present study evaluated the applicability of mixture models to stress gene inductions in Escherichia coli employing model toxicants with known modes of action in binary combinations. The results showed that even if the maximum of the dose–response curve is not known, making a classical ECx (concentration causing x% effect) approach impossible, mixture models can predict responses to the binary mixtures based on the single‐toxicant response curves. In most cases, the mode of action of the toxicants does not determine the optimal choice of model (i.e., CA, RA, or a deviation thereof).

Mixture, Escherichia coli, Stress gene, Concentration addition, Response addition

INTRODUCTION

Real‐life exposure scenarios of ecosystems and humans to toxicants almost exclusively involve mixtures of chemicals. Nonetheless, until recently, traditional ecotoxicology has concentrated on obtaining a better understanding of the mechanisms underlying the bioavailability of single compounds and their toxic effects on organisms. Today, most studies still focus on the impact of pure chemicals on the different levels of biological organization [1], sometimes taking into account the presence of factors modulating bioavailability [2–6], or they assess the impact of the whole environmental exposure without examining the activity of the individual compounds [7]. The former approach clearly adds to our understanding about the environmental impact of pollution, but it is not able to fully resolve the invariable complexity of environmental exposure, in which a multitude of physicochemical factors act together.

Mixture toxicity is a complex issue and remains poorly understood. This is illustrated by the fact that most of today's environmental legislation is based on risk assessment of single compounds or on the concentration addition (CA) model, which has only limited validity. Mixture toxicity can be influenced by different interactions of the compounds composing the mixture and the cellular or organismal system. Chemical reactions between different compounds can directly alter the toxicity of one or more constituents of the mixture, whereas interference with cellular systems, such as uptake, transport, and receptor binding, can cause toxicants in a mixture to react in, for example, an additive, synergistic, potentiating, or inhibiting manner.

It generally is accepted that two models, CA and independent action (IA; or response addition [RA]), have a broader application and are suited to predict mixture toxicity from the toxicities of the single toxicants. First described in 1926 [8], CA is applied when the chemicals in the mixture act through the same cellular mechanism on the same target (i.e., have the same mode of action). In this case, the effect of the mixture can be predicted from the known toxic units (i.e., the concentration of a compound divided by its x% effect concentration [ECx]) of all compounds in the mixture. Hence, given two compounds with the same toxicological mode of action, one can replace the other in the mixture without having an effect on the overall toxicity—that is, the sum of all toxic units in the mixture equals one (see Eqn. 1). Thus, the basis of the CA model is

where c_i equals the concentration of compound i and ECx_i represents the x% effect concentration for compound i.

The alternative model of IA [9] is valid for mixtures of toxicants with different modes of action and different targets, assuming no overlap or influence between each other. In this case, the overall effect of the mixture can be calculated from the effects of the individual toxicants at their respective concentrations (Eqn. 2). This, the basis of the IA model is

where E(C_mix) equals the total effect of the mixture and E(C_i) is the effect of compound i.

Note that both models assume the mixture under study is fully described in its chemical composition and that the (complete) dose‐response curves of all compounds in the mixture are known. Both models are extensively covered in the literature both theoretically [10] and as applied to different cases studies [11–15]. The scientific community has discussed which of both models would be the most widely applicable and how the best model can be chosen based on the specific traits of the mixture components only. De Zwart and Posthuma [16] evaluated a series of recent studies comparing the CA and RA models and concluded that although the CA model seems to be the most widely applicable, model selection based on the assumed toxicological mode of action suffices for valid prediction of the mixture toxicity, even in cases when the exact mechanism of toxicity is not known.

When using and interpreting either or both models, however, one must be aware that they are mere simplifications of a very complex reality. Several cases in the literature illustrate mixture effects that do not fit either model [17,18]. Recently, approaches have been put forward to evaluate deviations from either or both models [19]. Jonker et al. [20] consider both reference models as being two independent mathematical approaches to which mixture data can be compared. Within these reference approaches, those authors define four biologically relevant alternative models in a nested setup, and they present a detailed description of possible deviations and their mathematical expressions. First, in the no‐deviation model, the data fit the CA or IA model. In the second model, synergism or antagonism (SA), the measured mixture response is either above or below the predicted effect throughout the complete concentration range, assuming a perfect fit to either the CA or IA model. In the third, dose‐level dependency (DL), deviation from the reference model differs between the high and low concentration range, and in the fourth, dose‐ratio dependency, the deviation is determined by the actual ratio of the components in the mixture. Dose‐level or dose‐ratio deviation from the standard models directly implies that in function of the concentrations of the compounds in a mixture and/or their respective ratio, the mixture behavior can be different (e.g., antagonistic at the lower end of the concentration scale and synergistic at the higher end).

In the present study, a bacterial gene‐profiling assay was used. The assay consists of 14 bacterial, class 1 bioreporters, all carrying a different stress promoter fused to the complete β‐galactosidase gene (lacZ). The promoters are induced by different triggers, such as oxidative stress, osmolar imbalance, membrane and DNA damage, and protein perturbation. Additionally, growth inhibition is measured, thus allowing comparison of responses at different biological effect levels. The goal of the present study was to determine whether gene‐profiling assays can be used to study the toxicological mode of action of (binary) mixtures, whether effects at the gene‐induction level can be predicted from the single‐compound exposures, and if agreement with the higher‐level effect of growth inhibition is present.

Determination of comparable ECx levels is not possible for end points without a known maximum response level, as is the case for gene expression‐based systems (Fig. 1). Figure 1 illustrates that the median effect level can differ for different compounds even though the same end point is considered. Because the maximum induction level for every gene promoter is not only determined by its intrinsic regulatory mechanism but also is dependent on the nature of the inducing compound, a classical toxic unit approach is not applicable. As a consequence, the mixture models described before will be considered only as regression models to which data can be fitted; if a significant fit is obtained, the mixture can be considered to behave as described by that particular model. The values of the different regression factors, however, such as the median effect concentration (EC50), will have no true biological meaning and will not be considered as such.

Fig. 1.

Example of typical gene‐induction dose–response curves for three different compounds for the same promoter, illustrating different slope and different maximum responses and, therefore, different median effect levels (EC50s).

MATERIAL AND METHODS

Bacterial strains and stress gene assay

The bacterial strains and the complete stress gene assay have been described in detail by Dardenne et al. [21]. The present study includes an in‐depth assessment of the variability of the assay in terms of standard deviation, inter‐ and intraassay coefficients of variation, limits of detection, signal to noise ratio, and other classical performance characteristics of each promoter. All bacterial strains, except the SOS chromotest strain, are based on the Escherichia coli K12 derivative SF1 that contains the lac 4169 mutation, deleting the complete lac operon, and rpsL, rendering the strains resistant to streptomycin. All reporter constructs are promoter::lacZ fusions present as single copies on the bacterial chromosome. The PQ37 strain, as described for the SOS chromotest [22], is derived from E. coli GC4436 and carries the lacZ gene under control of the SfiA operon, which belongs to the SOS response system. This strain will be referred to as SfiA.

The stress gene promoters used in the present study respond to a variety of stress conditions, such as DNA damage, protein perturbation, oxidative stress, and growth arrest (see Table 1).

Chemicals

All chemicals used in the stress gene assays as toxicants (Table 2) were of at least analytical quality. Stock solutions of model toxicants were made up in ultrapure Milli‐Q® water (Millipore, Brussels, Belgium) except for pentachlorophenol (PCP; ethanol pro analysis). All PCP exposures contained 5% ethanol (actual exposure concentration), as did the nonexposed controls. All other chemicals were supplied by Merck (Darmstadt, Germany) or UCB (Brussels, Belgium) unless otherwise stated.

The present study used a number of model compounds with known mode of action in binary mixtures. All exposures were performed in half‐dilution series from the top concentration down. In total, seven doses were tested, plus an exposure blank (solvent only). The mixture composition and concentrations are given in Table 2. The top concentrations in every mixture equaled the top concentrations in the pure‐toxicant exposures. For practical reasons, the mixtures were restricted to a constant dose ratio. In other words, the ratio of both concentrations was constant in every exposure (Fig. 2); therefore, the dose–ratio deviation could not be tested for either base model. Testing for DL, however, remained possible.

Table 1.

Stress gene promoters fused to the lacZ reporter gene and their major inducers

Promoter	Gene product/Function	Responsive to	Reference(s)
KatG	Hydrogen peroxidase I	Oxidative stress	[25]
MicF	Antisense RNA to 5′ OmpF	Membrane integrity, osmotic stress	[26]
OsmY	Periplasmic Protein	Osmotic stress	[27]
UspA	Universal stress protein	Growth arrest	[28]
RecA	General recombination and DNA repair	SOS response	[29]
Zwf	Glucose‐6‐phosphate dehydrogenase	Oxidative stress	[30]
ClpB	Proteolytic activation of ClpP	Protein perturbation	[31]
UmuDC	DNA repair	Radiation and/or chemically induced DNA damage	[32]
MerR	Regulation of the mercury resistance operon (mer)	Heavy metals	[33]
Ada	Adaptive response to alkylation	DNA damage, mainly methyl adducts	[34]
DinD	Unknown function within the DNA damage inducible response	DNA damage	[35]
Soi28	Superoxide inducible gene	Superoxide radical generating agents	[36]
Nfo	Endonuc lease IV	Single‐stranded and double‐stranded DNA breaks, oxidative DNA damage	[37]
SfiA	Inhibitor of cell division	SOS response	[38]

Calculations

The enzyme activity and the fold‐induction (FI) at any given dose i for every strain were calculated using Equations 3 and 4 based on the optical densities (ODs) taken at several time points:

where PE is postexposure, PD is predose, SE is start of exposure, SA is start of β‐galactosidase assay, and EA is end of β‐galactosidase assay.

Introduction into the equation of the OD_600nm obtained before dosing corrects for possible color interference of the dosed toxicant. The equation assumes exponential growth during the exposure phase. Wherever the sum of fold‐inductions (FI_total) is used, Equation 5 applies:

where n is the number of significantly induced promoters.

Statistical evaluation

All calculations were performed using standard statistical algorithms provided in Excel® 2003 SP2 (Microsoft, Redmond, WA, USA). Inductions were considered to be significant based on the following evaluation: Signal significantly different from the blank (Dunnett's test, p < 0.05), r² > 0.5 (0.05 confidence level for the correlation coefficient at six degrees of freedom [n = 8] in a linear model), and positive nonzero slope (p < 0.05). The best‐fit model was chosen by a least‐squares regression and consequent comparison of model deviations versus the base model using a χ² test. Additionally, according to Occam's razor (i.e., the principle of parsimony), which states that the explanation of any phenomenon should make as few assumptions as possible, the least‐parameterized model was chosen as the optimal fit.

RESULTS

Table 3 lists the FI values of all pure‐compound and mixture exposures and the growth inhibition data dealt with in the text. All mixtures, except the NaN₃‐PCP mixture, induced all promoters that were induced by their constituents in the single‐compound exposures. A typical result is given in Figure 3, showing the induction profiles of paraquat (PQ) (Fig. 3A), hydrogen peroxide (H₂O₂; Fig. 3B), and the mixture of both (Fig. 3C). The mixture profile clearly shows induction of KatG, Zwf, UmuDC, Soi28, Nfo, and SfiA and, thus, reflects the inductions of the pure compounds.

Although in some mixtures the FI increases or decreases as compared to the individual exposures (see below), no mixture was observed to induce promoters that were not induced by the pure compounds. The inductions of the NaN₃‐PCP mixture were completely inhibited, whereas the constituents behaved as expected from their described modes of action, with NaN₃ inducing the membrane and protein perturbation reporters MicF and ClpB, respectively, and the UmuDC promoter, indicating DNA damage, and with PCP inducing MicF and ClpB.

Table 2.

Mixture compositions and their respective lesions. Numbers in parentheses: top concentrations used both for pure compound and mixture exposures. Exposures were performed in one half dilution series (seven dilutions) from the top down. Top concentrations in the mixtures equal the top concentrations in the pure toxicant exposures

Combination (Top concentration in exposure)		Lesion	Mode of action
Azothymine (25 ng/ml)	Methyl methanesulphonate (100 ng/ml)	DNA damage	Different
Azothymine (25 ng/ml)	Mitomycin C (250 ng/ml)	DNA damage	Similar
Azothymine (25 ng/ml)	Nalidixic Acid (500 ng/ml)	DNA damage	Similar
Paraquat(100 μg/ml)	Hydrogen peroxide (250 μM)	Oxidative stress	Different
Paraquat (100 μg/ml)	Cadmium (100 μg/ml)	Oxidative stress–Heavy metal toxicity	Different
Paraquat (100 μg/ml)	Pentachlorophenol (1.5 μg/ml)	Oxidative stress–Energy depletion	Different
CuCl2 (250 μg/ml)	Hydrogen peroxide (250 μM)	Oxidative stress–Heavy metal toxicity	Different
NaN3 (10 mM)	Pentachlorophenol (1.5 μg/ml)	Protein perturbation–Energy depletion	Different

Fig. 2.

Scheme of all tested mixtures. Axes denote the concentration of both compounds.

Fitting the data

All single‐compound and mixture responses were fit to both models and deviations independent of the theoretically predicted best (e.g., CA would be favored over IA in case of the same lesion and chemicals with the same mode of action). Table 4 lists the parameters of the best‐obtained fits within every base model for all mixtures; Figure 4 shows the correspondence of the measured data to the model predictions. Inspecting both the table and the graph shows clearly that in many cases, both the CA and IA models (albeit with or without deviation) are able to predict the mixture response from the pure‐toxicant data. The r² of fits obtained through the CA model are, in general terms, close to those of the IA model fit. In those cases, the selection of the best model was done by applying Occam's razor. This is clearly illustrated through the azothymine (AZT)–methyl methanesulfonate (MMS) mixture (FI_total in Table 4 and A1 in Fig. 4), for which within the CA model the more complex DL dependency fits best, whereas using the IA model, the base model attains the same r² (0.99); therefore, the latter is said to be the model of choice. Only a few cases do not fit either or both of the models. The RecA response in the AZT–mitomycin C (MitC) mixture does not fit the IA model at all. The signal does fit the CA model, but the relationship is poor. The PQ‐CdCl₂ mixture cannot be explained using the CA model or any of its deviations, but it can be predicted, at least in part, by using the IA DL model (see below for a more detailed discussion). Because the NaN₃‐PCP mixture does not show any responses and, consequently, cannot be fitted, this mixture is absent from the graphs.

Overall, only four of eight mixtures, or 10 of 19 end points measured, have their best fits within the same model. No mixture fits its different responses within the same functional form (CA or IA).

Same lesion, different modes of action

Azothymine and MMS both induce DNA damage, but they exert this activity through different mechanisms. Azothymine inhibits DNA synthesis as it is incorporated into the growing DNA strand, preventing elongation; MMS is a potent meth‐ylating agent. Azothymine exclusively induces the RecA and UmuDC promoters, whereas MMS also induces the adaptive response through its methylating activity. Table 3 shows the inductions of the RecA, UmuDC, and Ada promoters after exposure to MMS, AZT, and the mixture of both. Azothymine is determinative for both the RecA and UmuDC mixture responses. The lower UmuDC response in the mixture exposure (20.8 at the top dose vs 26.7 and 4.5 for AZT and MMS, respectively) seems to be compensated for by the Ada induction as the FI_total all significantly induced promoters is still well predicted by the pure‐compound inductions (A1 in Fig. 4). Both DNA‐damaging agents, as well as their mixture, did not provoke any significant growth inhibition during the exposure phase (data not shown).

The FI_total of AZT and MMS shows mathematically equally good fitting results to the CA DL model (r² = 0.9921, χ² = 0.0270) and the IA model (r² = 0.9900, p < 0.0001). Choosing the simplest model as the best, the IA model should be preferred. The same holds for the RecA induction with the CA SA model (r² = 0.9580, χ² = 0.0242) and the IA fit (r² = 0.9525, p < 0.0001). The UmuDC signal fits the CA DL model (r² = 0.9874, χ² = 0.0002) comparable to the IA SA model (r² = 0.9784, χ² = 0.0082); again, the latter is the least‐parameterized model and should have preference. Because throughout the dose range the mixture response is approximately 10% less than the response expected from the pure compounds in a nondeviating IA model, the response is defined as being antagonistic.

Paraquat and H₂O₂ induce oxidative stress through, respectively, redox cycling and direct generation of reactive oxygen species and consequent formation of radicals. Paraquat toxicity is counteracted by induction of the soxRS operon, inducing zwf, soi28, and nfo. Additionally, the SfiA promoter, a member of the SOS response pathway, is induced. Hydrogen peroxide induces the second oxidative stress response pathway in E. coli, the oxyR operon, represented here by the KatG promoter, and also provokes induction of UmuDC, which like SfiA is part of the E. coli SOS response. Because both compounds do not share any promoter induction within this assay, a direct comparison per promoter, as for MMS and AZT, is not meaningful. The sum of the induced promoters fits best to the CA DL model (r² 0.9963, p < 0.0001). Based on the r² (0.92 for CA SA and 0.95 for IA DL), the growth inhibition data fit best to the IA DL model; however, the difference was very small (F2 in Fig. 4). Therefore, the CA SA model is preferred, because it is less parameterized.

Copper(II) is an oxidative stressor inducing the soxRS operon (the Zwf, soi28, and Nfo promoters in our assay) in E. coli. In combination with H₂O₂ (inducing the oxyR regulon, represented here by the KatG promoter), a gene‐induction profile similar to the PQ‐H₂O₂ mixture is obtained. As a consequence, only the FI_total and growth inhibition fit are meaningful, because no common promoters are induced. The FI_total of the mixture (G1 in Table 4 and Fig. 4) fits best to the IA base model, whereas the growth‐inhibition data fit best to the CA base model (G2 in Table 4 and Fig. 4).

Same lesion, same mode of action

Azothymine and MitC both have DNA as their primary target and exhibit DNA damage through inhibition of the elongation of DNA synthesis. Their gene‐induction profiles are limited to induction of RecA and UmuDC (Fig. 5), as is the case for nalidixic acid (NAA). The latter inhibits DNA synthesis as well, yet it does so by inhibiting DNA gyrase, which is its primary target.

Table 3.

Fold inductions (FI) and/or growth inhibition (standard error) after exposure to the respective compounds and mixtures. FI_total equals the sum of all inductions corrected for the blank signal. The doses represent a one half dilution series from dose 7 down as described in the materials and methods section. Inhibition is the average growth inhibition for all strains versus the non exposed cells. AZT = azothymine; MMS = methyl methanesulfonate; MitC = mitomicin C; NAA = nalidixic acid; PQ = Paraquat; PCP = pentachlorophenol. Numbers in parentheses represent standard error

		Fold induction/growth inhibition
		Blank	Dose 1	Dose 2	Dose 3	Dose 4	Dose 5	Dose 6	Dose 7
AZT/MMS
RecA	AZT	1	5.3 (0.1)	4.3 (1.3)	5.7 (0.3)	6.0 (0.2)	6.3 (0.2)	6.8 (0.2)	7.9 (0.3)
	MMS	1	1.2 (0.0)	1.3 (0.1)	1.3 (0.1)	1.5 (0.1)	1.7 (0.1)	1.9 (0.3)	2.3 (0.5)
	Mixture	1	5.4 (0.1)	6.0 (0.3)	5.8 (0.2)	6.5 (0.4)	7.4 (0.3)	7.6 (0.1)	8.7 (0.2)
UmuDC	AZT	1	5.1 (0.1)	7.6 (0.2)	11.0 (0.7)	14.2 (0.9)	22.3 (2.3)	22.2 (1.1)	26.7 (1.3)
	MMS	1	1.2 (0.2)	1.4 (0.3)	1.6 (0.2)	1.7 (0.0)	2.2 (0.1)	3.7 (0.2)	4.5 (0.8)
	Mixture	1	5.1 (0.3)	7.0 (0.5)	10.5 (0.7)	13.8 (0.6)	20.8 (3.3)	20.2 (1.7)	20.8 (1.8)
Ada	AZT	1	1.1 (0.1)	0.9 (0.2)	1.0 (0.1)	1.0 (0.1)	1.1 (0.1)	1.2 (0.2)	1.0 (0.1)
	MMS	1	1.2 (0.2)	0.9 (0.3)	0.9 (0.2)	1.0 (0.3)	1.0 (0.3)	2.2 (1.3)	9.4 (1.9)
	Mixture	1	0.9 (0.1)	1.1 (0.1)	1.3 (0.1)	1.5 (0.1)	2.0 (0.1)	4.5 (0.5)	11.8 (1.1)
FItotal	Mixture	1	8.9 (0.4)	13.0 (0.7)	16.6 (0.8)	21.4 (0.8)	30.2 (3.3)	32.9 (1.8)	41.4 (2.3)
AZT/MitC
RecA	AZT	1	5.0 (0.1)	5.7 (0.1)	5.6 (0.3)	6.0 (0.3)	5.7 (0.1)	6.2 (0.0)	6.4 (0.2)
	MitC	1	1.1 (0.1)	1.2 (0.1)	1.3 (0.1)	1.7 (0.2)	2.3 (0.2)	3.4 (0.1)	4.8 (0.0)
	Mixture	1	4.7 (0.2)	5.2 (0.3)	5.3 (0.1)	6.5 (0.3)	6.3 (0.4)	5.3 (0.6)	6.4 (0.2)
UmuDC	AZT	1	5.0 (1.0)	6.7 (1.4)	9.8 (2.0)	14.2 (2.4)	19.0 (2.5)	24.9 (2.7)	31.9 (0.9)
	MitC	1	1.3 (0.0)	1.5 (0.0)	1.9 (0.1)	2.7 (0.0)	4.2 (0.1)	6.4 (0.1)	10.6 (0.2)
	Mixture	1	10.9 (0.6)	14.3 (0.6)	18.0 (0.6)	18.8 (1.8)	19.7 (1.6)	20.3 (0.8)	20.5 (0.7)
FItotal	Mixture	1	15.1 (0.7)	19.5 (0.7)	24.1 (0.7)	26.2 (1.8)	27.3 (1.7)	26.6 (1.0)	30.9 (1.6)
AZT/NAA
RecA	AZT	1	2.8 (0.0)	3.7 (0.1)	4.1 (0.0)	4.8 (0.2)	5.7 (0.3)	6.1 (0.2)	7.2 (0.3)
	NAA	1	1.0 (0.0)	1.2 (0.0)	1.4 (0.0)	1.6 (0.0)	1.6 (0.0)	2.1 (0.1)	3.5 (0.8)
	Mixture	1	2.9 (0.3)	3.4 (0.3)	3.6 (0.3)	3.5 (1.0)	4.7 (0.5)	6.3 (0.2)	7.2 (0.2)
UmuDC	AZT	1	3.2 (0.2)	4.8 (0.4)	7.1 (0.6)	10.9 (0.6)	12.7 (0.6)	13.0 (0.4)	13.1 (0.4)
	NAA	1	1.9 (0.0)	3.6 (0.1)	6.6 (0.1)	6.3 (0.1)	6.0 (0.1)	9.7 (0.2)	15.0 (0.3)
	Mixture	1	3.2 (0.2)	4.8 (0.4)	6.7 (0.5)	7.8 (0.7)	7.4 (0.3)	9.4 (0.4)	13.8 (1.7)
FItotal	Mixture	1	4.2 (0.4)	6.6 (0.6)	7.8 (0.7)	8.9 (1.3)	8.9 (0.7)	12.7 (0.6)	18.0 (1.7)
PQ/CdCl2
FItotal	PQ	1	7.7 (0.5)	12.0 (1.1)	15.6 (1.1)	19.9 (1.9)	21.8 (1.2)	24.6 (2.2)	26.1 (1.5)
	CdCl2	1	‐0.3 (0.2)	‐0.6 (0.2)	‐0.5 (0.2)	0.6 (0.2)	3.7 (0.4)	16.5 (1.2)	44.5 (3.9)
	Mixture	1	4.9 (0.1)	7.3 (0.3)	10.3 (0.2)	13.1 (0.3)	16.3 (0.4)	83.4 (0.9)	22.2 (10.0)
Inhibition	PQ	1	0.96 (0.001)	0.92 (0.001)	0.90 (0.001)	0.87 (0.001)	0.85 (0.001)	0.79 (0.001)	0.74 (0.001)
	CdCl2	1	0.87 (0.001)	0.87 (0.002)	0.81 (0.002)	0.76 (0.002)	0.76 (0.002)	0.66 (0.002)	0.43 (0.002)
	Mixture	1	0.89 (0.001)	0.84 (0.001)	0.77 (0.002)	0.67 (0.002)	0.56 (0.004)	0.30 (0.005)	0.18 (0.003)
PQ/PCP
FItotal	PQ	1	2.1 (0.2)	2.8 (0.2)	4.8 (0.2)	9.1 (0.3)	14.2 (0.9)	18.5 (0.9)	25.2 (1.0)
	PCP	1	1.5 (0.2)	1.4 (0.2)	1.7 (0.3)	2.2 (0.3)	4.1 (0.4)	7.4 (0.5)	7.8 (0.4)
	Mixture	1	2.4 (0.2)	3.0 (0.3)	3.6 (0.3)	5.8 (0.5)	5.0 (0.3)	6.1 (0.5)	3.4 (0.5)
Inhibition	PQ	1	1.01 (0.002)	0.94 (0.001)	0.88 (0.001)	0.86 (0.002)	0.92 (0.002)	0.92 (0.002)	0.89 (0.002)
	PCP	1	0.78 (0.005)	0.73 (0.004)	0.82 (0.005)	0.71 (0.005)	0.63 (0.004)	0.46 (0.003)	0.28 (0.002)
	Mixture	1	0.83 (0.002)	0.80 (0.002)	0.74 (0.002)	0.71 (0.002)	0.62 (0.002)	0.52 (0.002)	0.38 (0.002)
PQ/H2O2
FItotal	PQ	1	6.7 (0.3)	10.8 (0.5)	14.7 (0.7)	18.0 (0.6)	21.7 (1.4)	25.4 (1.4)	25.7 (1.5)
	H2O2	1	1.1 (0.2)	1.1 (0.1)	1.1 (0.2)	1.4 (0.2)	1.7 (0.2)	2.2 (0.3)	3.1 (0.2)
	Mixture	1	9.5 (0.6)	13.8 (1.1)	18.6 (1.7)	23.5 (1.5)	27.7 (1.7)	32.2 (1.4)	33.1 (2.4)
Inhibition	PQ	1	0.94 (0.002)	0.92 (0.003)	0.88 (0.002)	0.89 (0.001)	0.87 (0.001)	0.83 (0.002)	0.83 (0.002)
	H2O2	1	0.92 (0.001)	0.92 (0.001)	0.89 (0.001)	0.88 (0.001)	0.87 (0.001)	0.87 (0.001)	0.90 (0.001)
	Mixture	1	0.92 (0.002)	0.93 (0.003)	0.87 (0.003)	0.81 (0.003)	0.73 (0.003)	0.61 (0.002)	0.56 (0.001)
CuCl2/H2O2
FItotal	CuCl2	1	5.3 (0.2)	4.3 (0.2)	5.7 (0.3)	6.0 (0.3)	6.3 (0.3)	6.8 (0.5)	7.9 (0.6)
	H2O2	1	1.2 (0.4)	1.3 (0.3)	1.3 (0.4)	1.5 (0.7)	1.7 (0.3)	1.9 (0.2)	2.3 (0.3)
	Mixture	1	5.4 (0.2)	6.0 (0.2)	5.8 (0.2)	6.5 (0.1)	7.4 (0.3)	7.6 (0.2)	8.7 (0.4)
Inhibition	CuCl2	1	0.99 (0.002)	0.95 (0.002)	0.90 (0.002)	0.91 (0.002)	0.92 (0.002)	0.92 (0.002)	0.90 (0.001)
	H2O2	1	0.94 (0.001)	0.88 (0.001)	0.87 (0.002)	0.89 (0.002)	0.91 (0.002)	0.84 (0.002)	0.70 (0.001)
	Mixture	1	0.95 (0.001)	0.92 (0.001)	0.86 (0.001)	0.85 (0.001)	0.85 (0.002)	0.82 (0.001)	0.73 (0.001)
NaN3/PCP
FItotal	NaN3	1	2.9 (0.4)	7.2 (0.7)	9.1 (1.4)	10.6 (1.7)	11.8 (1.5)	12.3 (1.7)	14.3 (1.5)
	PCP	1	1.8 (0.2)	1.8 (0.3)	2.9 (0.3)	2.9 (0.5)	4.9 (0.4)	5.9 (0.5)	6.9 (0.5)
	Mixture	1	1.2 (0.2)	0.6 (0.2)	0.9 (0.3)	0.3 (0.5)	‐0.5 (0.3)	‐1.3 (0.4)	‐2.7 (0.2)
Inhibition	NaN3	1	0.75 (0.003)	0.63 (0.002)	0.44 (0.002)	0.37 (0.002)	0.32 (0.002)	0.29 (0.002)	0.19 (0.001)
	PCP	1	0.88 (0.007)	1.11 (0.012)	0.84 (0.009)	0.96 (0.008)	0.88 (0.008)	0.87 (0.009)	0.51 (0.004)
	Mixture	1	0.74 (0.002)	0.70 (0.003)	0.60 (0.002)	0.53 (0.002)	0.56 (0.002)	0.56 (0.003)	0.40 (0.002)

Fig. 3.

Bacterial gene‐profiling assay of paraquat (A), hydrogen peroxide (B), and the mixture of both (C). Respective mixture doses are equal to the pure‐compound concentrations. Note the difference in scale between A and B and between C and B because of the different maximum induction levels of the promoter. Standard errors are given in Table 4.

Neither the mixtures nor their constituents show significant growth inhibition during exposure. In general, the AZT‐MitC mixture (Table 3) exhibits lower gene‐expression responses at the higher doses on both the RecA and UmuDC promoters, as would be expected from the single‐compound exposures. This is illustrated by the UmuDC response, showing a FI of 31.9 and 10.6 for AZT and MitC, respectively, whereas the mixture only attains a FI of 20.5. The same trend is present throughout the dose range. The data fits of FI_total and UmuDC to the CA DL model show increasing deviation from the model prediction in the higher concentration range. The RecA signal is one of the few inductions that can be fitted to only one model (CA), although with a lower r² (0.92) as compared to other mixtures (B2 in Fig. 4).

The AZT‐NAA exposure (Table 3) is one of the four exhibiting the best‐fit results within one model (CA). Whereas UmuDC is strongly induced by both single‐component exposures showing a highest dose FI of 13.1 and 15.0 for AZT and NAA, respectively, the mixture response only at the highest dose exceeds the AZT induction (13.8). The NAA induction of UmuDC is only outcompeted at the low end of the dose range. The mixture response is antagonistic, as shown by the fits to the SA deviations of both the CA and IA models (CA SA is best).

Different lesion, different modes of action

Paraquat, a widely applied bipyridilium herbicide, is a redox cycling agent that causes oxidative stress and energy depletion. Escherichia coli tackles the adverse effects of PQ through induction of the soxRS operon. The cadmium ion is considered to be procarcinogenic and potentially induces a variety of lesions, but the primary response within our E. coli battery is induction of the Mer operon through upregulation of the MerR promoter. The FI_total of the mixture cannot be fitted by the CA model or any of its deviations and only gives a poor fit to the IA SA because of an extreme synergistic response at dose 6 (measured FI_total of 83.4 vs 49.9 as predicted by the model); at the lower dose range, the mixture behaves antagonistically. Remarkably, given the clear synergistic induction at the higher dose, the DL deviation does not show a significantly better fit.

Table 4.

Summary of fitting parameters. CA = concentration addition; IA = independent action; S/A = synergy ‐ antagonism; DL = dose level dependency; LDL = low dose level; HDL = high dose level. In case of best fit to one of the base models the p value is listed, in case of best fit to a deviation the χ² to the respective model is given. Inhibition = growth inhibition. Italic format represents the best‐fit model. A1 to H2 refer to the numbering in Figure 4

		Concentration addition				Independent action
		Deviation	r2	p/χ2	Functional form	Deviation	r2	p/χ2	Functional form
AZT/MMS
A1	FItotal	CA DL	0.99	0.0270	Synergism at LDL; antagonism at HDL	IA	0.99	0.0001	No deviation from base model
A2	RecA	CA S/A	0.96	0.0242	Synergism	IA	0.95	0.0001	No deviation from base model
A3	UmuDC	CA DL	0.99	0.0002	Antagonism at LDL; synergism at HDL	IA S/A	0.98	0.0082	Antagonism
AZT/MitC
B1	FItotal	CA DL	0.98	0.0001	Synergism at LDL; antagonism at HDL	IA S/A	0.93	<0.0001	Antagonism
B2	RecA	CA	0.94	0.0001	No deviation from base model				No valid fit
B3	UmuDC	CA DL	0.98	<0.000	Synergism at LDL; antagonism at HDL	IA DL	0.99	0.0001	Synergism at LDL; antagonism at HDL
AZT/NAA
C1	FItotal	CA S/A	0.96	0.0001	Antagonism	IA S/A	0.95	<0.0001	Antagonism
C2	RecA	CA	0.92	0.0001	No deviation from base model	IA S/A	0.92	0.0004	Antagonism
C3	UmuDC	CA S/A	0.94	0.0001	Antagonism	IA S/A	0.92	<0.0001	Antagonism
PQ/CdCl2
D1	FItotal				No valid fit	IA S/A	0.67	0.0002	Antagonism
D2	Inhibition	CA DL	1.00	0.0001	Synergism at LDL; antagonism at HDL	IA DL	0.98	0.0024	Synergism at LDL; antagonism at HDL
PQ/PCP
E1	FItotal	CA DL	0.98	0.0001	Antagonism at LDL; synergism at HDL	IA DL	0.98	0.0001	Antagonism at LDL; synergism at HDL
E2	Inhibition	CA	0.95	0.0001	No deviation from base model	IA S/A	0.96	0.0025	Antagonism
PQ/H2O2
F1	FItotal	DL	1.00	<0.0001	Synergism at LDL; antagonism at HDL	IA DL	0.99	<0.0001	Synergism at LDL; antagonism at HDL
F2	Inhibition	S/A	0.92	<0.0001	Synergism	IA DL	0.95	0.0011	Synergism at LDL; antagonism at HDL
CuCl2/H2O2
G1	FItotal	CA S/A	0.96	0.024	Synergism	IA	0.95	0.0001	No deviation from base model
G2	Inhibition	CA	0.88	0.0001	No deviation from base model	IA S/A	0.87	0.0160	Antagonism
NaN3/PCP
H1	FItotal				No gene inductions				No gene inductions
H2	Inhibition	DL	0.96	0.0400	Antagonism at LDL; synergism at HDL	IA DL	0.95	0.0018	Antagonism at LDL; synergism at HDL

Pentachlorophenol is an uncoupler of oxidative phosphor‐ylation, and in the bacterial gene‐profiling assay, it induces exclusively the MicF and ClpB promoters, indicating membrane instability and protein perturbation, respectively. The FI_total (E1 in Table 4 and Fig. 4) of PQ and PCP fits equally well to the CA and IA models (both within the DL deviation). The growth‐inhibition data are best explained by the CA base model (E2 in Table 4 and Fig. 4).

The DNA‐damaging agent NaN₃ is a chemically very reactive molecule causing protein denaturation effects. In E. coli, it induces ClpB, MicF, and UmuDC. The former two are in common with the PCP profile, but the latter is unique to NaN₃. The mixture of both is fully antagonistic, showing no gene induction at all, whereas growth inhibition fits the CA DL model best (H2 in Table 4 and Fig. 4).

DISCUSSION

Historically, the prediction of mixture toxicity from the effects of the individual compounds has been based on two assumptive models, CA and IA. The first model assumes that whenever toxicants have the same mode of action, they will be interchangeable in a mixture based on their respective toxic units. The latter model holds for toxicants with different modes of action and predicts the mixture toxicity based on the response of the individual constituents; thus, the alternative name is response addition (RA model). Both models were used successfully in a number of cases [11–13], but it generally is accepted that many mixtures show responses that deviate from predictions made using either the CA or IA model. Because of the increasing uncertainty in the low concentration range of the dose–response curve, this is even more so at low environmental exposure concentrations [23]. During the last several years, effort has been made to place both models at either side of a (continuous) scale, interconnected by deviation patterns; however, this approach has been unsuccessful so far.

Fig. 4.

Fits of all mixtures to the concentration addition (CA) and independent action (IA) models; the mixture identity and model (deviation) plotted are indicated in Table 3. The diagonal line represents the bisect of the graph (i.e., the theoretical perfect prediction).

Mixture toxicity studies based on quantitative gene‐expression data are, to our knowledge, absent from the literature. One of the reasons for this gap might be that an objective ECx value very often is impossible to obtain from gene‐induction measurements, because the maximum induction level of a pro‐moter::gene combination has no set maximum value common to all toxicants specific to that expression cassette. The maximum value attained is, among other factors, dependent on the cytotoxicity of the specific compound and the mechanism of the regulatory machinery of the expression cassette at hand; therefore, the EC50, in terms of the FI of a specific promoter, can differ between different toxicants, hampering ECx‐based approaches. Both the CA and IA models, however, depend specifically on these accurate ECx values. Looking at both models as being intrinsically different from one another, Jonker et al. [20] built a number of classical deviation patterns (e.g., synergism and DL) into the CA and IA model in a nested setup. Because of the nested approach, the deviations add increasingly more complex interactions to both the CA and IA models but simplify back to the base model when interactions are either not significant or absent. In the present study, a series of binary mixtures were assayed for the gene‐induction potential in a multi‐endpoint setup and the results fitted to the models of Jonker et al. [20]. The best‐fit model was chosen based on the regression parameters, and according to Occam's razor, in the case of equally good fits to two models, the least‐parameterized model is favored. All mixtures but two fitted to at least one model or deviation, showing r² values of greater than 0.90. The PQ‐CdCl₂ mixture shows a much poorer fit (r² = 0.67) because of the extreme merR response at higher cadmium levels, which clearly is potentiated by the presence of cadmium. At the higher level of biological organization, the growth inhibition data show a good fit to the CA model, deviating synergistically at a low dose level yet antagonistically at a high dose level, as opposed to the gene‐induction data. This result is in agreement with the findings of Cedergreen and Streibig [24], who showed that the choice of end point can determine the best‐fit model in exposures of Lemna minor to mixtures of herbicides. The NaN₃‐PCP combination, which was the second mixture not showing a good fit, behaved fully antagonistically on the gene‐induction level, silencing all responses; therefore, no model could be evaluated. The growth‐inhibition data fitted best to the CA DL model, indicating that the absence of inductions is not caused by complete complexation, reactivity, or any other mechanism making the chemicals unavailable or nontoxic to the bacterium. The exact reason for the silencing effect remains unclear.

Fig. 5.

Gene profile of mitomycin C, showing induction of RecA and UmuDC (SOS response). Standard errors are given in Table 4.

All other mixtures fit either model or one of the nested deviations with comparable regression parameters, making the choice for best model difficult and rather arbitrary. In some cases, the use of Occam's razor has a clear advantage. The FI_total of the AZT‐MMS mixture fits equally well to the CA DL deviation and to the IA base model. In that case, the simpler IA model clearly is the favored choice. For all three mixtures showing more then one induction data set (i.e., FI_total and induction of at least one promoter), the best fits are obtained within one model. In case of the UmuDC induction of the AZT‐MitC mixture, which fit best to the IA DL deviation, this seems not to be the case. The difference with the CA DL fit, however, is minimal and probably not significant.

The concept that the (non)equality of the mode of action of mixture constituents determines the model of choice, CA or IA, holds only in some cases, and as stated before, in our data set, the difference with the other model is minor. As a generalization based on the UmuDC signal in the AZT‐MitC mixture, all DNA‐damaging mixtures fit best to the model that would be predicted based on of the modes of action. Because none of these mixtures caused growth inhibition during the exposure phase, comparison to higher‐level effects was not possible. The constituents of the PQ‐CdCl₂ and the PQ‐PCP mixtures clearly have different molecular targets within the cell; thus, the IA model would be expected to best fit the data. Indeed, the data can be fitted using the IA model, but again, the CA model fits the data equally well (FI_total of PQ‐PCP) or better. The CuCl₂‐H₂O₂ mixture is subject to the same discussion, with best fits to the IA and CA model for the gene‐induction and growth‐inhibition data, respectively. Again, however, the alternative model also fits the data well.

Our data suggest that mixture‐induced responses on the gene‐expression level can be modeled within either the CA or IA model and that no clear preference based on mode of action exists. This unexpected finding could be caused by the inability to determine accurate and objective ECx levels, because one maximum induction level per promoter that would be valid for all toxicants is not known. Both mixture models and their respective deviations have been developed and evaluated regarding end points showing a generally applicable maximum (or minimum) level (e.g., mortality, immobility, and hatching). The differences in general toxicity and the specifics of the molecular transduction pathways inducing or repressing a given expression cassette, in terms of the particularity of every toxicant in a mixture, hampers the use of any model dependent on wide‐ranging maximum or minimum response levels. At present, we are evaluating the possibility of overcoming this obstacle by expressing FI levels in proportion to maximum FI curves obtained from a large set of gene‐expression data. The maximum expression level per promoter will be selected as the overall maximum and, consequently, all other dose–response curves will be scaled against this theoretical maximum. Hence, ECx values can be calculated and the rescaled curves fitted to the IA and CA models and their deviations to assess validity.

The present study showed that quantitative data regarding stress promoter induction by binary mixtures can be predicted accurately from the pure‐toxicant data by both the CA and IA models and their respective deviations. The choice of the best‐fit model, however, cannot objectively be made by evaluating the (non)equality of the toxicological mode of actions. Overall, both models fit the data well, as is the case for the growth‐inhibition data. This might be caused by the inability to calculate accurate ECx levels, and suggestions to overcome this obstacle associated with gene‐expression data are made. Independent of similar modes of action or the absence thereof, the mixture effects obtained show cases of no interaction, as indicated by fits to the base models, but synergistic, antagonistic, or dose‐level dependent combinations of both also were reported.