Research Cite this article: Baeder DY, Yu G, Hoze´ N, Rolff J, Regoes RR. 2016 Antimicrobial combinations: Bliss independence and Loewe additivity derived from mechanistic multi-hit models. Phil. Trans. R. Soc. B 371: 20150294. http://dx.doi.org/10.1098/rstb.2015.0294 Accepted: 10 March 2016 One contribution of 13 to a theme issue ‘Evolutionary ecology of arthropod antimicrobial peptides’. Subject Areas: theoretical biology, computational biology, immunology Keywords: antimicrobial peptide responses, antibiotics, synergy, antagonism, pharmacodynamic function, mathematical modelling

Antimicrobial combinations: Bliss independence and Loewe additivity derived from mechanistic multi-hit models Desiree Y. Baeder1, Guozhi Yu2, Nathanae¨l Hoze´1, Jens Rolff2,3 and Roland R. Regoes1 1

Institute of Integrative Biology, ETH Zurich, Universita¨tsstrße 16, 8092 Zurich, Switzerland Evolutionary Biology, Institut fu¨r Biologie, Freie Universita¨t Berlin, Ko¨nigin-Luise-Straße 1-3, 14195 Berlin, Germany 3 Berlin-Brandenburg Institute of Advanced Biodiversity Research (BBIB), Altensteinstraße 6, 14195, Berlin, Germany 2

DYB, 0000-0002-5585-6143; RRR, 0000-0001-8319-5293 Antimicrobial peptides (AMPs) and antibiotics reduce the net growth rate of bacterial populations they target. It is relevant to understand if effects of multiple antimicrobials are synergistic or antagonistic, in particular for AMP responses, because naturally occurring responses involve multiple AMPs. There are several competing proposals describing how multiple types of antimicrobials add up when applied in combination, such as Loewe additivity or Bliss independence. These additivity terms are defined ad hoc from abstract principles explaining the supposed interaction between the antimicrobials. Here, we link these ad hoc combination terms to a mathematical model that represents the dynamics of antimicrobial molecules hitting targets on bacterial cells. In this multi-hit model, bacteria are killed when a certain number of targets are hit by antimicrobials. Using this bottom-up approach reveals that Bliss independence should be the model of choice if no interaction between antimicrobial molecules is expected. Loewe additivity, on the other hand, describes scenarios in which antimicrobials affect the same components of the cell, i.e. are not acting independently. While our approach idealizes the dynamics of antimicrobials, it provides a conceptual underpinning of the additivity terms. The choice of the additivity term is essential to determine synergy or antagonism of antimicrobials. This article is part of the themed issue ‘Evolutionary ecology of arthropod antimicrobial peptides’.

Authors for correspondence: Desiree Y. Baeder e-mail: [email protected] Roland R. Regoes e-mail: [email protected]

1. Introduction

Electronic supplementary material is available at http://dx.doi.org/10.1098/rstb.2015.0294 or via http://rstb.royalsocietypublishing.org.

Antimicrobial peptides (AMPs) are produced by virtually all multicellular organisms, for example by plants [1], insects [2] and humans [3]. Moreover, multiple AMPs are expressed in concert [4–8]. The co-expression is hypothesized to be due to synergy between AMPs [9,10]. Multiple agents are said to synergize if they are more effective in combination than one would expect by adding their individual effects. Thus, to experimentally determine if AMPs synergize or antagonize one needs to compare their observed combined effect with a reference model that predicts how individual effects would add up independently. To define such a reference model is surprisingly challenging [11]. In most instances, there is no unique reference model and its choice sensitively impacts whether agents are found to synergize or not. In the case of pathogens, we are interested in how their growth can be inhibited. In this paper, we specifically define the effect of the inhibitor as the reduction of the pathogen’s growth rate. While we focus on the inhibiting effect of AMPs, the basic principles also apply to other important inhibitors, such as antibiotics and antibodies.

& 2016 The Author(s) Published by the Royal Society. All rights reserved.

(b)

net growth rate

independence

inhibitor 1 Loewe

Bliss no. molecules

no. molecules

Figure 1. The shape of the pharmacodynamic function determines the difference between predicted combination effects. Here, we show predicted effects for both Bliss independence and Loewe additivity (see §2) dependent on the number of inhibitor molecules. We plotted the combined effect for a 1 : 1 mixture of the two inhibitors, scaled by a factor 0.5 to adjust for the amount of inhibitors in combination. With this scaling, we are able to plot both the single effect curves and the combined effect curves using the same x-axis. (a) In the hypothetical case of linear pharmacodynamic functions describing single inhibitor effects of inhibitor 1 and 2, respectively, combined effects according to Bliss and Loewe are congruent (independence). Here, the effect in combination is equal to the sum of the two individual effects. The addition of the individual effects of inhibitor 1 and inhibitor 2 with n molecules each is equal to the effect of the combination of the inhibitors with 2n molecules. (b) For realistically sigmoid-shaped pharmacodynamic functions of single inhibitor effects of inhibitor 1 and inhibitor 2, Bliss independence and Loewe additivity models predict different effects of inhibitors in combination. The grey area marks all effects that are antagonistic according to Bliss independence and synergistic according to Loewe additivity. It has been shown that pathogen growth declines as a sigmoid function with the concentration of the inhibitor [9,12,13]. If the individual effects of inhibitors on pathogen growth were linear, instead of sigmoid, then they can simply be added to predict how the effects should combine without synergy (figure 1a). For nonlinear effect curves, however, the combined effect of two or more independent inhibitors is not necessarily obtained by a simple addition of the single effects. There are two widely used proposals that describe how independent effects combine and that serve as reference models to determine synergy or antagonism: Bliss independence and Loewe additivity. Bliss independence is based on probabilistic arguments: the combined effect of two inhibitors on the survival is defined as the probability of being affected by at least one of the inhibitors [14]. Applied to the specific case of pathogen growth rates, Bliss independence is equivalent to adding all single effects of the various inhibitors. Loewe additivity, in contrast, is built on the premise that an inhibitor cannot synergize or antagonize with itself [11,15]. Strictly speaking, inhibitor effects in combination are additive according to Loewe if two doses of the same inhibitor are added together. In combination, the dose of the first inhibitor is increased by the amount of the second inhibitor. Especially for nonlinear individual effect curves, this premise requires a sophisticated mathematical derivation. The complication arises from the fact that addition of inhibitors increases the effect by different amounts, depending on the amount of inhibitors already present. This is due to the slope of the nonlinear curve that is small for low and high inhibitor concentrations and large for inhibitor concentrations in between (figure 1b). As an example, addition of inhibitors at concentrations that by themselves already induce almost the maximum effect, i.e. at concentrations where the slope is small, do not increase the overall effect any more. While such behaviour could be interpreted as antagonism, according to Loewe it is defined as independent.

Figure 1b shows the combined effects of two inhibitors with nonlinear pharmacodynamic functions according to Bliss independence and Loewe additivity. The combined effect according to Bliss independence is the sum of the single effects of the two inhibitors on the growth rate, whereas according to Loewe additivity, it is by definition ‘in between’ the two individual effects (see §2). As a result, the maximum achievable effect with two inhibitors is predicted to be always higher according to Bliss independence than according to Loewe additivity. Of note, applied to linear effect curves, Loewe additivity and Bliss independence predict the same combined effects (figure 1a). To resolve these discrepancies between the ad hoc combination terms, we developed a mechanistic mathematical model that describes the interactions between inhibitors at the level of molecules. In our models, the two inhibitors either target two distinct mechanistic components and thus interact in an independent fashion (molecular independence), or they target a single component in which case they interfere with each other at the level of individual molecules (molecular dependence). Our approach is based on the multi-hit model framework originally introduced by Hedges [16] to examine the kinetics of colicin, an AMP produced by Escherichia coli. The multihit model idealizes the mechanistic pathways behind the inhibitor’s mode of action. According to this framework, a pathogen dies as soon as a certain threshold number of inhibitor molecules bind to or ‘hit’ pathogen receptors. We implemented independence by providing two distinct receptor types to which inhibitors can bind. In the alternative scenario of dependence, there is only one type of receptor for binding in our model. We found that our multi-hit model simulating independence of the molecule action predicts combined effects equivalent to Bliss independence. For dependence, the combined effect is consistent with Loewe additivity. Our results

Phil. Trans. R. Soc. B 371: 20150294

inhibitor 2

inhibitor 2

rstb.royalsocietypublishing.org

inhibitor 1

2

net growth rate

(a)

ð2:1Þ

Here, Emax denotes the maximum effect of inhibitor, EC50 is the concentration of the inhibitor resulting in 50% effect of Emax and k is the Hill coefficient, which describes the steepness of the curve (see electronic supplementary material, figure S1). Assuming that the bacterial population grows at a rate of r0 without inhibitors, the net growth rate in presence of an inhibitor, r(A), is rðAÞ ¼ r0 EðAÞ:

ð2:6Þ

Hereby, A1 isoeffective and A2 isoeffective are the concentrations of inhibitor 1 and 2, respectively, that are as effective as a combination of both inhibitors with the concentration A1 þ A2, 1=k1 ELoewe ðA1 , A2 Þ ð2:7Þ A1 isoeffective ¼ EC50 1 Emax 1 ELoewe ðA1 , A2 Þ

A2 isoeffective ¼ EC50 2

We describe the effect of a single inhibitor on the net growth rate of the bacterial population by a pharmacodynamic function (also known as Emax or Zhi model) [12,17,18]. Formally, the effect E(A) of the inhibitor present at the concentration A is given by Emax Ak : ECk50 þ Ak

A1 A2 þ : A1 isoeffective A2 isoeffective

and

(a) Mathematical description of single effects: the pharmacodynamic function

EðAÞ ¼

1¼

ð2:2Þ

ELoewe ðA1 , A2 Þ Emax 2 ELoewe ðA1 , A2 Þ

1=k2 ð2:8Þ

:

The combined effect under Loewe additivity rLoewe is then rLoewe ðA1 , A2 Þ ¼ r0 ELoewe ðA1 , A2 Þ ¼ r0 E1 ðA1 isoeffective Þ ¼ r0 E2 ðA2 isoeffective Þ:

ð2:9Þ ð2:10Þ ð2:11Þ

This definition is problematic for mixtures of inhibitors where the combined effect exceeds the maximum effect of one of the inhibitors [20]. In this range, Loewe additivity predicts suppression, i.e. more of an inhibitor is required in combination than alone to achieve the same effect (electronic supplementary material, figure S3).

(c) Multi-hit model (i) Model equations and parameter definitions

(b) Mathematical definition of Bliss independence and Loewe additivity To define the net growth rate in the presence of two different inhibitors, we need to consider two different pharmacodynamic functions, r1 ðA1 Þ ¼ r0 E1 ðA1 Þ

ð2:3Þ

and r2 ðA2 Þ ¼ r0 E2 ðA2 Þ:

ð2:4Þ

Here, A1 and A2 denote the concentrations of each inhibitor, and E1 and E2 their effects. We used the pharmacodynamic functions r1(A1) and r2(A2) as input for the two ad hoc reference models: Bliss independence and Loewe additivity. With both reference models, it is possible to determine independent effects of inhibitors in combination based on the individual effects of the inhibitors. Applying the reference models, we obtained the net growth rates predicted by each reference model, rBliss ðA1 , A2 Þ ¼ r0 EBliss ðA1 , A2 Þ and rLoewe ðA1 , A2 Þ ¼ r0 ELoewe ðA1 , A2 Þ as a function of the concentrations of both inhibitors, A1 and A2. We define Bliss independence [14] simply as the sum of the single effects. The net growth rate under Bliss independence, rBliss, is then rBliss ðA1 , A2 Þ ¼ r0 (E1 ðA1 Þ þ E2 ðA2 Þ):

ð2:5Þ

Bliss independence is defined as the product of effects [14], with the effect specified as normalized survival [11,14] or normalized growth rate [19]. In our opinion, an additive definition is more appropriate if effects on the net growth rate are considered. The net growth rate is the derivative of the population size of the pathogen. Using this relationship, multiplicative population size effect become additive when the population growth

In the multi-hit model, the total bacterial population Ntot is structured according to the number of inhibitor molecules attached to bacterial receptors [16]. The ‘pathogen receptors’ conceived by Hedges are not receptors in the biochemical sense of signal transmission, but represent any target site that the inhibitors in question can hit. A ‘hit’ comprises the mechanisms of action of the inhibitor, i.e. the travel to the target site, the attachment to the receptor and the subsequent biochemical reactions. Hereafter, we will use the terminology of Hedges who reduced a potentially complex hit to the simple attachment of the inhibitor molecule to a receptor. We built a multi-hit model with two inhibitor molecules (figure 2a). The number of molecules of each inhibitor type attached to cell receptors is denoted with i and j, respectively. How we limit the range of i and j is described below. For each number of attached inhibitors of type 1, i, and inhibitor of type 2, j, the temporal development of the subpopulation Ni,j is described by the following ordinary differential equation: dNi,j ¼ fðNi,j Þ di,j Ni,j dt þ a1,i,j A1 Ni1,j þ a2,i,j A2 Ni,j1 þ m1,iþ1,j ði þ 1ÞNiþ1,j þ m2,i,jþ1 ðj þ 1ÞNi,jþ1 a1,iþ1,j A1 Ni,j a2,i,jþ1 A2 Ni,j m1,i,j iNi,j m2,i,j jNi,j :

ð2:12Þ

Hereby, f(Ni,j ) is a term describing the bacterial replication. Based on the carpet model of AMP action [21,22], we assume that inhibitor molecules are distributed randomly on the pathogen cell surface and that the inhibitors stay attached when the cell is replicating. During replication, the inhibitors are redistributed to both daughter cells and each possible event of redistribution of i molecules of inhibitor 1 and j molecules of inhibitor 2 has a fixed probability 1/(i þ 1) 1/( j þ 1) (for detailed explanation, see

3

Phil. Trans. R. Soc. B 371: 20150294

2. Methods

rate is considered (for detailed derivation, see the electronic supplementary material, S1). Loewe & Muischnek stated that an inhibitor does not synergize with itself [15], which gives rise to the isobole equation [11] (electronic supplementary material, figure S2)

rstb.royalsocietypublishing.org

suggest using Bliss independence as a reference model when the inhibitors target distinct components. Loewe additivity is the reference model of choice if the inhibitors target the same component. In the specific case of AMPs, these results strongly suggest that Bliss independence should be used for AMPs that kill cells by forming pores. Loewe additivity, on the other hand, is the appropriate reference model for AMPs that destabilize the bacterial cell membrane by introducing positive charges (see §4).

(a)

inhibitor 2

(b)

4

rstb.royalsocietypublishing.org

¼b ¼b m1

a1

d

a2

a2

m2

m2 m1

a1

¼b

Figure 2. Diagrammatic representation of the multi-hit model. (a) Cells of the bacterial population Ntot are divided into classes according to how many inhibitors are attached to the cell surface. Ntot is the sum of all cells in all classes. The figure shows all classes for i 2, j 2; however, the matrix can be expanded. For reasons of simplicity, not all processes considered in the model are illustrated here and the attached inhibitors are not shown at the receptors of each cell. (b) Detailed depiction of all processes focusing on cell class N1,1, in which one inhibitor molecule of each type is attached to the cells’ surface. Note that the cell class N1,1 is in the middle of the figure with one inhibitor molecule of each type attached to the receptors. Cells from this class transit to class N2,1 at the rate a1,2,1 A1 N1,1 and to class N1,2 with the rate a2,1,2 A2 N1,1 : Cells from class N0,1 transit to N1,1 with the rate a1,1,1 A1 N0,1 and from class N1,0 with the rate a2,1,1 A1 N1,0 . Inhibitors detach from the cells in class N1,1 at the rates m1,1,1 1N1,1 and m2,1,1 1N1,1 , respectively. The class N1,1 gains cells through detachment of inhibitors from cells of higher classes with m1,2,1 2N2,1 in the case of detachment of inhibitor 1 and m2,1,2 2N1,2 in the case of detachment of inhibitor 2. The cells replicate at the rate b1,1N1,1; of all cells replicated, one-fourth stay in the class N1,1 whereas each one-fourth of the cells are categorized to the classes N0,1, N1,0 and N0,0, respectively. Lastly, cells can die at the rate d1,1 N1,1. Because we assumed the parameter to be the same independent of the class they are in, we use the following simplified expressions for the parameters: a1 instead of a1,1,1 and a1,2,1, a2 instead of a2,1,1 and a2,1,2, m1 instead of m1,1,1 and m1,2,1, m2 instead of m2,1,1 and m2,1,2, b instead of b1,1, and d instead of d1,1. the electronic supplementary material, S2)

fðNi,j Þ ¼

t1 X t2 X k¼i l¼j

1 1 bk,l Nk,l : k þ 1l þ 1

b ð2:13Þ

Note that the parameters t1 and t2 will be introduced in §2c(ii). Furthermore, di,j is the bacterial death rate constant. Lastly, a1,i,j and a2,i,j are the rates of attachment, and m1,i,j and m2,i,j are the rates of detachment of inhibitor 1 and 2, respectively. Both detachment rates describe the detachment rate per attached inhibitor molecule. A1 and A2 denote the concentrations of inhibitors of type 1 and 2, respectively, and are assumed to be constant through time. In the implementation of the model, we assumed that the attachment rates of an inhibitor are independent of the number of inhibitors already attached to the cell; therefore, a1,i,j ¼ a1 and a2,i,j ¼ a2 for all i and j. Similar assumptions were made for the detachment rates m1,i,j ¼ m1 and m2,i,j ¼ m2, the replication rate bi,j ¼ b and the death rate di,j ¼ d for all i and j. Figure 2b depicts the processes and rates incorporated into our model.

(ii) Killing mechanisms In the multi-hit model, bacterial cells are assumed to be killed when a certain number of inhibitor molecules are attached to them. In the case of one inhibitor, we denote by t the class for which this threshold is reached.

(iii) Zombi classes We extended the multi-hit model of Hedges [16] by zombi classes. These are necessary to keep the effect of the inhibitors from continuously increasing, which would be inconsistent with the maximum effect in the pharmacodynamic function. Cells enter into a zombi class when the respective killing threshold t1 or t2 is reached (figure 3).

z

d bz = 0 a2 m2

dz

Figure 3. Zombi class (grey box) of the multi-hit model. Here, t2 ¼ 3 and therefore, N1,2 enters the zombi class through attachment of one inhibitor 2 molecule (square-shaped molecules). In this class (N1,3), inhibitors do not detach any more (m2,1,3 ¼ 0), cells do not replicate any more (bZ ¼ b1,3 ¼ 0), and the cells are doomed to die with an increased death rate (dZ ¼ d1,3 . d1,2).

(d) Pharmacodynamic parameter estimation We solved the described differential equation (2.12) numerically to determine the net growth rate r(A1, A2) for any given combination of the two inhibitors. After a short phase during which the attachment classes converge towards a quasi-equilibrium, the growth dynamics is approximately exponential (with a possible negative growth rate). We calculate the net growth rate, r(A1, A2), from the rate of change of the total number of bacteria Ntot at the times t1 ¼ 80 min and t2 ¼ 120 min

rðA1 , A2 Þ ¼

logðNtot ðt2 ÞÞ logðNtot ðt1 ÞÞ : t2 t1

ð2:14Þ

From these estimates of the net growth rate for various combinations of inhibitor concentrations, we derived the pharmacodynamic parameters r0, Emax, EC50 and k. r0 is determined by evaluating the multi-hit model without inhibitors present (A1 ¼ A2 ¼ 0). Emax is calculated as the difference between r0 and the net growth for high inhibitor concentrations (A1 ! 1 and/or A2 ! 1). EC50 is estimated as the concentration

Phil. Trans. R. Soc. B 371: 20150294

inhibitor 1

¼b

Emax ¼ rðEC50 Þ: r0 2

ð2:15Þ

To obtain k for one of the inhibitors, the derivative of the pharmacodynamic function with respect to the concentration of this inhibitor at the EC50, dr(EC50)/dA, was solved for k

k¼

4ðdrðEC50 Þ=dAÞEC50 : Emax

ð2:16Þ

Using this relationship, k can be estimated from the rate of change of the net growth rate at the EC50 in our multi-hit simulations.

For only one type of inhibitor present, at a concentration A, the system of linear differential equations describing the multi-hit

0

b þm B b d aA 2 B B b B d aA m aA B 2 B B B 0 aA B MðAÞ ¼ B B B B 0 0 B B .. .. B B . . B @ 0 0

b 3 b þm 3

.

b d aA m 3

..

.

1 0 .. . 0

aA

.

.. .

..

.

b d aA m t1

0

0

..

.

aA

bZ dZ

0

C C C C C C C C C C: C C C C C C C C A

ð2:18Þ

(f ) Implementation The model equations are written in R (v. 3.1.3) [23] using RSTUDIO (v. 0.98.1103) [24]. We solved equation (2.12) numerically with the package deSolve [25]. The code is available upon request.

ð2:19Þ

N0 ðA ¼ 0, tÞ ¼ N0 ðA ¼ 0, t ¼ 0Þeðb0 d0 Þt

) ð2:20Þ

for 1 i t,

and when an infinite number of inhibitors of one type is present Ni ðA ! 1, tÞ ¼ 0

ð2:17Þ

is the (t þ 1)where NðtÞ ¼ ðN0 ðtÞ, N1 ðtÞ, . . . , Nt ðtÞÞ dimensional vector representing the number of cells in each class at time t and M(A) is the (t þ 1) (t þ 1) transition matrix representing the fluxes between the cell subpopulations Ni. The term M(A)N(t) in equation (2.17) is therefore the time increment of the cells in the classes N(t). M(A) is given by

..

where y is the eigenvector associated with r(A). Equation (2.19) can be solved explicitly only in the case of no inhibitor or when an infinite number of inhibitor molecules are present. In the case of no inhibitor molecules, we obtain the number of pathogen cells in each class

Ni ðA ¼ 0, tÞ ¼ 0

3 dN0 ðtÞ 6 dt 7 6 7 6 . 7 dNðtÞ ¼ MðAÞNðtÞ, 6 .. 7 ¼ 6 7 dt 4 dNt ðtÞ 5 dt

b t1 b þm t1

The exponential evolution of the number of cells as given by numerical solutions of the model allows solving the system of differential equations using a dominant eigenvalue approach. The maximal eigenvalue can be approximated by r(A) and is the solution to the eigenvalue equation MðAÞy ¼ rðAÞy,

2

b t1 .. .

... ..

5

for 0 i , t

Nt ðA ! 1, tÞ ¼ Nt ðA ! 1, t ¼ 0ÞeðbZ dZ Þt :

) ð2:21Þ

The net growth rate in the absence of inhibitors r0 is therefore only dependent on the birth and death rates of class i ¼ 0, which contains the cells without bound inhibitors: r0 ¼ b0 2 d0. In the presence of an infinite number of inhibitor molecules, the class where the threshold-hit value t is reached is dominating the system. For high inhibitor concentrations, the total number of bacteria Ntot is dependent on the parameters of the zombi class only: rA!1 ¼ bZ dZ : The maximum effect of the inhibitor Emax is the difference between r0 and rA! 1: Emax ¼ (b0 2 d0) 2 (bZ 2 dZ). No general formula for r(A) can be obtained analytically for intermediate concentrations. We could, however, obtain numerical solutions.

3. Results (a) Introducing additivity principles into the multi-hit framework To relate the pharmacodynamic functions to the level of individual molecules involved in the interaction between bacteria and inhibitors, we used the multi-hit model framework proposed by Hedges [16]. In this model, we described the binding of inhibitor molecules to bacterial receptors. Here, receptors represent any target sites that inhibitors hit and the term ‘hit’ describes the processes of an inhibitor travelling to the target site, binding to a receptor and the mechanisms of action following the binding. We structured the bacterial population by the number of inhibitor molecules that are attached to receptors on the bacterial cells. The binding is governed by the attachment and detachment rate constants a and m. Bacteria divide at a rate b0 and die at a rate d0. The multi-hit model owes its name to how the effect of the inhibitor is conceptualized: bacteria are assumed to die as soon as a certain number of inhibitors bind to or ‘hit’ them. This threshold number is denoted as t. We extended this framework by an additional zombi class. The zombi class is reached after t inhibitor molecules are attached to a bacterial cell and describes a state in which the cell is doomed to die. We assumed that zombi cells cannot divide any more (bZ ¼ 0) and that zombi cells die at an increased

Phil. Trans. R. Soc. B 371: 20150294

(e) Eigenvalue analysis

model (equation (2.12)) can be rewritten in the matrix form

rstb.royalsocietypublishing.org

that describes the following relationship best:

Using a dominant eigenvalue approach (see §2e), we derived that, in the absence of inhibitors, the bacterial population grows at a net rate of b0 2 d0 in the multi-hit model. This parameter combination corresponds to the pharmacodynamic parameter r0. Furthermore, we showed that the maximum effect in the multi-hit model is given by the death rate in the zombi class (b0 2 d0) 2 (bZ 2 dZ), which parallels the

(c) Multi-hit approach provides mechanistic underpinning of additivity reference models We compared the multi-hit model with the two ad hoc additivity reference models Bliss independence and Loewe additivity. The independent-hit model gives rise to a complex pattern of isoboles (lines of equal effect), convex for low while concave for high concentrations (figure 6a—note the linear scale for the inhibitor concentrations). Despite these complexities, the pattern agrees in great detail with the predictions of Bliss independence (figure 6b). There is a small deviation between the predictions from the two approaches around (and not just above) the EC50 (electronic supplementary material, figure S4a). According to Bliss independence, the maximum effect is the sum of the two individual maximum effects: Emax 1,2 ¼ Emax 1 þ Emax 2. This can be accomplished in the multi-hit model with molecular independence by choosing the death rates of the three zombi classes appropriately. An agreement with the prediction of Bliss independence is obtained by setting the death rate of the double zombi class to the sum of the death rates in the single zombi classes minus the death rate in the other classes: dZ 12 ¼ b0 2 d0 þ dZ 1 þ dZ 2 (for derivation, see electronic supplementary material, S3). Thus, to achieve a quantitative agreement with Bliss independence, the death rate parameter in the double zombi class Z12 must be higher than in the single zombi classes. This implies that even cells doomed to die from one of the inhibitors can be driven to death at an accelerated rate by the additional effect of the remaining inhibitor. Importantly, Bliss independence is consistent with only the very particular form of molecular independence in which the zombi classes are characterized by the equation above. We found that molecular dependence predicts linear isoboles when the net growth rate is plotted against both the inhibitor concentrations (figure 6c). Combining the individual effects of the two inhibitors according to Loewe, we also obtained linear isoboles (figure 6d). As expected for Loewe additivity, the maximum effect of two inhibitors combined is the same as that of a single inhibitor. On a more quantitative level, the predictions of the dependent-hit model agree well with Loewe additivity, except for a small range of concentrations above the EC50 (electronic supplementary material, figure S4b)—similarly to the deviation between the multi-hit model and pharmacodynamic function for a single inhibitor.

4. Discussion Bliss independence [14] and Loewe additivity [15], as well as the median effect equation for multiple drugs [26] are commonly used as reference models for the combined action of

6

Phil. Trans. R. Soc. B 371: 20150294

(b) Linking the multi-hit model to pharmacodynamic functions

pharmacodynamic parameter Emax. The number of inhibitor molecules t needed for bacterial cell killing in the multi-hit model is related to the pharmacodynamic parameters k— describing the steepness of the pharmacodynamic relationship—and EC50—the concentration which affects the net growth rate by 50% of Emax. The relationship between t, k and EC50 can be extracted only numerically, and is depicted in figure 5. The numerical solutions of equation (2.19) show that EC50 increases linearly with t and that the Hill coefficient k raises to an asymptotic value for large t. When t increases, the net growth rate curves thus keep a similar shape but are shifted to the right: the larger t, the higher the concentration of inhibitors has to be to obtain the same effect.

rstb.royalsocietypublishing.org

rate of dZ. This extension is necessary to obtain a finite maximum effect at very high inhibitor levels: without the zombi class, multi-hit models predict that bacterial growth rate declines for increasing inhibitor concentrations without boundary for increasing inhibitor concentrations. We expanded the multi-hit approach to describe the effect of two distinct inhibitors. To this end, we structured the bacterial population with respect to the numbers of inhibitor molecules of both types that are attached to receptors on the bacterial cells. As a result, the variables describing these bacterial subpopulations feature two indices (see §2), instead of just one as for a single inhibitor. To conceptualize killing in the two-inhibitor multi-hit model, we considered two scenarios: molecular independence and molecular dependence. In the first scenario, molecular independence, the inhibitors hit two different targets (figure 4a). Therefore, the effects of inhibitor 1 and inhibitor 2 are independent and consequently, two threshold numbers, t1 and t2, are required to characterize the effect of the inhibitors. We also defined three zombi classes with birth and death rates bZ 1, bZ 2, bZ 12 and dZ 1, dZ 2, dZ 12 (figure 4b). This scenario is equivalent to assuming that the two inhibitors attach to distinct receptor types on the surface of the bacterial cell. On a more abstract level, molecular independence is a simplistic description of inhibitors that affect different bacterial components. In the scenario of molecular dependence, we assumed that it does not matter which inhibitor molecule ‘hits’ the bacterial cell target. This scenario is characterized by a single threshold parameter t and a single zombi class (figure 4e) with birth and death rates bZ and dZ, and is equivalent to assuming that the two inhibitors attach to the same receptor type (figure 4d). To quantify the effects predicted by the two-inhibitor multi-hit model, we determined the total population size for two time points t1 and t2 under different inhibitor concentrations by numerical simulations of the model (see §2). All parameter values used are listed in table 1. The total population size is the sum of all living bacterial subpopulations, including zombi classes. The increase or decrease of the total population size between the two time points was used to estimate the net bacterial growth rate. We compared the net growth rates from the multi-hit model with the prediction of the ad hoc combination terms. To calculate the combination terms, the pharmacodynamic parameters characterizing each of the two inhibitors are required. To estimate these, we fitted the pharmacodynamic function to net growth rates obtained under single inhibitors in our multi-hit model. Figure 4c,f shows these fits. The functional form of the effect derived from the multi-hit model agrees surprisingly well with the predictions of the pharmacodynamic function, although there are subtle deviations for concentrations above the EC50 when the effect starts to saturate.

(a)

(d)

7

rstb.royalsocietypublishing.org

inhibitor 2

Phil. Trans. R. Soc. B 371: 20150294

(b)

inhibitor 2

(e)

inhibitor 1

inhibitor 1

z2

z

z12

z1

(f) 1

1

0

0 –1

–1

net growth rate

net growth rate

(c)

–2 –3

–2 –3 –4

–4

–5

–5

–6 109

1010

1011 A

1012

1013

109

1010

1011 A

1012

1013

Figure 4. Model assumptions and predictions of the independent and dependent-hit model. Receptors of the multi-hit models are different (a) and the same (d ) for the two inhibitors in the independent-hit and dependent-hit model, respectively. (b) In the independent-hit model, the number of hits of inhibitor 1 needed to put a cell into the zombi class t1 is independent of t2 and vice versa. (e) The number of hits needed to put bacteria in the zombi class Z is dependent on both inhibitor types in the dependent-hit model. The model output and the estimated pharmacodynamic curve in the case of single drug application is illustrated in (c) for the independent-hit model (filled square inhibitor 1, filled triangle inhibitor 2) and in (f ) for the dependent-hit model (open circle inhibitor 1 and 2). Note that single drug application in the dependent-hit model always gives the same output, regardless of the AMP type, whereas for the independent-hit model, the output is only the same if all parameters are the same (not shown here). All parameters used are listed in table 1. two independent inhibitors [11,27]. In this paper, we concentrated on Bliss independence and Loewe additivity because, unlike the median effect equation, these two reference models can be applied to inhibitors with different Hill

coefficients—a situation often encountered when the effect on the net growth rate of pathogens is considered. Bliss independence and Loewe additivity are defined ad hoc from abstract principles. Using the multi-hit framework,

(a)

(b)

8

rstb.royalsocietypublishing.org

4.5 2.5 × 1011 4.0 EC50

2.0 × 1011 3.5 k

1.5 × 1011

3.0 1.0 × 1011 2.5 2.0 2

4

6

2

8 10 12 14 t

4

6

8 10 12 14 t

Figure 5. The number of hits necessary to kill a cell t influences both EC50 and k positively. (a) The relationship between t and EC50 appears to be linear. (b) It is clearly a nonlinear relationship between t and k. Filled square inhibitor 1 (independent-hit model), filled triangle inhibitor 2 (independent-hit model), open circle inhibitor 1 and 2 (dependent-hit model). The results are the same for both inhibitor 1 and inhibitor 2 in the dependent-hit model due to the model assumptions. All parameters—except for t, which we varied according to the x-axis of the figures—are listed in table 1. Note that t1 ¼ t2 in the independent-hit framework. (a)

(b)

2.5 × 1011

2.5 × 1011

1.5 × 1011

1.5 × 1011

5.0 × 1010

5.0 × 1010

A2

5 × 1010 (c)

1.5 × 1011

2.5 × 1011 (d)

2.5 × 1011

5 × 1010

1.5 × 1011

2.5 × 1011

5.0 × 1010

1.5 × 1011 A1

2.5 × 1011

2.5 × 1011

1.5 × 1011

1.5 × 1011

5.0 × 1010

5.0 × 1010

A2

5.0 × 1010

1.5 × 1011 A1

2.5 × 1011

Figure 6. Contour plots of (a) the independent-hit model, (b) the Bliss independence reference model, (c) the dependent-hit model, and (d ) the Loewe additivity reference model. The results of the multi-hit models (a,c) for single inhibitor application are used to fit pharmacodynamic curves. These curves are then used to determine the predicted additive effect for all combinations of the two types of inhibitors in the ad hoc reference models (b,d). The isobole of r(A1, A2) ¼ 0 is marked as a bold line. All parameters used are listed in table 1. we linked these two common reference models to the mechanistic level on which the inhibitor molecules interact with molecular targets on the bacterial cell. The results presented in this paper provide guidance for choosing the reference model that is most appropriate for a specific pair of inhibitors: Bliss independence should be the model of choice if distinct components are targeted, whereas

Loewe additivity is more appropriate if the inhibitors target the same bacterial component. The choice of the reference model is essential to determine if two inhibitors synergize or antagonize. Two or more inhibitors synergize if their combined effect is greater than the effect predicted by the reference model. If the combined effect is smaller than the predicted, effect, then the combination

Phil. Trans. R. Soc. B 371: 20150294

5.0 × 1010

Table 1. Parameter values used for the independent- and dependent-hit model. Unless otherwise speciﬁed, the parameter values listed in this table were used to obtain the results. value

unit

description

universal model parameter 80 min t1

ﬁrst time point at which the multi-hit model is evaluated

t2 N0,0(t ¼ 0)

120 106

min 1

second time point at which the multi-hit model is evaluated number of cells in class with i ¼ 0 and j ¼ 0 at the time point t ¼ 0 h

Ni,j (t ¼ 0)

0.00

1

number of cells in all classes with i . 0 and j . 0 at the time point t ¼ 0 h

b d

0.71 0.01

h h21

birth rate of bacterial cells in all classes Ni,j with i , t1 and j , t2 death rate of bacterial cells in all classes Ni,j with i , t1 and j , t2

a1

10210

h21

attachment rate of inhibitor 1 molecules in all classes Ni,j with 0 , i t1 and 0 j , t2. For i ¼ 0, a1 ¼ 0 h21

a2

10210

h21

attachment rate of inhibitor 2 molecules in all classes Ni,j with 0 i , t1 and 0 , j t2. For j ¼ 0, a2 ¼ 0 h21

m1

1.00

h21

detachment rate of inhibitor 1 molecules from all classes Ni,j with 0 , i , t1 and 0 j , t2. For i ¼ 0,

h21

m1 ¼ 0 h21 detachment rate of inhibitor 2 molecules from all classes Ni,j with 0 i , t1 and 0 , j , t2. For j ¼ 0, m2 ¼ 0 h21

m2

1.00

independent-hit model parameter

t1 5 1 t2 3 1 0.00 h21 bZ 1 bZ 2 0.00 h21 bZ 12 0.00 h21 dZ 1 4.00 h21 dZ 2 3.00 h21 dZ 12 7.70 h21 aZ 1 10210 h21 aZ 2 10210 h21 mZ 1 1.00 h21 mZ 2 1.00 h21 dependent-hit model parameter t1 , t2 5 1 0.00 h21 bZ dZ 5.00 h21 aZ 1, aZ 2 0.00 h21 mZ 0.00 h21

number of inhibitor 1 molecules attached to the bacterial cell necessary to kill the cell number of inhibitor 2 molecules attached to the bacterial cell necessary to kill the cell birth rate of bacterial cells in zombi classes NZ,i,j with i ¼ t1 and j , t2 birth rate of bacterial cells in zombi classes NZ,i,j with i , t1 and j ¼ t2 birth rate of bacterial cells in zombi class NZ,i,j with i ¼ t1 and j ¼ t2 death rate of bacterial cells in zombi classes NZ,i,j with i ¼ t1 and j , t2 death rate of bacterial cells in zombi classes NZ,i,j with i , t1 and j ¼ t2 death rate of bacterial cells in zombi class NZ,i,j with i ¼ t1 and j ¼ t2 attachment rate of inhibitor 1 molecules in zombi classes Ni;t2 with 0 , i. For i ¼ 0, aZ1 ¼ 0 h21 attachment rate of inhibitor 2 molecules in zombi classes Nt1 ;j with 0 , j. For j ¼ 0, aZ2 ¼ 0 h21 detachment rate of inhibitor 1 molecules in the zombi classes Ni;t2 with 1 i t1. For i ¼ 0, mZ1 ¼ 0 h21 detachment rate of inhibitor 2 molecules in the zombi classes Nt1 ;j with 1 j t2. For j ¼ 0, mZ2 ¼ 0 h21 number of inhibitor 1 and inhibitor 2 molecules necessary to kill a bacterial cell birth rate of bacterial cells in zombi classes NZ,i,j with i ¼ t1 and/or j ¼ t2 death rate of bacterial cells in zombi classes NZ,i,j with i ¼ t1 and/or j ¼ t2 attachment rate of inhibitor 1 and inhibitor 2 molecules in all classes NZ,i,j with i ¼ t2 and/or j ¼ t2 detachment rate of inhibitor molecules from all classes Ni,j with i ¼ t1 and/or j ¼ t2

is called antagonistic. The deviation of the combined effect from the reference model provides a quantitative measure of the degree of synergy or antagonism based on either Bliss independence, such as in [28], or Loewe additivity, such as in [29 – 31]. Except for a few clear-cut cases in which one of the inhibitors is effective only in combination with the other [10,32], or in the case of suppression [19,33], the classification of a combination as synergistic or antagonistic depends on the reference model. For high concentrations of inhibitors, Bliss independence is the more conservative choice of a reference model if one seeks to determine synergistic effect. Loewe additivity is more conservative when determining antagonism.

Synergy and antagonism are deviations from the reference models, and conform neither to the molecular dependence nor the molecular independence scenarios of our multi-hit models. As an example, if an inhibitor targeting the cell membrane facilitates the travel of a second inhibitor to its molecular target within the cell, the rate at which the second inhibitor hits, a, depends on the number of hits the cell has received from the first AMP—a clear instance of synergy. This type of interaction has been observed in the case of combinations of AMPs [10,32] and antibiotics [19]. The reference models describe additive interactions, and our multi-hit models are only used to describe additivity of mixtures based on the two approaches introduced by Loewe and Bliss.

Phil. Trans. R. Soc. B 371: 20150294

21

rstb.royalsocietypublishing.org

parameter

9

Authors’ contributions. R.R.R. initiated this project. R.R.R. conceived the model. D.Y.B. wrote the took part in the analysis of the results. D.Y.B., drafted the manuscript. All authors revised gave final approval for publication.

D.Y.B., G.Y., J.R. and R script. All authors R.R.R., J.R. and N.H. the manuscript and

Competing interests. We have no competing interests. Funding. R.R.R. gratefully acknowledges the financial support of the Swiss National Science Foundation (grant number 31003A_149769), G.Y.’s research is funded by the China Scholarship Council and J.R. gratefully thanks the European Research Council (grant number 260986).

Acknowledgements. We thank Paul Schmid-Hempel and an anonymous reviewer for helpful comments. We also thank Alexey Mikaberidze for sharing his knowledge about Loewe Additivity with us.

References 1.

2.

3.

4.

Rahnamaeian M. 2011 Antimicrobial peptides. Modes of mechanism, modulation of defense responses. Plant Signal. Behav. 6, 1325 –1332. (doi:10.4161/psb.6.9.16319) Schmid-Hempel P. 2005 Evolutionary ecology of insect immune defenses. Annu. Rev. Entomol. 50, 529–551. (doi:10.1146/annurev.ento.50.071803. 130420) Harder J, Bartels J, Christophers E, Schro¨der JM. 1997 A peptide antibiotic from human skin. Nature 387, 861. (doi:10.1038/43088) Marxer M, Vollenweider V, Schmid-Hempel P. 2016 Insect antimicrobial peptides act synergistically to inhibit a trypanosome parasite. Phil. Trans. R. Soc. B 371, 20150302. (doi:10.1098/rstb. 2015.0302)

5.

6.

7.

8.

Johnston PR, Makarova O, Rolff J. 2014 Inducible defenses stay up late: temporal patterns of immune gene expression in Tenebrio molitor. G3 4, 947 –955. (doi:10.1534/g3.113.008516) Mak P, Zdybicka-Barabas A, Cytryn´ska M. 2010 A different repertoire of Galleria mellonella antimicrobial peptides in larvae challenged with bacteria and fungi. Dev. Comp. Immunol. 34, 1129 –1136. (doi:10.1016/j.dci.2010.06.005) Yeaman MR, Yount NY. 2003 Mechanisms of antimicrobial peptide action and resistance. Pharmacol. Rev. 55, 27 –55. (doi:10.1124/ pr.55.1.2) Zasloff M. 2002 Antimicrobial peptides of multicellular organisms. Nature 415, 389 –395. (doi:10.1038/415389a)

9.

Yu G, Baeder DY, Regoes RR, Rolff J. 2016 Combination effects of antimicrobial peptides. Antimicrob. Agents Chemother. 60, 1717–1724. (doi:10.1128/AAC.02434-15) 10. Rahnamaeian M et al. 2015 Insect antimicrobial peptides show potentiating functional interactions against Gram-negative bacteria. Proc. R. Soc. B 282, 20150293. (doi:10.1098/rspb. 2015.0293) 11. Greco WR, Bravo G, Parsons JC. 1995 The search for synergy: a critical review from a response surface perspective. Pharmacol. Rev. 47, 331–385. 12. Regoes RR, Wiuff C, Zappala RM, Garner KN, Baquero F, Levin BR. 2004 Pharmacodynamic functions: a multiparameter approach to the design

10

Phil. Trans. R. Soc. B 371: 20150294

found to be equivalent to Bliss independence. On the other hand, if we combine two AMPs that both hit the cell membrane, we would apply the model of molecular dependence and would use Loewe additivity as the reference model. Nevertheless, for a specific inhibitor pair, it might not always be easy to determine if their effects are truly independent or dependent, because our multi-hit models only describe molecular processes on an abstract level. Moreover, inhibitors may act in multiple ways which limit a straightforward categorization of modes of action. AMPs may use several mechanisms to kill target cells [39]. In such cases, we suggest considering the more conservative ad hoc term, which is either Bliss independence or Loewe additivity, depending on the individual pharmacodynamic curves and the concentrations of both inhibitors. The multi-hit model approach should be applied with care to antibiotics and antibody inhibition. The efficacy of antibiotics is dependent on pathogen density [40] and in combination antibiotics display complex interaction patterns [41], neither of which we addressed in this study. Moreover, we also considered that the cells die after a certain number of hits are reached. While this may be a valid assumption for AMPs [9], many antibiotics inhibit the replication of pathogens in addition to killing them. To use the approach to describe the action of antibiotics will thus require to extend the multi-hit models by an effect on the replication rates [42]. This applies even more to antibodies that primarily inhibit pathogen replication and induce death only through very specific mechanisms, such as antibody-dependent cell-mediated cytotoxicity.

rstb.royalsocietypublishing.org

The classification of mixtures as synergistic, additive or antagonistic is essential for successful treatment of pathogen infection. While it is obvious that the efficacy of synergistic mixtures is higher than for additive and antagonistic mixtures, synergistic combinations have been found to also favour emergence of resistant strains [34,35]. We generically referred to ‘inhibitors’, because our multi-hit model approach is applicable to a wide variety of effectors, such as AMPs, antibiotics and antibodies. Hedges [16] developed the multi-hit model for AMPs but did not consider inhibitor combinations. However, combinations of AMPs are common in host–microbe interactions [4–8]. The vast majority of AMPs are cationic, hence mostly targeting bacterial cell membranes that are negatively charged. They also often possess hydrophobic residues that result in interactions with fatty acids in the bacterial cell membrane. Most AMPs hence target the cell membranes and only very few AMPs enter the cells and target intercellular compounds [36]. Individual AMPs such as magainin can also interact with different targets on the cell surface [8], or, in the case of a-defensin and cathelicidin, mediate competition between bacteria [37]. Our multi-hit models reduce the molecular modes of action to a sequence of generic hits. Nevertheless, they can be applied to derive the reference model for a concrete pair of inhibitors. To this end, we need to assemble all the relevant molecular processes involved in the action of each inhibitor into a ‘hit’ and also need to gather information on how many hits are required for action (our parameter t). For protegrin peptides that target the cell membrane, for example, a ‘hit’ would comprise the travel to the cell membrane and attachment. This AMP forms pores consisting of eight AMP molecules [38], and 100 pores are required for lethal cell damage [38]. Therefore, the threshold number of hits needed to kill the cell, t, would be 800 for protegrin peptides. To derive the reference model of a combination of two AMPs, we need to assess if the threshold number of hits of one of the AMPs depends on the number of hits the bacterial cell has already received from the other AMP. If we combine an AMP that targets the cell membrane with an AMP that hit intracellular targets, the two thresholds are independent of the number of hits of the other AMP. We would therefore apply the scenario of molecular independence, which we

14.

15.

17.

18.

19.

20.

21.

22.

24.

25.

26.

27.

28.

29.

30. 31.

32.

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

resistance. Science 351, aad3292. (doi:10.1126/ science.aad3292) Michel JB, Yeh PJ, Chait R, Moellering RC, Kishony R. 2008 Drug interactions modulate the potential for evolution of resistance. Proc. Natl Acad. Sci. USA 105, 14 918–14 923. (doi:10.1073/pnas. 0800944105) Pena-Miller R, Laehnemann D, Jansen G, FuentesHernandez A, Rosenstiel P, Schulenburg H, Beardmore R. 2013 when the most potent combination of antibiotics selects for the greatest bacterial load: the smile-frown transition. PLoS Biol. 11, e1001540. (doi:10.1371/journal.pbio.1001540) Nicolas P. 2009 Multifunctional host defense peptides: intracellular targeting antimicrobial peptides. FEBS J. 276, 6483–6496. (doi:10.1111/j. 1742-4658.2009.07359.x) Habets MGJL, Rozen DE, Brockhurst MA. 2012 Variation in Streptococcus pneumoniae susceptibility to human antimicrobial peptides may mediate intraspecific competition. Proc. R. Soc. B 279, 3803– 3811. (doi:10.1098/rspb.2012.1118) Bolintineanu D, Hazrati E, Davis HT, Lehrer RI, Kaznessis YN. 2010 Antimicrobial mechanism of pore-forming protegrin peptides: 100 pores to kill E. coli. Peptides 31, 1– 8. (doi:10.1016/j.peptides. 2009.11.010) Nguyen LT, Haney EF, Vogel HJ. 2011 The expanding scope of antimicrobial peptide structures and their modes of action. Trends Biotechnol. 29, 464 –472. (doi:10.1016/j.tibtech.2011.05.001) Udekwu KI, Parrish N, Ankomah P, Baquero F, Levin BR. 2009 Functional relationship between bacterial cell density and the efficacy of antibiotics. J. Antimicrob. Chemother. 63, 745 –757. (doi:10. 1093/jac/dkn554) Ocampo PS et al. 2014 Antagonism between bacteriostatic and bactericidal antibiotics is prevalent. Antimicrob. Agents Chemother. 58, 4573– 4582. (doi:10.1128/AAC.02463-14) Abel zur Wiesch P et al. 2015 Classic reaction kinetics can explain complex patterns of antibiotic action. Sci. Transl. Med. 7, 287ra73. (doi:10.1126/ scitranslmed.aaa8760)

11

Phil. Trans. R. Soc. B 371: 20150294

16.

23.

ACS Chem. Biol. 9, 2003–2007. (doi:10.1021/ cb500426r) R Core Team. 2015 R: a language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. (http://www. Rproject.org) Racine JS. 2012 RStudio: a platform-independent IDE for R and Sweave. J. Appl. Econ. 27, 167– 172. (doi:10.1002/jae.1278) Soetaert K, Petzoldt T, Setzer RW. 2010 Solving differential equations in R: package deSolve. J. Stat. Softw. 33, 1–25. (doi:10.18637/jss.v033.i09) Chou TC, Talalay P. 2005 Generalized equations for the analysis of inhibitions of Michaelis – Menten and higher-order kinetic systems with two or more mutually exclusive and nonexclusive inhibitors. Eur. J. Biochem. 115, 207–216. (doi:10.1111/j. 1432-1033.1981.tb06218.x) Fitzgerald JB, Schoeberl B, Nielsen UB, Sorger PK. 2006 Systems biology and combination therapy in the quest for clinical efficacy. Nat. Chem. Biol. 2, 458 –466. (doi:10.1038/nchembio817) Ankomah P, Levin BR. 2012 Two-drug antimicrobial chemotherapy: a mathematical model and experiments with Mycobacterium marinum. PLoS Pathog. 8, e1002487. (doi:10.1371/journal.ppat.1002487) Berenbaum MC. 1977 Synergy, additivism and antagonism in immunosuppression. A critical review. Clin. Exp. Immunol. 28, 1– 18. Berenbaum MC. 1989 What is synergy? Pharmacol. Rev. 41, 93–141. (doi:10.1007/s13398-014-0173-7.2) Chou TC, Talalay P. 1984 Quantitative analysis of dose-effect relationships: the combined effects of multiple drugs or enzyme inhibitors. Adv. Enzyme Regul. 22, 27 –55. (doi:10.1016/0065-2571(84) 90007-4) Zdybicka-Barabas A, Mak P, Klys A, Skrzypiec K, Mendyk E, Fiolka MJ, Cytryn´ska M. 2012 Synergistic action of Galleria mellonella anionic peptide 2 and lysozyme against Gram-negative bacteria. Biochim. Biophys. Acta Biomembr. 1818, 2623–2635. (doi:10.1016/j.bbamem.2012.06.008) Baym M, Stone LK, Kishony R. 2016 Multidrug evolutionary strategies to reverse antibiotic

rstb.royalsocietypublishing.org

13.

of antibiotic treatment regimens. Antimicrob. Agents Chemother. 48, 3670– 3676. (doi:10.1128/AAC.48. 10.3670-3676.2004) Shen L et al. 2008 Dose –response curve slope sets class-specific limits on inhibitory potential of antiHIV drugs. Nat. Med. 14, 762–766. (doi:10.1038/ nm1777) Bliss CI. 1939 The toxicity of poisons applied jointly. Ann. Appl. Biol. 26, 585–615. (doi:10.1111/j.17447348.1939.tb06990.x) Loewe S, Muischnek H. 1926 U¨ber Kombinationswirkungen. Naunyn. Schmiedebergs. Arch. Pharmacol. 114, 313– 326. (doi:10.1007/ BF01952257) Hedges AJ. 1966 An examination of single-hit and multi-hit hypotheses in relation to the possible kinetics of colicin adsorption. J. Theor. Biol. 11, 383–410. Zhi J, Nightingale CH, Quintiliani R. 1986 A pharmacodynamic model for the activity of antibiotics against microorganisms under nonsaturable conditions. J. Pharm. Sci. 75, 1063–1067. (doi:10.1002/jps.2600751108) Nolting A, Dalla Costa T, Rand KH, Derendorf H. 1996 Pharmacokinetic-pharmacodynamic modeling of the antibiotic effect of piperacillin in vitro. Pharm. Res. 13, 91–96. (doi:10.1023/A:10160 85402278) Yeh P, Tschumi AI, Kishony R. 2006 Functional classification of drugs by properties of their pairwise interactions. Nat. Genet. 38, 489–494. (doi:10. 1038/ng1755) Mikaberidze A, McDonald BA, Bonhoeffer S. 2014 Can high-risk fungicides be used in mixtures without selecting for fungicide resistance? Phytopathology 104, 324–331. (doi:10.1094/ PHYTO-07-13-0204-R) Gazit E, Miller IR, Biggin PC, Sansom MS, Shai Y. 1996 Structure and orientation of the mammalian antibacterial peptide Cecropin p1 within phospholipid membranes. J. Mol. Biol. 258, 860–870. (doi:10.1006/jmbi.1996.0293) Roversi D, Luca V, Aureli S, Park Y, Mangoni ML, Stella L. 2014 How many antimicrobial peptide molecules kill a bacterium? The case of PMAP-23.