 Methodology article
 Open access
 Published:
A foundation for reference models for drug combinations with an application to Loewe’s reference model
BMC Bioinformatics volume 21, Article number: 460 (2020)
Abstract
Background
Treating patients with combinations of drugs that have synergistic effects has become widespread practice in the clinic. Drugs work synergistically when the observed effect of a drug combination is larger than the effect predicted by the reference model. The reference model is a theoretical null model that returns the combined effect of given doses of drugs under the assumption that these drugs do not interact. There is ongoing debate on what it means for drugs to not interact. The controversy transcends mathematical punctuality, as different noninteraction principles result in different reference models. A famous reference model that has been in existence for already a long time is Loewe’s reference model. Loewe’s vision on noninteraction was purely intuitive: two drugs do not interact if all combinations of doses that result in a certain given effect lie on a straight line.
Results
We show that Loewe’s reference model can be obtained from much more fundamental principles. First, we introduce the new notion of complementary dose. Secondly, we reformulate the existing concept of equivalent dose, whereby our formulation is more general than existing ones. Finally, a very general noninteraction principle is put forward. The proposed noninteraction principle represents a certain interplay between complementary and equivalent doses: drugs are noninteracting if complementarity is preserved under equivalence. It is then shown that Loewe’s reference model naturally follows from these principles by an appropriate choice of complementarity.
Conclusions
The presented work increases insight into Loewe’s reference model for drug combinations, which is realized by the introduction of a very general noninteraction principle that does not refer to any specific doseresponse curve, nor to any property of applicable doseresponse curves.
Background
For many diseases it has become standard practice to give drug combinations to patients. The biochemical rationale is that an effective drug combination may target multiple proteins or pathways, compared to monotherapies [1]. The effects resulting from a drug combination may even be larger than what would have been expected based on the properties of the doseresponse curves of the single drugs, in which case the effects are called synergistic. Diseases for which one has discovered drug combinations with synergistic effects include cancer [2], Parkinson’s disease [3], HIV [4,5,6] and asthma [7,8,9,10]. The impact of multitherapies on such severe diseases explains why combinatorial drug screening and synergy scoring are very hot research topics [11]. A rigorous assessment of synergy in drug combinations must be based on the comparison of the observed effect of the drug combination with a noninteractive reference model [12]. The reference model is a theoretical model that produces the effect of a combination of given drugs under the assumption that these drugs do not interact. There is still wide debate on the precise meaning of noninteraction between drugs. Several researchers have put forward different noninteraction principles, resulting in dissimilar reference models. A very well established noninteraction principle is the one introduced by Bliss who associates a random variable with each drug, thereby assimilating noninteraction between the drugs with statistical independence between the corresponding random variables [13]. His reference model is still used to detect drug synergies, e.g. in the web tool SynergyFinder [14]. Another very popular reference model, which is also implemented in the aforementioned web tool as well as in CImbinator [15], is the Loewe additivity reference model [16]. Since the work in this paper relates to Loewe’s reference model, we provide a brief background of this model in section "Related work: Loewe’s reference model".
Methods
In this paper we introduce the new concept of complementary dose (section "Complementary dose"). Two doses of the same drug are complementary with respect to a given effect if their sum produces that effect. The concept of equivalent dose is frequently used in the literature on drug combinations; it has been introduced to establish a certain similarity between doses of different drugs. In section "Equivalent dose" we introduce a general formulation of this notion. The formulation is radically different than existing ones: complementarity is associated with the linear transformation of basis vectors in twodimensional Euclidean space. We also give a concrete example of such a linear transformation, to which we refer as ’the special case’. In section "Noninteraction principle" a new and very general noninteraction principle is introduced. It is stated in terms of a certain interplay between complementarity and equivalent doses. Having constructed this noninteraction principle, the definition of a general reference model then naturally follows (section "Construction of reference model"). The reference model for the special case of equivalent doses turns out to be Loewe’s reference model. Thus the main result of the paper is that a fundamental and general noninteraction principles is introduced, based on constructed notions of complementary and equivalent doses, and that Loewe’s reference model results as one member of the class of reference models that rely on this noninteraction principle.
Notational conventions and assumptions
We consider two drugs, with corresponding doseresponse curves \(f_1\) and \(f_2\). The doseresponse curves are assumed to be given. An arbitrary dose of the ith drug is denoted by \(u_i\) with \(i \in \{1,2\}\). If another dose of the same drug is needed in the description, we denote this dose by \(u'_i\). An arbitrary effect is denoted by y. If any of the two drugs is considered, we refer to this drug as the ith drug, while the other drug is then referred to as the jth drug.
We make the assumption that the doseresponse curves are invertible. This assumption is common in the drug combinations literature, although it is often not explicitly stated, e.g. [17]. For simplicity, we assume that all doses are nonnegative, although this restriction can be dropped without altering any of the results below.
The effect produced by the reference model, which is constructed below, for given doses \(u_1\) and \(u_2\) will be denoted by \({\mathcal {F}}(u_1,u_2)\).
Related work: Loewe’s reference model
Loewe states that two drugs are noninteracting if all combinations of doses that have a certain effect y lie on a straight line. Given an effect y and the doses \(u^{\star }_1\) and \(u^{\star }_2\) for which it holds that \(f_1(u^{\star }_1)=f_2(u^{\star }_2)=y\), it is then the straight line from \((u^{\star }_1,0)\) to \((0,u^{\star }_2)\) that contains the pairs of doses \((u_1,u_2)\) that result in the effect y under the assumption of noninteraction. This straight line is given by the equation
Taking into account that \(f_i(u^{\star }_i)=y\) or, equivalently, that \(u^{\star }_i=f^{1}_i(y)\), we find that Loewe’s reference model, denoted by \({\mathcal {F}}_L\), is given by the following definition.
Definition 1
(Loewe’s reference model) The combined effect produced by Loewe’s reference model is given by \(y={\mathcal {F}}_L(u_1,u_2)\) with y the solution of
Complementary dose
We will introduce two concepts, namely complementary dose and equivalent dose. Both concepts apply to a given dose of a given drug. However, whereas the complementary dose returns a dose of the same drug, the equivalent dose returns a dose of another drug. The concept of complementary dose is introduced in this section, while the idea of equivalent dose is considered in the next section.
Definition 2
(Complementary dose) Given an effect y and a dose \(u_i\) with \(0 \le u_i \le f^{1}_i(y)\), we define \(h_{i,y}\) as
The dose \(h_{i,y}(u_i)\) is called the complementary dose of \(u_i\) with respect to y or, in short, the complement of \(u_i\) with respect to y.
The meaning of the definition is that if \(u_i\) is a dose for which \(0 \le u_i \le f^{1}_i(y)\), and \(u'_i\) is its complement with respect to y, then \(f_i(u_i+u'_i)=y\). Thus two doses are complementary with respect to y if added together they produce the effect y.
The following observations can be made from the given definition:

\(h_{i,y}\) is the straight line from \((0,f^{1}_i(y))\) to \((f^{1}_i(y),0)\).

The function \(h_{i,y}\) is an affine transformation.
Since \(h_{i,y}\) is an affine transformation, it can be described by a linear transformation. This requires to represent doses in twodimensional Euclidean space, which can be performed by denoting a dose \(u_i\) as the twodimensional vector \(U_i = (u_i,1)^T\).
The function \(h_{i,y}(u_i)\) can then be represented as the matrix \(H_{i,y}\):
Indeed, we have that
Equivalent dose
Related work
The concept of equivalent dose is used in several papers on drug combinations. The idea is that a dose of a certain drug is equivalent to a dose of another drug if these doses show some similarity. An intuitive definition of equivalent dose was introduced by Lederer et al. [17]: two doses \(u_i\) and \(u_j\) are equivalent if these doses result in the same effect, i.e. if \(f_i(u_i)=f_j(u_j)\). We use the notation \(\xi _{ij}(u_j)\) to denote the dose that is equivalent to \(u_j\). This means that according to the definition by Lederer et al., the dose \(u_i\) of the ith drug that is equivalent to dose \(u_j\) of the jth drug is given by
The main reason that the concept of equivalent dose is used in the literature is to express the reference model in terms of a fixed drug, say the ith drug. This is done by transforming the given dose of the jth drug into its equivalent dose of the ith drug, and then expressing the combined effect produced by the reference model as the effect produced by the ith doseresponse curve for the sum of \(u_i\) and the equivalent dose of \(u_j\). This means that one way to define a reference model is as
A new general formulation
We introduce a new formulation of the concept of equivalent dose. Our formulation deviates from existing notions of equivalent dose in two ways:

The dose that is equivalent to a given dose depends on a given effect. The reason is that below we will consider a certain interplay between complementary dose and equivalent dose, which requires that the equivalent dose depends on a given effect, just as the complementary dose depends on such a given effect (cf. section "Complementary dose").

The equivalent dose is interpreted as a change of basis in twodimensional Euclidean space.
This means that the transformation from a given dose to its equivalent dose can be represented by a general matrix
with \((a_{ij,y},c_{ij,y})^T\) and \((b_{ij,y},d_{ij,y})^T\) representing, in linear algebra terms, the new basis vectors, i.e. the transformations of the old basis vectors \((1,0)^T\) and \((0,1)^T\). We thus envisage the concept of equivalent dose as a changeofbasis linear transformation in twodimensional space.
Depending on the intended precise interpretation of equivalent dose, the components of \(E_{ij,y}\) may differ. Other values for these components imply other definitions of equivalent dose. It will be shown that Loewe’s reference model actually relies on a very simple changeofbasis transformation. We describe this transformation in the next subsection, while the use of it in the context of Loewe’s model is considered later.
A special case
The simplest transformation scales both units vectors by a constant \(\alpha _{ij,y}\), i.e.
with I denoting the identity matrix. Now, it is desired that for a given y and a given dose \(u_j\) for which \(f_j(u_j)=y\) it holds that the equivalent dose with respect to y is given by \(u_i\) for which \(f_i(u_i) = y\). In other words, that
where c is an unimportant constant. By virtue of (7), and taking into account that \(u_j=f^{1}_j(y)\), this means that
Combining this with (7) this implies that
Further reference to \(E_{ij,y}\) will denote this specific matrix, and not the general matrix given by (6), unless otherwise stated.
Noninteraction principle
In this section we introduce a new noninteraction principle. The noninteraction principle will rely on the introduced notions of complementary and equivalent dose. We first describe, in section "General case", the principle for two drugs in the general case, i.e. for any matrix \(E_{ij,y}\). The noninteraction principle is then specifically considered for the special case where \(E_{ij,y}\) is given by (8) (cf. section "Special case" below).
General case
Let us consider an effect y and a dose \(0 \le u_j \le f^{1}_j(y)\) of the jth drug. Consider its complementary dose which is, according to Definition 2, and taking into account (3), given by \(U'_j = H_{j,y} \, U_j\), where \(U_j=(u_j,1)^T\). Now consider the dose of the ith drug that is equivalent to \(U'_j\), given by
On the other hand, the equivalent dose to \(U_j\) is given by \(U_i = E_{ij,y} \, U_j\).
The noninteraction principle that we put forward then requires that \(U_i\) and \(U'_i\) are complementary doses, i.e. that it holds that
Simply put: Noninteraction means that complementarity is preserved under equivalence. Indeed, we have assumed that \(U'_j\) and \(U_j\) are complementary doses, i.e. \(U'_j = H_{j,y} \, U_j\). Furthermore, the equivalent dose to \(U'_j\) is \(U'_i = E_{ij,y} \, U'_j\), and the equivalent dose to \(U_j\) is \(U_i = E_{ij,y} \, U_j\). Saying that the complementarity between \(U'_j\) and \(U_j\) also applies to the corresponding equivalent doses \(U'_i\) and \(U_i\) is indeed what is expressed by (10).
The principle is illustrated in Fig. 1. We summarize this principle in the following definition.
Definition 3
(General noninteraction principle) Two drugs are noninteracting if complementary doses are preserved under equivalence. This means that
for any effect y.
Notice that (11) is indeed the same as (10), by taking into account (9) and the fact that \(U_i = E_{ij,y} \, U_j\). Equation (11) clearly shows that the noninteraction principle can also be interpreted as stating that the order of taking equivalent and complementary doses should not matter.
Special case
For the special case (cf. section "A special case"), it follows from (3) and (8) that the noninteraction principle (11) is given by
It follows that this basic case results in a very severe restriction on the doseresponse curves, which is summarized in the following definition.
Definition 4
(Noninteraction principle for the special case) Two drugs are noninteracting for the choice of E given by (8) if
Construction of reference model
Having established the noninteraction principle, we introduce a general reference model in section "General case". This is then specialized to the case where the matrix E is given by (8).
General case
Given are a dose \(u_i\) of the ith drug and a dose \(u_j\) of the jth drug. The question of constructing a reference model then amounts to describing the effect of combining doses \(u_i\) and \(u_j\) assuming that the two drugs obey the above described noninteraction principle. It is tempting to use the same formulation as for the explicit reference models by Lederer et al., cf. (5), where we simply replace their definition of equivalent dose by the one that we have introduced. This would result in:
where \(i=1\) and \(j=2\) or vice versa. For consistency reasons the choice should not matter, although this issue is not considered here. What is of interest here is that (13) does not describe a welldefined reference model as y is unknown, implying that \(\xi _{ij,y}(u_j)\) is undetermined.
To circumvent this matter, let us start from the effect y, whatever its value is exactly, and consider \({\mathcal {F}}_i\), given by (13). The reference model \({\mathcal {F}}_i\) can only produce the effect y by combining the doses \(u_i\) and \(u_j\) if the equivalent dose to \(u_j\) is complementary to \(u_i\). That is, when \(u_i = h_{i,y}\, \xi _{ij,y}\, u_j\). Indeed, it is only when this equality is true that it follows that the right hand side of (13) is given by
using Definition 2 of complementary doses.
In other words, we have found that y is determined by the equation \(u_i = h_{i,y}\, \xi _{ij,y} \, u_j\). The reference model \({\mathcal {F}}_i(u_i,u_j)\) is therefore described as follows:
with y the solution of
or in vectormatrix notation:
Special case
For the special case, y is, of course, also the solution of
but with \(E_{ij,y}\) given by (8). This means that the above equation reduces to
This results in two equations, but only the first one is of interest, since the second one gives the trivial identity \(1=1\). From (18) it follows that the first equation is given by
Applying the noninteraction principle for the special case, given by (12), the above equation turns out to be
In other words, we have inferred Loewe’s reference model, as the above equation is the same as (1).
Discussion
Summary of our work
Reference models rely on a noninteraction principle. It is generally thought that different reference models are distinguished by different noninteraction principles. In this paper we have shown that it is possible to construct a general noninteraction principle that may be accepted to apply to all nonreference models. The concept that may differ from reference model to reference model is the formulation of equivalent dose. Indeed, we have demonstrated that Loewe’s reference model follows from a specific instantiation of our general formulation of equivalent dose: equation (8) is a specific example of an equivalent dose definition that obeys the general formulation given by (6). In contrast, the general noninteraction principle that we introduced, given by Definition 3, does not require any additional specification: it applies to Loewe’s reference model, just as it may apply to any other reference model. Of course, because the noninteraction principle is described in terms of a certain interplay between complementary dose and equivalent dose, the specific noninteraction principle in the case of Loewe shall be different than for other reference models (which rely on a different instantiation of equivalent dose).
Novelties in the paper
One novelty of our work is that the introduced noninteraction principle (cf. Definition 3) is very general: it does not refer to any specific doseresponse curve, nor to any property of applicable doseresponse curves. It just applies to all doseresponse curves, provided they are invertible (because of the presence of the matrix H that relies on the inverse of the involved doseresponse curve).
Another novelty is that the noninteraction principle recognizes that the concept of equivalent dose, which is often employed in the literature in the construction of a noninteraction principle, is not sufficient to define noninteraction. An additional notion is needed, namely complementary dose. Noninteraction can then be defined in terms of a certain interplay between equivalence and complementarity. More specifically: drugs are noninteracting if complementarity is preserved under equivalence.
From the introduced noninteraction principle, which relies on the constructed notions of equivalence and complementarity, we were able to naturally derive Loewe’s reference model. That is, Loewe’s reference model is a consequence of the introduced noninteraction principle. In contrast to Loewe, we did not have to assume from the outset that noninteraction means that the reference model is a set of straight lines (one straight line for each given effect). In other words: the principles we introduced are more fundamental than the assumptions that Loewe made. The fact that one possible interpretation of noninteraction corresponds to straight lines does not have to be taken for granted, it simply follows from the introduced noninteraction principle for a certain, simple choice of equivalent dose.
Future work
The noninteraction principle we have introduced is very general, but it is still formulated in the rather restrictive framework of linear algebra. In particular, the concepts of complementary and equivalent doses are defined in terms of linear transformations. The class of reference models would be extended by allowing complementarity and equivalence to be defined in terms of nonlinear transformations. For example, the often used definition of equivalence as given by (4) is nonlinear.
Furthermore, it would be useful work to consider if other well known reference models could be deduced from the established framework.
Availability of data and materials
Not applicable.
References
Yadav B, Wennerberg K, Aittokallio T, Tang J. Searching for drug synergy in complex doseresponse landscapes using an interaction potency model. Comput. Struct. Biotechnol. J. 2015;13:504–13.
Preuer K, Lewis RPI, Hochreiter S, et al. DeepSynergy: predicting anticancer drug synergy with deep learning. Bioinformatics. 2018;34:1538–46.
Matsunaga S, Kishi T, Iwata N. Combination therapy with zonisamide and antiparkinson drugs for Parkinson’s disease: a metaanalysis. J. Alzheimers. Dis. 2017;56:1229–39.
Cihlar T, Fordyce M. Current status and prospects of HIV treatment. Curr. Opin. Virol. 2016;18:50–6.
Scourfield A, Waters L, Nelson M. Drug combinations for HIV: what’s new? Expert. Rev. Anti Infect. Ther. 2011;9:1001–11.
Wilson LE, Gallant JE, Mayer KH. The management of treatmentrxperienced HIVinfected patients: new drugs and drug combinations. Clin. Infect. Dis. 2009;48:214–21.
Breton M, Lelorier J, Forget A, Blais L. Use of combination therapy in asthma: are they prescribed according to guidelines. Respir. Med. 2007;101:1916–23.
Nelson HS. Combination therapy of bronchial asthma. Allergy Asthma Proc. 2001;22:217–20.
Ruggeri I, Bragato D, Colombo GL, et al. Cost and appropriateness of treating asthma with fixedcombination drugs in local health care units in Italy. Clinicoecon. Outcomes Res. 2012;4:375–82.
Saleh JA. Combination therapy in asthma: a review. Niger. J. Med. 2008;17:238–43.
Pemovska T, Bigenzahn JW, SupertiFurga G. Recent advances in combinatorial drug screening and synergy scoring. Curr. Opin. Pharmacol. 2018;42:102–10.
Schindler M. Theory of synergistic effects: Hilltype response surfaces as ‘nullinteraction’ models for mixtures. Theor. Biol. Med. Model. 2017;14:15.
Bliss CI. The toxicity of poisons applied jointly. Ann. Appl. Biol. 1939;26:585–615.
Ianevski A, He L, Aittokallio T, Tang J. SynergyFinder: a web application for analyzing drug combination doseresponse matrix data. Bioinformatics. 2017;33:2413–5.
Flobak Å, Vazquez M, Lægreid A, Valencia A. CImbinator: a webbased tool for drug synergy analysis in small and largescale datasets. Bioinformatics. 2017;33:2410–2.
Loewe S. Die quantitativen probleme der pharmakologie. Ergebnisse Physiol. 1928;27:47–187.
Lederer S, Dijkstra T, Heskes T. Additive dose response models: explicit formulation and the Loewe additivity consistency condition. Front. Pharmacol. 2018;9:31.
Funding
This work was supported by the Norwegian Research Council project COLOSYS, under the frame of ERACoSysMed1, the ERANet for Systems Medicine in clinical research and medical practice. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Author information
Authors and Affiliations
Contributions
WDM developed the framework. MK read and approved the final manuscript.
Corresponding author
Ethics declarations
Ethics approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
About this article
Cite this article
De Mulder, W., Kuiper, M. A foundation for reference models for drug combinations with an application to Loewe’s reference model. BMC Bioinformatics 21, 460 (2020). https://doi.org/10.1186/s12859020037714
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s12859020037714