1. Introduction

A growing body of scholarship shows that experiments play a variety of different roles in the biological sciences (see Weber Reference Weber and Zalta2012 for an overview). In addition to the epistemic roles traditionally associated with them—hypothesis testing and supporting causal inferences—experiments can sometimes also play exploratory or preparatory roles (e.g., Burian Reference Burian1997; Weber Reference Weber2005; Elliott Reference Elliott2007; O’Malley Reference O’Malley2007). Furthermore, new experimental approaches can have a dramatic impact on the practice of biology; they have even been shown to steer biological research into entirely new directions without the help of any new theories (Rheinberger Reference Rheinberger1997). In this article I want to examine the epistemic role of a kind of experimentation that is becoming more important in biological research but that has not been much discussed in the literature: experimental modeling.

To a first approximation, I mean by “experimental modeling” the construction and study of a specific kind of model system, namely, a model system that is composed of living organisms, not necessarily “model organisms” in the sense in which biologists use that term. Model organisms in the latter sense play a variety of different roles in biological research (Weber Reference Weber2005; Leonelli and Ankeny Reference Leonelli and Ankeny2011; Love and Travisano Reference Love and Travisano2013). Here I am interested in a very specific epistemic use of living organisms, namely, in order to re-create or simulate specific biological processes. These organisms need not be model organisms in the usual sense, even though they often are. I will examine the epistemic role of such experimental models and contrast it both with other forms of experimentation and with other forms of modeling. My goal is to show that the epistemic role of experimental models is quite unlike that of ordinary experiments and more akin to the role of theoretical models.

To give some classic examples of what I mean by experimental modeling, evolutionary biologists have been re-creating evolutionary processes in the laboratory by using fruit flies since as early as the 1930s (e.g., Timoféeff-Ressovsky Reference Timoféeff-Ressovsky1934). I present some more recent examples of experimental evolution in section 3.1. In a similar way, twentieth-century ecologists have staged interspecies competitions in artificial communities or microcosms (e.g., Park Reference Park1948). Today, as we shall see (sec. 3.2), experimental ecologists can engineer synthetic ecological systems with very specific properties in order to study ecological and evolutionary processes.

In the following section I try to locate experimental models in the conceptual landscape of models, experiments, and simulations. I show that a crucial difficulty lies in distinguishing the epistemic role of experimental models from that of ordinary experiments. In section 3 I introduce two examples of experimental modeling, both from the area of experimental evolution. Section 4 examines to what extent these experimental models provide causal knowledge, as the production of this kind of knowledge is traditionally seen as the primary role of experiments. In section 5 I further characterize the epistemic role of the experimental models by using the notions of stand-in and in vivo representation. Section 6 tries to answer the question of what epistemic advantages this approach offers, in particular as compared with the computational or “in silico” approach. Section 7 summarizes my conclusions.

2. Models, Experiments, and Simulations

As a number of authors have pointed out, the construction and study of material, biological systems may sometimes play a similar or even the same epistemic role as theorizing (e.g., Griesemer Reference Griesemer1990; Plutynski Reference Plutynski2001; Waters Reference Waters2012). On this view, Park’s (Reference Park1948) experimental work involving flour beetles should be considered to be analogous to Lotka’s (Reference Lotka1925) and Volterra’s (Reference Volterra1926) well-known mathematical models of interspecific competition; they are both forms of theorizing, but with different means. Park used living organisms, while Lotka and Volterra used calculus and differential equations to model competition in ecological communities. Both models are idealized; the conditions they create or assume do not exist in the wild. Also, both approaches are rigorous in their own way: Lotka’s and Volterra’s work was mathematically rigorous, while Park’s work was rigorous at the level of experimental design and control. Finally, much like mathematical models, experimental models allow scientists to explore the space of what is possible in nature.Footnote ¹ Thus, there are many important similarities between experimental models and theoretical models.

Much like experiments, scientific models can also play a variety of epistemic roles (Morgan and Morrison Reference Morgan and Morrison1999; Frigg and Hartmann Reference Frigg, Hartmann and Zalta2012). But it seems that a standard role of theoretical models in science is to selectively represent some aspects of a kind of system or process, which may be real or fictive (e.g., Giere Reference Giere2004; Frigg Reference Frigg2010; Weisberg Reference Weisberg2013). A distinctive feature of models seems to be that they represent their target in an indirect way (Weisberg Reference Weisberg2007). When engaging in theoretical modeling, scientists investigate a surrogate or stand-in for the systems they are interested in because that stand-in is epistemically more tractable. Another feature is that models come in different forms: some are abstract, consisting of mathematical equations of which some of the symbols are taken to represent physical magnitudes, while some are concrete, for example, the Watson-Crick model of the double helix as it was built from cardboard and wire in 1953.

If physical systems can function as theoretical models, then so can living organisms. Indeed, several authors have suggested that so-called model organisms such as Drosophila melanogaster, E. coli, or Arabidopsis thaliana, which have been extremely important in biological research over the past 100 years or so, are basically concrete theoretical models (e.g., Leonelli and Ankeny Reference Leonelli and Ankeny2011; Frigg and Hartmann Reference Frigg, Hartmann and Zalta2012; Weisberg Reference Weisberg2013). While I disagree with this view (as do Levy and Currie, Reference Levy and Currieforthcoming), the case is different for the experimental models under consideration here.Footnote ²

If we accept the claim that experimental models may serve a similar or even the same role as theoretical models, we should be able to say more clearly what distinguishes their epistemic role from that of experiments traditionally conceived. The existing works on experimental modeling cited above do not address this question. But this question is important because if we cannot answer it, we are driven toward the conclusion that all experiments are also theoretical models, a claim that strikes me as counterintuitive.Footnote ³ Note also that what is considered to be a distinctive feature of models, the indirect way in which they represent their target, may also be said about certain experiments. For example, ecologists who want to study the effect of increased atmospheric CO₂ levels on forests will not experiment on a whole forest but on a suitably fenced part of a forest.Footnote ⁴ Thus, experiments may also use some sort of a stand-in for their target.

To add to the confusion, at least some experimental models can also be viewed as simulations, a scientific category that also turned out to be difficult to distinguish conceptually from experiments. Recent discussion on this issue has centered on the question of the extent to which the “materiality” or the stuff of which the systems studied and of which the target systems are made matters conceptually and epistemically (e.g., Guala Reference Guala, Magnani and Nersessian2002; Morgan Reference Morgan and Radder2003, Reference Radder2005; Parker Reference Parker2009; Winsberg Reference Winsberg2009). In our present examples, it is difficult to decide what should count as relevantly the same stuff, but I argue later that in experimental modeling causal processes may be used to represent physically different processes (see sec. 5).

Winsberg (Reference Winsberg2009) has argued that the pertinent difference between simulations and experiments lies in the way in which their external validity (i.e., their representativeness with respect to the intended target systems) is justified. According to him, the external validity of a simulation is justified with recourse to the reliability of the principles used for constructing the simulation model. This is knowledge that concerns primarily the target. By contrast, the external validity of experiments is ascertained by an appeal to different kinds of knowledge, including knowledge that the objects of the experiment are properly isolated or shielded from disturbing influences and that the object belongs to the same (relevant) kind as the target.

According to Winsberg’s distinction, experimental models such as the ones considered here all qualify as experiments because much knowledge about the experimental objects is needed to argue for external validity. Thus, his proposal—whatever its merits with respect to the epistemology of simulations—does not help to distinguish between experimental models and ordinary experiments.

Before moving on, I should briefly clarify what I mean by an “ordinary” experiment. For the purposes of this article I suggest the following: A principal role of biological experiments is the discovery of specific causal relations (Weber Reference Weber and Zalta2012). In other words, ordinary experiments serve to provide premises for causal inference. To give an example, the main purpose of the legendary “PaJaMo” experiment in molecular biology—named after the three scientists who designed and performed it in 1959, Arthur Pardee, François Jacob, and Jacques Monod—was to establish that the induction of transcription in the E. coli lac operon is caused by the release of a repressor protein from a specific DNA binding site (see Weber Reference Weber2005, chap. 3 for details). I take this example to be paradigmatic for much experimental activity in biological research. What I want to show here is that experimental models play a somewhat different role, indeed a role that is more similar to theoretical models.

3. Two Examples of Experimental Modeling

4. What Kind of Knowledge Does Experimental Modeling Provide?

The evolution of multicellularity experiments (example A) established the causal relevance of gravity in the evolutionary transition from uni- to multicellularity in this system. However, it should be noted that this was not the main purpose of these experiments. The experimental model was designed to study an important evolutionary transition in general, not specifically the effect of gravity on the evolution of yeast cells. Thus, the experiment served not primarily to establish a claim of causal relevance. Rather, the system was already designed such that a certain relation of causal relevance holds, namely, a fitness relevance of gravity. To be sure, the experiment did show that rapid settling was a cause of multicellularity in this system. But the main point of the experiment was to show that, if multicellularity is favored by natural selection, mechanisms that realize multicellularity evolve within a few generations. These mechanisms include postdivision cell adhesion and the beginnings of soma–germ line distinction. Thus, the experiment probed primarily for the dynamics of an evolutionary transition, not for some specific kind of causal dependence.

The case is somewhat different in example B. There, like in example A, the system was also already designed to exhibit a set of causal dependencies, namely, the mutual dependence of the two species on each other as a source of metabolically available energy. It was clear that this mutual dependence would cause evolutionary change. The main question that the experiment sought to answer was how this evolutionary change would occur: Does one species adapt to the other, or do they both undergo adaptive change? Does the interaction always lead to a stable mutualism, or is it prone to instability and even to occasional extinction events? Does the stability increase over evolutionary time?

While these experiments also investigated the dynamics of evolutionary change, some of these questions can also be understood as questions about causality: Does the ecological interaction between two species cause adaptations in one or in both species? Does coevolution cause a state of greater stability over time? Thus, the experimental inquiry had an important causal aspect. But the point of this study was not merely to establish that an evolution toward greater stability occurred in this type of ecological interaction. The idea is that the evolutionary changes observed generalize to other kinds of ecological mutualistic interaction between species, interactions that are mediated by physically different kinds of processes. Indeed, quite similar behaviors appear to have been observed by Hillesland and Stahl with their symbiotic methanotrophic bacteria and by Shou et al. with their symbiotic yeast. But the nature of the ecological interaction was quite different: in the first case, it was a symbiosis involving energy-rich foodstuffs; in the second case, essential amino acids needed for protein synthesis.

Thus, we have identified an important difference between experimental models and ordinary experiments: while ordinary experiments seek to establish a causal dependence in a particular system, experimental models provide knowledge of causal processes that generalizes to systems where biologically and chemically quite different kinds of causes are at work.

In the following section I compare experimental models with other kinds of scientific models.

5. Models, Stand-Ins, and the Technique of In Vivo Representation

Scientific models come in different forms: some abstract, some concrete, some mathematical, others physical or computational. According to a number of authors, models represent fictive model systems (e.g., Frigg Reference Frigg2010). Typical model systems known from physics are the ideal gas, Bohr’s model of the atom, and the harmonic oscillator. Physicists invent such models and attribute certain properties to them (e.g., point masses in the case of the ideal gas, frictionless motion in the case of the harmonic oscillator). Then, they study what other properties result from the properties that they built into their fictive model systems, given certain assumptions about how physical properties are related (e.g., Newton’s laws). They use mathematical or sometimes computational or simulation methods in order to find out about these properties. Finally, physicists study how similar the model systems are to real physical systems.

I claim that experimental models share important features with such abstract, physical models. They also involve the construction of idealized model systems that do not exist in nature (e.g., an evolving isolated clone of bacteria in a homogeneous environment, an isolated two-species mutualism, or an isolated two-species predator-prey system). The model is endowed with certain properties that scientists deem important on the grounds of previous theoretical considerations. Then, the scientists study what other properties result from these properties that were built into the model systems. Finally, they compare how similar the model systems are to real systems in nature. Instead of mathematical derivations and computer simulations, they use living organisms. The model systems they study are not fictions; they are real, but no less idealized than classical model systems.

Another similarity in particular to computational models is the following. Experimental models use certain kinds of interactions to stand in for other biological interactions. For example, in the Travisano lab’s experiment on the evolution of multicellularity (sec. 3.1), the postdivision cell adhesion between yeast cells stands in for the adhesive interactions between cells of other multicellular organisms. Similarly, the gravity-induced settling stands in for whatever environmental condition favored multicellularity in our evolutionary past. By the same token, in Hillesland and Stahl’s work on mutualism, as well as in Shou et al. (sec. 3.2), the metabolic dependence of the two microbial strains used stands in for all kinds of mutualistic interactions between species.Footnote ⁵

Note that, in all of these cases, the interactions represented by the interaction that stands in for them are multiply realizable by different biological processes. This, I claim, is similar to computational modeling, where the physical processes occurring in a computer stand in for whatever interactions are being modeled. Computational algorithms are also multiply realizable by physical processes. Thus, experimental models bear many of the marks that are distinctive of theoretical models, marks that are not characteristic of what I call “ordinary” biological experiments.

As I already mentioned in the introduction, it could be objected that all experiments involve some kind of stand-in relation. I agree, but I maintain that in experimental models we have a different kind of standing in, namely, one where the process re-created in the laboratory can represent physically distinct kinds of processes. This is not the case in ordinary biological experiments, where whatever happens in the experiment is physically of the same kind as what happens in the target systems. Consider, for example, the PaJaMo experiment mentioned in section 2. There the genetic processes that occurred in the experimental sample were physically identical to the target processes—tokens of lac operon and repressor protein binding to it.Footnote ⁶

So, my answer to the question of what distinguishes experimental models is the following: they are distinguished by being bona fide cases of scientific modeling that provide in vivo representations of some process. By an “in vivo representation” I mean such a representation that shows the typical features of a model, such as being idealized, simplified, multiply realizable, and so on, but where the thing that does the representing is alive.

An important question is what it is that the model and its target have in common. I lack the space here to give this issue the treatment that it requires. To a first approximation, I suggest that what they have in common is some high-level functional properties such as being a predator. But this is an issue for further study.

6. Why Model In Vivo instead of In Silico?

Why do biologists build experimental models? Can they not model and simulate the behavior of simple systems such as the ones examined here by using mathematical and computational tools? The answer is, of course, that they can and do so often. In fact, some of the work I have examined here also involved the use of such models along with the experimental model. Hence, it seems that more ordinary forms of modeling go along with experimental modeling. So what is the added value of experimental modeling?

I suggest that the advantages differ from case to case. In example A, the experimental model offers avenues of research not available from mathematical models of evolutionary transitions simply for the lack of an appropriate theory. While there is a mathematically formulated theory to explain what the benefits of intercell cooperation are in evolutionary terms, there is no theory that would be able to predict whether yeast or any other cells can evolve the cellular mechanisms of multicellularity within a few generations. Evolutionary theory can at best predict that multicellularity will evolve if the proximate mechanisms necessary for its realization are available. But that they can become available rapidly in a eukaryote had to be shown experimentally.

It is more difficult to say what the advantage of experimental modeling is in example B because mathematical models that predict the behavior of two species engaged in a mutualistic interaction in ecological as well as evolutionary timescales are available. Here there are two kinds of fact not predictable from evolutionary theory. The first kind is similar to example A: that the necessary mutations become available for evolution to act on in both species is not predictable from evolutionary theory. This had to be shown empirically. The other kind of nonpredictable facts concerns the occurrence of other changes that accompany the changes that the system was designed to manifest. In example B, the yeast strains engineered for obligatory mutualism (Shou et al. Reference Shou, Ram and Vilar2007) showed a tendency to increase their ability to survive population bottlenecks. That this would occur is, to my knowledge, not predictable from evolutionary theory. In the example of an engineered two-species trophic mutualism (Hillesland and Stahl Reference Hillesland and Stahl2010), the occurrence of antagonistic effects surprised the scientists. That this would occur was also not predictable from theory.

A helpful question to ask is what kind of surprises experimental models give rise to. Morgan (Reference Morgan and Radder2003, Reference Radder2005) distinguishes between results that “surprise” and results that “confound” the researcher. As I understand her distinction, a surprise is an unexpected behavior that is predictable from existing theory in principle and thus merely unexpected owing to cognitive or computational limitations on the part of the modeler. Mathematical models, of course, can be full of surprises in this sense. By contrast, a result that confounds is “unexplainable within existing theory.” It seems to me that the unexpected results in the experimental models are of the second kind. However, it should be noted that, unlike in the economics examples discussed by Morgan, this unexplainability is not due to any defect in the existing theory. It is simply not the job of evolutionary theory to explain what kinds of mutations are available to a species. This, I suggest, is an important reason why experimental modeling has been so successful in evolutionary biology.

Finally, it should be noted that an important advantage of experimental models could be that the sources of potential error are different from those in, for example, mathematical models. Unlike the latter, experimental models do not have to rely on the truth of certain theoretical assumptions.Footnote ⁷ Instead, there are possible errors in the interpretation of data. Therefore, combining experimental with mathematical and computational modeling techniques, which was done in example B, could be an effective way of controlling for errors.

7. Conclusions

I hope to have established that there is a distinct kind of experimentation in biology that is best described as a form of scientific modeling. It consists of constructing model systems that are composed of living organisms (sometimes, but not necessarily, genetically modified) and that are used as in vivo representations of biological processes in such a way that some processes are used as stand-ins for other processes. Unlike the classical model systems known from physics, such as the ideal gas, the model systems that experimental biologists construct are concrete and real, not abstract and not fictive. Furthermore, they are alive. But apart from that, they exhibit the characteristic features of scientific models, in particular idealization and surrogacy.

Experimental models, even though they are also experiments, differ from ordinary biological experiments mainly in the kinds of claims they are used to support: their results are generalized to other kinds of systems where the causal interactions under study are realized by physically distinct processes.